Datasets:

timaeus

/

pile-arxiv

like 0

Follow

Timaeus 9

--- title: About quadratic differential operators --- [**SPECTRA AND SEMIGROUP SMOOTHING FOR NON-ELLIPTIC QUADRATIC OPERATORS**\ Michael <span style="font-variant:small-caps;">Hitrik</span>, Karel <span style="font-variant:small-caps;">Pravda-Starov</span>\ University of California, Los Angeles]{} \[subsection\] \[subsection\] \[subsection\] \[subsection\] **Abstract.** We study non-elliptic quadratic differential operators. Quadratic differential operators are non-selfadjoint operators defined in the Weyl quantization by complex-valued quadratic symbols. When the real part of their Weyl symbols is a non-positive quadratic form, we point out the existence of a particular linear subspace in the phase space intrinsically associated to their Weyl symbols, called a singular space, such that when the singular space has a symplectic structure, the associated heat semigroup is smoothing in every direction of its symplectic orthogonal space. When the Weyl symbol of such an operator is elliptic on the singular space, this space is always symplectic and we prove that the spectrum of the operator is discrete and can be described as in the case of global ellipticity. We also describe the large time behavior of contraction semigroups generated by these operators. **Key words.** quadratic differential operators, contraction semigroups, exponential decay, FBI-Bargmann transform, spectrum, semigroup smoothing. **2000 AMS Subject Classification.** 47A10, 47D06, 35P05. Introduction ============ Miscellaneous facts about quadratic differential operators ---------------------------------------------------------- Since the classical work by J. Sjöstrand [@sjostrand], the study of spectral properties of quadratic differential operators has played a basic rôle in the analysis of partial differential operators with double characteristics. Roughly speaking, if we have, say, a classical pseudodifferential operator $p(x,\xi)^w$ on ${\mathbb{R}}^n$ with the Weyl symbol $p(x,\xi)=p_m(x,\xi)+p_{m-1}(x,\xi)+\ldots$ of order $m$, and if $X_0=(x_0,\xi_0)\in {\mathbb{R}}^{2n}$ is a point where $p_m(X_0)=dp_m(X_0)=0$ then it is natural to consider the quadratic form $q$ which begins the Taylor expansion of $p_m$ at $X_0$. The study of a priori estimates for $p(x,\xi)^w$, such as hypoelliptic estimates of the form $$\norm{u}_{m-1}\leq C_K \left(\norm{p(x,\xi)^w u}_0+\norm{u}_{m-2}\right),\quad u\in C^{\infty}_0(K), \quad K\subset \subset {\mathbb{R}}^n,$$ then often depends on the spectral analysis of the quadratic operator $q(x,\xi)^w$. See also [@hypoelliptic], as well as Chapter 22 of [@hormander] together with further references given there. In [@sjostrand], the spectrum of a general quadratic differential operator has been determined, under the basic assumption of global ellipticity of the associated quadratic form. Now there exist many situations where one is naturally led to consider non-selfadjoint quadratic differential operators whose symbols are not elliptic but rather satisfy certain weaker conditions. An example particularly relevant to the following discussion is obtained if one considers the Kramers-Fokker-Planck operator with a quadratic potential [@HN]. The corresponding (complex-valued) symbol is not elliptic, but nevertheless, the operator has discrete spectrum and the associated heat semigroup is well behaved in the limit of large times — see [@HeSjSt]. The purpose of the present paper is to provide a proof of a number of fairly general results concerning the spectral and semigroup properties for the class of quadratic differential operators in the case when the global ellipticity fails. Specifically, and as alluded to above, we shall consider the class of pseudodifferential operators defined by the Weyl quantization formula, $$\label{3}\inc q(x,\xi)^w u(x) =\frac{1}{(2\pi)^n}\int_{{\mathbb{R}}^{2n}}{e^{i(x-y).\xi}q\Big(\frac{x+y}{2},\xi\Big)u(y)dyd\xi}, \num$$ for some symbols $q(x,\xi)$, where $(x,\xi) \in {\mathbb{R}}^{n} \times {\mathbb{R}}^n$ and $n \in {\mathbb{N}}^*$, which are complex-valued quadratic forms. Since the symbols are quadratic forms, the corresponding operators in (\[3\]) are in fact differential operators. Indeed, the Weyl quantization of the quadratic symbol $x^{\alpha} \xi^{\beta}$, with $(\alpha,\beta) \in {\mathbb{N}}^{2n}$ and $|\alpha+\beta| \leq 2$, is the differential operator $$\frac{x^{\alpha}D_x^{\beta}+D_x^{\beta} x^{\alpha}}{2}, \ D_x=i^{-1}\partial_x.$$ Let us also notice that since the Weyl symbols in (\[3\]) are complex-valued, the quadratic differential operators are a priori formally non-selfadjoint. In this paper, we shall first study the properties of contraction semigroups generated by quadratic differential operators whose Weyl symbols have a non-positive real part, $$\label{smm1}\inc \textrm{Re }q \leq 0. \num$$ Our first goal is to point out the existence of a linear subspace $S$ in ${\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n$, which will be called the singular space and which is defined in terms of the Hamilton map of the Weyl symbol $q$, such that when $S$ has a *symplectic structure*, the associated heat equation $$\label{sm1}\inc \left\lbrace \begin{array}{c} \displaystyle \frac{\partial u}{\partial t}(t,x)-q(x,\xi)^w u(t,x)=0 \\ u(t,\textrm{\textperiodcentered})|_{t=0}=u_0 \in L^2({\mathbb{R}}^n), \end{array} \right. \num$$ is smoothing in every direction of the orthogonal complement $S^{\sigma \perp}$ of $S$ with respect to the canonical symplectic form $\sigma$ on ${\mathbb{R}}^{2n}$, $$\label{11}\inc \sigma \big{(}(x,\xi),(y,\eta) \big{)}=\xi.y-x.\eta, \ (x,\xi) \in {\mathbb{R}}^{2n}, (y,\eta) \in {\mathbb{R}}^{2n}. \num$$ We shall also describe the large time behavior of contraction semigroups $$e^{tq(x,\xi)^w}, \ t \geq 0,$$ associated to (\[sm1\]). When the Weyl symbol $q$ satisfies (\[smm1\]) and an assumption of *partial* ellipticity, namely when $q$ is elliptic on the singular space $S$ in the sense that $$\label{sm2}\inc (x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0, \num$$ then $S$ is automatically symplectic, and we prove that the spectrum of the quadratic differential operator $q(x,\xi)^w$ is only composed of a countable number of eigenvalues of finite multiplicity, with its structure similar to the one known in the case of global ellipticity [@sjostrand]. It seems to us that the singular space $S$ introduced in this paper plays a basic rôle in the understanding of non-elliptic quadratic differential operators. Its study may therefore be also particularly relevant in the analysis of general pseudodifferential operators with double characteristics, when the ellipticity of their quadratic approximations fails. Before giving the precise statements of these results, let us begin by recalling some facts and notation about quadratic differential operators. Let $$\begin{aligned} q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n &\rightarrow& {\mathbb{C}}\\ (x,\xi) & \mapsto & q(x,\xi),\end{aligned}$$ be a complex-valued quadratic form with a non-positive real part, $$\label{inf1}\inc \textrm{Re }q(x,\xi) \leq 0, \ (x,\xi) \in {\mathbb{R}}^{2n}, n \in {\mathbb{N}}^*. \num$$ We know from [@mehler] (p.425) that the maximal closed realization of the operator $q(x,\xi)^w$, i.e., the operator on $L^2({\mathbb{R}}^n)$ with the domain $$\{u \in L^2({\mathbb{R}}^n) : q(x,\xi)^w u \in L^2({\mathbb{R}}^n)\},$$ coincides with the graph closure of its restriction to $\mathcal{S}({\mathbb{R}}^n)$, $$q(x,\xi)^w : \mathcal{S}({\mathbb{R}}^n) \rightarrow \mathcal{S}({\mathbb{R}}^n),$$ and that every quadratic differential operator whose Weyl symbol has a non-positive real part, generates a contraction semigroup. The Mehler formula proved by L. Hörmander in [@mehler] gives an explicit expression for the Weyl symbols of these contraction semigroups. Associated to the quadratic symbol $q$ is the numerical range $\Sigma(q)$ defined as the closure in the complex plane of all its values, $$\label{9}\inc \Sigma(q)=\overline{q({\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n)}. \num$$ We also recall [@hormander] that the Hamilton map $F \in M_{2n}({\mathbb{C}})$ associated to the quadratic form $q$ is the map uniquely defined by the identity $$\label{10}\inc q\big{(}(x,\xi);(y,\eta) \big{)}=\sigma \big{(}(x,\xi),F(y,\eta) \big{)}, \ (x,\xi) \in {\mathbb{R}}^{2n}, (y,\eta) \in {\mathbb{R}}^{2n}, \num$$ where $q\big{(}\textrm{\textperiodcentered};\textrm{\textperiodcentered} \big{)}$ stands for the polarized form associated to the quadratic form $q$. It follows directly from the definition of the Hamilton map $F$ that its real part $\textrm{Re } F$ and its imaginary part $\textrm{Im }F$ are the Hamilton maps associated to the quadratic forms $\textrm{Re } q$ and $\textrm{Im }q$, respectively. Next, (\[10\]) shows that a Hamilton map is always skew-symmetric with respect to $\sigma$. This is just a consequence of the properties of skew-symmetry of the symplectic form and symmetry of the polarized form, $$\label{12}\inc \forall X,Y \in {\mathbb{R}}^{2n}, \ \sigma(X,FY)=q(X;Y)=q(Y;X)=\sigma(Y,FX)=-\sigma(FX,Y).\num$$ Let us now consider the elliptic case, i.e., the case of quadratic differential operators whose Weyl symbols are *globally* elliptic in the sense that $$\label{sm3}\inc (x,\xi) \in {\mathbb{R}}^{2n}, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0. \num$$ In this case, the numerical range of a quadratic form can only take very particular shapes. J. Sjöstrand proved in [@sjostrand] (Lemma 3.1) that if $q$ is a complex-valued elliptic quadratic form on ${\mathbb{R}}^{2n}$, with $n \geq 2$, then there exists $z \in {\mathbb{C}}^*$ such that $\textrm{Re}(z q)$ is a positive definite quadratic form. If $n=1$, the same result is fulfilled if we assume besides that $\Sigma(q) \neq {\mathbb{C}}$. This shows that the numerical range of an elliptic quadratic form can only take two shapes. The first possible shape is when $\Sigma(q)$ is equal to the whole complex plane. This case can only occur in dimension $n=1$. The second possible shape is when $\Sigma(q)$ is equal to a closed angular sector with a vertex in $0$ and an aperture strictly less than $\pi$ (see [@mz] for more details). We also know that elliptic quadratic differential operators define Fredholm operators (see Lemma 3.1 in [@hypoelliptic] or Theorem 3.5 in [@sjostrand]), $$\label{14}\inc q(x,\xi)^w +z : B \rightarrow {L^2({\mathbb{R}}^n)}, \num$$ where $B$ is the Hilbert space $$\begin{aligned} \label{14.1} B= & \ \big\{ u \in {L^2({\mathbb{R}}^n)}: q(x,\xi)^wu \in {L^2({\mathbb{R}}^n)}\big\} \num \\ = & \ \big\{ u \in {L^2({\mathbb{R}}^n)}: x^{\alpha} D_x^{\beta} u \in {L^2({\mathbb{R}}^n)}\ \textrm{if} \ |\alpha+\beta| \leq 2\big\},\end{aligned}$$ with the norm $$\|u\|_B^2=\sum_{|\alpha+\beta| \leq 2}{\|x^{\alpha} D_x^{\beta} u\|_{{L^2({\mathbb{R}}^n)}}^2}.$$ Moreover, the index of the operator (\[14\]) is independent of $z$ and is equal to $0$ when $n \geq 2$. In the case $n=1$, the index can take the values $-2$, $0$ or $2$. It vanishes as soon as $\Sigma(q) \neq {\mathbb{C}}$. When $\Sigma(q) \neq {\mathbb{C}}$, J. Sjöstrand has proved in Theorem 3.5 of [@sjostrand] (see also Lemma $3.2$ and Theorem 3.3 in [@hypoelliptic]) that the spectrum of an elliptic quadratic differential operator $$q(x,\xi)^w : B \rightarrow {L^2({\mathbb{R}}^n)},$$ is only composed of eigenvalues with finite multiplicity, $$\label{15}\inc \sigma\big{(}q(x,\xi)^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F), \\ -i \lambda \in \Sigma(q)\setminus \{0\}} } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\}, \num$$ where $F$ is the Hamilton map associated to the quadratic form $q$ and $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F$ in ${\mathbb{C}}^{2n}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. Let us also recall the result proved in [@mz] about contraction semigroups generated by elliptic quadratic differential operators whose Weyl symbols have a non-positive real part. This result shows that, as soon as the real part of their Weyl symbols is a non-zero quadratic form, the norm of contraction semigroups generated by these operators decays exponentially in time. In this paper, we study the case when the ellipticity fails. Our second result (Theorem \[theorem2\]) extends the description of the spectra (\[15\]) to the case of quadratic differential operators whose Weyl symbols are partially elliptic, but not necessarily globally so. To get this result, we only require that these symbols have a non-positive real part and are elliptic on their associated singular spaces. We also prove a result on the exponential decay in time for the norm of contraction semigroups generated by non-elliptic quadratic differential operators. Let us now define this singular space. The *singular* space $S$ associated to the symbol $q$ is defined as the following intersection of the kernels, $$\label{h1}\inc S=\Big(\bigcap_{j=0}^{+\infty}\textrm{Ker}\big[\textrm{Re }F(\textrm{Im }F)^j \big]\Big) \cap {\mathbb{R}}^{2n}, \num$$ where the notation $\textrm{Re } F$ and $\textrm{Im }F$ stands respectively for the real part and the imaginary part of the Hamilton map associated to $q$. Notice that the Cayley-Hamilton theorem applied to $\textrm{Im }F$ shows that $$(\textrm{Im }F)^k X \in \textrm{Vect}\big(X,...,(\textrm{Im }F)^{2n-1}X\big), \ X \in {\mathbb{R}}^{2n}, \ k \in {\mathbb{N}},$$ where $\textrm{Vect}\big(X,...,(\textrm{Im }F)^{2n-1}X\big)$ is the vector space spanned by the vectors $X$, ..., $(\textrm{Im }F)^{2n-1}X$, and therefore the singular space is actually equal to the following finite intersection of the kernels, $$\label{h1bis}\inc S=\Big(\bigcap_{j=0}^{2n-1}\textrm{Ker}\big[\textrm{Re }F(\textrm{Im }F)^j \big]\Big) \cap {\mathbb{R}}^{2n}. \num$$ The subspace $S$ obviously satisfies the two following properties, $$\label{h3}\inc (\textrm{Re }F) S=\{0\} \textrm{ and } (\textrm{Im }F)S \subset S. \num$$ We can now give the statements of the main results contained in this paper. Statement of the main results ----------------------------- In the following statements, we consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}},$$ with a non-positive real part, $$\label{sm4}\inc \textrm{Re }q(x,\xi) \leq 0, \ (x,\xi) \in {\mathbb{R}}^{2n}, \ n \in {\mathbb{N}}^*, \num$$ and we denote by $S$ the singular space defined in (\[h1\]) or (\[h1bis\]). Our first result states that when the singular space $S$ has a symplectic structure, in the sense that the restriction of $\sigma$ to $S$ is nondegenerate, the heat equation (\[sm1\]) associated to the operator $q(x,\xi)^w$ is smoothing in every direction of its orthogonal complement $S^{\sigma \perp}$ with respect to the canonical symplectic form in ${\mathbb{R}}^{2n}$. \[theorem1\] Let us assume that the singular space $S$ has a symplectic structure. If $(x',\xi')$ are some linear symplectic coordinates on the symplectic space $S^{\sigma \perp}$, then for all $t>0$, $N \in {\mathbb{N}}$ and $u \in L^2({\mathbb{R}}^n)$, $$\label{sm4bis}\inc \big((1+|x'|^2+|\xi'|^2)^N\big)^we^{t q(x,\xi)^w}u \in L^2({\mathbb{R}}^n). \num$$ Let us mention that the assumption about the symplectic structure of $S$ is always fulfilled by any quadratic symbol $q$ elliptic on $S$, i.e., $$\label{sm5}\inc (x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0. \num$$ This assumption is therefore always fulfilled for elliptic quadratic differential operators. We will see that it is also the case for instance for the Kramers-Fokker-Planck operator with a quadratic potential, which is a non-elliptic operator. When $q$ is a complex-valued quadratic form with a non-positive real part verifying (\[sm5\]), we can give another description of the singular space in terms of the eigenspaces of $F$ associated to its real eigenvalues. Under these assumptions, the set of real eigenvalues of the Hamilton map $F$ can be written as $$\sigma(F) \cap {\mathbb{R}}=\{\lambda_1,...,\lambda_r,-\lambda_1,...,-\lambda_r\},$$ with $\lambda_j \neq 0$ and $\lambda_j \neq \pm \lambda_k$ if $j \neq k$. The singular space is then the direct sum of the symplectically orthogonal spaces $$\label{sm5bis}\inc S=S_{\lambda_1} \oplus^{\sigma \perp} S_{\lambda_2} \oplus^{\sigma \perp}... \oplus^{\sigma \perp} S_{\lambda_r}, \num$$ where $S_{\lambda_{j}}$ is the symplectic space $$\label{sm5bis1}\inc S_{\lambda_j}=\big(\textrm{Ker}(F -\lambda_j) \oplus \textrm{Ker}(F+\lambda_j) \big) \cap {\mathbb{R}}^{2n}. \num$$ These facts will be proved in section \[examples\]. Our second result deals with the structure of the spectra for non-elliptic quadratic differential operators. This result extends the description of the spectra (\[15\]) proved by J. Sjöstrand in [@sjostrand] (Theorem 3.5) for elliptic quadratic differential operators to the case of quadratic differential operators which are only partially elliptic. To get this description, we only require in addition to the assumption (\[sm4\]) the property of partial ellipticity (\[sm5\]) for their Weyl symbols. \[theorem2\] If $q$ is a complex-valued quadratic form with a non-positive real part and if $q$ is elliptic on $S$, $$(x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0,$$ then the spectrum of the quadratic differential operator $q(x,\xi)^w$ is only composed of eigenvalues of finite multiplicity, $$\label{sm6}\inc \sigma\big{(}q(x,\xi)^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F), \\ -i \lambda \in {\mathbb{C}}_- \cup (\Sigma(q|_S) \setminus \{0\}) } } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\}, \num$$ where $F$ is the Hamilton map associated to the quadratic form $q$, $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F$ in ${\mathbb{C}}^{2n}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$, $$\Sigma(q|_S)=\overline{q(S)} \textrm{ and } {\mathbb{C}}_-=\{z \in {\mathbb{C}}: \emph{\textrm{Re }}z<0\}.$$ Since the singular space $S$ is distinct from the whole phase space as soon as the real part of $q$ is not identically equal to zero, Theorem \[theorem2\] is a generalization of the result proved by J. Sjöstrand for elliptic quadratic differential operators. Finally, we give a result concerning the large time behavior of contraction semigroups generated by non-elliptic quadratic differential operators, which extends the result obtained by the second author in [@mz]. \[theorem3\] Let us consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part, such that its singular space $S$ has a symplectic structure. Then, the following assertions are equivalent: - The norm of the contraction semigroup generated by the operator $q(x,\xi)^w$ decays exponentially in time, $$\exists M>0, \exists a >0, \forall t \geq 0, \ \|e^{tq(x,\xi)^w}\|_{\mathcal{L}(L^2)} \leq M e^{-a t}.$$ - The real part of the symbol $q$ is a non-zero quadratic form $$\exists (x_0,\xi_0) \in {\mathbb{R}}^{2n}, \ \emph{\textrm{Re }}q(x_0,\xi_0) \neq 0.$$ - The singular space is distinct from the whole phase space $S \neq {\mathbb{R}}^{2n}$. Since the assumption about the symplectic structure of the singular space $S$ is always fulfilled when the symbol $q$ verifies (\[sm5\]), Theorem \[theorem3\] is a generalization of the result proved in [@mz] for elliptic quadratic differential operators. Let us also notice that we cannot drop completely the assumption about the symplectic structure of the singular space. Indeed, let us consider the quadratic differential operator defined in the Weyl quantization by the symbol $$q(x,\xi)=-x^2.$$ This operator is just the operator of multiplication by $-x^2$, which generates the contraction semigroup $$e^{t q(x,\xi)^w}u=e^{-t x^2}u, \ t \geq 0, \ u \in L^2({\mathbb{R}}^n),$$ whose norm is identically equal to 1, $$\|e^{t q(x,\xi)^w}\|_{\mathcal{L}(L^2)}=1, \ t \geq 0.$$ *Remark.* Let us mention that our proof will show in particular that when $q$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n}$, $n \geq 1$, with a non-positive real part and a zero singular space $S=\{0\}$, then $$e^{t q(x,\xi)^w}=e^{t q(x,\xi)^w} \Pi_a+\mathcal{O}_a(e^{-at}), \ t \geq 0,$$ in the space $\mathcal{L}(L^2)$ of bounded operators on $L^2({\mathbb{R}}^n)$, for any $a>0$ such that $$\begin{aligned} & \ \sigma\big(q(x,\xi)^w\big) \cap \{ z \in {\mathbb{C}}: {\textrm{Re }}z =-a\} \\ = & \ \Big\{ \sum_{\substack{\lambda \in \sigma(F), \\ \textrm{Re}(-i \lambda) <0 } } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\}\cap \{ z \in {\mathbb{C}}: {\textrm{Re }}z=-a \}=\emptyset,\end{aligned}$$ where $\Pi_a$ stands for the finite rank spectral projection associated to the following eigenvalues of the operator $q(x,\xi)^w$, $$\begin{aligned} & \ \sigma\big(q(x,\xi)^w\big) \cap \{ z \in {\mathbb{C}}: -a \leq {\textrm{Re }}z \} \\ = & \ \Big\{ \sum_{\substack{\lambda \in \sigma(F), \\ \textrm{Re}(-i \lambda) <0 } } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\}\cap \{ z \in {\mathbb{C}}: -a \leq {\textrm{Re }}z \},\end{aligned}$$ where $F$ is the Hamilton map associated to the quadratic form $q$ and $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F$ in ${\mathbb{C}}^{2n}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. Let us now explain the key arguments in our proofs of these theorems. Structure of the proof ---------------------- Our main assumption about the symplectic structure of the singular space $S$ fulfilled in the assumptions of all the three theorems allows us to find some symplectic coordinates $(x',\xi')$ in $S^{\sigma \perp}$ and $(x'',\xi'')$ in $S$ such that the complex-valued quadratic form $q$ verifying (\[sm4\]) can be written as the sum of two quadratic forms with a tensorization of the variables $(x',\xi')$ and $(x'',\xi'')$, $$q=q|_S+q|_{S^{\sigma \perp}}, \ (x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n},$$ where the first quadratic form $q|_S$ is equal to $$q|_S=i \tilde{q}|_S,$$ with $\tilde{q}|_S$ a real-valued quadratic form; and where the second quadratic form $q|_{S^{\sigma \perp}}$ is a complex-valued quadratic form with a non-positive real part. This real part is not in general negative definite (unless the real part of $q$ is). However, it follows from the definition of the singular space $S$ that the average of the real part of the quadratic form $q|_{S^{\sigma \perp}}$ by the flow generated by the Hamilton vector field of its imaginary part, $H_{\textrm{Im}q|_{S^{\sigma \perp}}}$, $$\langle \textrm{Re }q|_{S^{\sigma \perp}}\rangle_T(X')=\frac{1}{2T}\int_{-T}^T{\textrm{Re }q|_{S^{\sigma \perp}}(e^{tH_{\textrm{Im}q|_{S^{\sigma \perp}}}}X')dt}, \ T>0, \ X'=(x',\xi'),$$ is negative definite. Studying the contraction semigroup $$\label{tom1}\inc e^{tq|_{S^{\sigma \perp}}^w}, \ t \geq 0, \num$$ generated by the operator $q|_{S^{\sigma \perp}}^w$, on the FBI-Bargmann transform side, we prove, using the averaging property just mentioned, that (\[tom1\]) is compact and strongly regularizing for every $t >0$. This compactness result is really the key point in our proofs of the three theorems, and their complete statements then follow from a small additional amount of work. Some examples {#examples} ------------- In this section, we prove that if a quadratic symbol $q$ is elliptic on its singular space then the singular space always has a symplectic structure. We also check that this property is fulfilled for the Kramers-Fokker-Plank operator with a quadratic potential. ### Partially elliptic quadratic differential operators Let us consider the case of quadratic differential operators whose Weyl symbols are elliptic on their singular spaces. Let $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ be a complex-valued quadratic form, which is elliptic on its singular space $S$, $$\label{sm12bis}\inc (x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0. \num$$ We want to prove that $$S=\Big(\bigcap_{j=0}^{2n-1}\textrm{Ker}\big[\textrm{Re }F(\textrm{Im }F)^j \big]\Big) \cap {\mathbb{R}}^{2n},$$ has a symplectic structure. This fact follows from some arguments similar to those used in [@mz] (Lemma 3). We can assume that $S \neq \{0\}$ since the space $\{0\}$ is obviously symplectic. Let us therefore consider $X_0 \in S \setminus \{0\}$. We define $$\label{2.3.101}\inc \left\lbrace \begin{array}{c} e_1=X_0\\ \displaystyle {\varepsilon}_1=-\frac{1}{\textrm{Im } q(X_0)}\textrm{Im } F X_0. \end{array} \right. \num$$ This is possible, since from (\[10\]) and (\[h3\]), we have $$\textrm{Re } q(X_0)=\sigma(X_0,\textrm{Re }F X_0)=0,$$ and the ellipticity of $q$ on $S$ implies that $$\textrm{Im } q(X_0) \neq 0,$$ as $X_0 \in S \setminus \{0\}$. By using the skew-symmetry of the Hamilton map $\textrm{Im }F$ (see (\[12\])), it follows that $$\sigma({\varepsilon}_1,e_1)=\sigma\big{(}- (\textrm{Im } q(X_0))^{-1} \textrm{Im } F X_0,X_0\big{)}=(\textrm{Im } q(X_0))^{-1}\sigma(X_0,\textrm{Im } F X_0)=1,$$ which shows that the system $(e_1,{\varepsilon}_1)$ is symplectic. We get from (\[h3\]) and (\[2.3.101\]) that $$\textrm{Vect}(e_1,{\varepsilon}_1) \subset S.$$ If $S=\textrm{Vect}(e_1,{\varepsilon}_1)$, the singular space $S$ is symplectic. If it is not the case, so that, $$S \neq \textrm{Vect}(e_1,{\varepsilon}_1),$$ we can continue our construction of a symplectic basis for $S$ by considering $$X_1 \in S \setminus \textrm{Vect}(e_1,{\varepsilon}_1)$$ and $$\label{hh3}\inc \tilde{X}_1=X_1+\sigma(X_1,{\varepsilon}_1)e_1-\sigma(X_1,e_1){\varepsilon}_1 \in S \setminus \textrm{Vect}(e_1,{\varepsilon}_1). \num$$ Let us set $$\label{2.3.107}\inc \left\lbrace \begin{array}{c} e_2=\tilde{X}_1\\ \displaystyle {\varepsilon}_2=-\frac{1}{\textrm{Im } q(\tilde{X}_1)}\big( \textrm{Im } F \tilde{X}_1+\sigma(\textrm{Im }F \tilde{X}_1,{\varepsilon}_1)e_1-\sigma(\textrm{Im }F \tilde{X}_1,e_1){\varepsilon}_1\big), \end{array} \right. \num$$ which is again possible according to (\[h3\]) and the assumption of ellipticity on $S$ since $$\textrm{Re } q(\tilde{X}_1)=\sigma(\tilde{X}_1,\textrm{Re }F \tilde{X}_1)=0,$$ because $\tilde{X}_1 \in S \setminus \{0\}$. Then, we can directly verify by using (\[hh3\]) and (\[2.3.107\]) that $(e_1,e_2,{\varepsilon}_1,{\varepsilon}_2)$ is a symplectic system. By using (\[h3\]) again, we get $$\textrm{Vect}(e_1,e_2,{\varepsilon}_1,{\varepsilon}_2) \subset S.$$ If $S=\textrm{Vect}(e_1,e_2,{\varepsilon}_1,{\varepsilon}_2)$, then $S$ is symplectic. If it is not the case, then $$S \neq \textrm{Vect}(e_1,e_2,{\varepsilon}_1,{\varepsilon}_2),$$ we can again iterate the preceding construction. After a finite number of such iterations, we obtain with this process a symplectic basis of $S$, proving its symplectic structure. Let us now consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part $$\textrm{Re }q \leq 0,$$ such that (\[sm12bis\]) is fulfilled and denote by $F$ its Hamilton map. We know from Proposition 4.4 in [@mehler] that the kernel $\textrm{Ker}(F+\lambda)$ is the complex conjugate of the kernel $\textrm{Ker}(F-\lambda)$ for every $\lambda \in {\mathbb{R}}$, and that the spaces $$\textrm{Ker}(F-\lambda) \oplus \textrm{Ker}(F+\lambda),$$ where $\lambda \in {\mathbb{R}}^*$, and $\textrm{Ker }F$, are the complexifications of their intersections with ${\mathbb{R}}^{2n}$. Let us set $$\label{venice1}\inc S_0=(\textrm{Ker }F) \cap {\mathbb{R}}^{2n} \num$$ and $$\label{venice2}\inc S_{\lambda}=\big(\textrm{Ker}(F-\lambda) \oplus \textrm{Ker}(F+\lambda)\big) \cap {\mathbb{R}}^{2n}, \num$$ for $\lambda \in {\mathbb{R}}^*$. Proposition 4.4 in [@mehler] also shows that $$\textrm{Re } F \ \textrm{Ker}(F \pm \lambda)=\{0\},$$ for all $\lambda \in {\mathbb{R}}$. This implies that $$\label{venice3}\inc (\textrm{Re } F)S_{\lambda}=\{0\} \textrm{ and } (\textrm{Im } F)S_{\lambda} \subset S_{\lambda}, \num$$ and proves in view of (\[h1bis\]) that for all $\lambda \in {\mathbb{R}}$, $$\label{venice4}\inc S_{\lambda} \subset S. \num$$ If $0 \in \sigma(F) \cap {\mathbb{R}}$, this would imply that $S_0 \neq 0$. Since from (\[venice1\]), $$q(X)=\sigma(X,FX)=0,$$ for all $X \in S_0$, the inclusion (\[venice4\]) would then contradict our assumption of ellipticity on the singular space (\[sm12bis\]). This proves that the set of real eigenvalues of the Hamilton map $F$ can be written as $$\label{venice5}\inc \sigma(F) \cap {\mathbb{R}}=\{\lambda_1,...,\lambda_r,-\lambda_1,...,-\lambda_r\}, \num$$ with $\lambda_j \neq 0$ and $\lambda_j \neq \pm \lambda_k$ if $j \neq k$. Let us now check that the spaces $S_{\lambda_j}$, $j=1,...,r$, are symplectic. Let $X_0$ be in $S_{\lambda_j}$ such that for all $Y \in S_{\lambda_j}$, $$\sigma(X_0,Y)=0.$$ It follows that for all $Y$ and $Z$ in $S_{\lambda_j}$, $$\sigma(X_0,Y+iZ)=0,$$ which induces that $$\forall X \in \textrm{Ker}(F-\lambda_j) \oplus \textrm{Ker}(F+\lambda_j), \ \sigma(X_0,X)=0,$$ because $ \textrm{Ker}(F-\lambda_j) \oplus \textrm{Ker}(F+\lambda_j)$ is a complexification of $S_{\lambda_j}$. On the other hand, since $X_0 \in S_{\lambda_j}$, we have $F X_0 \in \textrm{Ker}(F-\lambda_j) \oplus \textrm{Ker}(F+\lambda_j)$, which implies that $$q(X_0)=\sigma(X_0,F X_0)=0.$$ We then deduce from the ellipticity of $q$ on the singular space (\[sm12bis\]) and (\[venice4\]) that $X_0=0$, which proves the symplectic structure of the space $S_{\lambda_j}$. Let us now assume that there exists another real eigenvalue $\lambda_k$ of $F$ distinct from $\lambda_j$ and $-\lambda_j$. We already know that this eigenvalue $\lambda_k$ is necessarily non-zero. Let $$X \in \textrm{Ker}(F-{\varepsilon}_1 \lambda_j) \textrm{ and } Y \in \textrm{Ker}(F-{\varepsilon}_2 \lambda_k),$$ with ${\varepsilon}_1, {\varepsilon}_2 \in \{\pm 1\}$, we obtain from the skew-symmetry property of the Hamilton map $F$ with respect to $\sigma$ that $$\sigma(X,Y)=\sigma(X,{\varepsilon}_2^{-1} \lambda_k^{-1} FY)=-\frac{1}{{\varepsilon}_2 \lambda_k}\sigma(FX,Y)=-\frac{{\varepsilon}_1 \lambda_j}{{\varepsilon}_2 \lambda_k} \sigma(X,Y).$$ Since $$\left|\frac{{\varepsilon}_1 \lambda_j}{{\varepsilon}_2 \lambda_k} \right| \neq 1,$$ because $\lambda_j$ and $\lambda_k$ are real numbers such that $\lambda_k \not\in \{\lambda_j,-\lambda_j\}$, we finally deduce that $$\sigma(X,Y) =0,$$ which proves that the spaces $S_{\lambda_j}$ and $S_{\lambda_k}$ are symplectically orthogonal, and we get from (\[venice4\]) and (\[venice5\]) that $$\label{venice6}\inc S_{\lambda_1} \oplus^{\sigma \perp} S_{\lambda_2} \oplus^{\sigma \perp} ... \oplus^{\sigma \perp} S_{\lambda_r} \subset S. \num$$ Let us prove that the singular space is actually exactly equal to this direct sum of symplectic spaces. We recall that from (\[h3\]), $$(\textrm{Re }F) S=\{0\} \textrm{ and } (\textrm{Im }F) S \subset S.$$ Since $$q(X)=\sigma(X,FX)=i \sigma(X,\textrm{Im }FX), \ X \in S,$$ we deduce from (\[sm12bis\]) and the lemma 18.6.4 in [@hormander] that we can find new symplectic basis $(\tilde{e}_1,...,\tilde{e}_m,\tilde{{\varepsilon}}_1,...,\tilde{{\varepsilon}}_m)$ in the symplectic space $S$ such that $$\label{venice6bis}\inc q(X)=i {\varepsilon}\sum_{j=1}^m{\mu_j(\tilde{\xi}_j^2+\tilde{x}_j^2)}, \ X=\tilde{x}_1 \tilde{e}_1+...+ \tilde{x}_m \tilde{e}_m+\tilde{\xi}_1 \tilde{{\varepsilon}}_1+...+\tilde{\xi}_m \tilde{{\varepsilon}}_m, \num$$ where ${\varepsilon}\in \{\pm 1\}$ and $\mu_j>0$ for all $j=1,...,m$. Indeed, this is linked to the fact that a real-valued elliptic quadratic form must be positive definite or negative definite. By computing $F$ from (\[venice6bis\]), we get that $$\label{venice7}\inc FX_j=-{\varepsilon}\mu_j X_j \textrm{ and } F \tilde{X}_j= {\varepsilon}\mu_j \tilde{X}_j, \num$$ if $X_j=\tilde{e}_j+i \tilde{{\varepsilon}}_j$ and $\tilde{X}_j=\tilde{e}_j-i\tilde{{\varepsilon}}_j$ for all $j=1,...,m$. The identities (\[venice7\]) prove that the singular space is actually equal to the direct sum of the symplectic spaces $S_{\lambda_j}$ defined in (\[venice2\]), $$S=S_{\lambda_1} \oplus^{\sigma \perp} S_{\lambda_2} \oplus^{\sigma \perp} ... \oplus^{\sigma \perp} S_{\lambda_r}.$$ ### Kramers-Fokker-Planck operator with a quadratic potential {#planck} Let us consider the Kramers-Fokker-Planck operator [@HN], $$K=-\Delta_v+\frac{v^2}{4}-\frac{1}{2}+v.\partial_x-\big(\partial_xV(x)\big).\partial_v, \ (x,v) \in {\mathbb{R}}^{2},$$ with a quadratic potential $$V(x)=\frac{1}{2}ax^2, \ a \in {\mathbb{R}}^*.$$ We can write $$K=-q(x,v,\xi,\eta)^w-\frac{1}{2},$$ with $$q(x,v,\xi,\eta)=-\eta^2-\frac{1}{4}v^2-i(v \xi-a x \eta).$$ This symbol $q$ is a non-elliptic complex-valued quadratic form with a non-positive real part and a numerical range equal to the half-plane $$\Sigma(q)=\{z \in {\mathbb{C}}: \textrm{Re }z \leq 0\}.$$ We can directly check that its Hamilton map $F$ for which $$q(x,v,\xi,\eta)=\sigma\big((x,v,\xi,\eta),F(x,v,\xi,\eta) \big),$$ is given by $$F= \left( \begin{array}{cccc} 0 & -\frac{1}{2}i & 0 & 0 \\ \frac{1}{2}ai& 0 & 0 & -1 \\ 0 & 0 & 0 & - \frac{1}{2}ai \\ 0 & \frac{1}{4} & \frac{1}{2}i & 0 \end{array} \right),$$ and that the singular space $$S= \Big(\bigcap_{j=0}^{3}\textrm{Ker}\big[\textrm{Re }F(\textrm{Im }F)^j \big]\Big) \cap {\mathbb{R}}^{4},$$ is reduced to the trivial symplectic space $\{0\}$. [**Acknowledgment**]{}. The research of the first author is supported in part by the National Science Foundation under grant DMS–0653275 and the Alfred P. Sloan Research Fellowship. He would also like to thank Joe Viola for a stimulating discussion. Symplectic decomposition of the symbol {#sympl} ====================================== In this section, we explain how the main assumption about the symplectic structure of the singular space $S$ fulfilled in the hypotheses of all our three theorems allows us to tensor the variables in the symbol $q$ by writing it as a sum of two quadratic forms where the first one is purely imaginary-valued and where the second one verifies the averaging property of its real part by the flow defined by the Hamilton vector field of its imaginary part. Let us consider a complex-valued quadratic form $q$ verifying (\[sm4\]) and let us assume that the singular space $S$ defined in (\[h1bis\]) is symplectic. Let us recall that it is well the case when $q$ verifies (\[sm5\]). Then, we can find $\chi$, a real linear symplectic transformation of ${\mathbb{R}}^{2n}$, such that $$\label{sm12}\inc (q \circ \chi)(x,\xi)= q_1(x',\xi')+ iq_2(x'',\xi''), \ (x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n}, \num$$ where $q_1$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n'}$ with a non-positive real part $$\label{sm13}\inc \textrm{Re }q_1 \leq 0, \num$$ and $q_2$ is a real-valued quadratic form verifying the following properties: \[proposition1\] The two quadratic forms $q_1$ and $q_2$ satisfy the following properties: - For all $T>0$, the average of the real part of the quadratic form $q_1$ by the flow defined by the Hamilton vector field of $\emph{\textrm{Im }}q_1$, $$\langle\emph{\textrm{Re }}q_1\rangle_{T}(X')=\frac{1}{2T}\int_{-T}^{T}{\emph{\textrm{Re }}q_1(e^{tH_{\emph{\textrm{Im}}q_1}}X')dt}, \ X'=(x',\xi') \in {\mathbb{R}}^{2n'},$$ is negative definite. - The quadratic form $$\sum_{j=0}^{2n-1}{\emph{\textrm{Re }} q_1\big((\emph{\textrm{Im }} F_1)^j X'\big)}, \ X'=(x',\xi') \in {\mathbb{R}}^{2n'},$$ where $F_1$ stands for the Hamilton map of $q_1$, is negative definite. - If the symbol $q$ fulfills an additional assumption of ellipticity on $S$, $$\label{la}\inc (x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0,\num$$ then we can assume that $$q_2(x'',\xi'')={\varepsilon}\sum_{j=1}^{n''}{\lambda_j(\xi_j''^2+x_j''^2)},$$ where ${\varepsilon}\in \{\pm 1\}$ and $\lambda_j>0$ for all $j=1,...,n''$. To prove these results, we begin by considering $S^{\sigma \perp}$, the orthogonal complement of $S$ in ${\mathbb{R}}^{2n}$ with respect to the symplectic form and $F$ the Hamilton map of $q$. The space $S^{\sigma \perp}$ is symplectic because it is the case for $S$. Moreover, since according to (\[h3\]), $S$ is stable by the maps $\textrm{Re } F$ and $\textrm{Im } F$, its orthogonal complement also fulfills these properties. Indeed, let $X$ be in $S^{\sigma \perp}$. By using (\[h3\]) and the skew-symmetry of any Hamilton map with respect to $\sigma$, we get for all $Y \in S$, $$\sigma(Y,\textrm{Re } F X)+i \sigma(Y,\textrm{Im } F X)=-\sigma(\textrm{Re } F Y,X)-i \sigma(\textrm{Im } F Y,X)=0,$$ because $(\textrm{Re } F) Y \in S$ and $(\textrm{Im } F) Y \in S$. This induces that for all $Y \in S$, $$\sigma(Y,\textrm{Re } F X)=\sigma(Y,\textrm{Im } F X)=0,$$ and proves that $(\textrm{Re } F) X \in S^{\sigma \perp }$ and $(\textrm{Im } F) X \in S^{\sigma \perp }$. We can then write the phase space ${\mathbb{R}}^{2n}$ as a direct sum of two symplectically orthogonal real symplectic spaces stable by the maps $\textrm{Re } F$ and $\textrm{Im } F$, $$\label{2.3.89}\inc {\mathbb{R}}^{2n}= S_1 \oplus^{\sigma \perp} S_2, \ (\textrm{Re } F)S_j \subset S_j, \ (\textrm{Im } F)S_j \subset S_j, \num$$ for $j \in \{1,2\}$ with $$\label{h5}\inc S_1=S^{\sigma \perp} \textrm{ and } S_2= S. \num$$ Let us now consider a symplectic basis $(e_{1,j}, ...,e_{N_j,j},{\varepsilon}_{1,j},...,{\varepsilon}_{N_j,j})$ of $S_j$. By collecting these two bases, we get a symplectic basis of ${\mathbb{R}}^{2n}$, which allows by using the stability and the orthogonality properties of the spaces $S_j$ to obtain the following decomposition of $q$, $$\begin{aligned} q(x,\xi) = & \ \sigma\Big{(} \sum_{\substack{1 \leq j \leq 2, \\ 1 \leq k \leq N_j}}{(x_{k,j} e_{k,j}+\xi_{k,j} {\varepsilon}_{k,j})}, F\big{(} \sum_{\substack{1 \leq j \leq 2,\\ 1 \leq k \leq N_j}}{(x_{k,j} e_{k,j}+\xi_{k,j} {\varepsilon}_{k,j})}\big)\Big)\\ = & \ \sum_{1 \leq j \leq 2}\sigma \Big{(}\sum_{1 \leq k \leq N_j}{(x_{k,j} e_{k,j}+\xi_{k,j} {\varepsilon}_{k,j})},F\big{(} \sum_{1 \leq k \leq N_j}{(x_{k,j} e_{k,j}+\xi_{k,j} {\varepsilon}_{k,j})}\big)\Big).\end{aligned}$$ This implies that we can find symplectic coordinates $$(x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n},$$ where $(x',\xi')$ and $(x'',\xi'')$ are some symplectic coordinates in $S^{\sigma \perp}$ and $S$ respectively, such that $$\label{inf11.5}\inc q(x,\xi)=q_1(x',\xi')+q_2(x'',\xi''), \num$$ with $$\label{inf12}\inc q_1(x',\xi')=\sigma\big((x',\xi'),F|_{S^{\sigma \perp} }(x',\xi')\big) \num$$ and $$\label{inf13}\inc q_2(x'',\xi'')= \sigma\big((x'',\xi''),F|_{S}(x'',\xi'')\big). \num$$ Since from (\[h3\]), $$(\textrm{Re }F)S=\{0\},$$ the quadratic form $q_2$ is purely imaginary-valued and can be written as $$\label{sm14}\inc q_2=i \tilde{q}_2, \num$$ where $\tilde{q}_2$ is the real-valued quadratic form $$\label{sm15}\inc \tilde{q}_2(x'',\xi'')= \sigma\big((x'',\xi''),\textrm{Im }F|_{S}(x'',\xi'')\big). \num$$ When the additional assumption (\[la\]) is fulfilled, this quadratic form $\tilde{q}_2$ must be elliptic on ${\mathbb{R}}^{2n''}$. Since a real-valued elliptic quadratic form is necessarily a positive definite or negative definite quadratic form, we deduce from the lemma 18.6.4 in [@hormander] that we can find new symplectic coordinates $(x'',\xi'')$ in $S$ and ${\varepsilon}\in \{\pm 1\}$ such that $$\label{sm16}\inc \tilde{q}_2(x'',\xi'')={\varepsilon}\sum_{j=1}^{n''}{\lambda_j(\xi_j''^2+x_j''^2)}, \num$$ where $\lambda_j>0$ for all $j=1,...,n''$. This proves $(iii)$ in Proposition \[proposition1\]. Let us now study the properties of the quadratic form $q_1$. We denote by $F_1$ its Hamilton map $$\label{h8}\inc F_1=F|_{S^{\sigma \perp}}, \num$$ and define the following quadratic form $$\label{2.3.98}\inc r(X')=\sum_{j=0}^{2n-1}{\textrm{Re } q_1\big((\textrm{Im } F_1)^j X'\big)}, \ X'=(x',\xi') \in S^{\sigma \perp}. \num$$ Since from (\[sm4\]), (\[inf11.5\]) and (\[sm14\]), $\textrm{Re } q_1$ is a non-positive quadratic form, we already know that $r$ is a non-positive quadratic form. We now prove that $r$ is actually a negative definite quadratic form. Let us consider $X_0' \in S^{\sigma \perp}$ such that $$r(X_0')=0.$$ The non-positivity of the quadratic form $\textrm{Re } q_1$ induces that for all $j=0,...,2n-1$, $$\label{2.3.99}\inc \textrm{Re }q_1\big((\textrm{Im } F_1)^j X_0'\big)=0. \num$$ Let us denote by $\textrm{Re }q_1(X';Y')$ the polarized form associated to $\textrm{Re }q_1$. We deduce from the Cauchy-Schwarz inequality, (\[10\]) and (\[2.3.99\]) that for all $j=0,...,2n-1$ and $Y' \in S^{\sigma \perp}$, $$\begin{aligned} |\textrm{Re }q_1\big(Y';(\textrm{Im }F_1)^j X_0'\big)|^2= & \ |\sigma\big(Y',\textrm{Re }F_1 (\textrm{Im }F_1)^j X_0'\big)|^2 \\ \leq & \ [-\textrm{Re }q_1(Y')] [-\textrm{Re }q_1\big((\textrm{Im } F_1)^j X_0'\big) ]=0.\end{aligned}$$ It follows that for all $j=0,...,2n-1$ and $Y' \in S^{\sigma \perp}$, $$\sigma\big(Y',\textrm{Re }F_1 (\textrm{Im }F_1)^j X_0'\big)=0,$$ which implies that for all $j=0,...,2n-1$, $$\label{2.3.100}\inc \textrm{Re }F_1(\textrm{Im }F_1)^j X_0'=0, \num$$ because from (\[2.3.89\]), (\[h5\]) and (\[h8\]), $\textrm{Re }F_1(\textrm{Im }F_1)^j X_0' \in S^{\sigma \perp}$ and $S^{\sigma \perp}$ is a symplectic vector space. Since $X_0' \in S^{\sigma \perp}$, we deduce from (\[h1bis\]), (\[2.3.89\]), (\[h5\]), (\[h8\]) and (\[2.3.100\]) that $X_0' \in S \cap S^{\sigma \perp}=\{0\}$, which proves that $r$ is a negative definite quadratic form. This proves $(ii)$ in Proposition \[proposition1\]. *Remark.* According to the previous proof, let us notice that the property $(ii)$ implies that for all $X \in {\mathbb{R}}^{2n'}$, $X \neq 0$, there exists $j_0 \in \{0,...,2n-1\}$ such that $$\label{sm17}\inc \forall \ 0 \leq j \leq j_0-1, \ \textrm{Re }F_1(\textrm{Im }F_1)^j X=0, \ \textrm{Re }F_1(\textrm{Im }F_1)^{j_0} X \neq 0. \num$$ Let us now prove that for all $T>0$, the average of the real part of the quadratic form $q_1$ by the flow defined by the Hamilton vector field of $\textrm{Im }q_1$, $$\langle\textrm{Re }q_1\rangle_{T}(X')=\frac{1}{2T}\int_{-T}^{T}{\textrm{Re }q_1(e^{tH_{\textrm{Im}q_1}}X')dt},$$ is negative definite. Let us notice that this flow is globally defined since the symbol $\textrm{Im }q_1$ is quadratic. Let us consider $X_0'$ in ${\mathbb{R}}^{2n'}$ such that $$\label{sm18}\inc \langle\textrm{Re }q_1\rangle_{T}(X_0')=0.\num$$ Since $\textrm{Re }q_1$ is a non-positive quadratic form, it follows from (\[sm18\]) that $$\textrm{Re }q_1(e^{tH_{\textrm{Im}q_1}}X_0')=0,$$ for all $-T \leq t \leq T$. This implies in particular that for all $k \in {\mathbb{N}}$, $$\label{h12}\inc \frac{d^k}{dt^k}\big(\textrm{Re }q_1(e^{tH_{\textrm{Im}q_1}}X_0')\big)\big|_{t=0}=H_{\textrm{Im} q_1}^k\textrm{Re }q_1(X_0')=0.\num$$ If $X_0' \neq 0$, we deduce from $(ii)$ that there exists $j_0 \in \{0,...,2n-1\}$ such that $$\label{st2}\inc \forall \ 0 \leq j \leq j_0-1, \ \textrm{Re } q_1\big((\textrm{Im } F_1)^{j} X_0'\big)=0, \ \textrm{Re } q_1\big((\textrm{Im } F_1)^{j_0} X_0'\big)<0. \num$$ Let us check that it would imply that $$\label{st4}\inc H_{\textrm{Im} q_1}^{2j_0} \textrm{Re } q_1(X_0') \neq 0, \num$$ and contradict (\[h12\]). To prove (\[st4\]), we use some arguments already used in [@mz] together with the following lemma also proved in [@mz]. \[ll2\] If $q_1$ and $q_2$ are two complex-valued quadratic forms on ${\mathbb{R}}^{2n}$, then the Hamilton map associated to the complex-valued quadratic form defined by the Poisson bracket $$\{q_1,q_2\}=\frac{\partial q_1}{\partial \xi}.\frac{\partial q_2}{\partial x}-\frac{\partial q_1}{\partial x}.\frac{\partial q_2}{\partial \xi},$$ is $-2[F_1,F_2]$ where $[F_1,F_2]$ stands for the commutator of $F_1$ and $F_2$, the Hamilton maps of $q_1$ and $q_2$. We deduce from the previous lemma that the Hamilton map associated to the quadratic form $H_{\textrm{Im} q_1}^{2j_0} \textrm{Re } q_1$ is $$\label{st4.1}\inc 4^{j_0}[\textrm{Im }F_1,[\textrm{Im }F_1,[...,[\textrm{Im }F_1,\textrm{Re }F_1]...], \num$$ with exactly $2j_0$ terms $\textrm{Im }F_1$ appearing in the formula. We can write $$\begin{aligned} \label{st4.5} & \ 4^{j_0}[\textrm{Im }F_1,[\textrm{Im }F_1,[...,[\textrm{Im }F_1,\textrm{Re }F_1]...] \num \\ = & \ \sum_{j=0}^{2j_0}{(-1)^j c_j (\textrm{Im }F_1)^j \textrm{Re }F_1 (\textrm{Im }F_1)^{2j_0-j}},\end{aligned}$$ with $c_j >0$ for all $j=0,...,2j_0$. Indeed, by using the following identity $$[P,[P,Q]]=P^2Q-2PQP+QP^2,$$ we can prove by induction that for all $n \in {\mathbb{N}}^*$, there exist some positive constants $d_{n,j}$, $j=0,...,2n$ such that $$[P,[P,[...,[P,Q]...]=\sum_{j=0}^{2n}{(-1)^j d_{n,j} P^j Q P^{2n-j}},$$ if there are exactly $2n$ terms $P$ in the left-hand side of the previous identity. It follows from (\[st4.1\]) and (\[st4.5\]) that $$\begin{aligned} \label{st5} & \ H_{\textrm{Im} q_1}^{2j_0} \textrm{Re } q_1(X_0')= (-1)^{j_0} c_{j_0}\sigma\big(X_0',(\textrm{Im }F_1)^{j_0}\textrm{Re }F_1(\textrm{Im }F_1)^{j_0}X_0' \big) \num \\ + & \ \sum_{j=0}^{j_0-1}(-1)^j c_j \sigma\big(X_0',(\textrm{Im }F_1)^j\textrm{Re }F_1(\textrm{Im }F_1)^{2j_0-j}X_0' \big)\\ + & \ \sum_{j=0}^{j_0-1}(-1)^{2j_0-j}c_{2j_0-j} \sigma\big(X_0',(\textrm{Im }F_1)^{2j_0-j}\textrm{Re }F_1(\textrm{Im }F_1)^{j}X_0' \big).\end{aligned}$$ Now on the one hand, $$\begin{aligned} \sigma\big(X_0',(\textrm{Im }F_1)^{j_0}\textrm{Re }F_1(\textrm{Im }F_1)^{j_0}X_0' \big)= & \ (-1)^{j_0}\sigma\big((\textrm{Im }F_1)^{j_0}X_0',\textrm{Re }F_1(\textrm{Im }F_1)^{j_0}X_0' \big)\\ = & \ (-1)^{j_0}\textrm{Re }q_1\big((\textrm{Im }F_1)^{j_0}X_0'\big),\end{aligned}$$ by the skew-symmetry of the Hamilton map $\textrm{Im }F_1$. On the other hand, using (\[10\]), (\[st2\]), and the Cauchy-Schwarz inequality, we get $$\begin{aligned} & \ |\sigma\big(X_0',(\textrm{Im }F_1)^j\textrm{Re }F_1(\textrm{Im }F_1)^{2j_0-j}X_0' \big)|\\ =& \ |\sigma\big((\textrm{Im }F_1)^jX_0',\textrm{Re }F_1(\textrm{Im }F_1)^{2j_0-j}X_0' \big)|\\ = & \ -\textrm{Re }q_1\big((\textrm{Im }F_1)^jX_0';(\textrm{Im }F_1)^{2j_0-j}X_0'\big) \\ \leq & \ [-\textrm{Re }q_1((\textrm{Im }F_1)^jX_0')]^{\frac{1}{2}} [-\textrm{Re }q_1((\textrm{Im }F_1)^{2j_0-j}X_0')]^{\frac{1}{2}}=0\end{aligned}$$ and $$\begin{aligned} & \ |\sigma\big(X_0',(\textrm{Im }F_1)^{2j_0-j}\textrm{Re }F_1(\textrm{Im }F_1)^{j}X_0' \big)|\\ =& \ |\sigma\big((\textrm{Im }F_1)^{2j_0-j}X_0',\textrm{Re }F_1(\textrm{Im }F_1)^{j}X_0' \big)|\\ = & \ -\textrm{Re }q_1\big((\textrm{Im }F_1)^{2j_0-j}X_0';(\textrm{Im }F_1)^{j}X_0'\big) \\ \leq & \ [-\textrm{Re }q_1((\textrm{Im }F_1)^{2j_0-j}X_0')]^{\frac{1}{2}} [-\textrm{Re }q_1((\textrm{Im }F_1)^{j}X_0')]^{\frac{1}{2}}=0\end{aligned}$$ if $j=0,...,j_0-1$. It follows from (\[st2\]) and (\[st5\]) that $$H_{\textrm{Im} q_1}^{2j_0} \textrm{Re } q_1(X_0')=c_{j_0}\textrm{Re }q_1\big((\textrm{Im }F_1)^{j_0}X_0'\big) < 0,$$ because $c_{j_0} >0$. This proves (\[st4\]) and ends the proof $(i)$. $\Box$ *Remark.* Let us notice that we have actually proved that the symbol $q_1$ has a finite order $\tau$, $$\label{st1}\inc 1 \leq \tau \leq 4n-2, \num$$ in every point of the set $q_1({\mathbb{R}}^{2n'}) \setminus \{0\}$. We recall that the order $k(x_0,\xi_0)$ of a symbol $p(x,\xi)$ at a point $(x_0,\xi_0) \in {\mathbb{R}}^{2n}$ (see section 27.2, chapter 27 in [@hormander]) is the element of ${\mathbb{N}}\cup \{+\infty\}$ defined by $$\label{t10.5}\inc k(x_0,\xi_0)=\sup\big{\{}j \in \mathbb{Z} : p_I(x_0,\xi_0)=0, \ \forall\ 1 \leq |I| \leq j\big{\}}, \num$$ where $I=(i_1,i_2,...,i_k) \in \{1,2\}^k$, $|I|=k$ and $p_I$ stands for the iterated Poisson brackets $$p_{I}=H_{p_{i_1}}H_{p_{i_2}}...H_{p_{i_{k-1}}}p_{i_k},$$ where $p_1$ and $p_2$ are respectively the real and the imaginary part of the symbol $p$, $p=p_1+ip_2$. The order of a symbol $q$ at a point $z$ is then defined as the maximal order of the symbol $p=q-z$ at every point $(x_0,\xi_0) \in {\mathbb{R}}^{2n}$ verifying $$p(x_0,\xi_0)=q(x_0,\xi_0)-z=0.$$ Proofs of the main results ========================== Heat semigroup smoothing for non-elliptic quadratic operators {#compact} ------------------------------------------------------------- In this section, we prove Theorem \[theorem1\]. Let us consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part $$\textrm{Re }q \leq 0,$$ such that its singular space $S$ has a symplectic structure. We recall that this assumption is actually fulfilled in the assumptions of Theorems \[theorem1\], \[theorem2\] and \[theorem3\] since this singular space is always symplectic when the symbol is elliptic on $S$. We can then use the symplectic decomposition of the symbol obtained in section \[sympl\]. We deduce from (\[sm12\]) and (\[sm13\]) that there exists $\chi$, a real linear symplectic transformation of ${\mathbb{R}}^{2n}$, such that $$\label{inv1}\inc (q \circ \chi)(x,\xi)= q_1(x',\xi')+ iq_2(x'',\xi''), \ (x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n}, \num$$ where $q_1$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n'}$ with a non-positive real part $$\label{inv2}\inc \textrm{Re }q_1 \leq 0, \num$$ and $q_2$ is a real-valued quadratic form verifying the properties stated in Proposition \[proposition1\]. The key point in our proof of Theorems \[theorem1\], \[theorem2\] and \[theorem3\] is to prove the following proposition. \[sperp\] If $n' \geq 1$, then the spectrum of the quadratic differential operator $q_1(x',\xi')^w$ is only composed of eigenvalues with finite multiplicity $$\sigma\big{(}q_1(x',\xi')^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F_1), \\ \emph{\textrm{Re}}(-i \lambda)<0} } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\},$$ where $F_1$ is the Hamilton map associated to the quadratic form $q_1$ and $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F_1$ in ${\mathbb{C}}^{2n'}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. Moreover, the operator $q_1(x',\xi')^w$ generates a contraction semigroup such that $$e^{tq_1(x',\xi')^w}u \in \mathcal{S}({\mathbb{R}}^{n'}),$$ for any $t>0$ and $u \in L^2({\mathbb{R}}^{n'})$. [*Remark.*]{} It will be clear from the proof that Proposition 3.1.1 extends to the vector-valued case, so that if ${\mathcal H}$ is a complex Hilbert space and $u \in L^2({\mathbb{R}}^{n'};{\mathcal H})$ then for any $t>0$ we have $e^{tq_1(x',\xi')^w}u \in \mathcal{S}({\mathbb{R}}^{n'};{\mathcal H})$. Theorem \[theorem1\] directly follows from Proposition 3.1.1 together with the preceding remark. Indeed, by using the symplectic invariance of the Weyl quantization given by the theorem 18.5.9 in [@hormander], we can find a metaplectic operator $U$, which is a unitary transformation on ${L^2({\mathbb{R}}^n)}$ and an automorphism of $\mathcal{S}({\mathbb{R}}^n)$ such that $$\label{inv2.5}\inc (q \circ \chi)(x,\xi)^w=U^{-1} q(x,\xi)^w U. \num$$ This implies at the level of the generated semigroups that $$\label{inv3}\inc e^{t(q \circ \chi)(x,\xi)^w}=U^{-1} e^{t q(x,\xi)^w} U, \ t \geq 0. \num$$ Since from the tensorization of the variables (\[inv1\]), $$e^{t(q \circ \chi)(x,\xi)^w}=e^{tq_1(x',\xi')^w} e^{it q_2(x'',\xi'')^w},$$ we directly deduce from (\[inf12\]), (\[inv3\]), Proposition \[sperp\] together with the following remark, and the symplectic invariance of the Weyl quantization that if $(x',\xi')$ are some symplectic coordinates on the symplectic space $S^{\sigma \perp}$ then for all $t >0$, $N \in {\mathbb{N}}$ and $u \in L^2({\mathbb{R}}^n)=L^2({\mathbb{R}}^{n'}; L^2({\mathbb{R}}^{n''}))$, we have $$\big((1+|x'|^2+|\xi'|^2)^N\big)^we^{t q(x,\xi)^w}u \in L^2({\mathbb{R}}^n),$$ which proves Theorem \[theorem1\]. Let us now prove Proposition \[sperp\]. For convenience, we drop the index and we consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part $$\label{kps2}\inc \textrm{Re }q \leq 0, \num$$ such that for all $T>0$, the average of the real part of the quadratic form $q$ by the flow defined by the Hamilton vector field of $\textrm{Im }q$, $$\label{inv5}\inc \langle\textrm{Re }q\rangle_{T}(X)=\frac{1}{2T}\int_{-T}^{T}{\textrm{Re }q(e^{tH_{\textrm{Im}q}}X)dt}, \ X=(x,\xi) \in {\mathbb{R}}^{2n}, \num$$ is negative definite. We also know from (\[sm17\]) that for all $X \in {\mathbb{R}}^{2n}$, $X \neq 0$, there exists $j_0 \in {\mathbb{N}}$ verifying $$\label{inv6}\inc \forall \ 0 \leq j \leq j_0-1, \ \textrm{Re }F(\textrm{Im }F)^j X=0, \ \textrm{Re }F(\textrm{Im }F)^{j_0} X \neq 0, \num$$ if $F$ stands for the Hamilton map of the quadratic form $q$. Let us denote by $$\inc \label{eq3} Q=q(x,\xi)^w,\num$$ the quadratic differential operator defined by the Weyl quantization of the symbol $q$. When proving Proposition \[sperp\], we shall work with the metaplectic FBI-Bargmann transform $$\inc \label{eq4} T u(x) = C \int_{{\mathbb{R}}^n} e^{i\varphi(x,y)} u(y)\,dy,\ x\in {\mathbb{C}}^n, \ C>0,\num$$ where we may choose $$\varphi(x,y)=\frac{i}{2} (x-y)^2,$$ as in the standard Bargmann transform. Other quadratic phase functions $\varphi$ such that ${\textrm{Im }}\varphi''_{y y}>0$ and $\det \varphi''_{x y}\neq 0$, are also possible (see section 1 of [@Sj95]). It is well known that for a suitable choice of $C>0$, $T$ defines a unitary transformation $$T: L^2({\mathbb{R}}^n)\rightarrow H_{\Phi_0}({\mathbb{C}}^n),$$ where $$\inc \label{eq4.1} H_{\Phi_0}({\mathbb{C}}^n)={\rm Hol}({\mathbb{C}}^n)\cap L^2\big({\mathbb{C}}^n, e^{-2\Phi_0(x)}L(dx)\big), \num$$ with $$\Phi_0(x)=\sup_{y \in {\mathbb{R}}^n} -{\textrm{Im }}\varphi(x,y)=\frac{1}{2} \left({\textrm{Im }}x\right)^2,$$ and $L(dx)$ being the Lebesgue measure in ${\mathbb{C}}^n$. [*Remark.*]{} Let us recall (see, e.g. section 3 of [@SjZw]) that the same definitions apply in the vector-valued case, so that we have a unitary operator $$T: L^2({\mathbb{R}}^n; {\mathcal H})\rightarrow H_{\Phi_0}({\mathbb{C}}^n; {\mathcal H}),$$ where ${\mathcal H}$ is a complex Hilbert space. We recall next from [@Sj95] that $$\inc \label{eq5} TQu=Q_0Tu,\ u\in \mathcal{S}({\mathbb{R}}^n), \num$$ where $Q_0$ is a quadratic differential operator on ${\mathbb{C}}^n$ whose Weyl symbol $q_0$ satisfies $$\inc \label{inv7} q_0\circ \kappa_T = q. \num$$ Here $$\inc \label{eq5.1} \kappa_T: {\mathbb{C}}^{2n}\ni \big(y,-\varphi'_y(x,y)\big) \mapsto \big(x,\varphi'_x(x,y)\big)\in {\mathbb{C}}^{2n}, \num$$ is the complex linear canonical transformation associated to $T$. From [@Sj95], we recall next that if we define $$\inc \label{inv8} \Lambda_{\Phi_0}=\Big\{\Big(x,\frac{2}{i}\frac{\partial \Phi_0}{\partial x}(x)\Big): x\in {\mathbb{C}}^n \Big\}, \num$$ then we have $$\inc \label{inv9} \Lambda_{\Phi_0}=\kappa_T({\mathbb{R}}^{2n}). \num$$ When $$\sigma = \sum_{j=1}^n d\xi_j \wedge dx_j,$$ is the complex symplectic (2,0)-form on ${\mathbb{C}}^{2n}={\mathbb{C}}^n_x\times {\mathbb{C}}^n_{\xi}$, then the restriction $\sigma_{\Lambda_{\Phi_0}}$ of $\sigma$ to $\Lambda_{\Phi_0}$ is real and nondegenerate. The map $\kappa_T$ in (\[eq5.1\]) can therefore be viewed as a canonical transformation between the real symplectic spaces ${\mathbb{R}}^{2n}$ and $\Lambda_{\Phi_0}$. Continuing to follow [@Sj95], let us recall next that when realizing $Q_0$ as an unbounded operator on $H_{\Phi_0}({\mathbb{C}}^n)$, we may use first the contour integral representation $$Q_0 u(x) = \frac{1}{(2\pi)^n}\int_{\theta=\frac{2}{i}\frac{\partial \Phi_0}{\partial x}\left(\frac{x+y}{2}\right)} e^{i(x-y)\cdot \theta} q_0\Big(\frac{x+y}{2},\theta\Big)u(y)\,dy\,d\theta,$$ and then, using that the symbol $q_0$ is holomorphic, by a contour deformation we obtain the following formula for $Q_0$ as an unbounded operator on $H_{\Phi_0}({\mathbb{C}}^n)$, $$\inc \label{eq5.5} Q_0 u(x) = \frac{1}{(2\pi)^n}\int_{\theta=\frac{2}{i}\frac{\partial \Phi_0}{\partial x}\left(\frac{x+y}{2}\right)+it\overline{(x-y)}} e^{i(x-y)\cdot \theta} q_0\Big(\frac{x+y}{2},\theta\Big)u(y)\,dy\,d\theta, \num$$ for any $t>0$. We shall now discuss certain IR-deformations of the real phase space ${\mathbb{R}}^{2n}$, where the averaging procedure along the flow defined by the Hamilton vector field of $\textrm{Im }q$ (see (\[inv5\])) plays an important rôle. To that end, let $G=G_T$ be a real-valued quadratic form on ${\mathbb{R}}^{2n}$ such that $$\inc \label{eq6} H_{\textrm{Im} q} G = -{\textrm{Re }}q+\langle{{\textrm{Re }}q}\rangle_T. \num$$ As in [@HeHiSj], we solve (\[eq6\]) by setting $$\inc \label{eq7} G(X)=\int_{{\mathbb{R}}} k_T(t) {\textrm{Re }}q(e^{tH_{\textrm{Im} q}}X)dt,\num$$ where $k_T(t)=k(t/2T)$ and $k\in C({\mathbb{R}}\backslash\{0\})$ is the odd function given by $$k(t)=0 \textrm{ for } |t| \geq \frac{1}{2} \textrm{ and } k'(t)=-1 \textrm{ for } 0<|t| < \frac{1}{2}.$$ Let us notice that $k$ and $k_T$ have a jump of size $1$ at the origin. Associated with $G$ there is a linear IR-manifold, defined for $0\leq {\varepsilon}\leq {\varepsilon}_0$, with ${\varepsilon}_0>0$ small enough, $$\inc \label{eq8} \Lambda_{{\varepsilon}G}=e^{i{\varepsilon}H_G}({\mathbb{R}}^{2n}) \subset {\mathbb{C}}^{2n},\num$$ where $e^{i{\varepsilon}H_G}$ stands for the flow generated by the linear Hamilton vector field $i{\varepsilon}H_G$ taken at the time 1. It is then well-known and easily checked (see, for instance, sections 3 and 5 in [@HeSjSt]), that $$\inc \label{eq9} \kappa_T(\Lambda_{{\varepsilon}G})=\Lambda_{\tilde{\Phi}_{{\varepsilon}}}:=\Big\{\Big(x,\frac{2}{i}\frac{\partial \tilde{\Phi}_{{\varepsilon}}}{\partial x}(x)\Big) : x\in {\mathbb{C}}^n\Big\},\num$$ where $\tilde{\Phi}_{{\varepsilon}}$ is a strictly plurisubharmonic quadratic form on ${\mathbb{C}}^n$, such that $$\inc\label{lin1} \tilde{\Phi}_{{\varepsilon}}(x)=\Phi_0(x)+{\varepsilon}G({\textrm{Re }}x, -{\textrm{Im }}x)+\mathcal{O}({\varepsilon}^2\abs{x}^2). \num$$ Associated with the function $\tilde{\Phi}_{{\varepsilon}}$ is the weighted space of holomorphic functions $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$ defined as in (\[eq4.1\]). The operator $Q_0$ can also be defined as an unbounded operator $$Q_0: H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n) \rightarrow H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n),$$ if we make a new contour deformation in (\[eq5.5\]) and set $$\inc \label{eq10} Q_0 u(x) = \frac{1}{(2\pi)^n}\int_{\theta=\frac{2}{i}\frac{\partial \tilde{\Phi}_{{\varepsilon}}}{\partial x}\left(\frac{x+y}{2}\right)+it\overline{(x-y)}} e^{i(x-y)\cdot \theta} q_0\Big(\frac{x+y}{2},\theta\Big)u(y)\,dy\,d\theta,\num$$ for any $t>0$. By coming back to the real side by the FBI-Bargmann transform, the operator $Q_0$ can be viewed as an unbounded operator on $L^2({\mathbb{R}}^n)$ with the Weyl symbol $$\inc\label{esk1} \widetilde{q}(X)=q\big(e^{i {\varepsilon}H_G}X\big), \num$$ and here the real part of this expression is easily seen to be equal to $$\textrm{Re }\widetilde{q}(X)=\textrm{Re }q(X) +{\varepsilon}H_{\textrm{Im}q} G(X)+\mathcal{O}({\varepsilon}^2\abs{X}^2).$$ It follows therefore from (\[kps2\]), (\[eq6\]) and the assumption that the quadratic form (\[inv5\]) is negative definite, that $$\inc \label{eq10.1} -\textrm{Re } \widetilde{q}(X)\geq \frac{{\varepsilon}}{C}\abs{X}^2, \ C>1,\ X \in {\mathbb{R}}^{2n}, \num$$ for $0<{\varepsilon}\ll 1$. We may therefore apply Theorem 3.5 of [@sjostrand] to the operator $Q_0$ viewed as an unbounded operator on $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$. \[spec\] Let us consider $Q_0$ as an unbounded operator on $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$, for $0<{\varepsilon}\leq {\varepsilon}_0$, with ${\varepsilon}_0>0$ sufficiently small. The spectrum of the operator $Q_0$ is only composed of eigenvalues with finite multiplicity $$\inc \label{eq11} \sigma(Q_0)=\Big\{ \sum_{\substack{\lambda\in \sigma(F), \\ \emph{\textrm{Re}}(-i\lambda)<0}} \left(r_{\lambda}+2 k_{\lambda}\right)(-i\lambda) : \ k_{\lambda}\in {\mathbb{N}}\Big\},\num$$ where $F$ is the Hamilton map associated to the quadratic form $q$ and $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F$ in ${\mathbb{C}}^{2n}$ belonging to the eigenvalue $\lambda\in {\mathbb{C}}$. To get the statement of this lemma, it suffices to combine Theorem 3.5 of [@sjostrand] together with the observation that the Hamilton maps $F$ and $\widetilde{F}$ of the quadratic forms $q$ and $\widetilde{q}$, respectively, are isospectral since from (\[esk1\]) the symbols $q$ and $\widetilde{q}$ are related by a canonical transformation. Having determined the spectrum of $Q_0$ in the weighted space $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$, $0<{\varepsilon}\ll 1$, we now come to the proof of Proposition \[sperp\]. In doing so, we shall study the spectral properties of the holomorphic quadratic differential operator $Q_0$ acting on $H_{\Phi_0}({\mathbb{C}}^n)$. We shall consider the heat evolution equation associated to the operator $Q_0$. Let us notice explicitly that we got this idea by studying Remark 11.7 in [@HeSjSt], and indeed, the following argument can be seen as a natural continuation of some ideas sketched in that remark. Using Fourier integral operators with quadratic phase in the complex domain, we may describe the heat semigroup $e^{tQ_0}$ for $0\leq t \leq t_0$, when $t_0>0$ is small enough. More precisely, we are interested in solving $$\left\lbrace \begin{array}{c} \displaystyle \frac{\partial u}{\partial t}(t,x)-Q_0 u(t,x)=0 \\ u(t,\textrm{\textperiodcentered})|_{t=0}=u_0 \in H_{\Phi_0}({\mathbb{C}}^n). \end{array} \right.$$ Let $\varphi(t,x,\eta)$ be the quadratic form in the variables $x$, $\eta$, depending smoothly on $t$, $0\leq t\leq t_0\ll 1$, and solving the Hamilton-Jacobi equation $$\left\lbrace \begin{array}{c} \displaystyle i\frac{\partial \varphi}{\partial t}(t,x,\eta)-q_0\Big(x,\frac{\partial \varphi}{\partial x}(t,x,\eta)\Big)=0 \\ \varphi(t,x,\eta)|_{t=0}=x\cdot \eta. \end{array} \right.$$ We know that for $0\leq t \leq t_0 \ll 1$, $\varphi(t,x,\eta)$ can be obtained as a generating function of the complex canonical transformation $$e^{itH_{q_0}} : \big(\varphi'_{\eta}(t,x,\eta),\eta\big)\mapsto \big(x,\varphi'_x(t,x,\eta)\big).$$ Then for $t\geq 0$ small enough, the operator $e^{tQ_0}$ acting on $H_{\Phi_0}({\mathbb{C}}^n)$ has the form $$e^{tQ_0} u = \frac{1}{(2\pi)^n}\int_{\Gamma_x} e^{i(\varphi(t,x,\eta)-y\cdot \eta)} a(t,x,y,\eta) u(y) dy\, d\eta,$$ where $a(t,x,y,\eta)$ is a suitable amplitude which we need not specify here, and, following the general theory of [@Sj82], we take $\Gamma_x$ to be a suitable contour passing through the critical point of the function $$(y,\eta)\mapsto -\textrm{Im}\big(\varphi(t,x,\eta)-y\cdot\eta\big)+\Phi_0(y).$$ We then know from the general theory that the operator $e^{tQ_0}$ is bounded $$\inc \label{invb1} e^{tQ_0}: H_{\Phi_0}({\mathbb{C}}^n) \rightarrow H_{\Phi_t}({\mathbb{C}}^n),\num$$ where $\Phi_t$ is a strictly plurisubharmonic quadratic form on ${\mathbb{C}}^n$, depending smoothly on $t$, such that if $$\inc\label{kps40} \Lambda_{\Phi_t}=\Big\{\Big(x,\frac{2}{i}\frac{\partial \Phi_t}{\partial x}(x)\Big) : x\in {\mathbb{C}}^n \Big\}, \num$$ then $$\inc \label{eq20} \Lambda_{\Phi_t} = {\mbox{\rm exp\,}}\big(t\widehat{H_{-\frac{1}{i}q_0}}\big)\left(\Lambda_{\Phi_0}\right). \num$$ Here, when $f$ is a holomorphic function on ${\mathbb{C}}^{2n}={\mathbb{C}}^n_x\times {\mathbb{C}}^n_{\xi}$, $H_f$ is the standard Hamilton field of $f$, of type (1,0), given by the usual formula $$H_f = \sum_{j=1}^n \Big(\frac{\partial f}{\partial \xi_j}\frac{\partial}{\partial x_j}-\frac{\partial f}{\partial x_j}\frac{\partial}{\partial \xi_j}\Big),$$ and $\widehat{H_f}=H_f+\overline{H_f}$ is the corresponding real vector field. It follows from the classical Hamilton-Jacobi theory applied with respect to the real symplectic form ${\textrm{Im }}\sigma$, where $$\sigma=\sum_{j=1}^n d\xi_j \wedge dx_j,$$ is the complex symplectic (2,0)-form on ${\mathbb{C}}^{2n}$, that the quadratic form $\Phi(t,x)=\Phi_t(x)$ introduced in (\[invb1\]), satisfies the eikonal equation $$\inc \label{eikonal} \left\lbrace \begin{array}{c} \displaystyle \frac{\partial \Phi}{\partial t}(t,x)- {\textrm{Re }}\Big[q_0\Big(x,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x)\Big)\Big]=0 \\ \Phi(t,\textrm{\textperiodcentered})|_{t=0}=\Phi_0. \end{array} \right. \num$$ See also [@SjHokk] for a detailed (and much more general) discussion of this point. Instrumental in the proof of Proposition \[sperp\] is the following result. \[phase\] For each $T_0>0$ small enough, there exists $\alpha=\alpha(T_0)>0$ such that $$\inc \label{eq20.1} \Phi_{T_0}(x)\leq \Phi_0(x) - \alpha \abs{x}^2,\ x\in {\mathbb{C}}^n. \num$$ Once Lemma \[phase\] has been established, it is easy to finish the proof of the first result in Proposition \[sperp\]. Indeed, elementary arguments together with (\[eq4.1\]) and (\[eq20.1\]) show that the natural embedding $H_{\Phi_t}({\mathbb{C}}^n) \rightarrow H_{\Phi_0}({\mathbb{C}}^n)$ is compact for $t>0$ small, and hence by using the semigroup property, we deduce that the semigroup $$\inc\label{kps41} e^{tQ_0}: H_{\Phi_0}({\mathbb{C}}^n) \rightarrow H_{\Phi_0}({\mathbb{C}}^n),\num$$ is compact for each $t>0$. An application of Theorem 2.20 in [@davies] then shows that the spectrum of the operator $Q_0$ acting on $H_{\Phi_0}({\mathbb{C}}^n)$ consists of a countable discrete set of eigenvalues each of finite multiplicity. When deriving the explicit description of the spectrum of $Q_0$ on $H_{\Phi_0}({\mathbb{C}}^n)$, we argue in the following way. Let us assume that $\lambda \in {\mathbb{C}}$ is an eigenvalue of $Q_0$ on $H_{\Phi_0}({\mathbb{C}}^n)$ and let $u_0 \in H_{\Phi_0}({\mathbb{C}}^n)$ be a corresponding eigenvector, $$Q_0 u_0 =\lambda u_0.$$ We deduce from (\[invb1\]) that $$e^{tQ_0} u_0 \in H_{\Phi_t}({\mathbb{C}}^n),$$ and since $$e^{tQ_0} u_0 = e^{t\lambda} u_0,$$ it follows from Lemma \[phase\] that $$u_0 \in H_{\Phi_0-\delta\abs{x}^2}({\mathbb{C}}^n),$$ for some $\delta>0$. In particular, we obtain from (\[lin1\]) that $u_0\in H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$ for ${\varepsilon}>0$ small enough, and hence that $\lambda$ is in the spectrum of the operator $Q_0$ acting on $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$, which has been described in Lemma \[spec\]. On the other hand, if $\lambda$ is in the spectrum of $Q_0$ acting on $H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$ and $u_0 \in H_{\tilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$ is a corresponding eigenvector, with ${\varepsilon}>0$ sufficiently small, then we have $$e^{tQ_0}u_0=e^{t\lambda}u_0\in H_{\widetilde{\Phi}_{{\varepsilon},t}}({\mathbb{C}}^n),$$ where $\widetilde{\Phi}_{{\varepsilon},t}$ is a quadratic form on ${\mathbb{C}}^n$ depending smoothly on $t\geq 0$ and ${\varepsilon}\geq 0$, for ${\varepsilon}$ sufficiently small, which satisfies the eikonal equation (\[eikonal\]) along with the initial condition $$\widetilde{\Phi}_{{\varepsilon},t}(x)|_{t=0}=\widetilde{\Phi}_{{\varepsilon}}(x).$$ It follows from (\[lin1\]) and (\[eikonal\]) that $$\widetilde{\Phi}_{{\varepsilon},t}(x)=\Phi_t(x)+\mathcal{O}({\varepsilon}\abs{x}^2),$$ where the implicit constant is uniform in $0\leq t\leq t_0$, for $t_0>0$ small enough. By taking $T_0>0$ small enough but fixed such that $0<T_0 \leq t_0$ and (\[eq20.1\]) holds, we get $$u_0= e^{-T_0\lambda} e^{T_0 Q} u_0 \in H_{\tilde{\Phi}_{{\varepsilon},T_0}}({\mathbb{C}}^n)= H_{\Phi_{T_0}+\mathcal{O}({\varepsilon}\abs{x}^2)}({\mathbb{C}}^n).$$ In view of (\[eq20.1\]), we can choose ${\varepsilon}_0>0$ small enough such that for all $0<{\varepsilon}\leq {\varepsilon}_0$, $$u_0 \in H_{\Phi_0-\tilde{\delta}|x|^2}({\mathbb{C}}^n) \subset H_{\Phi_0}({\mathbb{C}}^n),$$ where $\tilde{\delta}$ is a positive constant. It follows that $\lambda$ is also in the spectrum of the operator $Q_0$ acting on $H_{\Phi_0}({\mathbb{C}}^n)$. Altogether, this shows that the spectrum of $Q_0$ acting on $H_{\Phi_0}({\mathbb{C}}^n)$ is equal to the spectrum of $Q_0$ acting on $H_{\widetilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$, for ${\varepsilon}>0$ sufficiently small, and furthermore, that the algebraic multiplicities agree. We have therefore identified the spectrum of $Q_0$ on $H_{\Phi_0}({\mathbb{C}}^n)$ and also the spectrum of the operator $Q$ on $L^2({\mathbb{R}}^n)$ by coming back to the real side. [*Remark*]{}. In the argument above we have worked with the eigenfunctions of the operator $Q_0$ acting on $H_{\widetilde{\Phi}_{{\varepsilon}}}({\mathbb{C}}^n)$ for some sufficiently small but fixed ${\varepsilon}>0$. It may be interesting to notice that the (generalized) eigenfunctions of the operator $Q_0$ do not depend on ${\varepsilon}>0$, for $0<{\varepsilon}\leq {\varepsilon}_0$, with ${\varepsilon}_0>0$ small enough. See also Remark 11.7 in [@HeSjSt]. While refraining from providing a detailed proof of this statement, let us mention that its validity relies crucially upon the fact that the quadratic symbol $q_0$ is elliptic on $\Lambda_{\widetilde{\Phi}_{{\varepsilon}}}$, $0<{\varepsilon}\leq {\varepsilon}_0$. In the terminology of [@Sj82], this is an instance of the principle of non-characteristic deformations. We shall now prove Lemma \[phase\]. Integrating (\[eikonal\]) for $t=0$ to $t=T$, we obtain $$\Phi_T(x)-\Phi_0(x)=\int_0^{T} {\textrm{Re }}\Big[q_0\Big(x,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x)\Big)\Big]\,dt,\ T>0.$$ Here the integral in the right hand side is a real-valued quadratic form on ${\mathbb{C}}^n$, and Lemma \[phase\] is implied by the claim that it is negative definite. When proving the claim, we shall use the following general relation, explained for instance in [@MelinSj], $$\inc \label{eq21} \widehat{H_{if}} = H^{-{\rm Im}\sigma}_{-\textrm{Re}f},\num$$ valid for a holomorphic function $f$ on ${\mathbb{C}}^{2n}$. It follows from (\[eq21\]) that ${\textrm{Re }}q_0$ is constant along the flow of the Hamilton vector field $$\widehat{H_{-\frac{1}{i}q_0}}=H^{-\textrm{Im}\sigma}_{-\textrm{Re }q_0}.$$ We therefore deduce from (\[kps2\]), (\[inv7\]), (\[inv9\]) and (\[eq20\]) that $$\inc \label{kps30} \textrm{Re}\big[q_0|_{\Lambda_{\Phi_t}}\big] \leq 0. \num$$ According to (\[kps40\]), the equation (\[eikonal\]) implies therefore that $$\frac{\partial \Phi}{\partial t}(t,x)\leq 0,$$ so that the function $$t\mapsto \Phi_t(x),$$ is decreasing. Let us assume that there exists $x_0\in {\mathbb{C}}^n$, $x_0\neq 0$, such that $$\int_0^T {\textrm{Re }}\Big[q_0\Big(x_0,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x_0)\Big)\Big]\,dt=0.$$ Since from (\[kps40\]) and (\[kps30\]), the integrand is non-positive, it follows that for all $0\leq t\leq T$, $$\inc \label{kps31} {\textrm{Re }}\Big[ q_0\Big(x_0,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x_0)\Big)\Big]=0. \num$$ Therefore, in view of (\[eikonal\]), we get $$\frac{\partial \Phi}{\partial t}(t,x_0)=0,$$ for all $0\leq t \leq T$. Here, the quadratic form $$f_t(x)=\frac{\partial \Phi}{\partial t}(t,x)={\textrm{Re }}\Big[ q_0\Big(x,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x)\Big)\Big] \leq 0,$$ is non-positive and such that $f_t(x_0)=0$ for all $0 \leq t \leq T$. It follows that $$\nabla_{{\rm Re} x, {\rm Im} x}f_t(x_0)=0,$$ for all $0\leq t\leq T,$ and therefore $$\frac{\partial f_t}{\partial x}(x_0)=\frac{\partial^2 \Phi}{\partial x \partial t}(t,x_0)=0,$$ for all $0\leq t\leq T$. Hence the function $$t\mapsto \frac{\partial \Phi}{\partial x}(t,x_0),$$ does not depend on $t$ for $0\leq t\leq T$, so that $$\inc \label{kps32} \frac{\partial \Phi}{\partial x}(t,x_0)=\frac{\partial \Phi}{\partial x}(0,x_0)=\frac{\partial \Phi_0}{\partial x}(x_0), \num$$ for all $0\leq t\leq T$. Since from (\[kps40\]), the point $$\inc \label{kps33} \Big(x_0,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x_0)\Big)=\Big(x_0,\frac{2}{i}\frac{\partial \Phi_0}{\partial x}(x_0)\Big), \num$$ belongs to $\Lambda_{\Phi_t}$ for all $0\leq t\leq T$, we obtain from (\[inv8\]) and (\[eq20\]) that there exists $y_0(t)\in {\mathbb{C}}^n$ such that $$\inc \label{kps34} \Big(x_0,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x_0)\Big)={\mbox{\rm exp\,}}(t\widehat{H_{-\frac{1}{i}q_0}})\Big(y_0(t), \frac{2}{i}\frac{\partial \Phi_0}{\partial x}\big(y_0(t)\big)\Big). \num$$ It follows from (\[kps31\]) that for all $0\leq t\leq T$, $$\inc\label{kps35} \textrm{Re}\Big[q_0\Big(x_0,\frac{2}{i}\frac{\partial \Phi}{\partial x}(t,x_0)\Big)\Big]=\textrm{Re}\Big[q_0\Big(y_0(t),\frac{2}{i}\frac{\partial \Phi_0}{\partial x}\big(y_0(t)\big)\Big)\Big]=0, \num$$ because $\textrm{Re }q_0$ is constant along the flow of the Hamilton vector field $$\widehat{H_{-\frac{1}{i}q_0}} = H^{-{\rm Im}\sigma}_{-\textrm{Re}q_0}.$$ Let us now set $$\label{kps39}\inc L_0=\big\{\tilde{X} \in \Lambda_{\Phi_0} : \textrm{Re}[q_0(\tilde{X})]=0 \big\}.\num$$ We can notice by using similar arguments as in (\[2.3.99\]) and (\[2.3.100\]) that $$\{X \in {\mathbb{R}}^{2n}, \textrm{Re }\tilde{q}(X)=0\}=\textrm{Ker}(\textrm{Re }\tilde{F}) \cap {\mathbb{R}}^{2n},$$ for any complex-valued quadratic form $\tilde{q}$ with a non-positive real part if $\textrm{Re }\tilde{F}$ is the Hamilton map of the quadratic form $\textrm{Re }\tilde{q}$. We therefore deduce from (\[inv7\]) and (\[inv9\]) that $$\inc \label{kps40yyy} L_0=\kappa_T\big({\rm Ker}({\textrm{Re }}F)\cap {\mathbb{R}}^{2n}\big). \num$$ We get from (\[inv8\]), (\[kps33\]), (\[kps34\]), (\[kps35\]), (\[kps39\]) and (\[kps40yyy\]) that $$\label{kps35bis}\inc \Big(y_0(t),\frac{2}{i}\frac{\partial \Phi_0}{\partial x}\big(y_0(t)\big)\Big)={\mbox{\rm exp\,}}(t\widehat{H_{\frac{1}{i}q_0}})\Big(x_0,\frac{2}{i}\frac{\partial \Phi_0}{\partial x}(x_0)\Big) \in L_0, \num$$ for all $0\leq t\leq T$, and therefore $$\inc \label{eq25} {\textrm{Re }}F \Big(\kappa_T^{-1}\Big({\mbox{\rm exp\,}}(t\widehat{H_{\frac{1}{i}q_0}})\Big(x_0,\frac{2}{i}\frac{\partial \Phi_0}{\partial x}(x_0)\Big)\Big)\Big)=0, \num$$ for all $0\leq t\leq T$. In view of (\[inv8\]) and (\[inv9\]), we may write $$\inc \label{kps36} \Big(x_0,\frac{2}{i}\frac{\partial \Phi_0}{\partial x}(x_0)\Big)=\kappa_T(X_0), \num$$ for some $X_0\in {\mathbb{R}}^{2n}$, $X_0\neq 0$ because $x_0 \neq 0$. We can now deduce from (\[inv6\]) that there exists an integer $m \in {\mathbb{N}}$ such that $$\inc \label{eq26} {\textrm{Re }}F\left({\textrm{Im }}F\right)^j X_0=0,\quad 0\leq j<m, \num$$ while $$\inc \label{eq27} {\textrm{Re }}F\left({\textrm{Im }}F\right)^m X_0\neq 0.\num$$ On the other hand, we get from (\[eq25\]) and (\[kps36\]) that $$\inc \label{eq27.5} {\textrm{Re }}F\Big(\kappa_T^{-1} \big(\widehat{H_{\frac{1}{i}q_0}}\big)^j\kappa_T(X_0)\Big)=0, \num$$ for all $j \in {\mathbb{N}}$ because $q_0$ is a quadratic form. We shall establish the following result. \[lem3.1.3\] For all $0\leq j\leq m$, we have $$\big(\widehat{H_{\frac{1}{i}q_0}}\big)^j \kappa_T(X_0)=\big(H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0}\big)^j \kappa_T(X_0).$$ When proving this lemma, we shall argue by induction with respect to $j$, and start with the case $j=0$, which is of course fulfilled. Let us recall from (\[kps35bis\]) and (\[kps36\]) that $\kappa_T(X_0)\in L_0$, and notice that, as recalled for example in section 11 (Remark 11.7) in [@HeSjSt] (see also [@MelinSj]), we have at the points of $L_0$, $$\inc \label{eqSj} \widehat{H_{\frac{1}{i}q_0}}=H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0}. \num$$ Let us now check that for all $0\leq j\leq m-1$, $$\inc \label{eq28} \big(H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0}\big)^j \kappa_T(X_0)\in L_0.\num$$ Let us consider $0 \leq j \leq m-1$ and notice from (\[inv7\]) that $$\label{kps37}\inc \kappa_T(H_{\textrm{Im }q}) \kappa_T^{-1}= H_{\textrm{Im }q_0}^{\sigma_{\Lambda_{\Phi_0}}}.\num$$ Since a direct computation using (\[10\]) shows that $$\label{kps38}\inc H_{\textrm{Im }q} =2 \textrm{Im }F, \num$$ we obtain by using (\[10\]), (\[inv7\]), (\[eq26\]) and (\[kps37\]) that $$\begin{aligned} & \ \textrm{Re}\Big[q_0\big((H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0})^j \kappa_T(X_0)\big)\Big]=\textrm{Re}\Big[q_0\big(\kappa_T H_{{\textrm{Im }}q}^j (X_0)\big)\Big]\\ =& \ {\textrm{Re }}q\left(2^j ({\textrm{Im }}F)^j X_0\right) =2^{2j}\sigma\big((\textrm{Im }F)^jX_0, \textrm{Re }F(\textrm{Im }F)^jX_0 \big)=0,\end{aligned}$$ for any $0 \leq j \leq m-1$. Thus we have verified (\[eq28\]), and by an application of (\[eqSj\]), we get that for all $0\leq j\leq m-1$, $$\widehat{H_{\frac{1}{i}q_0}}\big((\widehat{H_{\frac{1}{i}q_0}})^j \kappa_T(X_0)\big)=H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0}\big((\widehat{H_{\frac{1}{i}q_0}})^j \kappa_T(X_0)\big)=(H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\, q_0})^{j+1}\kappa_T(X_0),$$ if $$(\widehat{H_{\frac{1}{i}q_0}})^j\kappa_T(X_0)=(H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\, q_0})^{j}\kappa_T(X_0).$$ This proves by induction Lemma \[lem3.1.3\]. It is now easy to finish the proof of Lemma \[phase\]. By using (\[eq27.5\]) when $j=m$, (\[kps37\]), (\[kps38\]) and applying Lemma \[lem3.1.3\], we get $$0={\textrm{Re }}F\big(\kappa_T^{-1} (H^{\sigma_{\Lambda_{\Phi_0}}}_{{\rm Im}\,q_0})^m \kappa_T(X_0)\big)={\textrm{Re }}F ( H_{{\rm Im}\,q}^m X_0)=2^m {\textrm{Re }}F({\textrm{Im }}F)^m X_0,$$ which contradicts (\[eq27\]) and completes the proof of Lemma \[phase\]. Let us finally notice that the semigroup $e^{tQ}$, $t>0$, is strongly regularizing on $L^2({\mathbb{R}}^n)$. We actually deduce from (\[invb1\]), (\[eq20.1\]) and the fundamental property of semigroups that for all $t>0$, there exists $\delta >0$ such that $$\forall u \in H_{\Phi_0}({\mathbb{C}}^n), \ e^{tQ_0} u \in H_{\Phi_0-\delta\abs{x}^2}({\mathbb{C}}^n),$$ on the FBI transform side. By using the fact that a holomorphic function $U$ on ${\mathbb{C}}^n$ is of the form $Tu$ for some $u\in \mathcal{S}({\mathbb{R}}^n)$, if and only if $$\forall N \in {\mathbb{N}}, \ \int_{{\mathbb{C}}^n} \abs{U(x)}^2 e^{-2\Phi_0(x)} \langle{x}\rangle^N\, L(dx)<+\infty,$$ (see for instance [@Sj95]) we finally obtain that $$\forall t>0, \forall u \in L^2({\mathbb{R}}^n), \ e^{tQ}u \in \mathcal{S}({\mathbb{R}}^n),$$ which ends the proof of Proposition \[sperp\]. Large time behavior of contraction semigroups {#beha} --------------------------------------------- In this section, we prove Theorem \[theorem3\]. Let us consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part $$\textrm{Re }q \leq 0,$$ such that its singular space $S$ has a symplectic structure. Let us assume that the real part of the symbol $q$ is a non-zero quadratic form $$\exists X_0 \in {\mathbb{R}}^{2n}, \ \textrm{Re }q(X_0) \neq 0.$$ This implies that the singular space is distinct from the whole phase space $S \neq {\mathbb{R}}^{2n}$ because from (\[10\]) and (\[h1\]), $$\forall X \in S, \ \textrm{Re }q(X)=\sigma(X,\textrm{Re }FX)=0.$$ It proves that $(ii)$ implies $(iii)$ in Theorem \[theorem3\]. Let us now assume that $S \neq {\mathbb{R}}^{2n}$ and prove $(i)$. We deduce from (\[sm12\]) and (\[inf12\]) that there exists $\chi$, a real linear symplectic transformation of ${\mathbb{R}}^{2n}$, such that $$\label{decay1}\inc (q \circ \chi)(x,\xi)= q_1(x',\xi')+ iq_2(x'',\xi''), \ (x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n}, \num$$ where $q_1$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n'}$, $n' \geq 1$, with a non-positive real part and $q_2$ is a real-valued quadratic form verifying the properties stated in Proposition \[proposition1\]. By using the symplectic invariance of the Weyl quantization given by the theorem 18.5.9 in [@hormander], we can find a metaplectic operator $U$, which is a unitary transformation on ${L^2({\mathbb{R}}^n)}$ and an automorphism of $\mathcal{S}({\mathbb{R}}^n)$ such that $$(q \circ \chi)(x,\xi)^w=U^{-1} q(x,\xi)^w U.$$ This implies at the level of the generated semigroups that $$e^{t(q \circ \chi)(x,\xi)^w}=U^{-1} e^{t q(x,\xi)^w} U, \ t \geq 0$$ and $$\label{decay2}\inc \|e^{t(q \circ \chi)(x,\xi)^w}\|_{\mathcal{L}(L^2)}=\|e^{t q(x,\xi)^w}\|_{\mathcal{L}(L^2)}, \ t \geq 0, \num$$ because $U$ is a unitary operator on $L^2({\mathbb{R}}^n)$. Since both operators $iq_2(x'',\xi'')^w$ and $-iq_2(x'',\xi'')^w$ generate contraction semigroups verifying $$\big(e^{t iq_2(x'',\xi'')^w}\big)^{-1}=e^{t (-iq_2(x'',\xi'')^w)}, \ t \geq 0,$$ the semigroup $e^{itq_2(x'',\xi'')^w}$ is unitary for all $t \geq 0$. It follows from the tensorization of the variables (\[decay1\]), $$e^{t(q \circ \chi)(x,\xi)^w}=e^{tq_1(x',\xi')^w}e^{itq_2(x'',\xi'')^w},$$ and (\[decay2\]) that $$\|e^{t q(x,\xi)^w}\|_{\mathcal{L}(L^2)}=\|e^{t q_1(x',\xi')^w}\|_{\mathcal{L}(L^2)}, \ t \geq 0.$$ For proving $(i)$, it is therefore sufficient to prove the exponential decay in time for the norm of the contraction semigroup generated by the operator $q_1(x',\xi')^w$. We have proved in Proposition \[sperp\] (see also (\[inf12\])) that the spectrum of the operator $q_1(x',\xi')^w$ is only composed of the following eigenvalues $$\label{decay2.5}\inc \sigma\big{(}q_1(x',\xi')^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F_1), \\ \textrm{Re}(-i \lambda) <0} } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\}, \num$$ where $F_1=F|_{S^{\sigma \perp}}$ is the Hamilton map associated to the quadratic form $q_1$ and $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F_1$ in ${\mathbb{C}}^{2n'}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. We have also seen in the proof of Proposition \[sperp\] that the contraction semigroup generated by the operator $q_1(x',\xi')^w$ is compact for all $t>0$. We proved this fact in (\[kps41\]) on the FBI transform side. This allows us to apply Theorem 2.20 in [@davies] to obtain the following description of the spectrum, $$\sigma(e^{t q_1(x',\xi')^w})=\{0\} \cup \big\{e^{t \mu} : \mu \in \sigma\big(q_1(x',\xi')^w \big)\big\}.$$ Its spectral radius is therefore given by $$\label{decay3}\inc \textrm{rad}\big(e^{t q_1(x',\xi')^w}\big)=e^{-at }, \num$$ where $$\label{decay4}\inc a=\textrm{inf}\Big\{ \sum_{\substack{\lambda \in \sigma(F_1), \\ \textrm{Re}(-i \lambda) <0} } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}\big(-\textrm{Re}(-i\lambda)\big) : k_{\lambda} \in {\mathbb{N}}} \Big\}. \num$$ It follows that the constant $a$ is positive. Since from Theorem 1.22 in [@davies], we have $$-a=\lim_{t \rightarrow +\infty}{\frac{1}{t} \log\|e^{tq_1(x',\xi')^w}\|_{\mathcal{L}(L^2)}},$$ we obtain that there exists $M>0$ such that $$\label{ocello0.1}\inc \|e^{tq_1(x',\xi')^w}\|_{\mathcal{L}(L^2)} \leq M e^{-\frac{a}{2}t}, \num$$ for all $t \geq 0$. This proves that $(iii)$ implies $(i)$. Finally, the fact that $(i)$ implies $(ii)$ is a consequence of a property that we have already mentioned, namely, when the real part of the symbol $q$ is identically equal to zero then the contraction semigroup $$e^{tq(x,\xi)^w},$$ is unitary for all $t \geq 0$. This ends the proof of Theorem \[theorem3\]. $\Box$ Spectra of non-elliptic quadratic operators ------------------------------------------- In this section, we prove Theorem \[theorem2\]. Let us consider a complex-valued quadratic form $$q : {\mathbb{R}}_x^n \times {\mathbb{R}}_{\xi}^n \rightarrow {\mathbb{C}}, \ n \in {\mathbb{N}}^*,$$ with a non-positive real part $$\textrm{Re }q \leq 0,$$ which is elliptic on its singular space $S$, $$\label{til1}\inc (x,\xi) \in S, \ q(x,\xi)=0 \Rightarrow (x,\xi)=0. \num$$ Let us recall that this assumption of partial ellipticity on the singular space ensures that the singular space has a symplectic structure. We can therefore resume the beginning of our reasoning explained in section \[compact\]: By using the symplectic decomposition of the symbol obtained in section \[sympl\], we deduce from (\[sm12\]) and (\[sm13\]) that there exists $\chi$, a real linear symplectic transformation of ${\mathbb{R}}^{2n}$, such that $$\label{til2}\inc (q \circ \chi)(x,\xi)= q_1(x',\xi')+ iq_2(x'',\xi''), \ (x,\xi)=(x',x'';\xi',\xi'') \in {\mathbb{R}}^{2n}, \num$$ where $q_1$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n'}$ with a non-positive real part $$\label{til3}\inc \textrm{Re }q_1 \leq 0, \num$$ and $q_2$ is a real-valued quadratic form verifying the properties stated in Propositions \[proposition1\] and \[sperp\]. To obtain the result of Theorem \[theorem2\], let us notice from Proposition \[proposition1\] that when the symbol $q$ is elliptic on $S$, we can assume that $$\label{til3.1}\inc q_2(x'',\xi'')={\varepsilon}\sum_{j=1}^{n''}{\lambda_j(\xi_j''^2+x_j''^2)},\num$$ where ${\varepsilon}\in \{\pm 1\}$ and $\lambda_j>0$ for all $j=1,...,n''$. By using again the symplectic invariance of the Weyl quantization given by the theorem 18.5.9 in [@hormander], we can find a metaplectic operator $U$, which is a unitary transformation on ${L^2({\mathbb{R}}^n)}$ and an automorphism of $\mathcal{S}({\mathbb{R}}^n)$ such that $$\label{til3.15}\inc (q \circ \chi)(x,\xi)^w=U^{-1} q(x,\xi)^w U. \num$$ Since the quadratic form $q_2$ is elliptic on ${\mathbb{R}}^{2n''}$, we deduce from the theorem 3.5 in [@sjostrand] that the spectrum of the operator $iq_2(x'',\xi'')^w$ is only composed of eigenvalues with finite multiplicity $$\label{til3.2}\inc \sigma\big{(}iq_2(x'',\xi'')^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(iF_2), \\ -i \lambda \in \Sigma(iq_2)\setminus \{0\}} } {\big{(}r_{\lambda}''+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\},\num$$ where $F_2$ is the Hamilton map associated to the quadratic form $q_2$ and $r_{\lambda}''$ is the dimension of the space of generalized eigenvectors of $i F_2$ in ${\mathbb{C}}^{2n''}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. We notice from (\[inf11.5\]), (\[inf12\]), (\[inf13\]) and (\[sm14\]) that $$\label{til4}\inc F_1=F|_{S^{\sigma \perp}} \textrm{ and } F_2=\frac{1}{i}F|_S. \num$$ Let us notice that if $\lambda$ is an eigenvalue of $F_1$, such that $\textrm{Re}(-i\lambda) \leq 0$, then we necessarily have $$\textrm{Re}(-i\lambda)<0,$$ because if we had $\textrm{Re}(-i\lambda)=0$, it would imply that the Hamilton map $F_1$ has a real eigenvalue and induce, as we saw in (\[venice4\]), that the singular space of the symbol $q_1$ is not reduced to $\{0\}$. However, this singular space is trivial by construction (see $(ii)$ in Proposition \[proposition1\]). This proves that $$\textrm{Re}(-i\lambda)<0.$$ By using now that when the numerical range of a quadratic form $\tilde{q}$ is contained in a closed angular sector $\Gamma$ with a vertex in $0$ and an aperture strictly less than $\pi$ then $\lambda$ is an eigenvalue of its Hamilton map $\tilde{F}$ if and only if $-\lambda$ is an eigenvalue of $\tilde{F}$, and $$-i\lambda \in \Gamma \textrm{ or } i \lambda \in \Gamma,$$ (see section 3 in [@hypoelliptic]), we obtain from (\[inf11.5\]), (\[inf12\]), (\[inf13\]) and (\[sm14\]) that $$\begin{aligned} \label{til4000} & \qquad \quad \big\{\lambda \in {\mathbb{C}}: \lambda \in \sigma(F), -i\lambda \in {\mathbb{C}}_- \cup (\Sigma(q|_S) \setminus \{0\})\big\} \num\\ =& \ \big\{\lambda \in {\mathbb{C}}: \lambda \in \sigma(F_1), \textrm{Re}(-i\lambda)<0\big\} \sqcup \big\{\lambda \in {\mathbb{C}}: \lambda \in \sigma(i F_2), -i\lambda \in \Sigma(iq_2) \setminus \{0\}\big\},\end{aligned}$$ where $F$ is the Hamilton map associated to the quadratic form $q$, $$\Sigma(q|_S)=\overline{q(S)} \textrm{ and } {\mathbb{C}}_-=\{z \in {\mathbb{C}}: \textrm{Re }z<0\}.$$ We shall now deduce from the tensorization of the variables (\[til2\]), (\[til3.15\]) and (\[til4\]) that the spectrum of the quadratic differential operator $q(x,\xi)^w$ is only composed of eigenvalues with finite multiplicity $$\label{ocello4}\inc \sigma\big{(}q(x,\xi)^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F), \\ -i \lambda \in {\mathbb{C}}_- \cup (\Sigma(q|_S) \setminus \{0\})} } {\big{(}r_{\lambda}+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\},\num$$ where $r_{\lambda}$ is the dimension of the space of generalized eigenvectors of $F$ in ${\mathbb{C}}^{2n}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. This result is trivial when the singular space is equal to the whole phase space because in that case the quadratic form $q_1$ is identically equal to 0 and $n''=n$. We therefore assume in the following that $$\label{ocello4.1}\inc S \neq {\mathbb{R}}^{2n}.\num$$ Let us begin by recalling that we know from (\[decay2.5\]) that the spectrum of the operator $q_1(x',\xi')^w$ is only composed of eigenvalues with finite multiplicity $$\label{ocello4.3}\inc \sigma\big{(}q_1(x',\xi')^w\big{)}=\Big\{ \sum_{\substack{\lambda \in \sigma(F_1), \\ \textrm{Re}(-i \lambda)<0} } {\big{(}r_{\lambda}'+2 k_{\lambda} \big{)}(-i\lambda) : k_{\lambda} \in {\mathbb{N}}} \Big\},\num$$ where $r_{\lambda}'$ is the dimension of the space of generalized eigenvectors of $F_1$ in ${\mathbb{C}}^{2n'}$ belonging to the eigenvalue $\lambda \in {\mathbb{C}}$. It follows from (\[ocello4.3\]) that when a positive constant $a$ verifies $$\label{ocello1}\inc \sigma\big(q_1(x',\xi')^w\big) \cap \{z \in {\mathbb{C}}: \textrm{Re }z=-a\} =\emptyset, \num$$ the operator $q_1(x',\xi')^w$ only has a finite number of eigenvalues in the half-plane $$\label{ocello2}\inc \{z \in {\mathbb{C}}: -a \leq \textrm{Re }z \}.\num$$ In the following, we besides assume that the singular space is not reduced to zero $$\label{ocello4.2}\inc S \neq \{0\},\num$$ because the description (\[ocello4\]) is then a direct consequence of (\[ocello4.3\]). We now need some estimates for the resolvent of the operator $q_1(x',\xi')^w$ to obtain the description (\[ocello4\]) for the spectrum of the operator $q(x,\xi)^w$. \[propolol\] For all $a>0$ such that $$\sigma\big(q_1(x',\xi')^w\big) \cap \{z \in {\mathbb{C}}: \emph{\textrm{Re }}z=-a\} =\emptyset,$$ there exists $C_a>0$ such that $$\label{myrtaeq3.1}\inc \big\|\big(z-q_1(x',\xi')^w\big)^{-1}\big\| \leq C_a, \num$$ for all $z \in {\mathbb{C}}$ with $-a < {\textrm{Re }}z $ and $\abs{{\textrm{Im }}z} \geq C_a$. Here the norm is the operator norm on $L^2$. When proving Proposition \[propolol\], we first recall from (\[ocello0.1\]) and (\[ocello4.1\]) that there exist $\tilde{a}>0$ and $M>0$ such that $$\label{ocello5.1}\inc \norm{e^{tq_1(x',\xi')^w}}\leq M e^{-\tilde{a} t},\ t \geq 0.\num$$ By using Theorem 2.8 in [@davies], we can write that for all $z \in {\mathbb{C}}$ such that $\textrm{Re }z >-\tilde{a}$, $$\big(z-q_1(x',\xi')^w\big)^{-1} = \int_0^{+\infty} e^{-z t}\,e^{tq_1(x',\xi')^w} dt,$$ and we deduce from (\[ocello5.1\]) that $$\label{myrtaeq3}\inc \big\|\big(z-q_1(x',\xi')^w\big)^{-1}\big\| \leq \frac{M}{\tilde{a}+{\textrm{Re }}z}, \num$$ for all $z \in {\mathbb{C}}$, $- \tilde{a} < {\textrm{Re }}z$. This proves the estimate (\[myrtaeq3.1\]) when the positive constant $a$ is small enough. To prove the result in the general case, we shall follow an argument used by L.S. Boulton in [@boulton]. Let us consider a positive constant $a$ verifying (\[ocello1\]). We have already seen that the operator $q_1(x',\xi')^w$ has only a finite number of eigenvalues with finite multiplicity in the half-plane $$\{z \in {\mathbb{C}}: -a \leq \textrm{Re }z \}.$$ We can therefore consider $\Pi_a$, the finite-rank spectral projection associated to the eigenvalues $$\sigma\big(q_1(x',\xi')^w\big) \cap \{ z \in {\mathbb{C}}: -a \leq {\textrm{Re }}z \},$$ and write for all $z \in {\mathbb{C}}$ with $z \not\in \sigma\big(q_1(x',\xi')^w\big)$ that $$\label{laet5}\inc \big(z-q_1(x',\xi')^w\big)^{-1} = \big(z-q_1(x',\xi')^w\big)^{-1} \Pi_a + \big(z-q_1(x',\xi')^w\big)^{-1}(1-\Pi_a). \num$$ Here $$\label{laet6}\inc \big(z-q_1(x',\xi')^w\big)^{-1}(1-\Pi_a) = \big(z-q_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}\big)^{-1} (1-\Pi_a), \num$$ and we can write by using Theorems 1.22 and 2.8 in [@davies] that $$\label{laet2}\inc \big(z-q_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}\big)^{-1}= \int_0^{+\infty} e^{-z t} e^{tq_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}} dt, \num$$ for all $z \in {\mathbb{C}}$ with $-b <\textrm{Re }z $ if we set $$\label{laet1}\inc -b=\lim_{t\rightarrow +\infty} \frac{1}{t} \log \|e^{tq_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}}\|.\num$$ Let us now notice that the contraction semigroup $$e^{tq_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}}=e^{tq_1(x',\xi')^w}(1-\Pi_a),$$ is compact for any $t>0$ since we have seen in section \[beha\] that the contraction semigroup generated by the operator $q_1(x',\xi')^w$ is compact for any $t>0$. Since on the other hand the spectrum of the operator $q_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}$ is equal to $$\sigma\big(q_1(x',\xi')^w\big) \cap \{ z \in {\mathbb{C}}: {\textrm{Re }}z \leq -a\},$$ we deduce from Theorems 1.22 and 2.20 in [@davies] that $-b< -a,$ which implies in view of (\[laet1\]) that there exists $\tilde{M}>0$ such that $$\label{laet3}\inc \|e^{tq_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}}\| \leq \tilde{M} e^{-a t},\ t \geq 0.\num$$ We then deduce from (\[laet2\]) and (\[laet3\]) that $$\label{laet4}\inc \big\|\big(z-q_1(x',\xi')^w|_{{\rm Ran}(1-\Pi_a)}\big)^{-1}(1-\Pi_a)\big\| \leq \frac{\tilde{M}}{a+{\textrm{Re }}z} \|1-\Pi_a\|,\num$$ for all $z \in {\mathbb{C}}$ with $-a < \textrm{Re }z$. Since on the other hand $$\big(z-q_1(x',\xi')^w\big)^{-1} \Pi_a = \big(z-q_1(x',\xi')^w|_{{\rm Ran}\,\Pi_a}\big)^{-1} \Pi_a,$$ and the vector space ${\rm Ran}\,\Pi_a$ is finite-dimensional, we therefore have $$\label{laet7}\inc \big\|\big(z-q_1(x',\xi')^w\big)^{-1} \Pi_a \big\|=\mathcal{O}_a(1),\num$$ for any $z \in {\mathbb{C}}$ when $|{\textrm{Im }}z|$ is large enough depending on $a$. We finally deduce the result of Proposition \[propolol\] from (\[laet5\]), (\[laet6\]), (\[laet4\]) and (\[laet7\]). *Remark.* Let us notice that the previous proof actually shows that when $q$ is a complex-valued quadratic form on ${\mathbb{R}}^{2n}$, $n \geq 1$, with non-positive real part and a zero singular space $$S=\{0\},$$ then $$e^{t q(x,\xi)^w}=e^{t q(x,\xi)^w} \Pi_a+\mathcal{O}_a(e^{-at}), \ t \geq 0,$$ in $\mathcal{L}(L^2)$ for any $a>0$ such that $$\sigma\big(q(x,\xi)^w\big) \cap \{z \in {\mathbb{C}}: \textrm{Re }z=-a\} =\emptyset.$$ We can now resume our proof of Theorem \[theorem2\]. In doing so, we recall that any quadratic differential operator $\tilde{q}(x,\xi)^w$ whose symbol has a non-positive real part, is defined by the maximal closed realization on $L^2({\mathbb{R}}^n)$ with the domain $$\{u \in L^2({\mathbb{R}}^n) : \tilde{q}(x,\xi)^w u \in L^2({\mathbb{R}}^n)\},$$ which coincides with the graph closure of its restriction to $\mathcal{S}({\mathbb{R}}^n)$, $$\tilde{q}(x,\xi)^w : \mathcal{S}({\mathbb{R}}^n) \rightarrow \mathcal{S}({\mathbb{R}}^n).$$ By noticing from (\[til3.15\]) that $$\label{laet9}\inc \sigma\big((q \circ \chi)(x,\xi)^w\big)=\sigma\big(q(x,\xi)^w\big),\num$$ we deduce from (\[til2\]), (\[til3.2\]), (\[til4000\]) and (\[ocello4.3\]) that it is sufficient for obtaining (\[ocello4\]) and ending the proof of Theorem \[theorem2\] to establish that $$\label{laet8}\inc \sigma\big(q(x,\xi)^w\big)=\sigma\big(q_1(x',\xi')^w\big) +\sigma\big(iq_2(x'',\xi'')^w\big). \num$$ We then notice that since the spectra of the operators $q_1(x',\xi')^w$ and $q_2(x'',\xi'')^w$ are only composed of eigenvalues, we directly get the first inclusion $$\sigma\big(q_1(x',\xi')^w\big) + \sigma\big(iq_2(x'',\xi'')^w\big) \subset \sigma\big(q(x,\xi)^w\big),$$ by considering the functions $$u(x',x'')=u_1(x')u_2(x'') \in L^2({\mathbb{R}}^n),$$ where $u_1$ and $u_2$ are respectively some eigenvectors of the operators $q_1(x',\xi')^w$ and $q_2(x'',\xi'')^w$, since these functions are eigenvectors for the operator $q(x,\xi)^w$. We shall now prove the opposite inclusion. Let us consider $z\in {\mathbb{C}}$ such that $$\label{ocello113}\inc z \not\in \sigma\big(q_1(x',\xi')^w\big) +\sigma\big(iq_2(x'',\xi'')^w\big). \num$$ In view of (\[laet9\]), it is sufficient to prove that the map $$(q \circ \chi)(x,\xi)^w-z : \{u \in L^2({\mathbb{R}}^n) : (q \circ \chi)(x,\xi)^w u \in L^2({\mathbb{R}}^n) \} \rightarrow L^2({\mathbb{R}}^n),$$ is bijective to obtain the second inclusion. We denote by $$\varphi_{\alpha}(x)= H_{\alpha}(x) e^{-x^2/2}, \ \alpha\in {\mathbb{N}}^n,$$ the orthonormal basis of $L^2({\mathbb{R}}^n)$ composed by Hermite functions. Here the Hermite polynomials $H_{\alpha}(x)$ satisfy $$H_{\alpha}(x)=\prod_{j=1}^n H_{\alpha_j}(x_j),$$ and therefore we write $$\label{laeteq8}\inc \varphi_{\alpha}(x)=\varphi_{\alpha'}(x')\varphi_{\alpha''}(x''),\ \alpha'\in {\mathbb{N}}^{n'},\ \alpha''\in {\mathbb{N}}^{n''}. \num$$ Let us consider the following equation with $u$ and $v$ in $L^2({\mathbb{R}}^n)$, $$\label{laeteq9}\inc (q \circ \chi)(x,\xi)^wu-z u=v. \num$$ We can write $$\label{laeteq10}\inc u(x)=\sum_{\alpha',\alpha''} a_{\alpha'\alpha''} \varphi_{\alpha'}(x')\varphi_{\alpha''}(x''),\ v(x)=\sum_{\alpha',\alpha''} b_{\alpha'\alpha''} \varphi_{\alpha'}(x')\varphi_{\alpha''}(x''), \num$$ where the two sums are taken for $(\alpha',\alpha'') \in {\mathbb{N}}^{n'} \times {\mathbb{N}}^{n''}$. By using from (\[til2\]) that $$(q \circ \chi)(x,\xi)^w=q_1(x',\xi')^w+i q_2(x'',\xi'')^w,$$ we obtain from (\[til3.1\]) that $$\begin{aligned} \label{laeteq11} & \ (q \circ \chi)(x,\xi)^wu-z u \\ =& \ \sum_{\alpha',\alpha''} a_{\alpha'\alpha''}\big[ q_1(x',\xi')^w \varphi_{\alpha'}(x')+ i \mu_{\alpha''}\varphi_{\alpha'}(x')-z \varphi_{\alpha'}(x') \big]\varphi_{\alpha''}(x''),\num\end{aligned}$$ with $$\mu_{\alpha''}={\varepsilon}\sum_{j=1}^{n''} \lambda_j \left(2\alpha''_j+1\right),$$ since $$q_2(x'',\xi'')^w \varphi_{\alpha''}(x'')=\mu_{\alpha''}\varphi_{\alpha''}(x'').$$ By setting for any $\alpha''\in {\mathbb{N}}^{n''}$, $$v_{\alpha''}(x')=\sum_{\alpha' \in {\mathbb{N}}^{n'}}b_{\alpha'\alpha''}\varphi_{\alpha'}(x') \in L^2({\mathbb{R}}^{n'}),$$ so that according to (\[laeteq10\]), $$\label{laeteq14}\inc v(x)=\sum_{\alpha'' \in {\mathbb{N}}^{n''}}v_{\alpha''}(x')\varphi_{\alpha''}(x''),\num$$ we deduce from (\[laeteq11\]) that for solving the equation (\[laeteq9\]), we have to solve all the equations $$\label{laeteq12}\inc \big(q_1(x',\xi')^w+i\mu_{\alpha''}-z\big)u_{\alpha''}(x')=v_{\alpha''}(x'), \ \alpha'' \in {\mathbb{N}}^{n''}, \num$$ where $$u_{\alpha''}(x')=\sum_{\alpha' \in {\mathbb{N}}^{n'}} a_{\alpha'\alpha''} \varphi_{\alpha'}(x') \in L^2({\mathbb{R}}^{n'}).$$ We deduce from (\[ocello113\]) that there is a unique solution $u_{\alpha''}(x')$ in $L^2({\mathbb{R}}^{n'})$ for each of the equations (\[laeteq12\]). This proves that for every $v \in L^2({\mathbb{R}}^n)$, there is at most one solution to the equation (\[laeteq9\]). Let us denote by $u_{\alpha''}$ the solutions to the equations (\[laeteq12\]) and $$\label{laeteq13}\inc u=\sum_{\alpha'' \in {\mathbb{N}}^{n''}} u_{\alpha''}(x')\varphi_{\alpha''}(x''). \num$$ The equation (\[laeteq9\]) will have a unique solution in $L^2({\mathbb{R}}^n)$ for every $v \in L^2({\mathbb{R}}^n)$ if we prove that the function $u$ defined in (\[laeteq13\]) actually belongs to $L^2({\mathbb{R}}^n)$. This is the case. Indeed, we obtain from (\[laeteq14\]) and (\[laeteq12\]) that $$\begin{gathered} \|u\|_{L^2({\mathbb{R}}^n)}^2 = \sum_{\alpha'' \in {\mathbb{N}}^{n''}} \|u_{\alpha''}\|_{L^2({\mathbb{R}}^{n'})}^2 = \sum_{\alpha'' \in {\mathbb{N}}^{n''}} \big\|\big(q_1(x',\xi')^w+i\mu_{\alpha''}-z\big)^{-1}v_{\alpha''} \big\|_{L^2({\mathbb{R}}^{n'})}^2 \\ \leq C \sum_{\alpha'' \in {\mathbb{N}}^{n''}} \|v_{\alpha''}\|_{L^2({\mathbb{R}}^{n'})}^2=C\|v\|_{L^2({\mathbb{R}}^n)}^2<+\infty,\end{gathered}$$ because we deduce from Proposition \[propolol\] and (\[ocello113\]) that the quantities $$\big\|\big(q_1(x',\xi')^w+i\mu_{\alpha''}-z\big)^{-1}\big\|,$$ are bounded with respect to the parameter $\alpha''$ in ${\mathbb{N}}^{n''}$. This ends our proof of Theorem \[theorem2\]. [aa]{} L.S. Boulton, *Non-self-adjoint harmonic oscillator semigroups and pseudospectra*, J. Operator Theory, **47**, 413-429 (2002). E.B.Davies, *One-Parameter Semigroups*, Academic Press, London (1980). B.Helffer, F.Nier, *Hypoelliptic estimates and spectral theory for Fokker-Planck operators and Witten laplacians*, SLN 1862, Springer Verlag, 2005. F.Hérau, J.Sjöstrand, C.Stolk, [*Semiclassical analysis for the Kramers-Fokker-Planck equation*]{}, Comm. PDE, **30**, no.4-6, 689-760 (2005). F.Hérau, M.Hitrik, J.Sjöstrand, [*Tunnel effect for Kramers-Fokker-Planck type operators*]{}, Ann. Henri Poincaré, to appear. L.Hörmander, *A class of hypoelliptic pseudodifferential operators with double characteristics*, Math. Ann., **217**, 165-188 (1975). L.Hörmander, *The analysis of linear partial differential operators* (vol. I,II,III,IV), Springer Verlag (1985). L.Hörmander, *Symplectic classification of quadratic forms, and general Mehler formulas*, Math. Z., **219**, 413-449 (1995). A.Melin, J.Sjöstrand, [*Determinants of pseudodifferential operators and complex deformations of phase space*]{}, Methods Appl. Anal., **9**, no. 2, 177-237 (2002). K.Pravda-Starov, *Contraction semigroups of elliptic quadratic differential operators*, to appear in Mathematische Zeitschrift (2007). J.Sjöstrand, *Parametrices for pseudodifferential operators with multiple characteristics*, Ark. för Mat., **12**, 85-130 (1974). J.Sjöstrand, [*Singularités analytiques microlocales*]{}, Astérisque, **95**, 1-166, Soc. Math. France, Paris (1982). J.Sjöstrand, [*Analytic wavefront sets and operators with multiple characteristics*]{}, Hokkaido Math. Journal, [**12**]{}, no. 3, part 2, 392-433 (1983). J.Sjöstrand, [*Function spaces associated to global I-Lagrangian manifolds*]{}, Structure of solutions of differential equations, Katata/Kyoto, 1995, World Sci. Publ., River Edge, NJ (1996). J.Sjöstrand, M.Zworski, [*The complex scaling method for scattering by strictly convex obstacles*]{}, Ark. för Mat., **33**, 135-172 (1995). <span style="font-variant:small-caps;">UCLA Department of Mathematics, Los Angeles, CA 90095-1555, USA</span>\ *E-mail address:* **hitrik@math.ucla.edu**, **karel@math.ucla.edu**

--- author: - | Paul Koerber[^1] and Alexander Sevrin\ Theoretische Natuurkunde, Vrije Universiteit Brussel\ Pleinlaan 2, B-1050 Brussels, Belgium\ E-mail: , title: 'The non-abelian D-brane effective action through order $\alpha''{}^4$' --- Introduction ============ The bosonic worldvolume degrees of freedom of a single D$p$-brane are $9-p$ scalar fields and a $U(1)$ gauge field in $p+1$ dimensions [@pol]. In the limit of slowly varying fields, the effective action which determines the dynamics of the D-brane at low energies, is known to all orders in $\alpha '$. It is the ten-dimensional supersymmetric Born-Infeld action, dimensionally reduced to $p+1$ dimensions [@BI]. Its supersymmetric extension was obtained in [@susynbi]. The knowledge of the full effective action was crucial for numerous applications. Once several, say $n$, D-branes coincide, the gauge group is enhanced from $U(1)$ to $U(n)$, [@witten]. The full non-abelian extension of the Born-Infeld theory is not known yet. This is mainly due to two (related) reasons. As all fields take their values in the adjoint representation of $U(n)$, an ordering prescription is needed. Besides this we also need to include derivative terms. Indeed the limit of (covariantly) constant fields is intrinsically related to the abelian limit as $D_c F_{ab}=0$ implies that $[F_{dc},F_{ab}]=0$. This can be seen in another way. In the abelian case, the effective action consists of terms of the form (in a very schematic notation) $g^{-2}\alpha'{}^{2m-2} \partial^{2n} F^{2m-n}$. Performing the following rescaling[^2], $x\rightarrow \varepsilon^{-1} \beta^{-1}x$, $A\rightarrow\beta A$, $\alpha'\rightarrow \varepsilon^{-1}\beta^{-2}\alpha'$, $g\rightarrow \varepsilon\beta^2g$, we find that $g^{-2}\alpha'{}^{2m-2} \partial^{2n} F^{2m-n}\rightarrow \varepsilon^n g^{-2}\alpha'{}^{2m-2} \partial^{2n} F^{2m-n}$. Sending $\varepsilon\rightarrow 0$, the derivative terms vanish and the Born-Infeld action remains. In the non-abelian case we still have that $g^{-2}\alpha'{}^{2m-2} D^{2n} F^{2m-n}\rightarrow \varepsilon^n g^{-2}\alpha'{}^{2m-2} D^{2n} F^{2m-n}$. However we also have that $D\cdot= \partial\cdot+[A,\cdot]\rightarrow\varepsilon\beta(\partial\cdot+\varepsilon^{-1}[A,\cdot])$ and $F= \partial A+[A,A]\rightarrow \varepsilon\beta^2(\partial A+\varepsilon^{-1}[A,A])$ which makes the $\varepsilon\rightarrow 0$ limit meaningless. Other arguments for the relevance of the derivative terms in both the abelian and non-abelian case were given in [@bilal1]. At this moment the non-abelian effective action is known through $ {\cal O}(\alpha'{}^3)$ including the terms quadratic in the gauginos. The leading order term of the effective action for $n$ coinciding D$p$-branes is the ten-dimensional $N=1$ supersymmetric $U(n)$ Yang-Mills theory dimensionally reduced to $p+1$ dimensions. There are no ${\cal O}(\alpha')$ corrections. The bosonic ${\cal O}(\alpha'{}^2)$ were first obtained in [@direct] and [@direct1] while the fermionic terms were obtained in [@goteborg] and [@bilal]. In [@goteborg] supersymmetry fixed the correction while in [@bilal] a direct calculation starting from four-point open superstring amplitudes was used. At higher orders a direct calculation becomes problematic[^3] and one has to rely on indirect methods. A very powerful strategy was proposed in [@lies][^4]. Central in the approach of [@lies], was the existence of higher-dimensional generalizations of instantons: stable holomorphic bundles. They solve the Yang-Mills equations of motion and in a D-brane context they correspond to BPS configurations. The effective action can be viewed as a deformation of the Yang-Mills action. Requiring that stable holomorphic bundles, or some deformation thereof, solve the equations of motion yields a recursive method to construct the effective action. While this approach becomes very tedious at higher orders, it has the great advantage that its algorithmic nature allows for a computerized approach. To this end a program in Java, an object oriented language based on the syntax of C, was developed [@paul]. In [@sk1] the method was successfully applied to determine the bosonic ${\cal O}(\alpha'{}^3)$ terms in the effective action. Very recently, in [@groningen], supersymmetry was used not only to confirm the results of [@sk1] but to construct the terms quadratic in the gauginos through this order as well. The full effective action through this order was tested in [@test] and there is no doubt left that the result is indeed correct. In the present paper we extend the results of [@sk1] and we obtain the bosonic terms in the effective action at order $ \alpha'{}^4$. Furthermore we will propose a very economic way to organize the terms in the effective action. Throughout the paper we will put $2\pi\alpha'=1$. General strategy {#strategy} ================ We consider a $U(n)$ Yang-Mills theory in $2p$ dimensions with a Euclidean signature. Its equations of motion are given by[^5], $$\begin{aligned} \label{eom0} D_a F_{ab}=0.\end{aligned}$$ Switching to complex coordinates $z^\alpha=(x^{2\alpha-1}+ix^{2\alpha})/\sqrt{2}$, $\bar z^{\bar\alpha}=(x^{2\alpha-1}-ix^{2\alpha})/\sqrt{2}$, eq. (\[eom0\]) reads, $$\begin{aligned} 0&=&D_{\bar\alpha }F_{\alpha \bar\beta}+D_{\alpha }F_{\bar\alpha \bar\beta}\nonumber\\ &=&D_{\bar\beta}F_{\alpha \bar\alpha }+2D_{\alpha }F_{\bar\alpha \bar\beta},\end{aligned}$$ where we used the Bianchi identities. One sees that imposing the following linear relations between the field-strengths, $$\begin{aligned} \label{hol} F_{\alpha\beta}=F_{\bar\alpha\bar\beta}=0,\end{aligned}$$ and $$\begin{aligned} \label{duy} \sum_\alpha F_{\alpha\bar\alpha}\equiv F_{\alpha\bar\alpha}=0,\end{aligned}$$ provides a solution to the equations of motion. In four dimensions, these are exactly the standard instanton equations. In general, eq. (\[hol\]) defines a holomorphic bundle and eq. (\[duy\]) requires it to be stable [@duy]. These solutions were discovered in [@mons]. Following [@sk1], we subsequently construct, order by order in $\alpha{}'$, all possible terms of the correct dimension which can be build out the field-strength and its derivatives. More concretely, at order $\alpha'{}^m$, we write down all possible terms having $2n$ covariant derivatives, with $n\in\{0,1,\cdots m\}$, and $m-n+2$ field-strength tensors, taking the cyclicity of the group trace into account. Each of these terms gets an arbitrary coupling constant. The same game has to be played for the most general deformation of eq. (\[duy\]), which through this order has the form $$\begin{aligned} \label{duy1} F_{\alpha\bar\alpha}=\sum_{n=1}^{m} {\cal F}_{(n)},\end{aligned}$$ where $ {\cal F}_{(n)}$ with $n<m$ were already determined at lower orders and $ {\cal F}_{(m)}$ is the most general polynomial of terms having $2n$ derivatives and $m-n+1$ field-strengths, where $n\in\{0,1\cdots,m\}$. Here again we leave the coupling constants free. Surprisingly, the coupling constants in [*both*]{} the lagrangian and the stability condition are determined by requiring that eqs. (\[hol\]) and (\[duy1\]) solve the equations of motion. The presence of derivative terms however, complicates the analysis considerably. Indeed when writing down the most general lagrangian at a certain order in $\alpha'$, the following points have to be taken into account, 1. Terms can be related through partial integration. 2. Terms can be related by Bianchi identities. 3. The $[D,D]\cdot=[F,\cdot]$ identities[^6] relate certain terms. 4. Certain terms can be eliminated by field redefinitions. Similar considerations hold, apart from the first point, for the stability condition. We discard the fourth point until we perform the final analysis of the resulting action. The reason for this is that certain field redefinitions affect the complex structure such that eq. (\[hol\]) gets modified and, as a consequence, our method is not entirely field redefinition independent. Because of the identities of points one, two and three, many terms are dependent. Handling this essentially boils down to choosing an appropriate basis of independent terms in the lagrangian as well as in the stability condition. The basis terms in the lagrangian must be such that there is a neat separation between the irrelevant field redefinition terms and the relevant non field redefinition terms. In concreto, when expressing the result in a certain basis, the coefficients of some of the basis terms can be shifted by field redefinitions where, again, great care is needed because the field redefinitions are also related by the identities of point two and three. The shifts are independent when the number of such field redefinition terms is minimal. In that case, we can bring their coefficients to zero. So only the projection of the result on the non field redefinition terms is relevant. Of course, there is still a lot of freedom in choosing a basis, so we need some kind of organizational principle to get rid of at least part of this freedom in an economic way. Organizational principle {#organization} ======================== As in the abelian case, it is possible to classify terms at a certain order in $\alpha'$ according to the number of derivatives. For this, we must get rid of the $[D,D]\cdot=[F,\cdot]$ identities. Consider any linear combination of terms. Now, start with the terms without derivatives and fully symmetrize in the fieldstrength tensors. In this process, we use the $[D,D]\cdot=[F,\cdot]$ identities to convert the introduced commutators of fieldstrengths to commutators of derivatives[^7]. Next, we turn to the terms with two derivatives an again fully symmetrize them, whereby a term of the form $DF$ or $D^2F$ is considered as a single entity. Again, terms with more derivatives are added as compensation. We proceed in this fashion order by order in the number of derivatives. Since the resulting terms are symmetrized in the fieldstrength tensors, all “non-abelianality” sits in the covariant derivatives. From now on all terms, whether they be part of the lagrangian, stability condition, field redefinitions or equations of motion, should be thought of as symmetrized. Note that this approach follows the spirit of [@arkady] and we could call it a “generalized symmetrized trace prescription”. A major advantage is that the non-abelian calculation follows the abelian one more closely, since symmetrized terms in the lagrangian lead to symmetrized terms in the equations of motion. Except for the obvious fact that the derivatives are non-commuting there are however some other differences: 1. There are new identities, because in symmetrizing we only used up part of the ${[}D,D{]} \cdot = {[}F,\cdot{]}$ identities. An example will clarify this. Only one of the identities[^8] $$\begin{aligned} {[}F_{12}, F_{34}{]} & = & {[}D_1, D_2{]} F_{34} \nonumber\\ {[}F_{34}, F_{12}{]} & = & {[}D_3, D_4{]} F_{12} \, ,\end{aligned}$$ is used to commute $F_{12}$ and $F_{34}$. The other one is left in the form: $$\begin{aligned} {[}D_1, D_2{]} F_{34} + {[}D_3, D_4{]} F_{12} = 0 \, .\end{aligned}$$ These kind of identities are related to antisymmetry, as in this case, or to Jacobi identities of fieldstrength commutators. In general they read: $$\begin{aligned} && {[}D_1, {[}D_2, \ldots ,{[}D_{m-2}, {[}D_{m-1},D_{m}{]}{]}\ldots{]}{]} D_{m+1} D_{m+2} \ldots D_{m+n-2} F_{m+n-1,m+n} + \nonumber\\ && {[}D_{m+1}, {[}D_{m+2}, \ldots ,{[}D_{m+n-2}, {[}D_{m+n-1},D_{m+n}{]}{]}\ldots {]}{]} D_1 D_2 \ldots D_{m-2} F_{m-1,m} = 0 \, , \nonumber\\\end{aligned}$$ and we call them [*antisymmetrized Bianchi identities*]{}. 2. Consider the stability condition[^9] eq. (\[duy1\]): $$\begin{aligned} F_{\alpha\bar{\alpha}} - {\cal F}_{(2)} - {\cal F}_{(3)} - \cdots = 0 \, .\end{aligned}$$ Somewhere in the equations of motions we will find $\mbox{Sym} \{F_{\alpha\bar{\alpha}} T\}$, where $\mbox{Sym}$ means “symmetrized in the fieldstrength tensors” and $T$ contains, say, $n$ fieldstrengths. For this to be zero, and thus for the stability condition to solve this piece of the equations of motion, the term $-\mbox{Sym} \{({\cal F}_{(2)} )T\}$ must also be present in the equations of motion at higher order. Here the inner brackets mean that the factors of ${\cal F}_{(2)}$ stay together, so we must further symmetrize this term by mixing those factors among the factors of $T$. Unlike in the abelian case, terms with more derivatives are thus introduced. Interestingly, the number of extra derivatives in these terms must be a multiple of four. Indeed, both $\mbox{Sym}\{({\cal F}_{(2)}) T\}$ and $\mbox{Sym} \{ {\cal F}_{(2)} T \}$ are symmetric under the reversion of all factors, so their difference is symmetric as well and hence must contain an [*even*]{} number of commutators. This is the only way in which terms without derivatives — and using also only contributions without derivatives to the stability condition — communicate with terms with derivatives and thus the mechanism that prevents the “ordinary” symmetrized trace to have BPS solutions. After we found the result in this way, we used a second organizing principle to simplify even further. We tried to use basis terms with as many “groups” of nested covariant derivative commutators as possible. Each group corresponds, using a $[D,D]\cdot=[F,\cdot]$ identity to a commutator of fieldstrengths or equivalently to an algebra structure constant which can be put in front. To clarify this correspondence we write result in different ways using $[D,D]\cdot=[F,\cdot]$ identities: $$\begin{split} [D_3,D_2] D_4 F_{51} D_5 [D_4,D_3] F_{12} & = \\ [F_{32},D_4 F_{51}]D_5 [F_{43},F_{12}] & = \\ [D_4,[D_5,D_1]]F_{32} D_5 [D_1,D_2]F_{43} & \, . \end{split}$$ The first and the third line show that the number of [*groups*]{} of commutators of derivatives is important rather than the number of commutators. Although we pushed the $[D,D]\cdot=[F,\cdot]$ identities to the right initially, in this way we can easily see which terms can be pushed how far in the other direction if necessary. This might be useful when comparing to results following from string amplitude calculations, as they tend to give results where commutators of derivatives are pushed into antisymmetric combinations of fieldstrength tensors. See figure \[alpha3\] and \[alpha4\] for the classification of the terms. Obviously, there is still basis freedom left, so that some terms in the result can move to the right. E.g. result can also be written as a sum of terms with $2$ and $3$ structure constants[^10], however not with terms with only $3$ structure constants. Hence the grey area in figure \[alpha3\]. The result {#result} ========== The purely bosonic part of the non-abelian effective action through $ {\cal O}(\alpha'{}^4)$ is of the form, $$\begin{aligned} {\cal L}=\frac{1}{g^2}\left({\cal L}_0+{\cal L}_2+{\cal L}_3+{\cal L}_4\right),\end{aligned}$$ where the leading term is simply[^11] $$\begin{aligned} \label{l0} {\cal L}_0=\;=\; - {{\rm Tr}}\, \left\{\frac{1}{4}F^2\right\}\,.\end{aligned}$$ Subsequently we have $$\begin{aligned} \label{l2} {\cal L}_2=\mbox{STr} \left\{\frac{1}{8} F^4 - \frac{1}{32} F^2F^2 \right\},\end{aligned}$$ where STr denotes the symmetrized trace prescription. At this point both the overall multiplicative factor in front of the action as well as the scale of the gauge fields got fixed [@sk1]. The next term is[^12] $$\begin{aligned} \label{l3} {\cal L}_3=\frac{\zeta (3)}{2\pi^3}{{\rm Tr}}\left\{[D_3,D_2] D_4 F_{51} \, D_5 [D_4,D_3] F_{12} \right\}.\end{aligned}$$ In fact our method did not fix the overall multiplicative factor in front of this term. As explained in [@sk1] and [@groningen], this was most fortunate as one expects from string theory that the coupling constants of terms which are of odd order in $\alpha'$ are irrational while our method yields, at this order, only rational coupling constants. We determined the coupling constants by comparing the relevant terms to the partial results in [@bilal1]. Note that eq. (\[l3\]) looks very different from the expression given in [@sk1]. The reason for this is clear. In [@sk1], we used precisely the opposite organizational strategy as the one proposed in section \[organization\]. Comparing eq. (\[l3\]) to the corresponding expression in [@sk1] (or [@groningen] where yet another basis was chosen) shows convincingly that the scheme proposed in section \[organization\] is the most economical. We now turn to the new result of this paper. The fourth order contribution to the effective action reads (see figure \[alpha4\]) $$\begin{aligned} {\cal L}_4 = {\cal L}_{4,0} + {\cal L}_{4,2} + {\cal L}_{4,4} \, ,\end{aligned}$$ with $$\begin{split} {\cal L}_{4,0} & = \mbox{STr} \left( \frac{1}{12} F_{12}F_{23}F_{34}F_{45}F_{56}F_{61} - \frac{1}{32} F_{12}F_{23}F_{34}F_{41}F_{56}F_{65} + \frac{1}{384} F_{12}F_{21}F_{34}F_{43}F_{56}F_{65} \right) , \\ {\cal L}_{4,2} & = \frac{1}{48} \mbox{STr} \Big( -2 \, F_{12}D_{1}D_{6}D_{5}F_{23}D_{6}F_{34}F_{45} - F_{12}D_{5}D_{6}F_{23}D_{6}D_{1}F_{34}F_{45} \\ & + 2 \, F_{12}\left[D_{6},D_{1}\right] D_{5}F_{23}F_{34}D_{4}F_{56} + 3 \, D_{4}D_{5}F_{12}F_{23}\left[D_{6},D_{1}\right]F_{34}F_{56} \\ & + 2 \, D_{6}\left[D_{4},D_{5}\right]F_{12}F_{23}D_{1}F_{34}F_{56} + 2 \, D_{6}D_{5}F_{12}\left[D_{6},D_{1}\right]F_{23}F_{34}F_{45} \\ & + 2 \, \left[D_{6},D_{1}\right]D_{3}D_{4}F_{12}F_{23}F_{45}F_{56} \\ & + \left[D_{6},D_{4}\right]F_{12}F_{23}\left[D_{3},D_{1}\right]F_{45}F_{56} \Big) , \\ {\cal L}_{4,4} & = \frac{1}{1440} \mbox{STr} \Big( D_6 [D_4,D_2]D_5 D_5 [D_1,D_3] D_6 F_{12} F_{34} + 4 \, D_2 D_6 [D_4,D_1][D_5,[D_6,D_3]] D_5 F_{12} F_{34} \\ & + 2 \, D_2 [D_6,D_4][D_6,D_1] D_5 [D_5,D_3] F_{12} F_{34} + 6 \, D_2 [D_6,D_4]D_5[D_6,D_1][D_5,D_3] F_{12} F_{34} \\ & + 4 \, D_6 D_5 [D_6,D_4][D_5,D_1][D_4,D_3] F_{12} F_{23} + 4 \, D_6 D_5 [D_4,D_2][D_6,D_1][D_5,D_3] F_{12} F_{34} \\ & + 4 \, D_6 [D_5,D_4][D_3,D_2][D_5,[D_6,D_1]] F_{12} F_{34} \\ & + 2 \, [D_6,D_1][D_2,D_6][D_5,D_4][D_5,D_3]F_{12}F_{34} \Big) \, . \end{split}$$ The terms with zero derivatives, ${\cal L}_{4,0}$, form of course the symmetrized trace Born-Infeld action, while the abelian limit of the terms with four derivatives, the first two terms of ${\cal L}_{4,2}$, can be shown to agree with [@abelian4derivative] [^13]. If we use our method in the abelian case at order $\alpha'{}^4$, these terms have an arbitrary constant, since terms with a different number of derivatives have no way of communicating in this case. In the non-abelian case however, the coefficient is fixed at precisely the right value ! Note that there are no terms with two derivatives nor with six derivatives. If there were, they would have had an arbitrary constant since, following the reasoning of point 2 in section \[organization\], the symmetrized trace does not communicate directly with terms in which the number of derivatives is not a multiple of four. Conclusions =========== In the present paper we calculated the bosonic $\alpha'{}^4$ contribution to the non-abelian effective D-brane action. As was already obvious from the $ {\cal O}(\alpha'{}^3)$ calculation, [@sk1], [@groningen], the inclusion of derivative terms is unavoidable. The “generalized symmetrized trace prescription” proposed in section \[organization\] gives rise to a very economical way of organizing the action which has the additional advantage that it closely mimics the abelian case. When determining the coupling constants in the effective action and the deformed stability condition, we had to solve 1816 equations for 546 unknowns. The fact that we found a unique (modulo field redefinitions) solution consists in itself a strong check on our calculation. An independent check would follow the lines proposed in [@HT], where the mass spectrum in the presence of constant magnetic background fields was calculated from the effective action and compared to the string theoretic result. A careful analysis in [@DST] showed that the symmetrized trace prescription without derivative terms deviated from the string theoretical results, starting at order $ \alpha'{}^4$. The derivative corrections which were obtained in this paper should cure this problem. We postpone this check to a future paper. In this context we should also mention the analysis in [@STT] where the mass spectrum in the presence of constant magnetic background fields combined with the requirement that the abelian limit was correctly reproduced, were used to determine the effective action as far as possible. Besides the fact that this method yielded a multi-parameter solution an important ansatz was made: only terms without derivatives were taken into account. Our present paper clearly indicates that this ansatz was too strong. The present result opens the way to all-order statements. E.g., we already noticed, within our generalized symmetrized trace prescription, that terms without derivatives only communicate with terms with 4, 8, ... derivatives. It follows that terms with 2, 6, 10, ... derivatives have arbitrary constants in our method. It might very well turn out that the constants for the terms with 2 derivatives vanish and that these terms are zero, just as in the abelian case [@abelian4derivative]. As a consequence, terms with $2+4n$, with $n\in\IN$, could vanish as well for the even orders of $\alpha'$. The all-order lessons we can draw from our method is presently under investigation and we will come back to this in a separate publication. Related to this, a careful analysis of the deformed stability condition has been started. Finally, we hope that our paper can shed some light on the remarkable results recently obtained in [@stieberger]. The authors are supported in part by the “FWO-Vlaanderen” through project G.0034.02, in part by the Federal Office for Scientific, Technical and Cultural Affairs through the Interuniversity Attraction Pole P5/27 and in part by the European Commission RTN programme HPRN-CT-2000-00131, in which the authors are associated to the University of Leuven. [99]{} J. Polchinski, [*Dirichlet-branes and Ramond-Ramond charges*]{}, , ; J. Dai, R. Leigh and J. Polchinski, [*New connections between string theories*]{}, . E.S. Fradkin and A.A. Tseytlin, [*Nonlinear electrodynamics from quantized strings*]{}, ; A. Abouelsaood, C.G. Callan, C.R. Nappi and S.A. Yost, [*Open strings in background gauge fields*]{}, ; R.G. Leigh, [*Dirac-Born-Infeld action from Dirichlet sigma model*]{}, . M. Cederwall, A. von Gussich, B.E.W. Nilsson and A. Westerberg, [*The Dirichlet super-three-brane in ten-dimensional type-IIB supergravity*]{}, , ; M. Aganagic, C. Popescu and J. H. Schwarz, [*D-brane actions with local kappa symmetry*]{}, , and [*Gauge-invariant and gauge-fixed D-brane actions*]{}, , ; M. Cederwall, A. von Gussich, B. E. W. Nilsson, P. Sundell and A. Westerberg, [*The Dirichlet super-p-branes in ten-dimensional type-IIA and -IIB supergravity*]{}, , ; E. Bergshoeff and P. K. Townsend, [*Super D-branes*]{}, , . E. Witten, [*Bound states of strings and p-branes*]{}, , . A. Bilal [*Higher derivative corrections to the non-abelian Born-Infeld action*]{}, , . D. J. Gross and E. Witten, [*Superstring modifications of Einstein’s equations*]{}, . A.A. Tseytlin, [*Vector field effective action in the open superstring theory*]{}, and . E. Bergshoeff, M. Rakowski and E. Sezgin, [*Higher-derivative super Yang-Mills theories*]{}, ; M. Cederwall, B.E.W. Nilsson and D. Tsimpis, [*The structure of maximally supersymmetric Yang-Mills theory: constraining higher-order corrections*]{}, , ; [*D=10 super Yang-Mills at $\ap{2}$*]{}, , . E. Bergshoeff, A. Bilal, M. de Roo and A. Sevrin, [*Supersymmetric non-abelian Born-Infeld revisited*]{}, , . L. De Fossé, P. Koerber and A. Sevrin, [*The uniqueness of the abelian Born-Infeld action*]{}, , . J.M. Drummond, P.S. Howe and U. Lindström, [*Kappa-symmetric non-abelian Born-Infeld actions in three dimensions*]{}, . P. Koerber, to appear. P. Koerber and A. Sevrin, [*The non-abelian open superstring effective action through order $\alpha'{}^3$*]{}, , . A. Collinucci, M. de Roo and M.G.C. Eenink, [*Supersymmetric Yang-Mills theory at order $\alpha'{}^3$*]{}, , . P. Koerber and A. Sevrin, [*Testing the ${\cal O}(\alpha'{}^3)$ term in the non-abelian open superstring effective action*]{}, , ; M. de Roo, M.G.C. Eenink, P. Koerber and A. Sevrin, [*Testing the fermionic terms in the non-abelian D-brane effective action through order $ \alpha'{}^3$*]{}, preprint, . K. Uhlenbeck and S.-T. Yau, [*On the existence of hermitian Yang-Mills connections on stable vectorbundles*]{}, [*Comm. Pure Appl. Math.*]{} [**39**]{} (1986) 257 and [*A note on our previous paper: on the existence of hermitian Yang-Mills connections on stable vectorbundles*]{}, [*Comm. Pure Appl. Math.*]{} [**42**]{} (1989) 703; S.K. Donaldson, [*Infinite determinants, stable bundles and curvature*]{}, [*Duke Math. J.*]{} [**54**]{} (1987) 231; see also chapter 15 in the second volume M.B. Green, J.H. Schwarz and E. Witten, [*Superstring theory*]{}, Cambridge University Press 1986. E. Corrigan, C. Devchand, D.B. Fairlie and J. Nuyts, [*First order equations for gauge fields in spaces of dimension greater than four*]{}, . A.A. Tseytlin, [*On non-abelian generalization of Born-Infeld action in string theory*]{}, , . O.D. Andreev and A.A. Tseytlin, [*Partition function representation for the open superstring effective action: cancellation of Möbius infinities and derivative corrections to Born-Infeld lagrangian*]{}, . N. Wyllard, [*Derivative corrections to D-brane actions with constant background fields*]{}, , . A. Hashimoto and W. Taylor, [*Fluctuation spectra of tilted and intersecting D-branes from the Born-Infeld action*]{}, , . F. Denef, A. Sevrin and J. Troost, [*Non-abelian Born-Infeld versus string theory*]{}, , . A. Sevrin, J. Troost and W. Troost, [*The non-abelian Born-Infeld action at order $F^6$*]{}, , . S. Stieberger and T.R. Taylor, [*Non-abelian Born-Infeld action and type I – heterotic duality (I): heterotic $F^6$ terms at two loops*]{}, . [^1]: Aspirant FWO [^2]: The parameter $\beta$ reflects the only scaling freedom left. It can e.g. conveniently be chosen such that $g$ and $\alpha'$ remain fixed in the limit $\varepsilon\rightarrow 0$. [^3]: We thank Stefan Stieberger for explaining this to us. [^4]: A very different approach was recently proposed in [@howe]. [^5]: Throughout this paper, we write all indices down and we sum over repeated (real and complex) indices. [^6]: In their most general form they read: $[D_1D_2\ldots D_{n-2}F_{n-1,n},D_{n+1}\ldots D_{n+m-2}F_{n+m-1,n+m}] = $ $[D_1,[D_2,\ldots [D_{n-2},[D_{n-1},D_n]]\ldots]]D_{n+1} D_{n+m-2}F_{n+m-1,n+m}$. But we will use the above shorthand in the rest of the paper. [^7]: This means that we push the $[D,D]\cdot=[F,\cdot]$ identities to the right. The reader could wonder why we do not push those identities to the left and use for instance symmetrized derivatives. This leads indeed to fewer terms and fewer identities, because there are no antisymmetrized Bianchi identities. On the other hand the remaining partial integration and Bianchi identities are more complicated and the final result is awful because we get no clear separation between field redefinition and non field redefinition terms [^8]: When writing down long equations or results in the rest of the paper, we will use a shorthand notation for the indices i.e.$1,2,3,\ldots$ instead of $a_1,a_2,a_3,\ldots$ Repeated indices are still summed over. [^9]: The order $\alpha'$ correction is zero. [^10]: See the result in [@groningen]. [^11]: We use the following notation: $F^m\equiv F_{a_1a_2}F_{a_2a_3}\cdots F_{a_ma_1}\equiv F_{12}F_{23}\cdots F_{m1}$. [^12]: All results are of course modulo field redefinition terms. [^13]: Although it is a little involved, this can still be shown by hand. See e.g. Appendix B of [@wyllard].

--- abstract: 'Experimental values of SdH and dHvA periods and cyclotron effective masses found by Rosenman and Doi et al. have been compared with the theoretical predictions derived in this work for a tetragonal narrow gap semiconductor. By the least square fit method the values of band parameters were obtained. It has been established that Cd$_3$As$_2$ has inverted band structure resembling HgTe under tensile stress.' author: - title: | Band structure of Cd$_3$As$_2$ from Shubnikov - de Haas and de Haas - van Alphen effects --- **Comment** The following paper on the band structure of Cd$_3$As$_2$ was published in a camera-ready form by my graduate student Jan Bodnar in the Proceedings of the 3rd International Conference on the Physics of Narrow Gap Semiconductors, edited by J. Rauluszkiewicz, M. Gorska and E. Kaczmarek, (PWN-Polish Scientific Publishers, Warszawa 1977), p. 311-316. A few months after the conference Jan died suddenly of a heart attack at the age of 27. Because lately Cd$_3$As$_2$ has become an important material in the field of topological matter, while the Proceedings of NGS Warsaw Conference are at present hardly available, I felt that Bodnar’s work should become accessible on arXiv. The paper appears without changes, some misprints have been corrected. I thank my colleagues Pawel Pfeffer and Grzegorz Mazur for friendly help with the preparation of electronic version.  \ Wlodek Zawadzki, \[zawad@ifpan.edu.pl\] \[sec:level1\]INTRODUCTION\ =========================== In heavily doped semiconductors, such as Cd$_3$As$_2$ (n$_{min}\approx 10^{18}$ cm$^{-3}$ \[1\]), measurements of SdH and dHvA effects serve as powerful tools in band structure investigations. In this work we analyse SdH and dHvA experiments in Cd$_3$As$_2$ \[2-4\] using a theory which presents a generalization of the Kane band model \[5\] for the case of a tetragonal semiconductor with small energy gap \[6\]. \[sec:level2\]THEORY\ ===================== An influence of the tetragonal field on the three-level Kane band model may be characterized by a parameter $\delta$, which represents the crystal field splitting of p-type levels. In both cases of the simple or inverted level ordering tetragonal field splits the fourfold degenerate state $\Gamma_8$ into twofold degenerate levels, thus giving a four-level scheme. Considering exactly the [[[**k$\cdot$p**]{} ]{}]{} coupling between these four levels we have obtained the following secular equation for the energies: $$\gamma({\cal E}) = f_2({\cal E}) \cdot (k^2_x + k^2_y) + f_1({\cal E}) \cdot k^2_z,$$ where $$\gamma({\cal E}) = ({\cal E}-{\cal E}_c)\left({\cal E}-{\cal E}_v-\frac{\Delta}{3}\right) \left[\left({\cal E}-{\cal E}_v+\frac{2 \Delta}{3}\right)\left({\cal E}-{\cal E}_v-\frac{\Delta}{3}\right)+\delta\left({\cal E}-{\cal E}_v+\frac{\Delta}{3}\right)\right],$$ with $$f_1({\cal E}) = P^2_{||}\left({\cal E}-{\cal E}_v-\frac{\Delta}{3}\right)\left({\cal E}-{\cal E}_v+\frac{\Delta}{3}\right),$$ $$f_2({\cal E}) = P^2_{\perp}\left[\left({\cal E}-{\cal E}_v-\frac{\Delta}{3}\right)\left({\cal E}-{\cal E}_v+\frac{\Delta}{3}\right) +\delta({\cal E}-{\cal E}_v)\right].$$ ${\cal E}$ is the energy of electrons or holes, P$_{\perp}$, P$_{||}$ and $\Delta$ are matrix elements of the momentum operator and of the spin-orbit interaction, respectively. ${\cal E}_c$ and ${\cal E}_v$ denote energies of the band edge states for the Hamiltonian without the spin-orbit coupling. These may be expressed by the energy gap ${\cal E}_g={\cal E}(\Gamma_{6s})-{\cal E}(\Gamma_{7p})$. Choosing suitable values of ${\cal E}_c$ and ${\cal E}_v$ we may describe with the help of Eq.(1) all possible four-band models of the tetragonal semiconductor with a small energy gap, i.e. simple or inverted band structure for $\delta >$0 and $\delta <$0. Equation of the type (1) was first derived by Kildal \[7\] for CdGeAs$_2$ under additional restrictions. It follows from Eq.(1) that the conduction band has simple ellipsoidal shape in k-space with the anisotropy depending, however, on the energy. Extreme cross-section of the Fermi ellipsoid in the plane perpendicular to the magnetic field **H** is given by $$S_m=\pi\gamma({\cal E})[f_2({\cal E})]^{-1/2}[f_2({\cal E})cos^2\Theta + f_1({\cal E})sin^2\Theta]^{-1/2},$$ where $\Theta$ is an angle between the magnetic field and the tetragonal axis. According to the semiclassical theory of Lifshitz and Kosevich \[8\] a period $\Delta(1/H)$ of SdH and dHvA oscillations and the cyclotron effective mass are related to S$_m$ as follows $$\Delta(1/H)=2\pi e/c\hbar S_m\;,\;\;\;\;\;\;\;\;\;\;m^*({\cal E},\Theta) = (\hbar^2/2\pi)(\delta S_m/\delta{\cal E}).$$ From expressions (6) we may determine the coefficient of anisotropy $$K = \Delta_{||}(\Theta = 0^{\circ})/\Delta_{\perp}(\Theta = 90^{\circ}) = [f_2({\cal E})/f_1({\cal E})]^{1/2}.$$ Equations (1)-(6) have been used to compare and fit the theoretical dependence of K and m$^*$ to available experimental data. \[sec:level2\]Numerical results\ ================================ By the least square method we have fitted simultaneously K as a function of $\Delta_{||}$ and m$^*(\Delta_{||})$ to the experimental data using five adjustable parameters: $P_{\perp}, P_{||}, {\cal E}_g, \delta, \Delta$ and obtained the following values: P$_{\perp} = (7.43 \pm 0.05)\times10^{-8}$ eVcm, P$_{||} = (7.21 \pm 0.05)\times10^{-8}$ eVcm, ${\cal E}_g = (-0.095 \pm 0.01)$ eV, $\delta = (0.085 \pm 0.01)$ eV, $\Delta = (0.27 \pm 0.03)$ eV. The negative value of ${\cal E}_g$ and positive value of $\delta$ indicate that Cd$_3$As$_2$ has the inverted band ordering with positive tetragonal field splitting. Theoretical dependence of K$^2$ and m${^*}^2$ upon 1/$\Delta_{||}$ (and the corresponding electron concentrations) obtained for our band model and the determined band parameters are compared with experimental data in Figs. 1 and 2 (curves 1). We show also theoretical predictions for simple level ordering with ${\cal E}_g$ = +0.025 eV (curve 2) and ${\cal E}_g$ = +0.14 eV (curve 3) with other parameters unchanged. It should be emphasized that a negative value of $\delta$ would give an increase of anisotropy ratio K$^2$ with concentration and thus it has been ruled out. In Fig. 3 we plot the directional dependence of m$^*$. It is worth noting that the theoretical anisotropy of m$^*$ ($\approx$ 1.10) is considerably smaller than the anisotropy of K$^2$ ($\approx$ 1.4) for the same sample, which is in agreement with Rosenman’s data \[2, 3\]. For a simple generalization of Kane’s model with anisotropy independent of the energy both quantities would have the same anisotropy. \[sec:level2\]Energy band model for C$_3$A$_2$ ============================================== Solutions of the secular equation (1) for our band parameters are shown in Fig. 4 for three directions of the wave vector **k**. The behavior of energy with **k** may suggest an overlap of the conduction and valence bands. However, the directional dependence of the energy with **k** clearly indicates that the conduction and the first valence band come into contact only at two **k** values and the energy gap becomes zero at these two points. Flat parts of the conduction and first valence band are a consequence of neglecting higher levels, which would have given a finite curvature of the bands, similar to those of HgTe under tensile stress \[9\]. ![\[fig:epsart\][${\cal E}$ (**k**) relation for the four-level model of Cd$_3$As$_2$. The wave vector **k** (in $10^6$ cm$^{-1}$ units) is directed: a) $0^{\circ}$, b) $5^{\circ}$), c) $90^{\circ}$ from the tetragonal axis .The energy is in 10$^{-1}$ eV units. ]{}[]{data-label="fig4th"}](Bodnar_Fig4) The energy band model of Cd$_3$As$_2$ has been a controversial problem in the last years. Wagner et al \[10\] first suggested that Cd$_3$As$_2$ has the inverted level ordering. This was subsequently supported by Caron \[11\] and more recently by Cisowski and Arushanov \[12\] by measurements of the thermoelectric power under pressure. As follows from the above analysis, the best fit to the SdH and dHvA data is obtained with the inverted band structure and positive crystal field splitting. Other models do not give good fit to the experiment with reasonable values of the band parameters. I am pleased to thank Dr W. Zawadzki for his continuous help and interest in this work. I am also indebted to Prof. W. Zdanowicz, Dr J. Cisowski and Dr E. Arushanov for informative discussions. [99]{}\[sec:TeXbooks\] W. Zdanowicz and L. Zdanowicz, Annual Review of Material Science **5**, 301 (1975). I. Rosenman, J. Phys. Chem. Solids **30**, 1385 (1969). I. Rosenman, PhD Thesis, Paris 1967. H. Doi et al. Sci. Reports. Res. Inst. Tohoku Univ. **A20**, 190 (1069). E. O. Kane, J. Phys. Chem. Solids **1**, 249 (1959). J. Bodnar, Reports of Institute of Physics, Polish Academy of Sciences **69**, 96 (1976). H. Kildal, Phys. Rev. B **10**, 5082 (1974). I. M. Lifshitz and A.M. Kosevich, J. Phys. Chem. Solids **4**, 1 (1958). L. Liu and W. Leung, Phys. Rev. B **12**, 2336 (1975). K. J. Wagner, E. D. Palik and E. M. Swiggard, J. Phys. Chem. Solids Suppl. **1**, 471 (1971). L. G. Caron, J. P. Jay-Gerin and M. J. Aubin, to be published. J. Cisowski and S. Arushanov, private communication.

--- abstract: | This paper is the first of a series addressed to the investigation of galaxy formation/evolution in small scale systems of galaxies (SSSGs) which are located in low density cosmic environments. Our algorithm for SSSG selection, includes galaxy systems of 2 or more galaxies lying within $\Delta$$cz$$\leq$ 1000 km s$^{-1}$ and a 200 $h_{100}^{-1}$ kpc radius volume. We present the analysis of the photometric and spectroscopic properties of 19 member galaxies belonging to a sample of 11 SSSGs. In the $\mu_e$ – $r_e$ plane, early–type members may be considered “ordinary”, not “bright” galaxies in the definition given by @cap92 with a significant fraction of galaxies having a disk or disky isophotes. We do not detect fine structure and signatures of recent interaction events in the early–type galaxy population, a picture also confirmed by the spectroscopy. At odd, there are several spiral members with open arm configurations as expected in interacting systems. At the same time, emission lines in the spectra of spiral members fall in the HII regions regime defined with diagnostic diagrams [@vei87]. None of the objects displays unambiguous indication of nuclear activity, although four spiral nuclei could be ascribed to the class of Seyferts. The star formation rate seems enhanced over the average expected in spiral galaxies only for poorer SSSGs in particular pairs ($\leq$50 M$_\odot$ yr$^{-1}$) but without being in the range of starburst systems. author: - 'L. Tanvuia' - 'B. Kelm, P. Focardi' - 'R. Rampazzo' - 'W.W. Zeilinger' title: | Small scale systems of galaxies. I.\ Photometric and spectroscopic properties of members[^1] --- Introduction ============ Among cosmic environments, small scale systems of galaxies (SSSGs) in the field are those in which not only galaxy–galaxy interactions but also the evolution of cosmic structures can be studied at a “cellular” level. To the SSSG class of cosmic structure we may ascribe galaxy systems with different richness and density characteristics spanning from compact groups to poorer configurations like galaxy pairs. Together with compact galaxy groups [@hic97 and references therein], investigated in definitely more detail, recent X-ray observations suggest the physical reality of some loose groups [@pon96; @mul00] and even pairs [@hen99; @tri01]. The X-ray diffuse component and the plethora of faint galaxies associated to the few dominant members is interpreted as due to the presence of a deep potential well [@mul00]. Are then SSSGs long lasting associations or, in the hierarchical evolution scenario, the debris of older, richer and sparse configurations? Are pairs the debris of a pristine group and consequently a way station toward isolated Es? NGC 1132, an isolated elliptical with extended X-ray diffuse emission [@mul99], could be a prototypical example of the evolution of such systems. Some scenarios [@dia94; @gov96] depict rich and compact galaxy structures, like Hickson compact groups, as substructures of larger ones, with different degree of equilibrium. @zab98 found typically 20 to 50 dwarf galaxies associated to their X-ray detected groups. In this context, SSSGs of different richness and degree of compactness have to be studied comparatively as “single” class of cosmic environments. The understanding of galaxy–galaxy interaction phases and the evolution of cosmic structures are deeply interconnected. Galaxy encounters in SSSGs are less frequent than in galaxy clusters, but the low velocity dispersion of the SSSG may lead to efficient merging episodes [@bar96]. Interactions could severely alter the properties of a galaxy up to modify the original morphological class [@bar96; @ken96] and then re-direct the evolution of a galaxy triggering various phenomena ranging from star formation [@lon99] to galaxy activity [@mon94; @raf95; @lau95; @kee96; @kel98; @coz00]. In this picture, the environment plays a key role since it dictates the predominant type of encounters [@moo96; @bar96]. A connection should then exist between the local environment and the global properties of galaxies inhabiting it. For rich galaxy systems, many aspects of the connection between environment and galaxy properties have emerged, from radio [@hay84] to X-ray properties [@for82]. The link between the environment and the galaxy evolution in SSSGs is still not fully understood [@pon96; @lon98a; @lon98b; @lon99; @lon00; @ram00; @mul00; @coz00] and partly suffers a shortage of information with respect to richer environments. In order to understand which parameters dominate in galaxy interactions, we started a study of SSSGs, defined in 3D redshift space. SSSGs have been selected in a low density environment, i.e. they appear as overdense systems with respect to the average galaxy distribution. The present paper collects photometric and spectrophotometric data for the main members of a sample of SSSGs. Within the picture described above, dominant galaxies in SSSG have potentially the power of revealing the history and evolutionary phase of the respective SSSGs. Elliptical galaxies may be at the center of the potential well of a SSSG as suggested by X-ray observations [@mul00] and could present fine structures [@sch96] reminiscent of past interaction events. Spiral galaxies in mixed pairs could be the last un–digested member of a pre–existing group [@ram92] and could reveal their on–going interaction not only through morphological distortion but also in their possibly triggered activity. The goals of the present paper are the following: 1) to obtain a redshift estimate which will provide an independent check of the systemic velocities of the SSSG members typically made available by large redshift surveys, 2) to investigate the structure of the member galaxies through a detailed surface photometry, 3) to analyze the possible induced activity using medium resolution spectroscopy and diagnostic models. The paper is organized as follows: The sample is defined in § 2. A description of the photometric and spectroscopic observations, data reduction and analysis is given in § 3, while results of individual objects are presented in § 4. The global morphological, photometric and spectroscopic properties of the present sample as a function of the SSSG and SSSG member properties are discussed in § 5. The sample definition ===================== The SSSG sample includes 11 systems of 2 or more galaxies lying at similar redshift ($\Delta$cz $\leq$ 1000 km/s) and within a 200 $h^{-1}$kpc radius area, for which new spectroscopic or photometric data have been acquired. These confirmed SSSGs constitute a small subset of a SSSGs candidate sample which includes likely misclassified isolated galaxies in ZCAT [@huc92]. The version of ZCAT used for selection of candidates contains 57536 entries. The SSSGs candidate sample was selected with a two-step procedure: first, isolated galaxies were identified in ZCAT with an automatic code defining as isolated all those galaxies presenting no neighbour(s) with known redshift within $\Delta$cz $\pm$ 1000 km/s and $\Delta$R = 1$h^{-1}$ Mpc. With this procedure, 3890 galaxies were selected in the the redshift interval $3000 < cz < 10000$ km s$^{-1}$ out of a sample of 18677 galaxies in total. ZCAT is essentially a redshift compilation rather than a complete sample and also includes a $\approx$25% fraction of galaxies with unknown redshift, Therefore, many of the selected isolated galaxies presented one or more projected companion(s), and are therefore likely misclassified isolated galaxies. To account for this, the code has identified, in the second step, among the isolated galaxies, those presenting nearby projected neighbours (i.e. neighbours whose redshift is unknown), which are clearly likely to represent in fact small scale galaxy systems. This resulted in 423 candidate SSSGs. All candidates lying within 1 Abell radius and $\Delta$cz $\leq$ 1000 km/s from ACO clusters [@str99] have been excluded from the sample. We have also inspected DSS images to select among the projected pairs and groups those appearing outside clusters and dense large groups, and are consequently suited for photometric and spectral inspection. Once systems inside clusters are rejected one can safely assume the SSSG sample includes only systems in low density environment. However, we expect that SSSGs reside in environments not as sparse as those typically associated to single galaxies, based on the result by @foc02, showing (in their figure 10) that compact groups display a significant excess of neighbours with respect to galaxies lacking a close neighbour. The sample of galaxy systems we present here is very small: it covers a large range in radial velocity and absolute magnitude, and is partially based on catalogues that are not complete. Further it has been defined making also use of non–automatic selection criteria, causing an intrinsic and hardly quantifiable bias. Therefore, we do not expect the sample to be representative of all galaxy systems in low density environment. It seems however, that the sample still provide enough information that some general conclusions might be drawn, in particular concerning the different behaviours of early–type and late–type galaxies in pairs and groups detected in low density environments. The main data for the 11 confirmed SSSGs investigated here are presented in \[table-1\]. We have checked the NED for SSSGs neighbours with known redshift out to a distance of 600$h_{100}^{-1}$kpc and 1$h_{100}^{-1}$ Mpc respectively (see Table \[table-1\]). Though obtained from a data base which is clearly not complete, this information is listed to allow a quantitative first order approximation of the galaxy density around each SSSG. All SSSGs are in environments including less than 10 galaxies within a projected area of 1h$_{100}^{-1}$Mpc and a redshift range of $\pm$1000 km s$^{-1}$ from the SSSG center. This is used to argue that SSSGs are indeed in a low density environment, ranging from completely isolated to that typical for loose groups. The observed sample contains additionally 2 SSSGs which have not been investigated in detail, namely CGCG 054-011 ($\alpha$(2000) 17 15 08.7 $\delta$(2000) 08 27 22) and CGCG 054-013 ($\alpha$(2000) 17 15 11.9 $\delta$(2000) 08 25 35). According to literature data, they satisfied our selection criteria. However, our spectroscopic investigation revealed that they are an optical alignment having V$_{hel}$=10001$\pm$11 km s$^{-1}$ and V$_{hel}$=6444$\pm$11 km s$^{-1}$ respectively. They are no further considered in the paper. Observation and reduction ========================= Grey-scale images of the SSSG sample obtained with the 0.91m Dutch telescope at ESO La Silla are presented in Figure \[figure-1\]. Spectroscopic observations and analysis --------------------------------------- Long-slit spectra were acquired at ESO 1.52m telescope at La Silla equipped with a Boller & Chivens Cassegrain spectrograph during an observing run in 1996. The spectra of the SSSG member galaxies were obtained in the wavelength region 3500 $\leq \lambda \leq$ 11000 Åwith a dispersion of 3.69 Å pixel$^{-1}$. The slit (slit width 2) has been oriented along the line connecting the nuclei of two of the SSSG members in order to optimize exposure times. The detector used was a FA 2048L UV–coated CCD (ESO CCD \#15.) The members which have been observed within each SSSG are reported in Table \[table-2\]. The observing log of the observations is given in Table \[table-3\]. The spectra were calibrated with the ESO-MIDAS[^2] software package using standard procedures for bias subtraction and flat field correction. Artifacts produced by cosmic ray events were removed by applying a filtering algorithm. Wavelength calibration was performed on the frames by fitting a third order polynomial, using as reference the helium–argon spectrum taken before each object spectrum. The spectra were flux calibrated with the IRAF[^3] package KPNOSLIT. This procedure included the airmass and extinction corrections using standard stars and the atmospheric extinction coefficients. The flux calibrated, de–redshifted spectra are presented in Figure \[figure-2\]. Heliocentric systemic velocities have been obtained through a cross–correlation technique using the IRAF task CROSSCOR or through the interpolation of emission (or absorption) lines with a single gaussian fit, when the cross–correlation yielded no results because of low signal–to–noise absorption line data. Our redshift measurements were compared with data available in the literature. We found an average systemic velocity difference of 84 km s$^{-1}$ with a standard deviation of 132 km s$^{-1}$. The integrated flux of the prominent emission lines (H$\beta$, \[OIII\] $\lambda$4959Å, $\lambda$5007Å, H$\alpha$, \[NII\] $\lambda$6548Å, and \[SII\] $\lambda$6717Å) was measured using an interactive gaussian fitting procedure and is reported in the Table \[table-4\]. Photometric observations and analysis ------------------------------------- Imaging was carried out with the 0.91m Dutch telescope at ESO, La Silla, Chile, during a single run in 1996. Bessel $R$ band images were obtained under homogeneous observing conditions. The detector employed was a SITe 512$\times$512 pixel CCD (ESO CCD \#33) with a scale of 0.442$''$ pixel$^{-1}$ yielding a field of view of 3.8$\times$3.8. The image cleaning, dark and bias subtraction, flat–fielding, cosmic ray removal and the final calibration and image manipulation were performed using the ESO-MIDAS software package. Standard stars used for calibration purposes were obtained in the same instrumental set–up in the fields of PG1633+099 (4 stars), SA 110 (4 stars), Markarian A (4 stars), T–Phe (3 stars). Each of these fields contains several calibration stars, avoiding the crowding problem typical of fields surrounding globular cluster areas. We used the standard photometric transformation equations tailored to the ESO (La Silla) system of extinction coefficients. Ellipses were fitted by weighted least squares to the isophotes of the SSSG member galaxies using the IRAF ISOPHOTE package [@jed87] within STSDAS. Surface brightness, position angle and ellipticity profiles together with the Fourier–coefficients to quantify the deviations of the isophotes from pure ellipses were derived for the early–type members. The $b_4$ coefficient is in particular used to analyze the boxiness/diskiness of the early-type galaxies. In the case of late–type disk galaxies only the surface brightness, ellipticity and position angle profiles are used for the further analysis. Furthermore, in order to avoid artifacts due to the presence of spiral arms and bright HII regions, the ellipse fitting procedure was repeated iteratively until a satisfactory representation of the stellar disk component could be achieved. Using the parameters extracted from the surface photometry, a smooth model of the galaxy has been produced and subtracted from the original image in order to evidence faint structures [@sch92] possibly connected to on–going/past interaction episodes. The profiles outside the seeing dominated central region are presented in Figures \[figure-3\] and \[figure-4\]. The observing log and the results from surface photometry are summarized in Table \[table-5\]. Further, the imaging data was used to establish morphological types for those galaxies, where either no classification is available in NED or the present data reveal a a morphology in conflict with NED. Results and Comments on individual objects ========================================== In this section we comment about particular features of the SSSGs in the sample and on individual galaxies in the SSSGs recovered from surface photometry and spectroscopy. The galaxies are members of WBL 637 [@whi99] which is formed by 4 objects of which SSSG1a and SSSG1b are WBL members 002 and 003, respectively. The WBL determination agrees with our redshift study of possible neighbours as shown in Table 1. Both galaxies have early–type morphology. No emission lines are detected in the spectra of both members. Surface photometry of SSSG 1a reveals an early–type disk system seen nearly edge–on. The stellar disk component appears prominent both in the residual image and the $b_{4}$ profile. SSSG 1b is an elliptical galaxy, which shows a moderate, $\approx$ 10$^\circ$, isophote twist, while the ellipticity increases up to $\approx$0.4. The Fourier coefficient $b_{4}$ reveals a small disk component in the range 4$\leq r \leq$ 10. The presence of the inner stellar disk is also visible in the surface brightness profile. A diffuse structure is present in the residual image. A plethora of dwarf galaxies is detected nearby SSSG 1b. According to the systemic velocities of SSSG 1a and SSSG 1b, we suggest that the objects are physically paired, although we do not notice obvious signatures of interaction. Spectra, shown in Figure \[figure-2\], are typical of their morphological class. This system is a chain of 4 bright galaxies with a systemic velocity difference less than 140 km s$^{-1}$. Galaxies in this poor association appear also in the WBL catalog (WBL 642). The chain, from the north to the south, is composed of a Sc face-on spiral with multiple arms (SSSG 2d), followed by two early–type galaxies (SSSG 2b, SSSG 2a). The fourth member (SSSG 2c) is a face–on Sc spiral with two prominent open arms but diffuse patchy areas, some of them marking incipient spiral arms. SSSG 2a (E/S0) shows a rising ellipticity profile and a positive (up to 4%) Fourier coefficient $b_{4}$ indicating the presence of an inner stellar disk. No emission lines are detected. Three faint objects (Figure \[figure-1\]b) are nearby to SSSG 2a in projection, two north and one south of it. The northern object of the two is compact, with faint and narrow emission lines. The H$\alpha$/\[NII\] $\lambda$6583Å line ratio indicates a Seyfert 1–like AGN. The radial velocity is 2159$\pm$54 km s$^{-1}$. The second object is blue, with faint and narrow lines. The radial velocity is 2165$\pm$262 km s$^{-1}$. The southern object is an edge–on disk galaxy with a radial velocity of 6717$\pm$59 km s$^{-1}$ suggesting a physical association with SSSG 2a. SSSG 2b (E/S0) shows a low ellipticity $\epsilon \approx$ 0.2 and a strong isophotal twist of $\approx$ 120$^\circ$. The Fourier coefficients $a_{4}$ and $b_{4}$ indicate the presence of a stellar disk component in the inner part of the galaxy. SSSG 2c is a face–on late–type spiral galaxy with multiple arms of which two dominate. Surface brightness and position angle profiles reveal the presence of a small exponential bar component in the region 4$'' \leq r \leq $ 13$''$. The residual image shows a number of smaller spiral arms. SSSG 2d is a late–type face–on spiral with flocculent arms. Some of them mark incipient spiral arms. A number of HII regions appears in the residual image after subtracting a smooth galaxy model. SSSG 2c and SSSG 2d show emission lines ratios characteristic of HII regions. A number of faint galaxies is noted in the immediate surroundings of the bright galaxies. Of particular interest is a faint disk galaxy (Figure. \[figure-1\]c) which is detected between SSSG 2b and SSSG 2d. The object shows strong and narrow emission features in the spectrum, with line ratios consistent with typical HII regions. The measured radial velocity is 6626$\pm$64 km s$^{-1}$ in agreement with the systemic velocity of SSSG 2. The isolated multiplet consists of 4 confirmed galaxies with 5 possible neighbours within 1 Mpc. The two early–type member galaxies of the multiplet are seen nearly edge–on. SSSG 3a has been classified as E6 according to NED. However, the surface brightness profile reveals a bulge and a stellar disk component suggesting therefore an E/S0 type. The presence of a stellar disk component is further confirmed by an overall positive $b_{4}$ coefficient and evidenciated in the residual image. SSSG 3b has a constant position angle profile and the ellipticity increases up to 0.6. The surface brightness profile also shows the presence of a bulge and a disk component supporting the classification of S0:Sp from NED. The latter is also revealed in the residual image. The galaxy is surrounded by faint objects possibly dwarf galaxies. No emission lines are detected in SSSG 3a and SSSG 3b as expected from their morphological class. There are no morphological signatures of interactions detected between the galaxies. . The pair also known as KPG 551 [@kar72] shows pronounced signatures of interaction. The southern object SSSG 4a appears strongly interacting and is of irregular type. The northern object SSSG 4b, which is classified as SAB pec, shows an arc–like tail/arm extending to the northeast of the galaxy. This structure contains numerous HII regions [@jun98]. We measured a virtually null systemic velocity difference as reported in the literature. Both members show prominent emission lines. The apparent peculiarities of this isolated pair prevent us to perform a detailed surface photometry. Between the SSSG 4a and SSSG 4b there is a faint object, with narrow emission lines. No absorption lines were detected. The measured radial velocity is 6954$\pm$127 km s$^{-1}$. The isolated pair ($\Delta V$=104 km s$^{-1}$) is composed of an edge-on lenticular and a grand design spiral galaxy. The S0 galaxy SSSG 5a shows a two component surface brightness profile consisting of a bulge and stellar disk component. The ellipticity profile rises up to 0.5 and the position angle profile is nearly constant at $\approx$ 80$^\circ$. The Fourier coefficient $b_{4}$ further evidences the disk component ($\approx$ 8%). No emission lines are detected in the spectrum. The presence of strong spiral arms in SSSG 5b prevented a detailed surface photometric analysis. The galaxy spectrum shows faint emission lines. The members of the multiplet appear to be quite separated from each other. SSSG 6 is composed of three late–type members: SSSG 6a, seen almost face–on, SSSG 6b and a third object, for which no spectroscopic data are available and which is seen almost edge–on. The derived H$\alpha$/\[NII\] $\lambda$6583Å line ratio for the nucleus of SSSG 6a suggests a Seyfert 1–like AGN. This galaxy is also detected by IRAS (IRAS 00000-0359). The presence of strong spiral arms in SSSG 6a prevent us to perform a detailed surface photometry. A number of HII regions appears in the residual image after substraction of a smooth model galaxy. SSSG 6b appears to be distorted with two inner main arms and outer arms completely decoupled. The difference of their systemic velocities is $\approx$200 km s$^{-1}$ suggesting a physical association. The pair is composed of an edge–on lenticular with a thick disk and an early spiral galaxy with very thin spiral arms showing patchy HII regions. The systemic velocity difference between the members is only $\approx$70 km s$^{-1}$. SSSG 7a shows a rising ellipticity and a constant position angle profile. The $b_{4}$ profile indicates a strong (8%) disk component. The photometric data are influenced by a bright star in the vicinity. No spiral arm structure is visible suggesting therefore an S0 type instead of the classification as S? in the NED. SSSG 7b shows faint spiral arms after substraction of a smooth model galaxy. These arms are more reminiscent of a sort of shell structure created by weak interaction [@tho91; @wei93] rather than typical spiral arms. The H$\alpha$ intensity is approximately similar to that of \[NII\] in the galaxy center which indicates the presence of a strong H$\alpha$ component in absorption. The H$\alpha$/\[NII\] $\lambda$6583Å line ratio indicates an AGN of Seyfert 1 type. The loose pair consists of two barred spirals seen almost face-on. The imaging data reveal a population of faint objects, possibly dwarf galaxies, surrounding the bright SSSG members. A spectrum was obtained only for the SSSG 8b, which reveals emission lines typical for HII regions. The triplet is composed of an unperturbed elliptical and two spiral galaxies. SSSG 9a is a face–on spiral showing multiple arm structure, a small bulge and a prominent disk component. The analysis of the surface brightness profile reveals the presence of an elongated component (possibly a bar component) in the range 10$\leq r \leq$ 20. The spiral arms are very prominent with a number of large HII regions particulary in the southern arm. SSSG 9b is an apparently undisturbed E0. A spectrum was obtained of the elliptical (SSSG 9b) and the brighter spiral galaxy (SSSG 9a). The difference of the systemic velocities between these two galaxies is 122 km s$^{-1}$. It is noticeable that the stellar and gas components in SSSG 9a differ in the systemic velocity of about 80 km s$^{-1}$. The systemic velocity of the gas component is closer to the systemic velocity of SSSG 9b. @pie00 performed a photometric study of the pair which appears completely damaged by the encounter. They report that a complex system of debris extend from the galaxy in south–east direction with intense knots. SSSG 10a is of irregular shape shows strong and narrow emission lines in the spectrum. SSSG 10b is a typical S0, with a dust lane in the center and appears slightly damaged by the encounter in the eastern outer part. The line ratio (H$\alpha$/\[NII\] $\lambda$6583Å)$<$1 of the SSSG 10b is consistent with that of typical Seyfert galaxy, no \[OIII\] $\lambda$5007[Å]{} emission is detected. The pair does not present evident signatures of interaction as also reported by @pie00. SSSG 11a is a late–type spiral seen edge–on. Strong emission lines are detected in the spectrum. Also absorption lines are present, but they appear very faint. The northern member, SSSG 11b, is a faint spiral. The spectral features are too faint to be further analyzed. Properties of the morphological classes ======================================= Properties of early–type galaxies --------------------------------- The early–type galaxies of the SSSG sample have been searched for the presence of fine structures, using the scheme developed by @sch92. However, no significant structures could be detected. This may indicate that either these galaxies remained undisturbed during recent interactions within the SSSG or the interactions occured only on small scale. Such events could be mass accretions of (gas–rich) dwarf galaxies resulting in kinematically decoupled components which are frequently observed in early–type galaxies. The presently available observational material however has not sufficient resolution to answer this question. The Hamabe–Kormendy relation HK87 [@ham87] is a useful tool to study distributions of early–type galaxies using the classification by @cap92, hereafter CCD92, in ordinary and bright classes. Bright galaxies are defined as those having M$_B$ $<$ -19.3 and R$_e$ $>$ 3 kpc (H$_0$=70 km s$^{-1}$ Mpc$^{-1}$) while $\mu_e$=2.94log R$_e$ + 20.75 traces the HK87 relation. CCD92 show that ordinary galaxies do not tend to distribute along the HK87 relation but fill the plane for R$_e$ $<$ 3 kpc. The ordinary galaxy class is considered as a sort of “genetic variety” since it contributes to generate bright galaxies through merging processes [@nav90]. In order to transform our data from R band to B band we adopt a (B–R) = 1.5, i.e. the color of a SSP of solar metallicity (Z=0.02) and of an age of 13–15 Gyr. Figure \[figure-5\] shows the $\mu_e$ – log R$_e$ plane. The vast majority of our objects resides in the range of ordinary galaxies. This is not the general case of galaxies in low density environments as shown by @ram00. They noticed also that there is a number of bright galaxies well below the relation. These latter have been interpreted as a transient phase since they include strongly interacting members. These galaxies completely lack in our sample. This may suggest that SSSGs are either accordant redshift unrelated galaxies or that they are still in a stage of pre-coalescence and have not yet produced the bright merger remnants. Properties of spiral galaxies ----------------------------- The late–type galaxies of the SSSG sample show clear signatures of on–going interaction in contrast to the early–types. Following the classification of @elm87 (see Table \[table-6\] for classification) most spiral galaxies belong to the grand design class typically found in the densest group environments. At the same time, a significant number of disk galaxies is found in comparatively low–density environments as suggested by the rather small percentage of barred galaxies present in the sample if compared with those detected in binary samples by @elm82 and @red95. A large fraction of spirals, which are not seen edge–on, show open arms indicating an on-going interaction. According to @nog86 models, open arms develop in the very early phases of an encounter. Star formation rate reaches the maximum value of about 8 times as large as the pre–encounter value at about $3 \times 10^8$ years after the perigalactic passage of the perturber. In this context we suggest that our mutiplet SSSGs could be early phases in the hierarchical assembling process. The emission line spectra were used to study the ionisation mechanisms. The emission line intensity ratios, as defined by @vei87, \[OIII\]/H$\beta$ vs. \[NII\]/H$\alpha$ and \[OIII\]/H$\beta$ vs. \[SII\]/H$\alpha$ of the sample galaxies are plotted in Figure \[figure-6\]. @vei87 defined the intrinsic flux ratio for HII region–like objects I(H$\alpha$)/I(H$\beta$) = 2.85 and adopt for the intrinsic ratio of AGNs I(H$\alpha$)/I(H$\beta$) = 3.1. The values of the flux ratio in the sample are found to be generally bigger and using the two definitions of @vei87, the value of the flux intensity of H$\beta$ is corrected. This is attributed to the fact that the H$\beta$ flux is underestimated due to the presence of an absorption line component. The new value of the flux intensity of H$\beta$ is used to compute new \[OIII\]/H$\beta$ vs. \[NII\]/H$\alpha$ and \[OIII\]/H$\beta$ vs. \[SII\]/H$\alpha$ ratios (see errorbars in Figure \[figure-6\]). Figure \[figure-6\] shows that late–type galaxies in the sample are dominated by photoionisation: all objects lie in the region of the diagram, which is dominated by HII–regions. This is the case also for morphologically peculiar galaxies, as e.g. SSSG 4b and SSSG 6b and even strongly distorted galaxies, e.g. SSSG 4a and SSSG10a. Star formation rates (SFR) were calculated, using the calibration between SFR and H$\alpha$ fluxes adopted by @ken98 (see Table \[table-5\]): $SFR (M_\odot yr^{-1})= 7.9 \times 10^{-42} \times L_{H_\alpha} (erg s^{-1})$ Table \[table-4\] (column 12) collects the SFR for the disk galaxies of the sample. Most disk galaxies (9 out of 14) can be considered “normal”, with star formation rates up to 8 M$_\odot$ yr$^{-1}$. The remaining 5 objects have a significantly higher star formation rate. However only one exceeds the value of 50 M$_\odot$ yr$^{-1}$, which indicate a major starburst phenomenon. It is worth remarking that the 5 disk galaxies presenting the higher SFR are hosted in the more isolated SSSGs of the sample (see Table \[table-1\]), and in SSSGs containing only spiral galaxies. Summary ======= In this paper we have studied 19 members in 11 SSSGs and presented their photometric and spectroscopic properties. Our selection criteria have included small systems ranking from isolated pairs (KPG 551, RR23 and RR45) with no neighbours of similar luminosity to poor groups in low density environment. The groups are high density configurations and some of them are catalogued in the literature as nearby poor clusters, as in the case of SSSG 1 (WBL 637) and SSSG 2 (WBL 642). Although the relatively small number of objects studied so far does not allow to draw statistically significant conclusions some trends can still be pointed out. Early–type galaxies in SSSGs, even those which are found in more compact multiplets, do not show relevant signatures of interaction. The lack of evidence of early–type galaxies being the result of major merger events is a feature in common with HCGs, which have a low fraction of merging candidates and no evidence of enhanced far–infrared emission [@zep91; @zep93; @ver98]. Further, the early–type galaxies of our sample do not show fine structure. This is not surprising, indeed @red96 found that e.g. shell structures are found four times less frequently in interacting pairs than in isolated objects of the same class. This may be attributed to two causes. First, none of our early–type members is suffering or has recently suffered a strong interaction episode. Most of our early–type members have a disk component or are disky according to the shape parameter $b_4$ which is often larger than 2%. Second, the galaxies suffer “weak” but multiple interactions that are likely to destroy shell structures [@tho90]. At variance with early–type galaxies spirals displaying patterns typical of ongoing interaction and high star formation rate are found in SSGSs. These SSGSs are the most isolated in the sample and additionally are the only ones displaying a total spiral population. This finding confirms previous results from studies of classical pairs in the @kar72 catalog [@xus91] and pairs catalogs in the southern hemisphere [@com94] reporting significant enhancements of star formation due to the infall of fresh gas triggered by the interaction. At the same time the lack of evidence for enhanced star formation in spirals in the densest SSSGs agrees with previous results on HCGs [@sul93; @ver98] and on UZC–CGs [@kel03]. Neither our SSSGs which are strongly interacting nor members in multiplets do show unambiguously the presence of nuclear activity. Further steps in the study of our sample of SSSGs are planned/on–going. Among these, to obtain X-ray imaging and a deeper mapping of SSSGs through wide field imaging data which will permit a better definition of their environment. The spectroscopic information are then a necessary information to understand the significance of the population of faint galaxies accompanying SSSGs in their evolution. This will allow to test directly small scale substructure formations theories. High resolution spectroscopy will give the necessary information (line–strength indices etc.) about single members evolution. We are deeply indebted to Dr. Luca Reduzzi which performed observations. RR acknowledges the kind hospitality of the Institut für Astronomie der Universität Wien during the preparation of the paper. LT and WWZ aknowledge the support of the Austrian Science Fund (project P14783). The research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with National Aeronautics and Space Administration. Barnes, J. 1996, Galaxies: Interactions and Induced Star Formation, Saas–Fee Advanced Course 26, 275 Combes, F., Prugniel, P., Rampazzo R. & Sulentic, J. W. 1994, , 281, 725 Capaccioli, M., Caon, N., D’Onofrio, M. 1992, Structure, Dynamics and Chemical Evolution of Early–Type Galaxies, ESO-EIPC Workshop, Danziger J. et al., 43 (CCD92) Coziol, R., Iovino, A. & de Carvalho, R. R. 2000, , 120, 47 Diaferio, A. Geller, M.J. & Ramella, M. 1994, , 107, 868 Elmegreen, D. M. & Elmegreen, B. G. 1982, , 201, 1021 Elmegreen, D. M. & Elmegreen, B. G. 1987, , 314, 3 Focardi, P. & Kelm, B. 2002, , 391, 35 Forman, W. & Jones, C. 1982, , 20, 547 Governato, F., Tozzi, P. & Cavaliere, A. 1996, , 458, 18 Haynes, M.P., Giovanelli, R. & Chincarini, G. 1984, , 22, 445 Hamabe, M. & Kormendy, J. 1987, Structure and Dynamics of elliptical Galaxies, IAU Symp. No. 127 (Princeton), ed. T. de Zeeuw, Dordrecht: Reidel, 379: (HK87) Henricksen M. & Cousineau, S. 1999, , 511, 595 Hickson P. 1997, , 35, 357 Huchra, J.P, Geller, M.J., Clemens C.M., Tokarz S.P. & Michel A. 1992, Bull. Inf. CDS, 41,31 Jedrzejewski, R. 1987, MNRAS 226, 747 Junqueira, S., de Mello, D. F. & Infante, L., 1998, , 129, 69 Karachentsev, I. D. 1972, Catalogue of isolated pair of galaxies in the northern hemisphere, Soob-shch. Spets. Astrofiz. Obs. 7,3 Keel, W. C. 1996, , 111, 696 Kennicutt, R. C. 1996, Galaxies: Interactions and Induced Star Formation, Saas–Fee Advanced Course 26, 1 Kennicutt, R.C. 1998, , 36, 189 Kelm, B., Focardi, P. & Palumbo, G.G.C. 1998, , 335, 912. Kelm, B., Focardi P. & Zampieri, A. 2003, Galaxy Evolution III: From simple Approaches to self consistent Models, ed. G. Hensler, Kluwer Academic Publisher, in press Laurikainen, E. & Salo, H. 1995, 293, 683 Longhetti, M., Rampazzo, R., Bressan, A. & Chiosi, C. 1998a, , 130, 251 Longhetti, M., Rampazzo, R., Bressan, A. & Chiosi, C. 1998b, , 130, 267 Longhetti, M., Bressan, A., Chiosi, C. & Rampazzo, R. 1999, , 345, 519 Longhetti, M., Bressan, A., Chiosi, C. & Rampazzo, R. 2000, , 353, 917 Monaco, P., Giuricin, G., Mardirossian, F., Mezzetti, M. et al. 1994, , 436, 576 Moore, B., Katz, N., Lake, G., Dressler, A. & Oemler, A., Jr. 1996, , 379, 613 Mulchaey, J. S. & Zabludoff, A. I. 1999, , 514, 33 Mulchaey, J. S. 2000, , 38, 289 Navarro, J. F. 1990, , 242, 311 Noguchi, M. & Ishibashi, S. 1986, , 219, 305 Pierfederici, F., Rampazzo, R. & Reduzzi, L. 2000, , Vol. 40, 84. Ponman, T. J., Bourner, P. D. J., Ebeling, H. & Böhringer, H. 1996, , 283, 690 Rafanelli, P., Violato, M. & Baruffolo, A. 1995, , 109, 1546 Rampazzo, R. & Sulentic, J. W. 1992, 259, 43 Rampazzo, R., D’Onofrio, M., Bonfanti, P., Longhetti, M. & Reduzzi, L. 2000, , Vol. 40, 63 Reduzzi, L. & Rampazzo, R. 1995, , Vol. 30., 1: RR95 Reduzzi, L., Longhetti, M. & Rampazzo, R. 1996, , 282, 149 Schweizer, F. 1992, Structure, Dynamics and Chemical Evolution of Early–type Galaxies, ESO–EIPC Workshop, Danziger et al., 651 Schweizer F. 1996 Galaxies: Interactions and Induced Star Formation, Saas–Fee Advanced Course 26, 105 Struble, M. F. & Rood, H. J. 1999, , 125, 35 Sulentic J.W. & de Mello Rabaca, D. 1993, , 410 520 Thomson, R. C. & Wright, A. E. 1990, , 247, 122 Thomson, R. C. 1991, , 253, 256 Trinchieri, G. & Rampazzo, R. 2001, , 374, 454 Veilleux, S. & Osterbrock D. E. 1987, , 63, 295 Verdes–Montenegro, L., Yun, M. S., Perea, J., del Olmo, A. & Ho, P. T. P. 1998, , 497, 89 Weil, M.L. & Hernquist, L. 1993, , 405, 142 Xu, C. & Sulentic J. W. 1991, , 374, 407 White, R. A., Blinton, M., Bhavsar, S. P. et al. 1999, , 118, 2014 Zabludoff, A. & Mulchaey, J. 1998, , 498, L5 Zepf, S. E. & Whitmore, B. C. 1991, , 383, 542 Zepf, S. E. 1993, , 407, 448 [rccccrl]{} 1 & 6588 & 4 & 1 & 5 & 172 & WBL 637\ 2 & 6688 & 4 & 2 & 3 & 113 & WBL 642\ 3 & 4578 & 4 & 0 & 5 & 168 &\ 4 & 5769 & 2 & 0 & 0 & 2 & KPG 551\ 5 & 8526 & 2 & 0 & 1 & 104 &\ 6 & 6297 & 3 & 0 & 1 & 132 &\ 7 & 8824 & 2 & 0 & 0 & 74 &\ 8 & 5433 & 2 & 1 & 3 & 88 &\ 9 & 4739 & 3 & 1 & 4 & 108 &\ 10 & 3630 & 2 & 0 & 0 & 9 & RR23\ 11 & 6081 & 2 & 1 & 4 & 489 & RR45\ [rrrlcl]{} 1a & 17 17 25.2 & 07 41 43 & CGCG 054-019 & S0 & 6730$\pm$12\ b & 17 17 33.4 & 07 39 43 & CGCG 054-020 & E & 6368$\pm$13\ & 17 17 13.3 & 07 44 29 & CGCG 054-018 & & 6536$^{4)}$\ & 17 17 44.9 & 07 36 30 & UGC 10789 & & 6719$^{4)}$\ 2a & 17 31 55.0 & 06 29 00 & CGCG 055-003 & E/S0 & 6846$\pm$13\ b & 17 31 54.2 & 06 30 07 & CGCG 055-004 & E/S0 & 6687$\pm$15\ c & 17 31 57.0 & 06 28 09 & CGCG 055-005 & SBc & 6632$\pm$174$^{2)}$\ d & 17 31 58.8 & 06 31 56 & CGCG 055-006 & Sc & 6587$\pm$34$^{2)}$\ 3a & 20 32 36.7 & 09 53 02 & NGC 6927A & E/S0 & 4568$\pm$191$^{2)}$\ b & 20 32 38.2 & 09 54 59 & NGC 6927 & S0:Sp$^{1)}$ & 4344$\pm$206$^{3)}$\ & 20 32 50.2 & 09 55 38 & NGC 6928 & SB(s)ab$^{1)}$ & 4707 $^{4)}$\ & 20 32 58.8 & 09 52 28 & NGC 6930 & SB(s)ab?$^{1)}$ & 4694$^{4)}$\ 4a & 20 59 46.9 &-01 53 16 & UGC 11657 & Irr & 5768$\pm$45$^{2)}$\ b & 20 59 48.3 &-01 52 23 & UGC 11658 & SAB(rs)pec?$^{1)}$ & 5770$\pm$7$^{2)}$\ 5a & 22 01 01.7 & 08 06 34 &A2158+0752 & S0 & 8474$\pm$11\ b & 22 01 10.8 & 08 07 32 &A2158+0753 & S & 8578$\pm$6$^{2)}$\ 6a & 00 02 34.8 &-03 42 38 &MCG -01-01-024 & SB(s)bc?$^{1)}$ & 6448$\pm$30$^{2)}$\ b & 00 02 38.5 &-03 37 51 &MGC -01-01-025 & S & 6242$\pm$12$^{2)}$\ & 00 02 48.7 &-03 36 21 &MCG -01-01-026 & S & 6202$^{4)}$\ 7a & 00 03 22.3 &-10 46 14 &MGC -02-01-012 & S0 & 8861$\pm$11\ b & 00 03 32.1 &-10 44 41 & NGC 7808 &(R’)SA0$^{1)}$ & 8787$\pm$19\ 8a & 01 51 34.0 &-08 23 56 &MCG -02-05-065 & SB & 5389$\pm$2$^{2)}$\ b & 01 51 27.0 &-08 30 20 & NGC 0707 & (R’)SAB(s)0$^{1)}$ & 5477$\pm$2$^{2)}$\ 9a & 02 37 34.7 &-11 01 34 & NGC 1010 & SB & 4588$\pm$11\ b & 02 37 38.9 &-11 00 20 & NGC 1011 & E0 & 4754$\pm$17\ & 02 37 49.8 &-11 00 39 & NGC 1017 & S & 4876$^{4)}$\ 10a & 01 14 20.1 &-55 24 02 & NGC 0454 NED01 & Irr & 3626$\pm$2$^{2)}$\ b & 01 14 25.2 &-55 23 47 & NGC 0454 NED02 & Irr & 3635$\pm$2$^{2)}$\ 11a & 02 06 22.2 &-36 18 01 & ESO 354- G 036 & Sc$^{1)}$ & 6325$\pm$17$^{2)}$\ b & 02 06 53.1 &-36 27 08 & NGC 824 & SB(RS)b$^{1)}$ & 5836$\pm$10$^{2)}$\ [rrcl|rrcl|rrcl]{} & 134 & 30 & a, b & [**5**]{} & 67 & 30 & a, b & [**9**]{} & 41 & 30 & a, b\ [**2**]{} & 147 & 30 & a, c & [**6**]{} & 4 & 30 & a & [**10**]{} & & 30 & a, b\ [**2**]{} & 30 & 30 & b, d & [**6**]{} & 98 & 30 & b & [**11**]{} & & 30 & a\ [**3**]{} & 11 & 30 & a, b & [**7**]{} & 56 & 30 & b & & & &\ [**4**]{} & 22 & 30 & a, b & [**8**]{} & 62 & 30 & b & & & &\ [rrrrrrrcccrr]{} 2c & $^3$ & & & 67.6 & 37.6 &   & & –0.26 & & & 8\ 2d & $^3$ & & & 25.6 & 8.8 & 7.5 & & –0.47 & –0.73 & & 3\ 4a & 120.0 & 81.5 & 250.0 & 495.0 & 58.0 & 54.6 & 0.32 & –0.93 & –1.03 & 4.13 & 43\ 4b & 85.0 & 17.5 & 37.0 & 348.0 & 89.4 & 41.4 & –0.36 & –0.59 & –0.92 & 4.09 & 30\ 5b &   & & & 25.9 & 17.1 &   & & –0.18 & & & 5\ 6a & $^3$ & & & 23.2 & 49.9 & 16.8 & & 0.33 & –0.14 & & 3\ 6b & 22.8 & 11.5 & 20.0 & 110.0 & 24.9 & 37.7 & –0.06 & –0.65 & –0.47 & 4.82 & 11\ 7b & $^3$ & & & 14.8 & 16.5 & $^3$& & 0.05 & & & 3\ 8b &   & & 36.1 & 23.4 & 21.0 & 3.1 & & –0.05 & –0.88 & & 2\ 9a & 29.7 & 2.0 & 2.6 & 137.0 & 45.2 & 23.1 & –1.06 & –0.48 & –0.77 &4.61 & 8\ 9b & $^3$ & & & 30.3 & 2.1 &   & & –1.15 & & & 2\ 10a & 575.0 & 272.0 & 769.0 & 1490.0 &204.0 & 105.0 & 0.13 & –0.86 & –1.15 & 2.6 & 51\ 10b & $^3$ & & 82.8 & 562.0 &566.0 & 54.6 & & 0.003 & –1.01 & & 19\ 11a & 15.7 & 27.6 & 28.7 & 63.5 & 13.5 & 13.4 & 0.26 & –0.67 & –0.68 & 4.04 & 6\ [rrrccccccl]{} 1a&1.4 & 30 & 13.13 & -20.89 & 12.30 & 22.02 & 18.73\ b& & & 13.25 & -20.84 & 7.00 & 20.66 & 18.67\ 2a&1.2/1.4 &2$\times$30 & 13.96 & -20.13 & 4.11 & 20.65 & 18.82\ b& & & 14.21 & -19.92 & 3.08 & 20.47 & 18.60\ c& & & 14.15 & -19.96 & & & 20.29\ d& & & 14.30 & -19.79 & & & 20.70\ 3a&1.4 & 30 & 15.02 & -18.21 & 2.31 & 19.55 & 19.02\ b& & & 13.91 & -19.25 & 8.39 & 21.31 & 18.56\ 4a&1.8 &3$\times$20 & 13.75 & -20.05 & & & 20.14\ b& & & 13.07 & -20.74 & & & 20.21\ 5a&1.2 & 30 & 14.83 & -19.78 & 3.74 & 21.12 & 19.30\ b& & & 15.52 & -19.15 & & & 20.44\ 7a&1.4 & 30 & 13.19 & -21.55 & 5.47 & 20.40 & 18.21\ b& & & 12.54 & -22.17 & 12.30 & 21.89 & 18.49\ 8a&1.4 & 40 & 14.14 & -21.00 & & & 19.72\ b& & & 13.73 & -19.83 & & & 19.61\ 9a&1.4 & 30 & 12.82 & -20.54 & 12.29 & 21.40 & 19.67\ b& & & 13.35 & -20.01 & 6.02 & 21.00 & 18.69\ [lcc]{} 2c & 9 & bar\ 2d & 3 &\ 5b & 12 &\ 7b & 8 &\ 8b & 12 & bar\ 9a & 6 & bar\ [^1]: Based on observations obtained at the European Southern Observatory, La Silla, Chile (Programme Nr. 57.B–036) [^2]: ESO-MIDAS is developed and maintained by the European Southern Observatory. [^3]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy Inc., under cooperative agreement with the National Science Foundation.

--- author: - 'C. Abate' - 'O. R. Pols' - 'R. G. Izzard' - 'A. I. Karakas' date: 'Received ...; accepted ...' subtitle: 'II. Statistical analysis of a sample of 67 CEMP-$s$ stars.' title: | Carbon-enhanced metal-poor stars: a window\ on AGB nucleosynthesis and binary evolution. --- Introduction {#intro} ============ The very metal-poor stars observed in the Galactic halo are of low mass and exhibit abundances of iron of approximately $[\mathrm{Fe}/\mathrm{H}]\lesssim-2.0$. Very metal-poor stars carry the fingerprints of the early stages of evolution of the Milky Way and therefore have been extensively studied by different surveys in the past two decades, with particular focus on chemically peculiar stars [e.g.: @BeersAJ1992; @Christlieb2001; @Frebel2006; @Yanny2009]. Among these, carbon-enhanced metal-poor (CEMP) stars are very metal-poor stars enriched in carbon. The observed fraction of CEMP stars in the halo increases with increasing galactic latitude and with decreasing metallicity, and varies between approximately $9\%$ at $[{\mathrm{Fe}}/{\mathrm{H}}]<-2$ and about $25\%$ at $[{\mathrm{Fe}}/{\mathrm{H}}]<-3$ [@Cohen2005; @Marsteller2005; @Frebel2006; @Lucatello2006; @Lee2013; @Yong2013II]. In the literature CEMP stars are generally defined by the carbon excess $[{\mathrm{C}}/{\mathrm{Fe}}]>1.0$, and are classified in groups according to the observed abundances of barium and europium, two heavy elements produced by the slow ($s$-) and the rapid ($r$-) neutron-capture process, respectively. The exact definitions vary between different authors [e.g. @BeersChristlieb2005; @Jonsell2006; @Aoki2007; @Masseron2010], and in this work we adopt the following classification scheme. - CEMP-$s$ stars are CEMP stars that satisfy the criteria $[{\mathrm{Ba}}/{\mathrm{Fe}}] > 0.5$ and $[{\mathrm{Ba}}/{\mathrm{Eu}}] > 0$. - CEMP-$s/r$ stars are CEMP-$s$ stars enriched in europium, i.e. $[{\mathrm{Eu}}/{\mathrm{Fe}}] > 1$. - CEMP-$r$ stars have $[{\mathrm{Eu}}/{\mathrm{Fe}}] > 1$ and $[{\mathrm{Ba}}/{\mathrm{Eu}}] < 0$. - CEMP-no stars do not exhibit enhanced abundance of barium, i.e. $[{\mathrm{Ba}}/{\mathrm{Fe}}] < 0.5$. It has been suggested that CEMP stars owe their abundances to mass transfer of carbon-rich material in the past from a thermally-pulsing asymptotic giant branch (TP-AGB) primary star that today is an unseen white dwarf [e.g. @Wallerstein1998; @Preston2001; @BeersChristlieb2005; @Ryan2005]. In this work we focus on CEMP-$s$ stars, for which there are stronger arguments in favour of the binary mass-transfer scenario. Qualitatively, in this scenario the primary star evolves to the AGB phase, produces carbon and heavy elements and subsequently transfers part of this material by wind mass transfer to the low-mass companion star, the star observed today. A quantitative model of this process depends on many aspects of stellar evolution, AGB nucleosynthesis and binary interaction that are not well understood. In AGB stars many uncertainties are related to the physics of mixing, which determines the stellar structure and chemical composition and which in turn influence the radius of the star, and hence its luminosity and mass loss rate [@Herwig2005; @Constantino2014; @Fishlock2014]. The material that is expelled by the AGB star carries away angular momentum from the system and its fate depends on the mass transfer mechanism. Therefore the study of CEMP-$s$ stars provides us with a powerful tool to improve our understanding of AGB nucleosynthesis at low metallicity and of the mass-transfer process in binary systems. In our previous paper [@Abate2015-2 hereinafter Paper I] we analyse a sample of 15 CEMP binary stars with known orbital periods: through the comparison with the observed abundances while matching the measured periods we put new constraints on our models of binary stellar evolution and AGB nucleosynthesis. In most of the systems a combination of large mass accretion and efficient angular momentum loss is necessary to reproduce at the same time the observed chemical abundances and orbital periods. About half of the stars of the sample are not accurately reproduced by any of our models, regardless of the assumptions made about the dynamical evolution of the systems, and this points to the limitations in our AGB nucleosynthesis model, particularly on the maximum abundances of the heavy $s$-process elements and on the element-to-element ratios that are produced. In this work we extend the analysis of Paper I to a larger sample of $67$ CEMP-$s$ stars that includes systems without information about the orbital period. We model a grid of about $400,\!000$ binary stars with different masses and separations and we compare the modelled abundances with the observations. For each star in the sample we determine the model star that best matches the observed abundances, with the same procedure of ${\chi^2}$ minimization followed in Paper I. In every star for each observed element we compute the residual as the difference between the observed and the modelled abundance and subsequently we analyse the distributions of the residuals for every element individually. The purpose of this analysis is to investigate if statistically our model reproduces the observations, even though discrepancies may occur for some elements in individual stars, and thus to test our models with a study complementary to that of Paper I. The paper is organised as follows. In Sect. \[data\] we describe our observational sample, and in Sect. \[model\] we present our model of binary stellar evolution and nucleosynthesis. In Sect. \[VMP\] we discuss the set of initial abundances adopted in our model. In Sect. \[CEMPs\] we present the results of the analysis of our sample CEMP-$s$ stars, we compare the modelled and observed abundances of every element individually and we discuss the distribution of the residuals. In Sect. \[init-param\] we investigate the confidence limits on the initial parameters of our best fits and we study the distribution of these parameters. In Sect. \[discussion\] we discuss our results while Sect. \[conclusions\] concludes. ![$[{\mathrm{C}}/{\mathrm{Fe}}]$ vs $[{\mathrm{Fe}}/{\mathrm{H}}]$ in the stars of our observed sample. Grey plus signs () are carbon-normal, very metal-poor stars; CEMP-$s$ stars are classified in three groups according to the abundance of europium, as described in the text: CEMP-$s/nr$ stars (), CEMP-$s/ur$ stars ($\times$) and CEMP-$s/r$ stars (). Purple triangles () and grey dots () are CEMP-no stars and CEMP stars without measured abundances of heavy elements, respectively. The dotted line indicates $[{\mathrm{C}}/{\mathrm{Fe}}]=1$.[]{data-label="fig:datasample"}](Fig-1.pdf){width="48.80000%"} Data sample {#data} =========== Our database of observed very metal-poor stars is based on 580 stars catalogued in the SAGA observational database [@Suda2008; @Suda2011 last updated in January 2015] with iron abundance $-2.8\le[{\mathrm{Fe}}/{\mathrm{H}}]\le -1.8$. Among these objects we select the stars with observed abundances of carbon and barium and we ignore stars with only upper or lower limits. In some stars measurements of element abundances or stellar parameters are available from multiple sources. In most cases the measurements are consistent within the observational uncertainties and for the purpose of our study we use the arithmetic mean of the logarithm of the observed abundances and we adopt the largest observed error as the uncertainty on the measure. In case two measurements differ by more than the declared observational uncertainty and there is no obvious criterium to prefer one value, we compute the average and we adopt as the uncertainty half the difference between the two values. To this sample we add four CEMP stars studied by [@Masseron2010] with metallicity in the above range that were not present in the SAGA database. This selection leaves us with a sample of $378$ very metal-poor stars, $67$ of which are classified as CEMP-$s$ stars, $8$ as CEMP-no stars, $46$ are CEMP stars with no information about the abundances of neutron-capture elements, and the remaining $257$ are carbon-normal very metal-poor stars. In Fig. \[fig:datasample\] we show the carbon abundances of the stars in our observed sample as a function of $[{\mathrm{Fe}}/{\mathrm{H}}]$. CEMP-no stars are indicated as grey triangles. The dotted line represents the threshold carbon abundance $[{\mathrm{C}}/{\mathrm{Fe}}]=1$ above which the stars are defined CEMP stars. We classify CEMP-$s$ stars in three groups based on the abundance of europium: - In CEMP-$s/r$ stars the abundance of europium is enhanced, $[{\mathrm{Eu}}/{\mathrm{Fe}}]>1$ (open squares). - In CEMP-$s/nr$ stars the europium abundance is $[{\mathrm{Eu}}/{\mathrm{Fe}}]\le1$ (filled circles). - In CEMP-$s/ur$ stars the abundance of europium has not been determined (crosses), for example because the spectra have low signal-to-noise ratios, or the europium lines are blended. Table \[tab:obs\] summarises the surface gravities, temperatures, and abundances of iron, carbon, barium and europium (when available) of the $67$ CEMP-$s$ stars in our observed sample. For three stars in which barium is not observed and lanthanum is enhanced (HD$13826$, HD$198269$ and HD$201626$) the abundance of lanthanum is listed. The minimum uncertainty assumed in the chemical abundances and surface gravities is $0.1$ dex. We adopt an uncertainty of $100$ K in effective temperature unless differently stated. Models of binary evolution and nucleosynthesis {#model} ============================================== As in Paper I, in this study we use the code `binary_c/nucsyn` that couples algorithms to compute the evolution of stars in binary systems with a model of stellar nucleosynthesis. The details of our code and the prescriptions used for the binary stellar evolution and nucleosynthesis are extensively discussed by [@Izzard2004; @Izzard2006; @Izzard2009]. Our default input physics is the same as in Paper I, and we refer to this for a more complete description. In this section we summarise some important parameters adopted in our model, related to the wind mass-transfer process and the nucleosynthesis in the AGB phase (Sect. \[wind\] and \[nucsyn\]), and we describe the basic characteristics of our grid of model stars (Sect. \[grid\]). Wind-accretion rate and angular momentum loss {#wind} --------------------------------------------- In this work we compare the results derived with three different model sets based on different assumptions about the wind-accretion rate and the mechanism of angular momentum loss. The model sets are listed in Table \[tab:models\]. Model sets A and B are the same as in Paper I. In our default model set A we calculate the wind-accretion rate according to the prescription for wind Roche-lobe overflow (WRLOF) with a dependence on the mass ratio, as presented by @Abate2013 [Eq. 9]. We compute the angular momentum carried away by the wind material assuming a spherically symmetric wind [Eq. 4 of @Abate2013]. The results of Paper I show that a mechanism of wind accretion that is much more efficient at relatively short separations is necessary to reproduce the abundances observed in two thirds of the analysed CEMP-$s$ stars. To take this into account, in model set B we adopt the Bondi-Hoyle-Lyttleton (BHL) prescription as in Eq. (6) of [@BoffinJorissen1988] with $\alpha_{\mathrm{BHL}}=10$ (instead of $1\le\alpha_{\mathrm{BHL}}\le2$) to simulate a very efficient mass-transfer process. Model set B also adopts a prescription of efficient angular momentum loss, in which the material lost from the binary system carries away a multiple $\gamma = 2$ of the average specific orbital angular momentum [as in Eq. 2 of Izzard et al., 2010, and Eq. 10 of @Abate2013]. Model set C combines the WRLOF accretion rate as in model set A with the efficient angular momentum loss as in model set B. In the analysis performed by [@Abate2013] this is the model set that predicts the largest fraction of CEMP stars. AGB nucleosynthesis {#nucsyn} ------------------- We refer to the review papers by [@Busso1999] and [@Herwig2005] for a description of the nucleosynthesis process in the interior of AGB stars, the role of the ${^{13}\mathrm{C}}$-pocket at the top of the intershell region as a source of free neutrons for the production of slow neutron-capture elements ($s$-elements) in low-mass AGB stars, and the effect of the third dredge-up process (TDU hereinafter) that mixes the products of internal nucleosynthesis to the stellar surface. In the models of [@Karakas2010] a ${^{13}\mathrm{C}}$-pocket is created by including a partial mixing zone (PMZ) at the deepest extent of each TDU, where protons are mixed in the intershell region and subsequently captured by the $^{12}{\mathrm{C}}$ nuclei to form a layer rich in ${^{13}\mathrm{C}}$. The mass of the PMZ is a free parameter in the models, and [@Lugaro2012] study the effects of different masses of the PMZ on the surface abundances in AGB stars with different initial masses. We describe in Paper I how in our models the amount of material dredged-up from the intershell region is determined in order to reproduce the evolution predicted in the detailed models of [@Karakas2010] and [@Lugaro2012]. The chemical composition of the intershell region is saved in a table as a function of three parameters: the mass of the star at the beginning of the TP-AGB phase, the thermal-pulse number, and the mass of the PMZ, ${M_{\mathrm{PMZ}}}$. The surface abundances of an AGB star of mass $M_*$ evolve in time and are recalculated at every TDU, when material with the chemical composition of the intershell region taken from our table is mixed into the convective envelope. Grid of models {#grid} -------------- Our simulations are based on the same grid of models as in Paper I, with $N$ binary-evolution models distributed in the $M_1 - M_2 - \log_{10}a - {M_{\mathrm{PMZ}}}$ parameter space, where $M_{1,\,2}$ are the initial masses of the primary and secondary stars, respectively, $a$ is the initial separation of the system, and ${M_{\mathrm{PMZ}}}$ is the mass of the partial mixing zone of any star of the binary system that becomes an AGB star. The grid resolution $N=N_{\mathrm{M1}}\times N_{\mathrm{M2}}\times N_a \times N_{\mathrm{PMZ}}$, where we choose $N_{\mathrm{M1}} = 34$, $N_{\mathrm{M2}} = 28$, $N_a = 30$, and $N_{\mathrm{PMZ}} = 10$. The initial parameters are chosen as follows: - $M_1$ varies in the range \[$0.9,6.0$\] ${M_{\odot}}$. The grid spacing is $\Delta {M_{1,\mathrm{i}}}= 0.1\,{M_{\odot}}$ up to $3\,{M_{\odot}}$, and $\Delta {M_{1,\mathrm{i}}}= 0.25\,{M_{\odot}}$ otherwise. - $M_2$ is equally spaced in the range \[$0.2,0.9$\] ${M_{\odot}}$, and by definition ${M_{2,\mathrm{i}}}\le {M_{1,\mathrm{i}}}$. - $a$ varies between $10^2$ ${\mathrm{R}_{\odot}}$ and $10^5$ ${\mathrm{R}_{\odot}}$. The distribution of separations is flat in $\log_{10} a$. In the mass range considered here, at shorter separations the evolution of primary stars is interrupted before the AGB, because the systems undergo a common-envelope phase, whereas stars at wider separation do not interact in our models. The eccentricity is always zero. - When ${M_{1,\mathrm{i}}}< 3\,{M_{\odot}}$ we adopt $10$ different values for ${M_{\mathrm{PMZ}}}$, namely $0$, $10^{-4}$, $2\times10^{-4}$, $5\times10^{-4}$, $6.66\times10^{-4}$, $10^{-3}$, $1.5\times10^{-3}$, $2\times10^{-3}$, $3\times10^{-3}$, and $4\times10^{-3}\,{M_{\odot}}$. ${M_{\mathrm{PMZ}}}$ is zero otherwise, in accordance with the detailed models of AGB nucleosynthesis of [@Karakas2010]. ----------- --------------------------------- --------------------------------------------------- Model set wind accretion angular momentum efficiency loss A WRLOF spherically symmetric wind B BHL, $\alpha_{\mathrm{BHL}}=10$ $\Delta J/J = \gamma\,(\Delta M/M)$ ,  $\gamma=2$ C WRLOF $\Delta J/J = \gamma\,(\Delta M/M)$ ,  $\gamma=2$ ----------- --------------------------------- --------------------------------------------------- : Models of the wind-accretion efficiency and angular momentum loss used in this study.[]{data-label="tab:models"} {width="99.00000%"} All stars in our grid are formed in binary systems. In reality single stars or binaries in wider orbits may exist, and carbon-rich very metal-poor AGB stars have recently been observed [@Boyer2015II]. To determine whether the abundances observed in the stars of our sample can be reproduced by a single-star model, we add to our grid $200$ synthetic single stars with masses uniformly spaced in the range $[0.4,\,1.0]\,{M_{\odot}}$, and ${M_{\mathrm{PMZ}}}$ chosen as above. We assume that the accreted material is instantaneously mixed throughout the entire secondary star, as in the default model of Paper I. This approximation mimics the effect of efficient non-convective mixing processes, such as thermohaline mixing, and is reasonable in low-mass stars [@Stancliffe2007]. In our model we assume the same metallicity as in the detailed models of [@Lugaro2012], $Z=10^{-4}$, that corresponds approximately to $[{\mathrm{Fe}}/{\mathrm{H}}]\approx-2.3$. As initial chemical composition of the stars in our grid we adopt the abundances predicted in the one-zone Galactic chemical-evolution model of [@Kobayashi2011] at $[{\mathrm{Fe}}/{\mathrm{H}}]=-2.3$ that includes the contributions of core-collapse and type Ia supernovae and AGB stars. Their results extend up to $^{76}$Ge and for heavier isotopes we adopt the solar abundance distribution derived by [@Asplund2009] scaled to metallicity $Z=10^{-4}$. Different assumptions on the initial composition of the stars negligibly affect AGB nucleosynthesis, as discussed by [@Lugaro2012]. In the next section we discuss how representative our set of initial abundances is of the chemical composition of the very metal-poor stars in our sample. Comparison of the model initial abundances with carbon-normal metal-poor stars {#VMP} ============================================================================== In this section we focus on the $257$ “carbon-normal” very metal-poor stars in our sample, i.e. with $[{\mathrm{C}}/{\mathrm{Fe}}]<1$, called C-normal stars hereinafter. C-normal stars do not exhibit evidence of duplicity and are not expected to have changed their initial surface composition. Consequently, their abundances are expected to follow the distribution predicted by Galactic chemical-evolution models at metallicity $Z\approx10^{-4}$. Some abundance variations may be introduced by internal mixing processes, such as rotational mixing, gravitational settling of heavy nuclei or thermohaline mixing. The only significant change in the abundances of stars with small surface gravity, ${\log_{10}(g/\mathrm{cm}\,\mathrm{s}^{-2})}\lesssim4$, is because of the first dredge-up that reduces the carbon-to-nitrogen ratio but leaves essentially unaltered the abundances of the other elements. Our purpose is to compare the set of initial abundances adopted in our models with the abundances observed in C-normal stars and find discrepancies that are relevant to our study of the chemical composition of CEMP-$s$ stars. In Fig. \[fig:vmp-abunds\] we show the initial abundances adopted in our models (solid line) and the abundances observed in the C-normal stars. For every observed element we compute the mean of the logarithm abundance relative to iron, \[El/Fe\] (filled circles and plus signs). Only elements observed in more than one star are shown. Elements observed in fewer than seven stars are plotted with dashed bars, which connect the minimum and the maximum observed value. For each element observed in at least seven stars we calculate the median of the values of \[El/Fe\] and we select the group of approximately $68\%$ stars, half of which have \[El/Fe\] larger than the median and the other half have smaller values. The solid bars in Fig. \[fig:vmp-abunds\] indicate the interval of \[El/Fe\] that encloses the stars in this group. In the absence of significant nucleosythesis in C-normal stars, we expect that: (1) the mean and the median of the observed abundances coincide and are equal to the initial value predicted by the models of Galactic chemical evolution; (2) the deviations from the mean have a Gaussian distribution; (3) the solid bars are symmetrical with respect to the mean and correspond to the standard deviation in the observed abundances. However, Fig. \[fig:vmp-abunds\] shows that these three statements are not always true, and several discrepancies occur between the predictions of the Galactic chemical evolution model and the observations. We briefly analyse the discrepancies that are most relevant for our purposes. The differences between the model abundances and the observations of carbon and $s$-elements (e.g. Sr, Y, Zr, Ba, La, Ce, Pb) are on average within a factor of two ($0.3$ dex), although in some cases the observations have a large spread, for example carbon and barium ($0.8$ dex), cerium ($0.9$ dex) and lead ($1.1$ dex). These uncertainties are expected to have a small effect on our study of CEMP-$s$ stars because typically in low-mass AGB stars ($M_*\le3{M_{\odot}}$) the amount of carbon increases by more than two orders of magnitude and the abundances of $s$-elements by more than a factor of ten. The average abundance of nitrogen relative to iron is $1.5$ dex larger than the initial abundance assumed in our model, with a spread of $1$ dex. This discrepancy may alter the results of our best-fitting models. In stellar nucleosynthesis, nitrogen is made in the CN cycle when protons are mixed into regions of the star where carbon is abundant. For example, the mixing of protons can occur at the bottom of the convective envelope of red giants and AGB stars during first dredge-up and TDU, respectively. The increase in nitrogen abundance depends on the extent of the mixing, which is very uncertain [e.g.: @Charbonnel1998; @Boothroyd1999; @Karakas2010-1]. However, the total amount of carbon plus nitrogen is conserved in the CN cycle and therefore when both the elements are observed it is convenient to compare the predictions of our models with the combined abundance C+N. In C-normal stars the average C+N (right panel of Fig. \[fig:vmp-abunds\]) is underestimated by $0.5$ dex (approximately a factor of three). This offset should have a small effect on our fits of CEMP-$s$ stars, because the abundance of C+N increases by more than a factor of a hundred in AGB stars. Oxygen is underpredicted on average by a factor of three. The amount of oxygen increases by less than one order of magnitude during the AGB phase, hence an initial offset can substantially modify our results in the study for CEMP-$s$ stars. The observed abundances of sodium are underpredicted on average by $0.3$ dex and are spread over $0.7$ dex. A better model of the initial abundance of sodium is desirable because the amount of sodium produced in AGB nucleosynthesis depends strongly on the stellar mass and on the mass of the PMZ, and a reliable fit of its abundance can, in principle, provide constraints on these two parameters. The elements between atomic numbers $13$ (aluminium) and $30$ (zinc) are not always consistent with our initial assumptions. These elements are not modified by AGB nucleosynthesis, with the exception of aluminium that is partly produced in massive AGB stars, hence the offset observed in C-normal stars is expected to be an issue also in CEMP-$s$ stars. The elements with atomic numbers 62–71 and 75–80 are underestimated by $0.4$–$1$ dex, in many cases with large spreads in the observed abundances. In the solar system, more than $60\%$ of the total abundance of these elements is produced by the $r$-process, for example $69\%$ of samarium, $94\%$ of europium and $86\%$ of gadolinium [@Arlandini1999; @Bisterzo2011]. For brevity, these elements are hereinafter referred to as “$r$-elements”, although some fraction of their total amount is produced by the $s$-process. However, during AGB nucleosynthesis their abundances generally increase by less than $1$–$1.5$ dex, and therefore the initial offset may affect our analysis of the CEMP-$s$ stars. Analysis of the abundances in carbon-enhanced metal-poor stars {#CEMPs} ============================================================== We compare our binary evolution and nucleosynthesis models with the abundances of the $67$ CEMP-$s$ stars in our observational sample. We also include $14$ of the $15$ binary stars with known orbital periods studied in Paper I (those with $-2.8\le[{\mathrm{Fe}}/{\mathrm{H}}]\le1.8$) but in the present work, for comparison, we focus only on the chemical abundances, and we ignore the constraints on the period of the systems. In general we find initial parameters of the best-fit model different from Paper I. These differences give an estimate of the uncertainties in our present results caused by the lack of information on the orbital periods. For each star in our sample we determine the model in our grid that best reproduces the observed chemical abundances with the same procedure as in Paper I. Initially, to constrain the evolutionary stage, we select model stars that reproduce the measured surface gravity within the observational uncertainty, $\sigma_g$, at an age $10\le t\le13.7$ Gyr. Subsequently, for the stars that pass this selection, we compute the ${\chi^2}$ of each model from $${\chi^2}= \sum_i\frac{(A_{i,\mathrm{obs}} - A_{i,\mathrm{mod}})^2}{{\sigma_{i,\,\mathrm{obs}}}^2}~~.\label{eq:chi}$$ In Eq. (\[eq:chi\]) every element $i$ has observed abundance $A_{i,\mathrm{obs}}=12+\log_{10} (N_{i,\mathrm{obs}}/N_{{\mathrm{H}}})$, where $N_{i,\mathrm{obs}}$ and $N_{{\mathrm{H}}}$ are the number densities of $i$ and hydrogen, respectively, ${\sigma_{i,\,\mathrm{obs}}}$ is the observational error associated with $A_{i,\mathrm{obs}}$ and $A_{i,\mathrm{mod}}$ is the abundance predicted by the model. The minimum value of ${\chi^2}$, ${\chi^2_{\mathrm{min}}}$, determines the best model. In this procedure we do not include any constraint on the effective temperature, ${T_{\mathrm{eff}}}$, because ${T_{\mathrm{eff}}}$ depends strongly on the observed metallicity which varies by up to a factor of three in the stars of our sample compared to the value adopted in our model, $Z=10^{-4}\,$. All other parameters being equal, if the observed metallicity is lower (higher) than in our model we expect to find a model ${T_{\mathrm{eff}}}$ lower (higher) than the observed. As we noted in Paper I, because in our study we adopt a fixed metallicity for all our systems we ignore the scatter in observed \[Fe/H\] values. Qualitatively, a decrease in metallicity has two effects. First, the abundance relative to iron of the elements produced in AGB nucleosynthesis increases, because the initial amount of iron is lower, and iron is not produced in AGB stars. Second, because the production of neutrons is primary in AGB stars, the smaller the abundance of iron, the larger is the neutron-to-iron ratio, and consequently the abundances of the more neutron-rich elements (e.g. barium, lead) are enhanced. Opposite effects are caused by increasing metallicity. We restricted our sample to a $0.5$ dex range around $[{\mathrm{Fe}}/{\mathrm{H}}]=-2.3$ to have a sufficient number of stars for our analysis. It is important to consider that our approximation probably introduces bigger errors the larger the difference in \[Fe/H\] between observations and our model. An attempt to quantify this error is discussed in Sect. \[disc-1\]. We focus on the elements produced by nucleosynthesis in AGB stars. Therefore in Eq. (\[eq:chi\]) we take into account C, N, O, F, Ne, Na, Mg, and all the heavy neutron-capture elements with atomic number in the range $[31,\, 82]$, including the light-$s$ elements (or ls, namely strontium, yttrium and zirconium), the heavy-s elements (or hs, namely barium, lanthanum and cerium) and lead. The abundance of nitrogen produced in AGB stars is uncertain. In AGB stars of mass above about $3{M_{\odot}}$, carbon is efficiently converted to nitrogen at the base of the convective envelope [hot bottom burning, @Lattanzio1991]. At lower mass, some form of deep mixing of the envelope material operates down to regions where hydrogen burning occurs, as discussed in Sect. \[VMP\], but the exact amount depends on the extent of mixing, which is very uncertain [e.g. @Hollowell1988; @Gallino1998; @Goriely2000; @Stancliffe2010; @Lugaro2012]. However, the total amount of C+N is conserved, therefore when both elements are measured we consider their combined abundances. We exclude from Eq. (\[eq:chi\]) the elements from aluminium to zinc because they are not involved in AGB nucleosynthesis (aluminium is produced only in massive AGB stars), and the differences between models and observations reflect a discrepancy with our set of initial abundances, as we noticed in the sample of C-normal stars. For comparison, in some CEMP-$s/r$ stars we take into account a smaller set of elements, i.e. only those that are mostly produced in AGB nucleosynthesis (C, Mg, Sr, Y, Zr, Ba, La, Ce, Pb), as will be explained in Sect. \[CEMP-rs\]. Tables \[tab:best-WRLOFq-ssw\], \[tab:best-BoHo10-gamma2\] and \[tab:best-WRLOFq-gamma2\] in Appendix \[app:A\] summarise our results (computed with model sets A, B and C, respectively) for the best-fitting models of 43 CEMP-$s$ stars with number of degrees of freedom $\nu\ge2$. The number of degrees of freedom is calculated as $\nu=N_{\mathrm{obs}}-4$, where $N_{\mathrm{obs}}$ is the number of observed elements used in Eq. (\[eq:chi\]) and $4$ is the number of fitted parameters. Our results for the CEMP-$s$ stars with $\nu\le1$ are shown in Tables \[tab:few-WRLOFq-ssw\]–\[tab:few-WRLOFq-gamma2\]. Some stars fail to match one or more elements within their observed uncertainty. To quantify the discrepancy between the observations and the model predictions, we calculate the residuals, $$R_i = A_{i,\mathrm{obs}} - A_{i,\mathrm{mod}}~, \label{eq:residuals}$$ where $A_{i,\mathrm{obs}}$ and $A_{i,\mathrm{mod}}$ are the observed and modelled number abundance of element $i$ of the best fitting model, as in Eq. (\[eq:chi\]). For every element $i$, if our models reproduce correctly the nucleosynthesis and mass-transfer process, under the assumption that the measurements are only affected by Gaussian errors, the distribution of $R_i$ should resemble a Gaussian function centred on zero with a standard deviation approximately equal to the average observational error. ![Residual distributions in the model-A CEMP-$s$ stars of our sample with $\nu\ge2$. The number of stars in which the element is measured is indicated in brackets. The red-hatched, blue-filled and grey-dotted histograms indicate the residual distributions for CEMP-$s/nr$, CEMP-$s/r$ and CEMP-$s/ur$ stars, respectively. In the top-left panel the residual distribution of carbon is shown for those stars without a nitrogen measurement.[]{data-label="fig:histo_resid"}](Fig-3.pdf){width="48.80000%"} In Fig. \[fig:histo\_resid\] we plot the distributions of $R_i$ computed with the default model set A and binned in intervals of width $0.1$ for model stars with $\nu\ge2$. From the top left to the bottom right we show the distributions of the residuals computed for the abundances of carbon (for those stars without a nitrogen measurement), C+N, sodium, magnesium, strontium, barium, europium and lead. The subgroups of CEMP-$s/nr$, CEMP-$s/r$ and CEMP-$s/ur$ stars are indicated with hatched, filled and dotted histograms, respectively. In brackets we indicate the number of stars in which the element is measured. In Fig. \[fig:residCEMP-s\] we plot, for each star with $\nu\ge2$, the residuals $R_i$ of every observed element. The elements considered in Eq. (\[eq:chi\]) are shown as black filled circles, while the other elements are shown as grey open circles. The average values of the residuals of each element are shown as crosses. Panels (a)–(c) show the results obtained with our default model set A for CEMP-$s/nr$, CEMP-$s/r$ and CEMP-$s/ur$ stars, respectively. {width="99.00000%"} {width="99.00000%"} {width="99.00000%"} CEMP-s/nr stars {#CEMP-s/nr} --------------- Ten of the $67$ sampled CEMP-$s$ stars are classified as CEMP-$s/nr$, with $[{\mathrm{Ba}}/{\mathrm{Fe}}]>0.5$ and $[{\mathrm{Eu}}/{\mathrm{Fe}}]\le1$. For all the stars in this group we find good fits to the abundances. Four stars of this group, CS22942–019, CS22964–161A,B and HD198269 belong to binary systems with known orbital periods. In Paper I we find the models that best reproduce, at the same time, the observed abundances and the orbital periods of these stars. In the present work we ignore the constraint on the period in our models and we find essentially the same input parameters ${M_{1,\mathrm{i}}}$, ${M_{2,\mathrm{i}}}$ and ${M_{\mathrm{PMZ}}}$ as in the best models of Paper I, whereas there are large differences in the initial and final periods. These differences are due to the amount of mass $\Delta{M_{\mathrm{acc}}}$ that the secondary star has to accrete to reproduce the observed ${\log g}$ and surface abundances. For example, the model star CS22942–019 accretes $\Delta{M_{\mathrm{acc}}}\approx0.3\,{M_{\odot}}$ with all the model sets and consequently the ${\chi^2_{\mathrm{min}}}$ of the fit is the same (${\chi^2_{\mathrm{min}}}=21$). On the other hand, the initial periods vary by a factor of four (${P_{\mathrm{i}}}=6.4\times10^4$ days with model sets A and C, ${P_{\mathrm{i}}}=1.6\times10^4$ with model set B), while the final periods are ${P_{\mathrm{f}}}=1.0\times10^5,\,3.5\times10^3,\,{P_{\mathrm{f}}}=1.5\times10^4$ days with model set A, B and C, respectively (observed ${P_{\mathrm{orb}}}=2800$ days). Hence, only model set B reproduces the observed period within the uncertainty of our grid of models. For binary star CS22964–161A,B, which probably formed in a triple system [cf. @Thompson2008 and Paper I], we find the same results as in Paper I without significant differences between model sets A, B and C. A binary system of approximately $1.5\,{M_{\odot}}$ accretes a small amount of material ($\Delta{M_{\mathrm{acc}}}\approx0.06\,{M_{\odot}}$) from an initially $1.6\,{M_{\odot}}$ primary star in a very wide orbit (${P_{\mathrm{i}}}=2.7\times10^5,\,1.3\times10^5,\,2.6\times10^6$ days according to model sets A, B and C, respectively). Note that the observed orbital period in Table \[tab:obs\] does not correspond to the periods found in our models, which is the period of the unseen third star. The fit to the abundances observed in HD198269 improves compared to the result of Paper I if we ignore the orbital period (${P_{\mathrm{orb}}}=1295$ days). With essentially the same assumptions on the masses and longer orbital periods (see Tables \[tab:best-WRLOFq-ssw\]–\[tab:best-WRLOFq-gamma2\]) our model stars accrete $\Delta{M_{\mathrm{acc}}}\approx0.09\,{M_{\odot}}$ and we obtain ${\chi^2_{\mathrm{min}}}=9$, whereas in our best-fit model in Paper I (with model set B) we find $\Delta{M_{\mathrm{acc}}}=0.13\,{M_{\odot}}$ and ${\chi^2_{\mathrm{min}}}=18$. Fig. \[fig:residCEMP-s\]a shows that the abundances of almost all the elements involved in AGB nucleosynthesis are well reproduced and typically the mean of the residuals is close to zero. The abundance of sodium is well reproduced in three out of four CEMP-$s/nr$ stars, while it is overestimated by $0.5$ dex in CS22964–161B (in which $\sigma_{{\mathrm{Na}}}=0.3$ dex). The mean residual of yttrium (atomic number $39$) is approximately $-0.15$ dex, lower than for strontium and zirconium (${\overline{R}}\approx0$). The difference is smaller than the average error in the measurement of yttrium (${\overline{\sigma}_{\mathrm{obs}}}=0.20$ dex) but indicates that the abundance of yttrium is systematically overestimated compared to strontium and zirconium. The abundances of erbium (atomic number $68$) are underestimated by $0.6$ and $0.2$ dex for stars CS22880–074 and HD196944, respectively, while the observational uncertainty is approximately $0.2$ dex in both stars. An offset of about $0.6$ dex is also found by [@Bisterzo2012] in their analysis of CS22880–074. In addition, for this star our best-fitting model and the models of [@Bisterzo2012] well reproduce other elements traditionally associated with the $r$-process, such as europium and dysprosium. This evidence possibly suggests that the abundance determination of erbium in CS22880–074 is affected by observational bias. The abundances of lead are underestimated on average by $0.18$ dex, approximately the mean observational uncertainty of this element in CEMP-$s$ stars (${\overline{\sigma}_{\mathrm{obs}}}=0.20$ dex). The distribution of the residuals in Fig. \[fig:histo\_resid\] (red-hatched histogram) shows that the residuals of lead are always within $2{\overline{\sigma}_{\mathrm{obs}}}$. The distributions in Fig. \[fig:histo\_resid\] show that the abundances of the other elements are typically reproduced within $2{\overline{\sigma}_{\mathrm{obs}}}$. This indicates that on average our models of binary evolution and AGB nucleosynthesis reproduce reasonably well the abundances of CEMP-$s/nr$ stars, although the results should be interpreted with care because our sample is small and we do not constrain the orbital periods of the modelled systems. CEMP-s/r stars {#CEMP-rs} -------------- $20$ stars in our sample are classified as CEMP-$s/r$ stars. Two of these stars, BS$16080-175$ and BS$17436-058$, are not in Table \[tab:best-WRLOFq-ssw\] because only $5$ elements are observed and therefore $\nu=1$. Four CEMP-$s/r$ stars (CS22948–027, CS29497–030, HD224959 and LP625–44) are analysed in Paper I, and we find that to reproduce the large enhancements observed in most of the neutron-capture elements the secondary star has to accrete a large amount of mass, $\Delta{M_{\mathrm{acc}}}>0.2\,{M_{\odot}}$. At the observed orbital periods only model set B predicts efficient mass accretion, whereas without the constraint on the period also model sets A and C predict large mass accretion in much wider orbits, as shown in Tables \[tab:best-WRLOFq-ssw\] and \[tab:best-WRLOFq-gamma2\]. Our best-fitting models have rather high ${\chi^2_{\mathrm{min}}}$ despite the large accretion efficiency, partly because we fail to reproduce the enhanced abundances of the $r$-elements, and this points to a limitation in our nucleosynthesis model, as noted also in Paper I. For several CEMP-$s/r$ stars we find a particularly poor fit, with large reduced ${\chi^2}$ (${\chi^2_{\mathrm{min}}}/\nu\gtrsim5$, whereas at a visual inspection models appear to fit the observed abundances well if ${\chi^2_{\mathrm{min}}}/\nu\le3$). To verify if these results are caused only by the discrepancy in the $r$-elements we perform the ${\chi^2}$ analysis taking into account only nine elements that are produced in large amounts by our AGB nucleosynthesis models, and that are frequently detected in our observed sample: C, Mg, Sr, Y, Zr, Ba, La, Ce, Pb. This choice generally leads to a better fit of the abundances of these elements without significantly altering the results for the other elements. However, even with this choice, in most model stars the reduced ${\chi^2}$ still exceeds $3$, with the exceptions of CS22948–027 and CS29497–030. The best-fitting models of these two stars, calculated taking into account only nine elements, have the same primary and PMZ masses as in the models determined including all the observed elements (for both stars, ${M_{1,\mathrm{i}}}=1.5\,{M_{\odot}}$ and ${M_{\mathrm{PMZ}}}=2\times10^{-3}\,{M_{\odot}}$, with model set A). The secondary stars are slightly more massive (by approximately $0.05\,{M_{\odot}}$ for both stars with all model sets). In the other CEMP-$s/r$ stars large ${\chi^2_{\mathrm{min}}}$ are determined because in most of our models there is also an issue in reproducing the observed element-to-element ratios. In particular, it is difficult to reconcile the large enhancements of heavy-$s$ elements and lead with the relatively small abundances of lighter elements (carbon, sodium, magnesium, light-$s$ elements). As an example, we compare the modelled and observed abundances in star CS22898–027 in Fig. \[fig:CS22898-027\]. A model of a $0.5\,{M_{\odot}}$ secondary star that accretes $\Delta{M_{\mathrm{acc}}}\approx0.3\,{M_{\odot}}$ from a $2\,{M_{\odot}}$ primary star with ${M_{\mathrm{PMZ}}}=4\times10^{-3}\,{M_{\odot}}$ (dotted line) reproduces the large enhancement of the heavy-$s$ elements and lead, but overestimates carbon by $0.8$ dex, sodium and magnesium by more than $1$ dex and the light-$s$ elements by at least $0.5$ dex (consequently, ${\chi^2_{\mathrm{min}}}=632$). The best compromise between the light and heavy elements is the model with ${M_{1,\mathrm{i}}}= 1.5\,{M_{\odot}}$ and ${M_{\mathrm{PMZ}}}= 6.6\times10^{-4}\,{M_{\odot}}$, shown as a solid line in Fig. \[fig:CS22898-027\], where all elements up to zirconium are overestimated, the neutron-rich elements between barium and lead are underestimated and ${\chi^2_{\mathrm{min}}}= 100\,(\nu=11)$. Similar results are found with all model sets. ![CEMP-$s/r$ star CS22898–027. [*Points with error bars*]{}: observed abundances, in black the elements adopted in Eq. (\[eq:chi\]), in grey the other elements. [*Red solid line*]{}: the best-fitting model found with model set A, ${M_{1,\mathrm{i}}}=1.5\,{M_{\odot}}$ and ${M_{\mathrm{PMZ}}}=6.66\times10^{-4}\,{M_{\odot}}$. [*Blue dashed line*]{}: alternative model with ${M_{1,\mathrm{i}}}=2\,{M_{\odot}}$ and ${M_{\mathrm{PMZ}}}=4\times10^{-3}\,{M_{\odot}}$. [*Lower panel*]{}: the residuals of the two models, computed with Eq. (\[eq:residuals\]), are shown as red plus signs with error bars and blue triangles, respectively.[]{data-label="fig:CS22898-027"}](Fig-5.pdf){width="48.80000%"} In Fig. \[fig:hsls\] we plot the hs-to-ls ratio versus the abundance of europium observed in our sample CEMP-$s/nr$ and CEMP-$s/r$ stars. The abundances of hs and ls are defined, respectively, as $[{\mathrm{hs}}/{\mathrm{Fe}}] = ([{\mathrm{Ba}}/{\mathrm{Fe}}]+[{\mathrm{La}}/{\mathrm{Fe}}]+[{\mathrm{Ce}}/{\mathrm{Fe}}])/3$ and $[{\mathrm{ls}}/{\mathrm{Fe}}] = ([{\mathrm{Sr}}/{\mathrm{Fe}}]+[{\mathrm{Y}}/{\mathrm{Fe}}]+[{\mathrm{Zr}}/{\mathrm{Fe}}])/3$. If one of these abundances is not available we average among the elements present in our database. Orange circles and black crosses indicate, respectively, the observed stars for which we find a fit with small reduced ${\chi^2}$ (${\chi^2_{\mathrm{min}}}/\nu\le3$) and those for which our best fit has ${\chi^2_{\mathrm{min}}}/\nu>3$. The dotted line indicates the maximum value of the hs-to-ls ratio predicted in any of our models, $[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}=0.88$. We notice that: - $[{\mathrm{hs}}/{\mathrm{ls}}]$ is generally larger for increasingly large \[Eu/Fe\], because the abundances of heavy-$s$ elements and europium correlate in metal-poor stars [e.g. @Jonsell2006], whereas the abundances of light-$s$ elements are essentially independent of europium. - $11$ stars have observed $[{\mathrm{hs}}/{\mathrm{ls}}]<[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}$, and for all we find a model with ${\chi^2_{\mathrm{min}}}/\nu\le3$. Only $2$ out of $11$ stars are CEMP-$s/r$ stars (CS22881–036 and HD187861), and these have relatively low europium enhancements, $[{\mathrm{Eu}}/{\mathrm{Fe}}]<1.4$. - $17$ stars have observed $[{\mathrm{hs}}/{\mathrm{ls}}]>[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}$; for only two of these stars we find a model with ${\chi^2_{\mathrm{min}}}/\nu\le3$ and one of these is a CEMP-$s/nr$ star (HD$198269$). - $16$ out of $18$ CEMP-$s/r$ stars have $[{\mathrm{hs}}/{\mathrm{ls}}]>[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}$ (only the CEMP-$s/r$ stars with $\nu\ge2$ are shown in Fig. \[fig:hsls\]). The value $[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}$ is found in a model AGB star of mass ${M_{1,\mathrm{i}}}\approx1.5\,{M_{\odot}}$: consequently, an initial primary mass ${M_{1,\mathrm{i}}}=1.5\,{M_{\odot}}$ is selected in almost all our model CEMP-$s/r$ stars. However, only three model CEMP-$s/r$ stars reproduce the observed abundances well (${\chi^2_{\mathrm{min}}}/\nu\le3$). The distributions of the residuals of CEMP-$s/r$ stars in Fig. \[fig:residCEMP-s\]b reflect the difficulty in predicting the correct element-to-element ratios in most CEMP-$s/r$ stars. The abundances of carbon, sodium, strontium, yttrium, and zirconium are on average overestimated. The mean residual of carbon is ${\overline{R}}=-0.3$ dex, whereas the mean residual of C+N is only ${\overline{R}}=-0.1$ dex. This may be interpreted as the effect of extra mixing of protons in AGB stars converting carbon to nitrogen. However, the residuals of C+N have a large dispersion, therefore it is not possible to derive strong conclusions about the efficiency of the mixing process. Oxygen is underestimated on average by $0.6$ dex, approximately the same as in C-normal metal-poor stars. Because the abundance of oxygen is not much affected by AGB nucleosynthesis, this result suggests that the discrepancy could be reduced by adopting a larger initial abundance of oxygen in our models. The abundance of sodium is always overestimated, on average by $0.5$ dex. In stars HE0338–3945 and HE1305+0007 the residuals of sodium are approximately $-1$ while the observational uncertainties are $0.1$ and $0.2$ dex, respectively. The abundance of sodium is increasingly larger the larger the mass of the PMZ. In our model CEMP-$s/r$ stars we mostly assume ${M_{\mathrm{PMZ}}}\ge2\times10^{-3}\,{M_{\odot}}$ to reproduce the observed abundances of heavy-$s$ elements and lead; consequently we overestimate sodium. An offset of approximately $0.8$ dex in the sodium abundance of HE1305+0007 is also found by [@Bisterzo2012] with models that reproduce the large enhancements of heavy-$s$ elements and lead. In contrast, the best fit of [@Bisterzo2011] matches the sodium abundance in HE0338–3945 within the observational uncertainty. This is probably because the abundances of all $r$-process isotopes are initially pre-enriched, $[r/{\mathrm{Fe}}]^{\mathrm{ini}}=2$ dex, and no dilution of the accreted material on the secondary star is considered. As a consequence, even a low-mass AGB model ($M^{\mathrm{AGB}}_{\mathrm{ini}}=1.3\,{M_{\odot}}$) predicts large abundances of heavy elements, while it does not produce much sodium. Magnesium is the only light element with mean residual approximately zero in our models. As shown in Fig. \[fig:histo\_resid\], the maximum discrepancy between the observed abundances and our models is $0.4$ dex, within two times the average observational uncertainty (${\overline{\sigma}_{\mathrm{obs}}}=0.2$ dex). ![\[hs/ls\] vs \[Eu/Fe\] in our sample CEMP-$s/nr$ (those with $[{\mathrm{Eu}}/{\mathrm{Fe}}]<1$) and CEMP-$s/r$ stars. Orange filled circles indicate the stars for which we find a good fit (${\chi^2_{\mathrm{min}}}/\nu\le3$). Black crosses indicate the stars poorly reproduced (${\chi^2_{\mathrm{min}}}/\nu>3$). The dotted line indicates the maximum hs-to-ls ratio in our models, $[{\mathrm{hs}}/{\mathrm{ls}}]_{\mathrm{max}}=0.88$.[]{data-label="fig:hsls"}](Fig-6.pdf){width="48.80000%"} The computed residuals of ls, hs and lead confirm the results obtained for the example star CS22898–027: the elements of the first $s$-peak are overestimated on average by $0.2-0.5$ dex, whereas lead and the elements of the second $s$-peak are underestimated by $0.1-0.2$ dex, with the exception of cerium and lanthanum (${\overline{R}}\approx0$). A systematic discrepancy, negative for strontium and positive for barium and lead, is present also in the distributions of the residuals in Fig. \[fig:histo\_resid\] (blue histograms). Because none of the model stars reproduces the large observed hs-to-ls ratio, to minimise ${\chi^2}$ our model predicts larger abundances of strontium, and smaller abundances of barium and lead compared to the observations. The yttrium residuals are typically lower than strontium and zirconium, analogously to the results obtained for the sample of CEMP-$s/nr$ stars. The abundance of niobium (atomic number $41$) is observed only in star CS29497–030 and is underestimated by $0.34$ dex (with observational uncertainty $0.2$ dex). This is peculiar because according to the models niobium is formed from the radioactive decay of $^{93}{\mathrm{Zr}}$, and the abundance of zirconium is well reproduced in this stars ($R\approx0$). This discrepancy is possibly related to the uncertain neutron-capture cross section of $^{93}{\mathrm{Zr}}$, as we discuss in Sect. \[disc-1\]. The $r$-elements are systematically underproduced, perhaps not surprisingly because in our nucleosynthesis model the $r$-process is not included. Our models typically produce $[{\mathrm{Ba}}/{\mathrm{Eu}}]$ close to 1, and the maximum europium enhancement is approximately $[{\mathrm{Eu}}/{\mathrm{Fe}}]=1.5$, whereas in most CEMP-$s/r$ stars the observed $[{\mathrm{Ba}}/{\mathrm{Eu}}]$ is below $0.6$ and $[{\mathrm{Eu}}/{\mathrm{Fe}}]>1.5$. CEMP-s/ur stars {#CEMP-s/ur} --------------- $37$ stars in our sample do not have an observed abundance of europium, and therefore are classified as CEMP-$s/ur$ stars, $15$ of which are listed in Tables \[tab:best-WRLOFq-ssw\]–\[tab:best-WRLOFq-gamma2\] because at least $6$ elements have been observed. Three of these $15$ stars have measured orbital periods and are discussed in Paper I. One of them, BD$+04^{\circ}2466$, has a very wide orbit (${P_{\mathrm{orb}}}\approx4600$ days), and even if we do not consider the period constraint we find the same input parameters as in Paper I. On the contrary, to reproduce the period of HD$201626$ ($407$ days), our model binary stars experience a common-envelope phase which shrinks the orbit. Without the period constraint the model progenitor system of HD$201626$ has a $1.4\,{M_{\odot}}$ primary star that transfers $\Delta{M_{\mathrm{acc}}}\approx0.18\,{M_{\odot}}$ to its companion star in a wide orbit (cf. Tables \[tab:best-WRLOFq-ssw\]–\[tab:best-WRLOFq-gamma2\]). Consequently, the systems do not experience a common-envelope phase and therefore the final periods derived in the models are $20$ to $600$ times longer than that observed. Without the period constraint we find ${\chi^2}\approx8$ ($\nu=5$), whereas in our best-fit model in Paper I we find ${\chi^2}\approx14$ ($\nu=6$). In Paper I we find that the model for star HE0024–2523 needs to experience inefficient common-envelope ejection to reproduce the large enhancements of the heavy-$s$ elements. If we relax the period constraint the system does not enter a common-envelope phase and the final period of the system is $475$ days (with model set A), two orders of magnitude longer than the observed period of $3.14$ days. {height="36.00000%"} Fig. \[fig:residCEMP-s\]c shows that carbon, magnesium, zirconium and the heavy-$s$ elements are reproduced within $0.1$ dex on average. The elements with the largest discrepancy with the observations are C+N (${\overline{R}}=-0.45$), sodium (${\overline{R}}=-0.25$), strontium and yttrium (${\overline{R}}=-0.2$) and lead (${\overline{R}}=0.25$). The abundance of nitrogen is observed only in five systems (BD$+04^{\circ}2466$, HD$13826$, HD$187216$, HD$201626$, HD$5223$) and the average results are greatly affected by the poor fit of star HD$187216$, which is possibly biased by large observational errors [@Kipper1994], and in which our model cannot reconcile the large enhancements of the $s$-elements with the solar ratios observed for nitrogen, sodium and magnesium. Our best fit reproduces the $s$-elements but overestimates C+N and sodium by $1.1$ and $0.7$ dex, respectively. The average residuals of C+N and sodium in CEMP-$s/ur$ stars decrease if we remove HD187216 from the sample (${\overline{R}}=-0.24$ and $0.18$, respectively). The distribution of strontium is greatly affected by the poor fit of star HE0212–0557, where the observed heavy-$s$ elements are enhanced by 2 dex whereas $[{\mathrm{Sr}}/{\mathrm{Fe}}]\approx0$, that is $1.2$ dex lower than predicted by our best model. Our model also overestimates the abundance of yttrium ($[{\mathrm{Y}}/{\mathrm{Fe}}]=0.7$) by $0.6$ dex, and the abundances of carbon and magnesium by $0.5$ dex. The models of [@Bisterzo2012] have similar discrepancies for these elements. If we do not consider this star, the mean residuals of strontium and yttrium are $-0.1$ and $-0.15$, respectively. Although the differences are small, it may indicate the same systematic effect found in CEMP-$s/nr$ and CEMP-$s/r$ stars. Fig. \[fig:histo\_resid\] shows that the abundances of barium and lead are reproduced within $0.3$ dex, that is $1.5$ times the average observational uncertainty (${\overline{\sigma}_{\mathrm{obs}}}\approx0.2$), with the exception of the model for star HE0024–2523 that underestimates the observed $[{\mathrm{Pb}}/{\mathrm{Fe}}]=3.2$ by almost 1 dex. Confidence limits on the initials parameters {#init-param} ============================================ The results presented in Sect. \[CEMPs\] are based on the best-fitting models found from our ${\chi^2}$-minimization procedure. The confidence intervals of the input parameters of our models (i.e. the initial primary and secondary masses, the initial period and the PMZ mass) are determined with the same procedure as described in Paper I. In biref, on our grid of binary models and associated ${\chi^2}$, we fix one of the input parameters $p$ and for each grid value of $p$ we determine the minimum ${\chi^2}$ with respect to the other parameters, defining a function ${\chi^2}(p)$. A confidence region is an interval of $p$ for which $\Delta{\chi^2}={\chi^2}(p)-{\chi^2_{\mathrm{min}}}$ is below a certain threshold, where ${\chi^2_{\mathrm{min}}}$ is the ${\chi^2}$ of the best fit. If the observational errors are Gaussian then the thresholds $\Delta{\chi^2}=1$, $4$ and $9$ correspond to the confidence intervals with, respectively, $68.3\%$, $95.4\%$ and $99.7\%$ probability that the actual $p$ is in this interval. As we mention in Paper I, the observational errors in our sample are not necessarily Gaussian, and they may in some cases be affected by systematic effects. Consequently, the thresholds $\Delta{\chi^2}=1,\,4,\,9$ should not be used to calculate the theoretical Gaussian probabilities, and provide only an indication of the confidence levels. At a visual inspection models with ${\chi^2}$ below the threshold $\Delta{\chi^2}= 4$ predict very similar abundances to the best-fitting models. On the other hand, models with larger ${\chi^2}$ are clearly distinct from the best model and the observations are not well reproduced. Fig. \[fig:init\_param\] shows the input parameters of the $23$ model CEMP-$s$ stars with ${\chi^2_{\mathrm{min}}}/\nu<3$ and $\nu\ge2$ as determined with model set A of a spherically symmetric wind with WRLOF wind-accretion efficiency. CEMP-$s/nr$, CEMP-$s/r$ and CEMP-$s/ur$ stars are indicated as filled circles, crosses and open diamonds, respectively. The horizontal bars represent the confidence intervals determined with the threshold $\Delta{\chi^2}<4$. Model stars with ${\chi^2_{\mathrm{min}}}/\nu>3$ are not included because they do not reproduce the observed abundances well enough, while stars with $\nu<2$ do not provide strong constraints on the models. The confidence intervals shown in Fig. \[fig:init\_param\] are summarised in Table \[tab:confi\_WRLOFq-ssw\]. In some stars we find multiple local minima in the ${\chi^2}$ distribution. In these stars different combinations of initial parameters result in models with similar surface abundances and hence almost equal ${\chi^2}$. As an example, for star CS22880–074 two local minima of ${\chi^2}$ are found for models with primary mass between $0.9$ and $1.1{M_{\odot}}$ and initial period between $600$ and $1.6\times10^4$ days, and another around ${M_{1,\mathrm{i}}}=1.5{M_{\odot}}$ and ${P_{\mathrm{i}}}=[1.6,\,3.5]\times10^5$ days. Most of the best-fitting models have primary mass below $2\,{M_{\odot}}$. AGB models with masses in this range produce a significant amount of $s$-process elements regardless of the mass of the PMZ because of proton-ingestion events that occur in the first thermal pulses [as described by @Lugaro2012 which they refer to as regime $4$ of neutron capture]. As a consequence, the mass of the PMZ in our model stars is not well constrained, as indicated by the large confidence intervals in Fig. \[fig:init\_param\]b. Most model stars have ${M_{\mathrm{PMZ}}}$ larger than $10^{-3}\,{M_{\odot}}$ because in this range low-mass AGB models (${M_{1,\mathrm{i}}}<3\,{M_{\odot}}$) produce positive \[hs/ls\] and \[Pb/hs\] as observed in the majority of our sample CEMP-$s$ stars. {height="40.00000%"} {height="40.00000%"} Stars of mass around $1.5\,{M_{\odot}}$ and with ${M_{\mathrm{PMZ}}}\ge2\times 10^{-3}\,{M_{\odot}}$ produce the largest $[{\mathrm{hs}}/{\mathrm{ls}}]$ value and $[{\mathrm{Pb}}/{\mathrm{hs}}]>0$. As a consequence, most observed stars with very enhanced heavy-$s$ elements and lead are best modelled with primary stars of masses between ${M_{1,\mathrm{i}}}=1.4\,{M_{\odot}}$ and ${M_{1,\mathrm{i}}}=1.6\,{M_{\odot}}$. CEMP-$s/r$ stars are mostly observed with $[{\mathrm{hs}}/{\mathrm{ls}}]$ close to one or higher, as discussed in Sect. \[CEMP-rs\], and therefore our best fits are found with ${M_{1,\mathrm{i}}}=1.5\,{M_{\odot}}$. Two CEMP-$s/nr$ stars are modelled with ${M_{1,\mathrm{i}}}$ in the mass range $[1.75,\,3]\,{M_{\odot}}$, in which the $[{\mathrm{hs}}/{\mathrm{ls}}]$ ratio varies approximately between $-0.5$ and $0.5$ for ${M_{\mathrm{PMZ}}}$ between $0$ and $2\times10^{-3}\,{M_{\odot}}$ while \[Pb/hs\] is always positive [regime 2 of @Lugaro2012]. The secondary mass (Fig. \[fig:init\_param\]c) in the models depends on the observed enhancements. The default assumption in our models, as in Paper I, is that accreted material is efficiently mixed by thermohaline mixing. As a consequence, a large degree of accretion is required when the observed abundances are greatly enhanced. Hence, the secondary star has to be initially less massive to reach a mass of approximately $0.8-0.9\,{M_{\odot}}$ after accretion, and still be visible at $t>10$ Gyr. The input parameters ${M_{1,\mathrm{i}}}$, ${M_{\mathrm{PMZ}}}$ and ${M_{2,\mathrm{i}}}$ are similar in all model sets. On the contrary, there are several differences between the initial and final periods, as shown in the upper and lower panels of Fig. \[fig:peri-Pf\]. In model set C the binary systems typically start in wider orbits compared to the other model sets, because the dependence of the WRLOF process on the separation favours wider orbits compared to model set B and because of the more efficient angular momentum loss compared to model set A. The final periods computed with model set A and C are similar in many stars because the mass-transfer algorithm is the same in the two models, and therefore the binary stars need to have approximately the same separation during accretion to transfer the same amount of material. The majority of stars in model sets A and C have final periods longer than $4,\!600$ days, which is approximately the longest period measured in our sample. On the contrary, in model set B mass transfer is more efficient in close orbits, and consequently the majority of the systems are predicted at periods shorter than $10,\!000$ days. This is the only model set that predicts significant mass accretion in close orbits, as shown also in Paper I, and ten modelled stars have final periods less than $4,\!600$ days. The stars that need to accrete a small amount of mass to reproduce the observed abundances have longer periods because of the enhanced wind-accretion rate the and efficient angular momentum loss. Discussion ========== Comparison between observed and modelled abundances {#disc-1} --------------------------------------------------- Our best-fitting models are found with initial primary mass in the range between $0.9{M_{\odot}}$ and $3{M_{\odot}}$. In this mass range the neutron source is the ${^{13}\mathrm{C}}(\alpha,\, n)^{16}{\mathrm{O}}$ reaction that operates on ${^{13}\mathrm{C}}$ produced by the inclusion of a PMZ or by ingestion of protons in the stellar interior during the thermal pulses [regimes $2-4$ in @Lugaro2012]. The abundances of $s$-elements predicted in the models depend on the mass of the AGB star and of its PMZ. We note that: ($i$) \[Ba/Eu\] is always close to 1 dex, ($ii$) the value of \[hs/ls\] varies from $[{\mathrm{hs}}/{\mathrm{ls}}]\approx-0.5$ for ${M_{\mathrm{PMZ}}}=0\,{M_{\odot}}$ up to $[{\mathrm{hs}}/{\mathrm{ls}}]\approx0.9$ for ${M_{\mathrm{PMZ}}}\ge2\times10^{-3}{M_{\odot}}$ and ${M_{1,\mathrm{i}}}=1.5{M_{\odot}}$, and ($iii$) \[Pb/hs\] is always positive, and in some cases greater than one, for ${M_{1,\mathrm{i}}}$ above approximately $2{M_{\odot}}$. The success of our models in reproducing the observed abundances varies significantly for different classes of stars. CEMP-$s/nr$ stars typically exhibit $[{\mathrm{hs}}/{\mathrm{ls}}]\approx0.5$ dex and \[Ba/Eu\] between $0.5$ and $1$ dex, therefore our models generally well reproduce the abundances of the $s$-elements and also the $r$-elements. On the other hand, most CEMP-$s/r$ stars have $[{\mathrm{hs}}/{\mathrm{ls}}]\gtrsim1$, and therefore our models are not able to reproduce the light-$s$ peak and the heavy-$s$ peak simultaneously. In most cases the model with the minimum ${\chi^2}$ systematically overestimates the light-$s$ elements, on average by $0.3$ dex, and underestimates the heavy-$s$ elements, on average by $0.1$ dex. Furthermore, generally $[{\mathrm{Ba}}/{\mathrm{Eu}}]\lesssim1$ and the abundances of all $r$-elements are underestimated by a factor of two to up to one hundred. In our CEMP-$s/ur$ stars the abundances of the heavy-$s$ elements are typically reproduced within the observational uncertainty, as in CEMP-$s/nr$ stars, whereas the light-$s$ elements are systematically overestimated on average by $0.15$ dex, similar to CEMP-$s/r$ stars. These results are probably related to the reason why the abundance of europium is indeterminate. This group is likely to contain a mixture of CEMP-$s/nr$ stars and CEMP-$s/r$ stars. Hence, in some stars the abundance of europium is actually low, whereas in other stars europium is enhanced but it is not detected, for example because the europium lines are blended. The abundance ratio between carbon and heavy-$s$ elements is generally positive in our AGB-nucleosynthesis model, $[{\mathrm{C}}/{\mathrm{hs}}]\gtrsim0$. Consequently, in our synthetic stars high abundances of heavy-$s$ elements are associated with large enhancements of carbon. A similar correlation is also found in our observed sample: compared to CEMP-$s/nr$ stars, CEMP-$s/r$ stars typically exhibit higher abundances of carbon (Fig. \[fig:datasample\]) and heavy-$s$ elements (as indicated by the increase of \[hs/ls\] with \[Eu/Fe\] in Fig. \[fig:hsls\]). However, our models overestimate the carbon-to-hs ratio in CEMP-$s/r$ stars. As a consequence, in the majority of these stars the carbon abundance is overestimated by our models with high \[hs/Fe\]. A similar problem is also encountered by @Izzard2009 [e.g. Fig. 13b] and in some models of [@Bisterzo2011; @Bisterzo2012]. This probably indicates that the AGB-nucleosynthesis models should be able to produce higher hs abundances for a given amount of carbon, as we also discuss in a forthcoming paper [@Abate2015-3 *in press*]. Some discrepancies between models and observations are found in all three classes of CEMP-$s$ stars. The abundance of nitrogen is underestimated in $15$ out of $16$ stars by $0.1$ to $1.8$ dex. The difficulty in reproducing the observed abundance of nitrogen is a well known issue of AGB nucleosynthesis models. In AGB stars of mass below approximately $3{M_{\odot}}$ some nitrogen is produced from carbon by a deep mixing process that operates at the bottom of the convective envelope, but the exact amount is very uncertain [@Milam2009; @Stancliffe2010; @Karakas2010-1]. If we take into account the combined abundance of carbon and nitrogen our results improve significantly, and the observations are always reproduced within at most twice the observational uncertainty. Oxygen is underestimated in almost all stars of our sample, on average by $0.4$ dex, approximately the same amount as in C-normal stars (cf. Sect. \[VMP\]). The abundance of oxygen does not change much during AGB nucleosynthesis, and most likely the discrepancy could be solved by adopting a larger initial abundance without significantly affecting the abundances of the other elements. The abundance of sodium is overpredicted in all our model stars, and this points to a general issue in our nucleosynthesis model. Sodium is produced in the intershell region of the AGB star due to proton and neutron capture on the abundant $^{22}{\mathrm{Ne}}$. The abundance of sodium increases rapidly with increasing PMZ mass and therefore models with a smaller PMZ predict lower abundances of sodium at each stellar mass. However, also the abundance of magnesium is sensitive to the mass of the PMZ and a large PMZ is required in most stars to reproduce its abundance, as well as the abundance of heavy-$s$ elements. [@Lugaro2012] suggest three effects that may help lower the predicted sodium abundance: ($i$) in the detailed models of AGB nucleosynthesis the density profile of protons introduced to make the ${^{13}\mathrm{C}}$-pocket could decrease more rapidly with depth [e.g., @Goriely2000], ($ii$) a smaller $^{22}{\mathrm{Ne}}(p,\,\gamma)^{23}{\mathrm{Na}}$ reaction rate or ($iii$) a larger $^{23}{\mathrm{Ne}}(p,\,\alpha)^{20}{\mathrm{Ne}}$ reaction rate, both of which are very uncertain [@Iliadis2010]. In many CEMP-$s$ stars the abundance of yttrium is systematically lower than strontium and zirconium by $0.1-0.2$ dex, whereas in our models these elements are generally produced in approximately the same amount for every stellar mass and PMZ. This suggest that less yttrium should be produced in the models. Yttrium is produced by the reaction $^{88}{\mathrm{Sr}}(n,\,\gamma)^{89}{\mathrm{Sr}}$ and the subsequent decay of $^{89}{\mathrm{Sr}}$ to $^{89}{\mathrm{Y}}$ with a half life of $51$ days. The neutron-capture cross section of $^{89}{\mathrm{Sr}}$ is uncertain by a factor of two [@Bao2000]. Qualitatively, a larger cross section implies that before the decay of $^{89}{\mathrm{Sr}}$ the reaction $^{89}{\mathrm{Sr}}(n,\,\gamma)^{90}{\mathrm{Sr}}$ may occur and after two consecutive $\beta$-decays $^{90}{\mathrm{Zr}}$ is produced while $^{89}{\mathrm{Y}}$ is skipped and this effect may help to reduce the discrepancy. Niobium is observed in the stars CS29497–030 and HD$187216$ and it is underestimated by $0.35$ dex and $0.69$ dex, respectively [although the abundances of HD$187216$ are very uncertain, @Kipper1994]. However, $^{93}{\mathrm{Nb}}$ is formed by $\beta$-decay of $^{93}{\mathrm{Zr}}$ and in the two stars zirconium is well reproduced. A similar problem is found by [@Kashiv2010] in the context of presolar SiC grains, in which the Nb/Zr ratio is systematically higher than in CI chondrites. These authors suggest that a cross section of $^{93}{\mathrm{Zr}}$ reduced by a factor of two removes the discrepancy and we note that this solution would help to reduce the difference observed in the two CEMP-$s$ stars. New measurements of the $^{93}{\mathrm{Zr}}$ cross section are currently being analysed and will provide a more accurate determination of the reaction rate [@Lugaro2014]. The abundances of barium are underestimated by our models in all but two CEMP-$s/r$ stars and in the majority of the CEMP-$s/ur$ stars. However, we note that the observed barium abundances might be affected by large spectroscopic uncertainties because most of the observed barium lines are strong resonance lines, often saturated and sensitive to non-LTE effects [@Busso1995; @Andrievsky2009; @Masseron2010]. The observed abundance of lead in most CEMP-$s$ stars is larger than predicted by the models. The lead abundance increases with increasing stellar mass (\[Pb/Fe\] is maximum for ${M_{1,\mathrm{i}}}$ between $2{M_{\odot}}$ and $3{M_{\odot}}$) and PMZ mass. In most of the stars in our sample the choice of a more massive primary star implies that the abundances of sodium, magnesium, and the light-$s$ elements are overestimated. This suggest that nucleosynthesis models of low-mass stars (${M_{1,\mathrm{i}}}<2{M_{\odot}}$) should produce higher \[Pb/Na\], \[Pb/Mg\] and \[Pb/ls\]. On the other hand, the determination of lead abundances generally rely on one single line blended by CH-lines, therefore the observed lead abundances may be affected by systematic errors [@Masseron2010]. Lead is one of the most neutron-rich elements produced by AGB stars, and therefore its abundance is strongly metallicity-dependent [@Travaglio2001]. To assess the effect of metallicity on our results, we use the theoretical models presented in Fig. 1 of [@Travaglio2001] to construct a table of correction factors which account for the differences in lead abundance predicted by their models in the range $-2.8\le[{\mathrm{Fe}}/{\mathrm{H}}]\le-1.8$ and our model at $[{\mathrm{Fe}}/{\mathrm{H}}]=-2.3$. We correct the lead abundances predicted in our models by interpolating in this table the iron abundances observed in CEMP-$s$ stars. However, the distribution of the residuals of \[Pb/Fe\] does not improve with this analysis, even if we vary the correction factors based on different models of [@Travaglio2001]. The residuals of $[{\mathrm{Pb}}/{\mathrm{Fe}}]$ computed with our models do not exhibit a dependence on $[{\mathrm{Fe}}/{\mathrm{H}}]$. Consequently, even correcting for different metallicities, the discrepancies between our model predictions and the observations are not reduced. Hence, as yet it is not possible to quantify the errors introduced in our analysis by the assumption of models with fixed metallicity. A larger sample of reliable abundances of lead (and other $s$-elements) in CEMP-$s$ stars will help to address this issue. Comparison with previous results -------------------------------- We compare our results with the analysis performed by [@Bisterzo2012] on $94$ CEMP-$s$ stars in the range of metallicity $-3.5 \le [{\mathrm{Fe}}/{\mathrm{H}}] \le -1.0\,$, $60$ of which have $-2.8 \le [{\mathrm{Fe}}/{\mathrm{H}}] \le -1.8\,$ and are in common with our sample. One CEMP-$s/r$ star (HE2258–6358, analysed by [@Placco2013]) and six CEMP-$s/ur$ stars are only included in our sample. In five of these stars the abundances of only a few elements have been determined (hence $\nu\le1$), whereas HD187216 was not included in the study of [@Bisterzo2012] because of the possibly large observational biases. [@Bisterzo2012] compare the observations with the results of AGB nucleosynthesis models of single stars with masses between $1.2\,{M_{\odot}}$ and $2\,{M_{\odot}}$ in a range of metallicities and ${^{13}\mathrm{C}}$-pockets. A quantitative comparison with their results is not straighforward because of the intrinsic differences in the methods used in the two studies. For example, the treatment of the ${^{13}\mathrm{C}}$-pocket/PMZ is not the same in the AGB models by [@Bisterzo2012] and by @Karakas2010 [on which our model is based]. In the former, the parameter that determines the efficiency of the $s$-process is the abundance of ${^{13}\mathrm{C}}$ in the pocket, which is added at the top of the intershell region at each third dredge-up. In the latter, the size of the PMZ determines the mass of a layer of free protons that are mixed with the intershell. The protons react with the elements in the intershell and the mass fractions of ${^{13}\mathrm{C}}$, ${^{14}\mathrm{N}}$ and other isotopes are calculated consistently. The standard mass of the ${^{13}\mathrm{C}}$-pocket in the models by [@Bisterzo2012] corresponds roughly to ${M_{\mathrm{PMZ}}}=2\times10^{-3}{M_{\odot}}$ in our models. Another important difference of our study is that we explicitly compute the binary evolution, wind mass-transfer process and the mixing effects while in the paper by [@Bisterzo2012] these aspects are mimicked by including a dilution factor, i.e. the ratio between the amount of accreted material and the envelope mass of the accreting star. Despite these differences, in $46$ of the $60$ CEMP-$s$ stars common to both studies we find primary masses that are consistent with the results of [@Bisterzo2012] within the confidence intervals of the models. Among these stars are $21$ of the $23$ stars for which we find a good fit (with ${\chi^2_{\mathrm{min}}}/\nu\le3$ and $\nu\ge2$). The two exceptions are HD5223 and HE2232–0603. Compared to the results of [@Bisterzo2012], in these CEMP-$s/ur$ stars our models with low primary mass ($0.9-1.1\,{M_{\odot}}$) better match the observed carbon and lead (in both stars), fluorine, sodium and magnesium (in HD$5223$), whereas larger ${M_{1,\mathrm{i}}}$ overestimate the observed abundances. Of the stars for which there is no agreement with the masses predicted by [@Bisterzo2012], ten are CEMP-$s/r$ stars, while the other four are CEMP-$s/ur$ stars. One possible explanation of these discrepancies is in the way [@Bisterzo2012] account for the contribution of the $r$-process in these stars. The authors assume that CEMP-$s/r$ stars were formed from gas polluted by the explosion of a Type II supernova, and therefore adopt in their models an initial enrichment of the $r$-process component of the elements heavier than iron, between $[r/{\mathrm{Fe}}]^{\mathrm{ini}}=0.5$ and $[r/{\mathrm{Fe}}]^{\mathrm{ini}}=2$. In CEMP-$s/r$ stars this initial enhancement is chosen to reproduce the observed $[{\mathrm{Eu}}/{\mathrm{Fe}}]$. The enhanced abundance of the $r$-process component does not alter significantly the abundance of the $s$-process elements, except in low-mass stars (${M_{1,\mathrm{i}}}\lesssim1.7{M_{\odot}}$) with $[r/{\mathrm{Fe}}]^{\mathrm{ini}}=2$, in which case the heavy-$s$ elements are enriched by an extra factor of a few tenths of a dex [e.g., Fig. 6 of @Lugaro2012]. As a consequence, the $r$-enriched models of [@Bisterzo2012] with low initial masses ($M=1.2-1.4{M_{\odot}}$) and relatively small ${^{13}\mathrm{C}}$-pockets, corresponding roughly to ${M_{\mathrm{PMZ}}}\le10^{-3}{M_{\odot}}$, reproduce the abundance of the heavy-$s$ elements, $r$-elements and lead without greatly overestimating the abundances of the light-$s$ elements. In contrast, in our models neutron-capture elements are entirely produced by the $s$-process during the AGB phase. Consequently, our models of CEMP-$s/r$ stars are constrained towards relatively massive primary stars, $1.5{M_{\odot}}\le{M_{1,\mathrm{i}}}\le2.9\,{M_{\odot}}$, with ${M_{\mathrm{PMZ}}}\ge2\times10^{-3}\,{M_{\odot}}$, which predict large abundances of elements heavier than barium but typically overestimate the observed abundances of lighter elements. The origin of the $r$-elements in CEMP-$s/r$ stars is unclear. Many hypotheses have been suggested that can be classified in three main classes. In the first class of scenarios the binary systems were formed in a molecular cloud already enriched in $r$-elements by the nearby explosion of one or more Type II supernovae [@Cohen2003; @Jonsell2006; @Sneden2008; @Bisterzo2011; @Bisterzo2012]. In the second class of scenarios, the primary star of the binary system is relatively massive and undergoes the AGB phase, producing the $s$-elements, and then explodes as an electron-capture or a Type 1.5 supernova, providing the $r$-elements [@Zijlstra2004; @Wanajo2006; @Jonsell2006]. If the $s$-process and the $r$-process enrichments are independent, as suggested in the first class of scenarios, it is difficult to explain the correlation between the abundances of barium and europium that is observed in CEMP-$s/nr$ stars and in CEMP-$s/r$ stars. Moreover, the massive AGB stars invoked in the second class of scenarios are rare in a solar-neighbourhood initial mass function [@Pols2012]. Also, massive AGB stars produce more nitrogen than carbon [e.g. @Karakas2007] and would therefore produce nitrogen-enhanced metal-poor stars rather than CEMP stars. For a thorough discussion of these scenarios we refer to [@Jonsell2006] and [@Lugaro2009]. [@Lugaro2009] suggest a third speculative scenario in which $r$- and $s$-elements are both formed in low-mass extremely metal-poor AGB stars ($[{\mathrm{Fe}}/{\mathrm{H}}]<-3$), that may be able to produce very large densities of free neutrons when protons are ingested in the region of the He-flash [@Campbell2007; @Campbell2008; @Herwig2011; @Reifarth2014]. Constraints on binary mass transfer ----------------------------------- We considered several model sets with different combinations of assumptions for the models of wind mass-accretion rate and angular momentum loss. Because in this study we do not consider constraints on the orbital period there are no significant differences between the results obtained with the three model sets, with the exception of the distributions of initial and final orbital periods of the binary systems. Without the constraint on the orbital period our fitting procedure mostly selects model systems with periods longer than $5,\!000$ days, which is more than the longest period currently observed. In our default model set A relatively long periods are necessary to accrete large amounts of mass. Therefore, when large mass accretion is required to reproduce the observed abundances we find modelled CEMP-$s$ stars with periods approximately $20-500$ times longer than predicted in Paper I, in which the observed orbital periods are used to constrain our models (e.g. the stars CS$22942-019$ and CS$29497-030$). Shorter periods are found in a few cases if the observed abundances are reproduced by a model with a low-mass primary star (${M_{1,\mathrm{i}}}\le1.1{M_{\odot}}$) and small accreted mass. For these stars we compute orbital periods of a few thousand days, consistent with the results of Paper I (e.g. the stars BD$+04^{\circ}2466$ and HD$198269$). With our model set A we find that our best-fitting model systems have orbital periods between $600$ days and $400,\!000$ days with an average period of $90,\!000$ days. With model sets B and C we obtain on average ${P_{\mathrm{f}}}=17,\!000$ and ${P_{\mathrm{f}}}=60,\!000$ days, respectively. On the other hand, the average orbital period of the systems analysed in Paper I is approximately $1,\!500$ days. Furthermore, a recent statistical study of currently available radial velocity variations in CEMP-$s$ stars indicates that these stars are expected to have a maximum period of about $10,\!000$ days and an average period of $400$ days [@Starkenburg2014]. This suggests that a large proportion of the CEMP-$s$ stars in our sample have periods of less than a few thousand days, i.e. in closer orbits than we find in our best fits. If this hypothesis is correct, then AGB stars in binary systems have to transfer wind material with great efficiency at relatively short separations and this process is currently possible in our models only with some ad hoc assumptions, namely highly efficient mass-transfer and efficient angular momentum loss. This hypothesis is supported by the period distribution derived in barium stars, the equivalent of CEMP-$s$ stars at solar metallicity [e.g. @Jorissen1998], and by observations of ongoing mass transfer in post-AGB binaries [e.g. @Gorlova2012]. Summary and conclusions {#conclusions} ======================= In this work we have used our model of binary evolution and nucleosynthesis to calculate the best-fitting models to the surface abundances observed in a sample of $67$ CEMP-$s$ stars. We distinguish three classes of CEMP-$s$ stars based on the europium abundance. In CEMP-$s/nr$ stars the abundance of europium relative to iron is up to ten times as high as in the Sun. In CEMP-$s/r$ stars the europium-to-iron ratio is higher than ten times the solar ratio. In CEMP-$s/ur$ stars the abundance of europium is indeterminate. From the comparison between the observed and modelled abundances of these stars we draw the following conclusions. - The model stars that provide the best fit to the abundances observed in CEMP-$s$ stars have low initial mass (up to $2.5{M_{\odot}}$). In this range, the main neutron source for the $s$-process is the ${^{13}\mathrm{C}}(\alpha,\,n)^{16}{\mathrm{O}}$ reaction. - The chemical properties observed in CEMP-$s/nr$ and CEMP-$s/r$ stars are fundamentally different. The results of our models are consistent with the abundances observed in CEMP-$s/nr$ stars. On the contrary, most of the model CEMP-$s/r$ stars overestimate the observed abundances of light-$s$ elements and underestimate the abundances of heavy-$s$ elements, $r$-elements and lead. In CEMP-$s/r$ stars the ratio of heavy-$s$ elements to light-$s$ elements is too high to be reproduced in our models. This result points to a different nucleosynthesis process at the origin of the abundances in CEMP-$s/r$ stars, that enhances the heaviest elements without affecting the abundances of elements such as sodium, magnesium and light-$s$ elements. - The abundance of sodium is always overpredicted in our models. This discrepancy likely points to a problem in the numerical treatment of the partial mixing zone. A proton profile in the partial mixing zone weighted towards a low proton abundance may contribute to reducing this discrepancy. - The orbital periods predicted in this work are significantly longer than the results of Paper I, in which our models are constrained to reproduce the observed orbital periods of the systems. This indicates that either the currently known periods of CEMP-$s$ stars are biased towards close orbits and are not representative of the whole population, or our model of wind accretion in binary stars needs to efficiently transfer mass and lose angular momentum in close orbits. We are grateful to Dr. T. Suda for his kind support and help to access the SAGA database. We thank Dr. M. Lugaro, Dr. R.J. Stancliffe, Prof. R. Gallino and Dr. T. Masseron for many useful comments and constructive criticism of our work. CA is grateful for financial support from the Netherlands Organisation for Scientific Research (NWO) under grant 614.000.901. RGI thanks the Humboldt Foundation and the Science and Technology Facilities Council (STFC) for funding support. We thank the referee for her/his helpful comments that improved the clarity of this paper. [72]{} natexlab\#1[\#1]{} , C., [Pols]{}, O. R., [Izzard]{}, R. G., [Mohamed]{}, S. S., & [de Mink]{}, S. E. 2013, , 552, A26 , C., [Pols]{}, O. R., [Karakas]{}, A. I., & [Izzard]{}, R. G. 2015, , 576, A118 , C., [Pols]{}, O. R., Stancliffe, R. J., [et al.]{} 2015, , [*in press*]{} , S. M., [Spite]{}, M., [Korotin]{}, S. A., [et al.]{} 2009, , 494, 1083 , W., [Beers]{}, T. C., [Christlieb]{}, N., [et al.]{} 2007, , 655, 492 , C., [K[ä]{}ppeler]{}, F., [Wisshak]{}, K., [et al.]{} 1999, , 525, 886 , M., [Grevesse]{}, N., [Sauval]{}, A. J., & [Scott]{}, P. 2009, , 47, 481 , Z. Y., [Beer]{}, H., [K[ä]{}ppeler]{}, F., [et al.]{} 2000, Atomic Data and Nuclear Data Tables, 76, 70 , T. C. & [Christlieb]{}, N. 2005, , 43, 531 , T. C., [Preston]{}, G. W., & [Shectman]{}, S. A. 1992, , 103, 1987 , S., [Gallino]{}, R., [Straniero]{}, O., [Cristallo]{}, S., & [K[ä]{}ppeler]{}, F. 2011, , 418, 284 , S., [Gallino]{}, R., [Straniero]{}, O., [Cristallo]{}, S., & [K[ä]{}ppeler]{}, F. 2012, , 422, 849 , H. M. J. & [Jorissen]{}, A. 1988, , 205, 155 , A. I. & [Sackmann]{}, I.-J. 1999, , 510, 232 , M. L., [McQuinn]{}, K. B. W., [Barmby]{}, P., [et al.]{} 2015, , 800, 51 , M., [Gallino]{}, R., & [Wasserburg]{}, G. J. 1999, , 37, 239 , M., [Lambert]{}, D. L., [Beglio]{}, L., [et al.]{} 1995, , 446, 775 , S. W. 2007, PhD thesis, Monash University , S. W. & [Lattanzio]{}, J. C. 2008, , 490, 769 , C., [Brown]{}, J. A., & [Wallerstein]{}, G. 1998, , 332, 204 , N., [Green]{}, P. J., [Wisotzki]{}, L., & [Reimers]{}, D. 2001, , 375, 366 , J. G., [Christlieb]{}, N., [Qian]{}, Y.-Z., & [Wasserburg]{}, G. J. 2003, , 588, 1082 , J. G., [Shectman]{}, S., [Thompson]{}, I., [et al.]{} 2005, , 633, L109 , T., [Campbell]{}, S., [Gil-Pons]{}, P., & [Lattanzio]{}, J. 2014, , 784, 56 , C. K., [Karakas]{}, A. I., & [Stancliffe]{}, R. J. 2014, , 438, 1741 , A., [Christlieb]{}, N., [Norris]{}, J. E., [et al.]{} 2006, , 652, 1585 , R., [Arlandini]{}, C., [Busso]{}, M., [et al.]{} 1998, , 497, 388 , S. & [Mowlavi]{}, N. 2000, , 362, 599 , N., [Van Winckel]{}, H., & [Jorissen]{}, A. 2012, Baltic Astronomy, 21, 165 , F. 2005, , 43, 435 , F., [Pignatari]{}, M., [Woodward]{}, P. R., [et al.]{} 2011, , 727, 89 , D. & [Iben]{}, Jr., I. 1988, , 333, L25 , C., [Longland]{}, R., [Champagne]{}, A. E., & [Coc]{}, A. 2010, Nuclear Physics A, 841, 323 , R. G., [Dermine]{}, T., & [Church]{}, R. P. 2010, , 523, A10+ , R. G., [Dray]{}, L. M., [Karakas]{}, A. I., [Lugaro]{}, M., & [Tout]{}, C. A. 2006, , 460, 565 , R. G., [Glebbeek]{}, E., [Stancliffe]{}, R. J., & [Pols]{}, O. R. 2009, , 508, 1359 , R. G., [Tout]{}, C. A., [Karakas]{}, A. I., & [Pols]{}, O. R. 2004, , 350, 407 , K., [Barklem]{}, P. S., [Gustafsson]{}, B., [et al.]{} 2006, , 451, 651 , A., [Van Eck]{}, S., [Mayor]{}, M., & [Udry]{}, S. 1998, , 332, 877 , A. & [Lattanzio]{}, J. C. 2007, , 24, 103 , A. I. 2010, , 403, 1413 , A. I., [Campbell]{}, S. W., & [Stancliffe]{}, R. J. 2010, , 713, 374 , Y., [Davis]{}, A. M., [Gallino]{}, R., [et al.]{} 2010, , 713, 212 , T. & [Jorgensen]{}, U. G. 1994, , 290, 148 , C., [Karakas]{}, A. I., & [Umeda]{}, H. 2011, , 414, 3231 , J. C. 1991, Meteoritics and Planetary Science, 47, 1998 , Y. S., [Beers]{}, T. C., [Masseron]{}, T., [et al.]{} 2013, , 146, 132 , S., [Beers]{}, T. C., [Christlieb]{}, N., [et al.]{} 2006, , 652, L37 , M., [Campbell]{}, S. W., & [de Mink]{}, S. E. 2009, , 26, 322 , M., [Karakas]{}, A. I., [Stancliffe]{}, R. J., & [Rijs]{}, C. 2012, , 47, 1998 , M., [Tagliente]{}, G., [Karakas]{}, A. I., [et al.]{} 2014, , 780, 95 , B., [Beers]{}, T. C., [Rossi]{}, S., [et al.]{} 2005, Nuclear Physics A, 758, 312 , T., [Johnson]{}, J. A., [Plez]{}, B., [et al.]{} 2010, , 509, A93 , S. N., [Woolf]{}, N. J., & [Ziurys]{}, L. M. 2009, , 690, 837 , V. M., [Frebel]{}, A., [Beers]{}, T. C., [et al.]{} 2013, , 770, 104 , O. R., [Izzard]{}, R. G., [Stancliffe]{}, R. J., & [Glebbeek]{}, E. 2012, , 547, A76 , G. W. & [Sneden]{}, C. 2001, , 122, 1545 , R., [Lederer]{}, C., & [K[ä]{}ppeler]{}, F. 2014, Journal of Physics G Nuclear Physics, 41, 053101 , S. G., [Aoki]{}, W., [Norris]{}, J. E., & [Beers]{}, T. C. 2005, , 635, 349 , C., [Cowan]{}, J. J., & [Gallino]{}, R. 2008, , 46, 241 , R. J. 2010, , 403, 505 , R. J., [Glebbeek]{}, E., [Izzard]{}, R. G., & [Pols]{}, O. R. 2007, , 464, L57 , E., [Shetrone]{}, M. D., [McConnachie]{}, A. W., & [Venn]{}, K. A. 2014, , 441, 1217 , T., [Katsuta]{}, Y., [Yamada]{}, S., [et al.]{} 2008, , 60, 1159 , T., [Yamada]{}, S., [Katsuta]{}, Y., [et al.]{} 2011, , 412, 843 , I. B., [Ivans]{}, I. I., [Bisterzo]{}, S., [et al.]{} 2008, , 677, 556 , C., [Gallino]{}, R., [Busso]{}, M., & [Gratton]{}, R. 2001, , 549, 346 , G. & [Knapp]{}, G. R. 1998, , 36, 369 , S., [Nomoto]{}, K., [Iwamoto]{}, N., [Ishimaru]{}, Y., & [Beers]{}, T. C. 2006, , 636, 842 , B., [Rockosi]{}, C., [Newberg]{}, H. J., [et al.]{} 2009, , 137, 4377 , D., [Norris]{}, J. E., [Bessell]{}, M. S., [et al.]{} 2013, , 762, 26 , A. A. 2004, , 348, L23 Tables {#app:A} ====== In this section we summarise the observed and modelled properties of the $67$ CEMP-$s$ stars in our sample. Table \[tab:obs\] shows the observed temperatures, surface gravities and abundances of selected elements. The parameters of the best-fitting models to the CEMP-$s$ stars with at least of two degrees of freedom ($\nu\ge2$), as computed with model sets A, B and C, are shown in Tables \[tab:best-WRLOFq-ssw\]–\[tab:best-WRLOFq-gamma2\], as follows. - Columns 2–5: the fitted parameters ${M_{1,\mathrm{i}}}$, ${M_{\mathrm{PMZ}}}$, ${M_{2,\mathrm{i}}}$ and ${P_{\mathrm{i}}}$. - Columns 6–8: the mass accreted by the secondary star, $\Delta{M_{\mathrm{acc}}}$, the orbital period of the binary when the secondary star best reproduces the observed ${\log g}$ and surface abundances, ${P_{\mathrm{f}}}$, and for stars with observed orbital periods, the final period determined in Paper I. - Columns 9–11: ${\chi^2_{\mathrm{min}}}$, the number of degrees of freedom of the fit, $\nu$, and the reduced ${\chi^2}$ of the best fit, i.e. ${\chi^2_{\mathrm{min}}}/\nu={\chi^2}/\nu$. In Tables \[tab:few-WRLOFq-ssw\]–\[tab:few-WRLOFq-gamma2\] the physical parameters of the best-fitting models of stars with $\nu\le1$ are shown.The confidence intervals of the input parameters of our models with ${\chi^2_{\mathrm{min}}}/\nu\le3$ and $\nu\ge2$ computed with model set A are shown in Table \[tab:confi\_WRLOFq-ssw\]. ----------------------- ----------------------------------------------- --------------------------------- ------------------------------------ ------------------- -------------------------------- -------------------------------- --------------------- --------------------------------- ID ${\log_{10}(g/\mathrm{cm}\,\mathrm{s}^{-2})}$ ${T_{\mathrm{eff}}}/\mathrm{K}$ ${P_{\mathrm{orb}}}/\mathrm{days}$ number of $[{\mathrm{Fe}}/{\mathrm{H}}]$ $[{\mathrm{C}}/{\mathrm{Fe}}]$ $[s/{\mathrm{Fe}}]$ $[{\mathrm{Eu}}/{\mathrm{Fe}}]$ observed elements CEMP-$s/nr$ stars CS22880$-$074 $3.9\pm0.13$ $5917$ $ $ $21$ $-1.9$ $1.5$ $1.3$ $0.6$ CS22942$-$019 $2.2\pm0.4$ $4967$ $2800$ $18$ $-2.7$ $2.2$ $1.8$ $0.8$ CS22964$-$161A $3.7\pm0.2$ $6050$ $252.481$ $21$ $-2.4$ $1.6$ $1.4$ $0.7$ CS22964$-$161B $4.1\pm0.4$ $5850$ $252.481$ $22$ $-2.4$ $1.4$ $1.3$ $0.7$ CS30301$-$015 $0.8\pm0.1$ $4750$ $ $ $17$ $-2.7$ $1.7$ $1.5$ $0.3$ HD196944 $1.8\pm0.1$ $5234$ $ $ $27$ $-2.4$ $1.4$ $1.2$ $0.3$ HD198269 $1.3\pm0.25$ $4800$ $1295$ $11$ $-2.2$ $1.7$ $1.7$ $0.9$ HE0202$-$2204 $1.6\pm0.1$ $5280$ $ $ $21$ $-2.0$ $1.2$ $1.4$ $0.5$ HE1135$+$0139 $1.8\pm0.1$ $5487$ $ $ $20$ $-2.4$ $1.1$ $1.1$ $0.4$ HE2158$-$0348 $2.5\pm0.1$ $5215$ $ $ $18$ $-2.8$ $2.1$ $1.6$ $0.9$ CEMP-$s/r$ stars BS16080$-$175 $3.7\pm0.2$ $6240$ $ $ $6$ $-1.9$ $1.8$ $1.6$ $1.1$ BS17436$-$058 $2.7\pm0.2$ $5690$ $ $ $7$ $-1.8$ $1.6$ $1.7$ $1.2$ CS22881$-$036 $4.0\pm0.1$ $6200$ $ $ $14$ $-2.1$ $2.1$ $1.9$ $1.0$ CS22898$-$027 $3.7\pm0.28$ $6110$ $ $ $22$ $-2.3$ $2.0$ $2.3$ $2.0$ CS22948$-$027 $1.8\pm0.4$ $4800$ $426.5$ $21$ $-2.5$ $2.4$ $2.4$ $1.9$ CS29497$-$030 $4.0\pm0.5$ $6966$ $344$ $33$ $-2.5$ $2.4$ $2.3$ $1.7$ CS29526$-$110 $3.2\pm0.1$ $6500$ $ $ $18$ $-2.4$ $2.3$ $2.1$ $1.8$ CS31062$-$012 $4.2\pm0.38$ $6099$ $ $ $24$ $-2.8$ $2.3$ $2.1$ $1.6$ CS31062$-$050 $2.9\pm0.24$ $5489$ $ $ $37$ $-2.5$ $1.9$ $2.4$ $2.0$ HD187861 $2.0\pm0.35$ $4960$ $ $ $14$ $-2.4$ $2.0$ $1.9$ $1.3$ HD224959 $1.9\pm0.25$ $5050$ $1273$ $14$ $-2.1$ $1.8$ $2.2$ $1.7$ HE0131$-$3953 $3.8\pm0.1$ $5928$ $ $ $16$ $-2.7$ $2.5$ $2.2$ $1.7$ HE0143$-$0441 $4.0\pm0.35$ $6305$ $ $ $22$ $-2.4$ $2.0$ $2.4$ $1.7$ HE0338$-$3945 $4.1\pm0.33$ $6161$ $ $ $32$ $-2.5$ $2.1$ $2.4$ $2.0$ HE1105$+$0027 $3.5\pm0.1$ $6132$ $ $ $16$ $-2.5$ $2.0$ $2.4$ $1.9$ HE1305$+$0007 $1.5\pm0.5$ $4655$ $ $ $21$ $-2.2$ $2.1$ $2.6$ $2.2$ HE2148$-$1247 $3.9\pm0.1$ $6380$ $ $ $25$ $-2.4$ $2.0$ $2.4$ $2.0$ HE2258$-$6358 $1.6\pm0.1$ $4900$ $ $ $31$ $-2.7$ $2.4$ $2.3$ $1.7$ LP625$-$44 $2.6\pm0.3$ $5500$ $>4383$ $31$ $-2.8$ $2.3$ $2.8$ $1.9$ SDSSJ0912$+$0216 $4.5\pm0.1$ $6500$ $ $ $28$ $-2.6$ $2.3$ $1.6$ $1.3$ CEMP-$s/ur$ stars BD$+$04$^{\circ}$2466 $1.8\pm0.2$ $5032$ $4592.7$ $20$ $-2.1$ $1.3$ $1.6$ $ $ CS22945$-$017 $3.8\pm0.2$ $6400$ $ $ $6$ $-2.5$ $2.3$ $0.6$ $ $ CS22956$-$028 $3.9\pm0.1$ $6900$ $1290$ $11$ $-2.1$ $1.9$ $0.4$ $ $ CS29509$-$027 $4.2\pm0.1$ $7050$ $196$ $5$ $-2.1$ $1.5$ $1.3$ $ $ CS30338$-$089 $2.1\pm0.1$ $5000$ $ $ $10$ $-2.5$ $2.1$ $2.3$ $ $ HD13826 $0.1\pm0.35$ $3540$ $ $ $9$ $-2.5$ $1.6$ $1.4$ $ $ HD187216 $0.4\pm0.4$ $3500$ $ $ $14$ $-2.5$ $1.4$ $2.3$ $ $ HD201626 $2.2\pm0.25$ $5190$ $407$ $11$ $-2.1$ $2.1$ $2.0$ $ $ HD5223 $1.0\pm0.25$ $4500$ $ $ $18$ $-2.1$ $1.6$ $1.9$ $ $ HE0012$-$1441 $3.5\pm0.1$ $5730$ $ $ $7$ $-2.7$ $1.9$ $1.3$ $ $ HE0024$-$2523 $4.3\pm0.1$ $6625$ $3.14$ $17$ $-2.7$ $2.1$ $1.6$ $ $ HE0206$-$1916 $2.7\pm0.1$ $5200$ $ $ $8$ $-2.2$ $2.1$ $2.0$ $ $ HE0212$-$0557 $2.1\pm0.4$ $5075$ $ $ $15$ $-2.3$ $1.9$ $2.2$ $ $ HE0231$-$4016 $3.6\pm0.1$ $5972$ $ $ $17$ $-2.1$ $1.3$ $1.5$ $ $ HE0322$-$1504 $0.8\pm0.1$ $4460$ $ $ $5$ $-2.0$ $2.4$ $2.8$ $ $ HE0430$-$4404 $4.3\pm0.1$ $6214$ $ $ $15$ $-2.1$ $1.3$ $1.6$ $ $ HE0441$-$0652 $1.4\pm0.1$ $4900$ $ $ $11$ $-2.6$ $1.4$ $1.2$ $ $ HE0507$-$1430 $0.8\pm0.1$ $4560$ $446$ $5$ $-2.4$ $2.7$ $1.3$ $ $ HE0534$-$4548 $1.5\pm0.3$ $4250$ $ $ $4$ $-1.8$ $1.5$ $0.6$ $ $ HE1045$-$1434 $1.8\pm0.1$ $4950$ $ $ $5$ $-2.5$ $3.3$ $3.0$ $ $ HE1157$-$0518 $2.0\pm0.1$ $4900$ $ $ $8$ $-2.4$ $2.2$ $2.2$ $ $ HE1429$-$0551 $1.5\pm0.1$ $4700$ $ $ $7$ $-2.6$ $2.3$ $1.7$ $ $ HE1430$-$1123 $3.8\pm0.1$ $5915$ $ $ $15$ $-2.7$ $1.8$ $1.8$ $ $ HE1434$-$1442 $3.1\pm0.4$ $5420$ $ $ $15$ $-2.5$ $2.1$ $1.3$ $ $ HE1443$+$0113 $1.9\pm0.1$ $4945$ $ $ $5$ $-2.2$ $2.0$ $1.4$ $ $ HE1447$+$0102 $1.7\pm0.1$ $5100$ $ $ $7$ $-2.5$ $2.5$ $2.8$ $ $ HE1523$-$1155 $1.6\pm0.1$ $4800$ $ $ $6$ $-2.2$ $1.9$ $1.8$ $ $ HE1528$-$0409 $1.8\pm0.1$ $5000$ $ $ $6$ $-2.7$ $2.4$ $2.4$ $ $ HE2150$-$0825 $3.7\pm0.1$ $5960$ $ $ $18$ $-2.0$ $1.4$ $1.7$ $ $ HE2221$-$0453 $0.4\pm0.1$ $4400$ $ $ $7$ $-2.3$ $1.8$ $1.8$ $ $ HE2227$-$4044 $3.9\pm0.1$ $5811$ $ $ $14$ $-2.4$ $1.7$ $1.4$ $ $ HE2228$-$0706 $2.6\pm0.1$ $5100$ $ $ $8$ $-2.5$ $2.3$ $2.6$ $ $ HE2232$-$0603 $3.5\pm0.1$ $5750$ $ $ $17$ $-2.1$ $1.6$ $1.6$ $ $ HE2240$-$0412 $4.3\pm0.1$ $5852$ $ $ $13$ $-2.2$ $1.4$ $1.4$ $ $ SDSS0924$+$40 $4.0\pm0.3$ $6200$ $ $ $12$ $-2.6$ $2.7$ $1.9$ $ $ SDSS1707$+$58 $4.2\pm0.3$ $6700$ $ $ $8$ $-2.6$ $2.1$ $3.5$ $ $ SDSS2047$+$00 $4.5\pm0.3$ $6600$ $ $ $12$ $-2.1$ $2.0$ $1.6$ $ $ ----------------------- ----------------------------------------------- --------------------------------- ------------------------------------ ------------------- -------------------------------- -------------------------------- --------------------- --------------------------------- ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${P_{\mathrm{f}}}-$Paper I ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ----------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- ---------------------------- --------------------------- ------- ------------------------------- CEMP-$s/nr$ stars CS22880$-$074 $1.1$ $1.00 \times 10^{-3}$ $0.74$ $4.54 \times 10^{3}$ $0.09$ $6.03 \times 10^{3}$ $16.8$ $10$ 1.7 CS22942$-$019 $1.8$ $2.00 \times 10^{-4}$ $0.54$ $6.38 \times 10^{4}$ $0.30$ $1.03 \times 10^{5}$ $ 2.87\times 10^3 $ $21.1$ $9$ 2.3 CS22964$-$161A $1.6$ $2.00 \times 10^{-3}$ $0.79$ $2.74 \times 10^{5}$ $0.04$ $5.60 \times 10^{5}$ $ 5.60\times 10^5 $ $19.5$ $8$ 2.5 CS22964$-$161B $1.6$ $2.00 \times 10^{-3}$ $0.71$ $2.74 \times 10^{5}$ $0.02$ $5.60 \times 10^{5}$ $ 5.60\times 10^5 $ $11.8$ $9$ 1.3 CS30301$-$015 $2.5$ $4.00 \times 10^{-3}$ $0.76$ $4.81 \times 10^{3}$ $0.06$ $1.68 \times 10^{4}$ $11.9$ $8$ 1.5 HD196944 $1.4$ $1.00 \times 10^{-3}$ $0.81$ $5.84 \times 10^{3}$ $0.06$ $1.14 \times 10^{4}$ $13.8$ $12$ 1.2 HD198269 $1.0$ $4.00 \times 10^{-3}$ $0.79$ $1.63 \times 10^{3}$ $0.09$ $1.97 \times 10^{3}$ $ 1.43\times 10^3 $ $9.2$ $5$ 1.8 HE0202$-$2204 $0.9$ $3.00 \times 10^{-3}$ $0.79$ $1.19 \times 10^{3}$ $0.06$ $1.43 \times 10^{3}$ $5.2$ $7$ 0.7 HE1135$+$0139 $0.9$ $1.00 \times 10^{-3}$ $0.79$ $3.35 \times 10^{3}$ $0.08$ $4.01 \times 10^{3}$ $2.8$ $6$ 0.5 HE2158$-$0348 $1.5$ $3.00 \times 10^{-3}$ $0.69$ $1.32 \times 10^{5}$ $0.17$ $2.52 \times 10^{5}$ $16.4$ $7$ 2.3 CEMP-$s/r$ stars CS22881$-$036 $1.5$ $6.66 \times 10^{-4}$ $0.66$ $1.32 \times 10^{5}$ $0.15$ $2.57 \times 10^{5}$ $10.3$ $6$ 1.7 CS22898$-$027 $1.5$ $6.66 \times 10^{-4}$ $0.51$ $2.44 \times 10^{4}$ $0.31$ $2.89 \times 10^{4}$ $99.8$ $11$ 9.1 CS22948$-$027 $1.5$ $2.00 \times 10^{-3}$ $0.64$ $9.42 \times 10^{4}$ $0.24$ $1.58 \times 10^{5}$ $ 4.76\times 10^2 $ $29.5$ $9$ 3.3 CS29497$-$030 $1.5$ $2.00 \times 10^{-3}$ $0.46$ $2.47 \times 10^{4}$ $0.28$ $2.87 \times 10^{4}$ $ 4.04\times 10^2 $ $126.0$ $16$ 7.9 CS29526$-$110 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $2.42 \times 10^{4}$ $0.32$ $2.92 \times 10^{4}$ $22.4$ $6$ 3.7 CS31062$-$012 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $6.79 \times 10^{4}$ $0.25$ $1.00 \times 10^{5}$ $17.5$ $8$ 2.2 CS31062$-$050 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $2.42 \times 10^{4}$ $0.32$ $2.92 \times 10^{4}$ $96.6$ $20$ 4.8 HD187861 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.49 \times 10^{4}$ $16.6$ $8$ 2.1 HD224959 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $2.42 \times 10^{4}$ $0.32$ $2.96 \times 10^{4}$ $ 1.42\times 10^3 $ $64.6$ $8$ 8.1 HE0131$-$3953 $1.5$ $2.00 \times 10^{-3}$ $0.61$ $6.71 \times 10^{4}$ $0.28$ $9.88 \times 10^{4}$ $13.2$ $4$ 3.3 HE0143$-$0441 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.43 \times 10^{4}$ $37.2$ $7$ 5.3 HE0338$-$3945 $1.5$ $6.66 \times 10^{-4}$ $0.49$ $2.45 \times 10^{4}$ $0.30$ $2.88 \times 10^{4}$ $183.7$ $17$ 10.8 HE1105$+$0027 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $3.40 \times 10^{4}$ $0.32$ $4.28 \times 10^{4}$ $23.0$ $4$ 5.8 HE1305$+$0007 $1.5$ $6.66 \times 10^{-4}$ $0.54$ $2.42 \times 10^{4}$ $0.32$ $2.97 \times 10^{4}$ $56.0$ $10$ 5.6 HE2148$-$1247 $1.5$ $2.00 \times 10^{-3}$ $0.51$ $2.44 \times 10^{4}$ $0.31$ $2.89 \times 10^{4}$ $204.7$ $11$ 18.6 HE2258$-$6358 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.62 \times 10^{4}$ $293.5$ $18$ 16.3 LP625$-$44 $1.7$ $4.00 \times 10^{-3}$ $0.49$ $2.34 \times 10^{4}$ $0.34$ $2.86 \times 10^{4}$ $ 4.03\times 10^3 $ $173.2$ $17$ 10.2 SDSSJ0912$+$0216 $1.5$ $2.00 \times 10^{-3}$ $0.46$ $2.47 \times 10^{4}$ $0.28$ $2.87 \times 10^{4}$ $425.0$ $14$ 30.4 CEMP-$s/ur$ stars BD$+$04$^{\circ}$2466 $1.1$ $2.00 \times 10^{-3}$ $0.76$ $3.19 \times 10^{3}$ $0.09$ $4.27 \times 10^{3}$ $ 4.29\times 10^3 $ $13.4$ $7$ 1.9 HD13826 $1.7$ $0$ $0.79$ $3.90 \times 10^{3}$ $0.07$ $8.85 \times 10^{3}$ $3.7$ $3$ 1.2 HD187216 $1.5$ $5.00 \times 10^{-4}$ $0.56$ $1.71 \times 10^{4}$ $0.32$ $2.33 \times 10^{4}$ $28.7$ $4$ 7.2 HD201626 $1.4$ $4.00 \times 10^{-3}$ $0.71$ $1.34 \times 10^{5}$ $0.18$ $2.33 \times 10^{5}$ $ 4.75\times 10^2 $ $7.5$ $5$ 1.5 HD5223 $1.1$ $1.50 \times 10^{-3}$ $0.76$ $1.79 \times 10^{4}$ $0.17$ $2.44 \times 10^{4}$ $15.0$ $9$ 1.7 HE0024$-$2523 $0.9$ $2.00 \times 10^{-3}$ $0.81$ $4.19 \times 10^{2}$ $0.12$ $4.75 \times 10^{2}$ $ 3.33 $ $35.4$ $4$ 8.8 HE0212$-$0557 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.51 \times 10^{4}$ $67.9$ $5$ 13.6 HE0231$-$4016 $0.9$ $4.00 \times 10^{-3}$ $0.79$ $8.42 \times 10^{2}$ $0.07$ $8.81 \times 10^{2}$ $1.4$ $4$ 0.4 HE0430$-$4404 $0.9$ $4.00 \times 10^{-3}$ $0.84$ $5.88 \times 10^{2}$ $0.09$ $6.25 \times 10^{2}$ $1.0$ $2$ 0.5 HE1430$-$1123 $1.5$ $2.00 \times 10^{-3}$ $0.76$ $1.83 \times 10^{5}$ $0.08$ $3.97 \times 10^{5}$ $4.6$ $2$ 2.3 HE1434$-$1442 $1.0$ $3.00 \times 10^{-3}$ $0.79$ $2.06 \times 10^{5}$ $0.05$ $2.73 \times 10^{5}$ $17.7$ $3$ 5.9 HE2150$-$0825 $0.9$ $1.50 \times 10^{-3}$ $0.84$ $5.88 \times 10^{2}$ $0.09$ $6.25 \times 10^{2}$ $1.1$ $5$ 0.2 HE2232$-$0603 $0.9$ $1.00 \times 10^{-3}$ $0.84$ $5.88 \times 10^{2}$ $0.09$ $6.25 \times 10^{2}$ $9.7$ $4$ 2.4 SDSS0924$+$40 $1.5$ $2.00 \times 10^{-3}$ $0.51$ $2.44 \times 10^{4}$ $0.31$ $2.89 \times 10^{4}$ $14.0$ $2$ 7.0 SDSS2047$+$00 $1.5$ $6.66 \times 10^{-4}$ $0.54$ $1.36 \times 10^{5}$ $0.07$ $3.05 \times 10^{5}$ $0.7$ $3$ 0.2 ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${P_{\mathrm{f}}}-$Paper I ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ----------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- ---------------------------- --------------------------- ------- ------------------------------- CEMP-$s/nr$ stars CS22880$-$074 $1.1$ $1.00 \times 10^{-3}$ $0.79$ $7.10 \times 10^{4}$ $0.05$ $5.42 \times 10^{4}$ $16.6$ $10$ 1.7 CS22942$-$019 $1.8$ $2.00 \times 10^{-4}$ $0.56$ $1.59 \times 10^{4}$ $0.32$ $3.48 \times 10^{3}$ $ 3.23\times 10^3 $ $21.0$ $9$ 2.3 CS22964$-$161A $1.6$ $2.00 \times 10^{-3}$ $0.79$ $1.30 \times 10^{5}$ $0.04$ $1.56 \times 10^{5}$ $ 1.30\times 10^5 $ $19.6$ $8$ 2.5 CS22964$-$161B $1.6$ $2.00 \times 10^{-3}$ $0.71$ $1.30 \times 10^{5}$ $0.02$ $1.56 \times 10^{5}$ $ 1.30\times 10^5 $ $11.8$ $9$ 1.3 CS30301$-$015 $2.4$ $4.00 \times 10^{-3}$ $0.79$ $2.17 \times 10^{5}$ $0.04$ $4.09 \times 10^{4}$ $11.0$ $8$ 1.4 HD196944 $1.2$ $1.00 \times 10^{-3}$ $0.79$ $6.92 \times 10^{4}$ $0.04$ $4.63 \times 10^{4}$ $13.6$ $12$ 1.1 HD198269 $1.5$ $3.00 \times 10^{-3}$ $0.79$ $4.56 \times 10^{4}$ $0.08$ $2.29 \times 10^{4}$ $ 1.17\times 10^3 $ $9.4$ $5$ 1.9 HE0202$-$2204 $0.9$ $4.00 \times 10^{-3}$ $0.79$ $1.68 \times 10^{3}$ $0.07$ $1.51 \times 10^{3}$ $5.1$ $7$ 0.7 HE1135$+$0139 $0.9$ $6.66 \times 10^{-4}$ $0.81$ $1.05 \times 10^{5}$ $0.03$ $9.47 \times 10^{4}$ $3.2$ $6$ 0.5 HE2158$-$0348 $1.5$ $3.00 \times 10^{-3}$ $0.74$ $2.31 \times 10^{4}$ $0.15$ $1.17 \times 10^{4}$ $16.5$ $7$ 2.4 CEMP-$s/r$ stars CS22881$-$036 $1.5$ $6.66 \times 10^{-4}$ $0.71$ $2.93 \times 10^{3}$ $0.14$ $6.86 \times 10^{1}$ $9.4$ $6$ 1.6 CS22898$-$027 $1.1$ $1.00 \times 10^{-3}$ $0.64$ $3.30 \times 10^{3}$ $0.22$ $2.24 \times 10^{3}$ $114.0$ $11$ 10.4 CS22948$-$027 $1.5$ $2.00 \times 10^{-3}$ $0.61$ $5.98 \times 10^{3}$ $0.25$ $2.56 \times 10^{2}$ $ 3.69\times 10^2 $ $29.5$ $9$ 3.3 CS29497$-$030 $1.6$ $2.00 \times 10^{-3}$ $0.51$ $1.19 \times 10^{4}$ $0.30$ $2.31 \times 10^{3}$ $ 3.00\times 10^2 $ $145.8$ $16$ 9.1 CS29526$-$110 $1.6$ $2.00 \times 10^{-3}$ $0.54$ $1.19 \times 10^{4}$ $0.30$ $3.07 \times 10^{3}$ $25.2$ $6$ 4.2 CS31062$-$012 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $8.60 \times 10^{3}$ $0.23$ $4.69 \times 10^{2}$ $17.6$ $8$ 2.2 CS31062$-$050 $1.6$ $2.00 \times 10^{-3}$ $0.54$ $1.19 \times 10^{4}$ $0.30$ $3.07 \times 10^{3}$ $115.9$ $20$ 5.8 HD187861 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $1.20 \times 10^{4}$ $0.26$ $4.35 \times 10^{3}$ $17.3$ $8$ 2.2 HD224959 $1.8$ $4.00 \times 10^{-3}$ $0.54$ $1.60 \times 10^{4}$ $0.31$ $2.59 \times 10^{3}$ $ 1.41\times 10^3 $ $70.5$ $8$ 8.8 HE0131$-$3953 $1.5$ $2.00 \times 10^{-3}$ $0.61$ $5.98 \times 10^{3}$ $0.25$ $2.59 \times 10^{2}$ $13.6$ $4$ 3.4 HE0143$-$0441 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $1.21 \times 10^{4}$ $0.24$ $3.65 \times 10^{3}$ $39.5$ $7$ 5.6 HE0338$-$3945 $1.1$ $1.50 \times 10^{-3}$ $0.59$ $3.35 \times 10^{3}$ $0.22$ $2.05 \times 10^{3}$ $211.8$ $17$ 12.5 HE1105$+$0027 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $1.20 \times 10^{4}$ $0.26$ $4.38 \times 10^{3}$ $24.2$ $4$ 6.0 HE1305$+$0007 $1.1$ $1.00 \times 10^{-3}$ $0.66$ $3.28 \times 10^{3}$ $0.22$ $2.30 \times 10^{3}$ $64.2$ $10$ 6.4 HE2148$-$1247 $1.6$ $2.00 \times 10^{-3}$ $0.54$ $1.19 \times 10^{4}$ $0.30$ $3.07 \times 10^{3}$ $243.8$ $11$ 22.2 HE2258$-$6358 $1.5$ $1.00 \times 10^{-3}$ $0.59$ $1.20 \times 10^{4}$ $0.26$ $4.28 \times 10^{3}$ $301.9$ $18$ 16.8 LP625$-$44 $1.8$ $4.00 \times 10^{-3}$ $0.54$ $1.60 \times 10^{4}$ $0.31$ $2.61 \times 10^{3}$ $ 3.65\times 10^3 $ $188.9$ $17$ 11.1 SDSSJ0912$+$0216 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $8.60 \times 10^{3}$ $0.23$ $4.69 \times 10^{2}$ $454.2$ $14$ 32.4 CEMP-$s/ur$ stars BD$+$04$^{\circ}$2466 $0.9$ $2.00 \times 10^{-3}$ $0.76$ $8.49 \times 10^{2}$ $0.12$ $6.64 \times 10^{2}$ $ 4.25\times 10^3 $ $13.5$ $7$ 1.9 HD13826 $1.7$ $0$ $0.76$ $8.78 \times 10^{4}$ $0.05$ $2.95 \times 10^{4}$ $4.0$ $3$ 1.3 HD187216 $1.1$ $2.00 \times 10^{-4}$ $0.66$ $3.28 \times 10^{3}$ $0.22$ $2.07 \times 10^{3}$ $23.8$ $4$ 5.9 HD201626 $1.4$ $4.00 \times 10^{-3}$ $0.66$ $1.71 \times 10^{4}$ $0.16$ $8.35 \times 10^{3}$ $ 3.55\times 10^2 $ $7.8$ $5$ 1.6 HD5223 $1.1$ $1.50 \times 10^{-3}$ $0.74$ $1.81 \times 10^{4}$ $0.12$ $1.29 \times 10^{4}$ $14.6$ $9$ 1.6 HE0024$-$2523 $0.9$ $4.00 \times 10^{-3}$ $0.76$ $6.01 \times 10^{2}$ $0.16$ $5.23 \times 10^{2}$ $ 3.48 $ $37.0$ $4$ 9.3 HE0212$-$0557 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $1.20 \times 10^{4}$ $0.26$ $4.34 \times 10^{3}$ $70.1$ $5$ 14.0 HE0231$-$4016 $0.9$ $1.50 \times 10^{-3}$ $0.76$ $8.49 \times 10^{2}$ $0.12$ $6.80 \times 10^{2}$ $1.1$ $4$ 0.3 HE0430$-$4404 $0.9$ $3.00 \times 10^{-3}$ $0.74$ $8.55 \times 10^{2}$ $0.11$ $6.73 \times 10^{2}$ $0.7$ $2$ 0.4 HE1430$-$1123 $1.0$ $4.00 \times 10^{-3}$ $0.81$ $2.57 \times 10^{4}$ $0.09$ $2.31 \times 10^{4}$ $4.5$ $2$ 2.2 HE1434$-$1442 $1.0$ $3.00 \times 10^{-3}$ $0.84$ $7.19 \times 10^{4}$ $0.05$ $6.34 \times 10^{4}$ $18.3$ $3$ 6.1 HE2150$-$0825 $0.9$ $4.00 \times 10^{-3}$ $0.79$ $4.22 \times 10^{2}$ $0.20$ $4.08 \times 10^{2}$ $0.8$ $5$ 0.2 HE2232$-$0603 $0.9$ $6.66 \times 10^{-4}$ $0.76$ $6.01 \times 10^{2}$ $0.16$ $5.23 \times 10^{2}$ $8.8$ $4$ 2.2 SDSS0924$+$40 $1.6$ $2.00 \times 10^{-3}$ $0.51$ $1.19 \times 10^{4}$ $0.30$ $2.31 \times 10^{3}$ $16.0$ $2$ 8.0 SDSS2047$+$00 $1.1$ $1.50 \times 10^{-3}$ $0.51$ $3.85 \times 10^{4}$ $0.05$ $2.04 \times 10^{4}$ $0.8$ $3$ 0.3 ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ----------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- --------------------------- ------- ------------------------------- CEMP-$s/nr$ stars CS22880$-$074 $1.1$ $1.00 \times 10^{-3}$ $0.79$ $6.33 \times 10^{3}$ $0.09$ $4.94 \times 10^{3}$ $16.8$ $10$ 1.7 CS22942$-$019 $1.8$ $2.00 \times 10^{-4}$ $0.54$ $6.38 \times 10^{4}$ $0.29$ $1.47 \times 10^{4}$ $20.9$ $9$ 2.3 CS22964$-$161A $1.6$ $2.00 \times 10^{-3}$ $0.79$ $2.60 \times 10^{6}$ $0.04$ $3.05 \times 10^{6}$ $19.7$ $8$ 2.5 CS22964$-$161B $1.6$ $2.00 \times 10^{-3}$ $0.71$ $2.60 \times 10^{6}$ $0.03$ $3.05 \times 10^{6}$ $11.5$ $9$ 1.3 CS30301$-$015 $2.4$ $2.00 \times 10^{-3}$ $0.81$ $1.37 \times 10^{4}$ $0.05$ $2.28 \times 10^{2}$ $11.0$ $8$ 1.4 HD196944 $1.2$ $1.00 \times 10^{-3}$ $0.79$ $8.71 \times 10^{3}$ $0.07$ $5.89 \times 10^{3}$ $13.9$ $12$ 1.2 HD198269 $1.0$ $2.00 \times 10^{-3}$ $0.84$ $2.03 \times 10^{5}$ $0.07$ $1.77 \times 10^{5}$ $9.2$ $5$ 1.8 HE0202$-$2204 $0.9$ $2.00 \times 10^{-3}$ $0.81$ $1.67 \times 10^{3}$ $0.07$ $1.51 \times 10^{3}$ $5.1$ $7$ 0.7 HE1135$+$0139 $0.9$ $1.00 \times 10^{-3}$ $0.79$ $4.74 \times 10^{3}$ $0.09$ $4.43 \times 10^{3}$ $2.9$ $6$ 0.5 HE2158$-$0348 $1.5$ $3.00 \times 10^{-3}$ $0.71$ $3.69 \times 10^{5}$ $0.15$ $1.76 \times 10^{5}$ $15.4$ $7$ 2.2 CEMP-$s/r$ stars CS22881$-$036 $1.5$ $6.66 \times 10^{-4}$ $0.71$ $3.69 \times 10^{5}$ $0.15$ $1.77 \times 10^{5}$ $10.1$ $6$ 1.7 CS22898$-$027 $1.5$ $6.66 \times 10^{-4}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $101.6$ $11$ 9.2 CS22948$-$027 $1.5$ $2.00 \times 10^{-3}$ $0.64$ $1.88 \times 10^{5}$ $0.24$ $8.29 \times 10^{4}$ $29.6$ $9$ 3.3 CS29497$-$030 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $129.6$ $16$ 8.1 CS29526$-$110 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $22.5$ $6$ 3.8 CS31062$-$012 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $1.35 \times 10^{5}$ $0.26$ $5.39 \times 10^{4}$ $17.5$ $8$ 2.2 CS31062$-$050 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $98.3$ $20$ 4.9 HD187861 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $9.59 \times 10^{4}$ $0.28$ $3.63 \times 10^{4}$ $16.6$ $8$ 2.1 HD224959 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.74 \times 10^{4}$ $65.6$ $8$ 8.2 HE0131$-$3953 $1.5$ $2.00 \times 10^{-3}$ $0.61$ $1.34 \times 10^{5}$ $0.27$ $5.79 \times 10^{4}$ $13.3$ $4$ 3.3 HE0143$-$0441 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $9.59 \times 10^{4}$ $0.28$ $3.65 \times 10^{4}$ $37.2$ $7$ 5.3 HE0338$-$3945 $1.5$ $6.66 \times 10^{-4}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $188.6$ $17$ 11.1 HE1105$+$0027 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $6.79 \times 10^{4}$ $0.31$ $2.66 \times 10^{4}$ $23.0$ $4$ 5.8 HE1305$+$0007 $1.5$ $6.66 \times 10^{-4}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.73 \times 10^{4}$ $56.7$ $10$ 5.7 HE2148$-$1247 $1.5$ $2.00 \times 10^{-3}$ $0.56$ $4.81 \times 10^{4}$ $0.32$ $1.92 \times 10^{4}$ $208.9$ $11$ 19.0 HE2258$-$6358 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $6.75 \times 10^{4}$ $0.32$ $2.81 \times 10^{4}$ $293.4$ $18$ 16.3 LP625$-$44 $1.7$ $4.00 \times 10^{-3}$ $0.54$ $6.52 \times 10^{4}$ $0.35$ $1.86 \times 10^{4}$ $177.6$ $17$ 10.4 SDSSJ0912$+$0216 $1.5$ $2.00 \times 10^{-3}$ $0.49$ $6.92 \times 10^{4}$ $0.27$ $2.03 \times 10^{4}$ $429.1$ $14$ 30.7 CEMP-$s/ur$ stars BD$+$04$^{\circ}$2466 $1.1$ $1.50 \times 10^{-3}$ $0.74$ $4.54 \times 10^{3}$ $0.10$ $3.23 \times 10^{3}$ $13.2$ $7$ 1.9 HD13826 $1.7$ $0$ $0.76$ $1.11 \times 10^{4}$ $0.06$ $3.38 \times 10^{3}$ $3.8$ $3$ 1.3 HD187216 $1.5$ $5.00 \times 10^{-4}$ $0.56$ $4.81 \times 10^{4}$ $0.32$ $1.70 \times 10^{4}$ $28.7$ $4$ 7.2 HD201626 $1.4$ $4.00 \times 10^{-3}$ $0.71$ $2.67 \times 10^{5}$ $0.18$ $1.49 \times 10^{5}$ $7.5$ $5$ 1.5 HD5223 $1.1$ $1.50 \times 10^{-3}$ $0.76$ $2.54 \times 10^{4}$ $0.16$ $1.99 \times 10^{4}$ $15.1$ $9$ 1.7 HE0024$-$2523 $1.1$ $0$ $0.69$ $7.29 \times 10^{4}$ $0.14$ $5.39 \times 10^{4}$ $47.7$ $4$ 11.9 HE0212$-$0557 $1.5$ $2.00 \times 10^{-3}$ $0.59$ $9.53 \times 10^{4}$ $0.29$ $3.91 \times 10^{4}$ $67.9$ $5$ 13.6 HE0231$-$4016 $0.9$ $1.50 \times 10^{-3}$ $0.79$ $8.42 \times 10^{2}$ $0.08$ $6.62 \times 10^{2}$ $1.1$ $4$ 0.3 HE0430$-$4404 $0.9$ $4.00 \times 10^{-3}$ $0.79$ $8.42 \times 10^{2}$ $0.08$ $6.62 \times 10^{2}$ $0.7$ $2$ 0.3 HE1430$-$1123 $1.5$ $2.00 \times 10^{-3}$ $0.74$ $5.18 \times 10^{5}$ $0.09$ $2.45 \times 10^{5}$ $4.5$ $2$ 2.2 HE1434$-$1442 $1.0$ $4.00 \times 10^{-3}$ $0.79$ $2.90 \times 10^{5}$ $0.05$ $2.46 \times 10^{5}$ $17.8$ $3$ 5.9 HE2150$-$0825 $0.9$ $3.00 \times 10^{-3}$ $0.84$ $4.16 \times 10^{2}$ $0.14$ $3.95 \times 10^{2}$ $2.2$ $5$ 0.4 HE2232$-$0603 $0.9$ $6.66 \times 10^{-4}$ $0.84$ $5.88 \times 10^{2}$ $0.11$ $5.20 \times 10^{2}$ $10.1$ $4$ 2.5 SDSS0924$+$40 $1.5$ $2.00 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.75 \times 10^{4}$ $14.4$ $2$ 7.2 SDSS2047$+$00 $1.5$ $6.66 \times 10^{-4}$ $0.61$ $5.33 \times 10^{5}$ $0.08$ $1.90 \times 10^{5}$ $0.7$ $3$ 0.2 ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${P_{\mathrm{f}}}-$Paper I ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- ---------------------------- --------------------------- ------- ------------------------------- CEMP-$s/r$ stars BS16080$-$175 $2.9$ $4.00 \times 10^{-3}$ $0.59$ $9.29 \times 10^{3}$ $0.26$ $2.55 \times 10^{4}$ $20.7$ $1$ 20.7 BS17436$-$058 $1.1$ $4.00 \times 10^{-3}$ $0.74$ $3.61 \times 10^{4}$ $0.15$ $4.73 \times 10^{4}$ $18.4$ $1$ 18.4 CEMP-$s/ur$ stars CS22945$-$017 $2.0$ $2.00 \times 10^{-4}$ $0.74$ $1.66 \times 10^{5}$ $0.13$ $4.50 \times 10^{5}$ $3.7$ $-1$ CS22956$-$028 $1.3$ $0$ $0.64$ $1.25 \times 10^{4}$ $0.20$ $1.77 \times 10^{4}$ $ 1.37\times 10^3 $ $14.8$ $1$ 14.8 CS29509$-$027 $2.9$ $1.50 \times 10^{-3}$ $0.69$ $3.25 \times 10^{3}$ $0.11$ $5.28 \times 10^{3}$ $ 1.65\times 10^2 $ $0.8$ $-1$ CS30338$-$089 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.53 \times 10^{4}$ $6.1$ $-1$ HE0012$-$1441 $2.6$ $4.00 \times 10^{-3}$ $0.76$ $4.74 \times 10^{3}$ $0.06$ $1.58 \times 10^{4}$ $0.2$ $-1$ HE0206$-$1916 $1.5$ $6.66 \times 10^{-4}$ $0.64$ $9.42 \times 10^{4}$ $0.24$ $1.57 \times 10^{5}$ $2.5$ $-1$ HE0322$-$1504 $1.8$ $4.00 \times 10^{-3}$ $0.49$ $2.29 \times 10^{4}$ $0.36$ $3.13 \times 10^{4}$ $1.7$ $-1$ HE0441$-$0652 $1.2$ $3.00 \times 10^{-3}$ $0.79$ $1.55 \times 10^{3}$ $0.05$ $1.90 \times 10^{2}$ $0.2$ $-1$ HE0507$-$1430 $6.0$ $0$ $0.56$ $6.77 \times 10^{3}$ $0.26$ $3.83 \times 10^{4}$ $ 5.16\times 10^2 $ $0.9$ $-1$ HE0534$-$4548 $2.3$ $1.00 \times 10^{-3}$ $0.81$ $2.20 \times 10^{5}$ $0.02$ $9.10 \times 10^{5}$ $0.0$ $-2$ HE1045$-$1434 $6.0$ $0$ $0.49$ $1.36 \times 10^{4}$ $0.37$ $5.56 \times 10^{4}$ $17.1$ $-1$ HE1157$-$0518 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $4.78 \times 10^{4}$ $0.31$ $6.55 \times 10^{4}$ $1.7$ $-1$ HE1429$-$0551 $1.3$ $1.00 \times 10^{-3}$ $0.71$ $1.37 \times 10^{5}$ $0.17$ $2.29 \times 10^{5}$ $0.5$ $-1$ HE1443$+$0113 $1.7$ $0$ $0.74$ $2.79 \times 10^{3}$ $0.09$ $4.67 \times 10^{3}$ $0.1$ $-1$ HE1447$+$0102 $1.7$ $6.66 \times 10^{-4}$ $0.49$ $2.34 \times 10^{4}$ $0.34$ $2.94 \times 10^{4}$ $73.7$ $-1$ HE1523$-$1155 $1.7$ $4.00 \times 10^{-3}$ $0.76$ $1.75 \times 10^{5}$ $0.11$ $4.35 \times 10^{5}$ $1.2$ $-1$ HE1528$-$0409 $1.5$ $1.50 \times 10^{-3}$ $0.54$ $2.42 \times 10^{4}$ $0.32$ $2.99 \times 10^{4}$ $12.2$ $-1$ HE2221$-$0453 $1.8$ $4.00 \times 10^{-3}$ $0.74$ $1.73 \times 10^{5}$ $0.11$ $5.33 \times 10^{5}$ $1.3$ $-1$ HE2227$-$4044 $1.0$ $3.00 \times 10^{-3}$ $0.81$ $4.57 \times 10^{3}$ $0.10$ $5.55 \times 10^{3}$ $0.3$ $1$ 0.3 HE2228$-$0706 $1.7$ $2.00 \times 10^{-3}$ $0.51$ $2.33 \times 10^{4}$ $0.36$ $2.87 \times 10^{4}$ $7.6$ $-1$ HE2240$-$0412 $1.0$ $3.00 \times 10^{-3}$ $0.79$ $6.50 \times 10^{3}$ $0.09$ $7.96 \times 10^{3}$ $1.2$ $-1$ SDSS1707$+$58 $1.8$ $4.00 \times 10^{-3}$ $0.49$ $2.29 \times 10^{4}$ $0.36$ $2.87 \times 10^{4}$ $22.3$ $1$ 22.3 ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${P_{\mathrm{f}}}-$Paper I ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- ---------------------------- --------------------------- ------- ------------------------------- CEMP-$s/r$ stars BS16080$-$175 $3.0$ $0$ $0.59$ $3.64 \times 10^{4}$ $0.23$ $2.86 \times 10^{2}$ $20.8$ $1$ 20.8 BS17436$-$058 $1.5$ $1.50 \times 10^{-3}$ $0.74$ $2.91 \times 10^{3}$ $0.15$ $7.30 \times 10^{1}$ $18.4$ $1$ 18.4 CEMP-$s/ur$ stars CS22945$-$017 $2.0$ $2.00 \times 10^{-4}$ $0.71$ $3.74 \times 10^{3}$ $0.16$ $3.65 \times 10^{1}$ $3.4$ $-1$ CS22956$-$028 $1.3$ $0$ $0.69$ $1.23 \times 10^{4}$ $0.19$ $7.36 \times 10^{3}$ $ 1.05\times 10^3 $ $14.3$ $1$ 14.3 CS29509$-$027 $3.0$ $0$ $0.66$ $6.41 \times 10^{3}$ $0.15$ $2.08 \times 10^{1}$ $ 1.67\times 10^2 $ $0.4$ $-1$ CS30338$-$089 $1.1$ $1.50 \times 10^{-3}$ $0.66$ $4.63 \times 10^{3}$ $0.22$ $3.35 \times 10^{3}$ $5.8$ $-1$ HE0012$-$1441 $2.6$ $4.00 \times 10^{-3}$ $0.76$ $2.12 \times 10^{5}$ $0.05$ $3.38 \times 10^{4}$ $0.3$ $-1$ HE0206$-$1916 $1.1$ $1.50 \times 10^{-3}$ $0.74$ $6.41 \times 10^{3}$ $0.21$ $5.29 \times 10^{3}$ $1.9$ $-1$ HE0322$-$1504 $1.8$ $4.00 \times 10^{-3}$ $0.54$ $1.60 \times 10^{4}$ $0.31$ $2.40 \times 10^{3}$ $2.3$ $-1$ HE0441$-$0652 $1.5$ $3.00 \times 10^{-3}$ $0.79$ $1.29 \times 10^{5}$ $0.03$ $6.07 \times 10^{4}$ $0.1$ $-1$ HE0507$-$1430 $6.0$ $0$ $0.59$ $2.14 \times 10^{5}$ $0.27$ $2.51 \times 10^{2}$ $ 3.58\times 10^2 $ $1.0$ $-1$ HE0534$-$4548 $2.8$ $6.66 \times 10^{-4}$ $0.81$ $5.76 \times 10^{5}$ $0.02$ $8.47 \times 10^{4}$ $0.0$ $-2$ HE1045$-$1434 $6.0$ $0$ $0.56$ $3.02 \times 10^{5}$ $0.29$ $2.61 \times 10^{3}$ $20.9$ $-1$ HE1157$-$0518 $1.1$ $1.50 \times 10^{-3}$ $0.66$ $3.28 \times 10^{3}$ $0.22$ $2.30 \times 10^{3}$ $1.7$ $-1$ HE1429$-$0551 $1.3$ $1.00 \times 10^{-3}$ $0.66$ $1.24 \times 10^{4}$ $0.18$ $6.78 \times 10^{3}$ $0.5$ $-1$ HE1443$+$0113 $1.3$ $6.66 \times 10^{-4}$ $0.74$ $2.42 \times 10^{4}$ $0.11$ $1.45 \times 10^{4}$ $0.3$ $-1$ HE1447$+$0102 $1.2$ $2.00 \times 10^{-3}$ $0.61$ $3.24 \times 10^{3}$ $0.35$ $1.39 \times 10^{3}$ $79.1$ $-1$ HE1523$-$1155 $1.7$ $4.00 \times 10^{-3}$ $0.76$ $2.78 \times 10^{3}$ $0.12$ $3.76 \times 10^{1}$ $1.2$ $-1$ HE1528$-$0409 $1.6$ $1.00 \times 10^{-3}$ $0.54$ $1.19 \times 10^{4}$ $0.30$ $3.01 \times 10^{3}$ $13.1$ $-1$ HE2221$-$0453 $1.8$ $4.00 \times 10^{-3}$ $0.74$ $4.33 \times 10^{4}$ $0.12$ $1.34 \times 10^{4}$ $1.3$ $-1$ HE2227$-$4044 $1.0$ $1.50 \times 10^{-3}$ $0.81$ $7.24 \times 10^{4}$ $0.05$ $6.26 \times 10^{4}$ $0.2$ $1$ 0.2 HE2228$-$0706 $1.7$ $3.00 \times 10^{-3}$ $0.59$ $1.62 \times 10^{4}$ $0.26$ $4.49 \times 10^{3}$ $9.7$ $-1$ HE2240$-$0412 $1.5$ $2.00 \times 10^{-3}$ $0.74$ $2.06 \times 10^{3}$ $0.04$ $3.67 \times 10^{1}$ $0.8$ $-1$ SDSS1707$+$58 $1.8$ $4.00 \times 10^{-3}$ $0.54$ $1.60 \times 10^{4}$ $0.31$ $2.61 \times 10^{3}$ $24.9$ $1$ 24.9 ID ${M_{1,\mathrm{i}}}$ ${M_{\mathrm{PMZ}}}$ ${M_{2,\mathrm{i}}}$ ${P_{\mathrm{i}}}$ $\Delta{M_{\mathrm{acc}}}$ ${P_{\mathrm{f}}}$ ${\chi^2_{\mathrm{min}}}$ $\nu$ ${\chi^2_{\mathrm{min}}}/\nu$ ------------------- ---------------------- ----------------------- ---------------------- ---------------------- ---------------------------- ---------------------- --------------------------- ------- ------------------------------- CEMP-$s/r$ stars BS16080$-$175 $2.9$ $1.00 \times 10^{-3}$ $0.56$ $2.09 \times 10^{5}$ $0.26$ $1.57 \times 10^{4}$ $20.7$ $1$ 20.7 BS17436$-$058 $1.1$ $4.00 \times 10^{-3}$ $0.71$ $2.57 \times 10^{4}$ $0.16$ $2.00 \times 10^{4}$ $18.4$ $1$ 18.4 CEMP-$s/ur$ stars CS22945$-$017 $2.0$ $2.00 \times 10^{-4}$ $0.69$ $2.98 \times 10^{4}$ $0.15$ $7.49 \times 10^{3}$ $4.0$ $-1$ CS22956$-$028 $1.3$ $0$ $0.64$ $2.49 \times 10^{4}$ $0.19$ $1.37 \times 10^{4}$ $14.8$ $1$ 14.8 CS29509$-$027 $3.0$ $0$ $0.69$ $3.60 \times 10^{4}$ $0.11$ $5.09 \times 10^{2}$ $0.5$ $-1$ CS30338$-$089 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $9.53 \times 10^{4}$ $0.29$ $3.90 \times 10^{4}$ $6.1$ $-1$ HE0012$-$1441 $2.6$ $4.00 \times 10^{-3}$ $0.74$ $1.89 \times 10^{4}$ $0.07$ $2.55 \times 10^{2}$ $0.3$ $-1$ HE0206$-$1916 $1.5$ $1.00 \times 10^{-3}$ $0.66$ $2.35 \times 10^{4}$ $0.18$ $1.04 \times 10^{4}$ $2.4$ $-1$ HE0322$-$1504 $1.8$ $4.00 \times 10^{-3}$ $0.54$ $9.01 \times 10^{4}$ $0.36$ $2.09 \times 10^{4}$ $1.9$ $-1$ HE0441$-$0652 $1.5$ $3.00 \times 10^{-3}$ $0.81$ $2.86 \times 10^{3}$ $0.04$ $7.12 \times 10^{1}$ $0.1$ $-1$ HE0507$-$1430 $4.0$ $0$ $0.54$ $1.02 \times 10^{6}$ $0.32$ $2.76 \times 10^{4}$ $1.1$ $-1$ HE0534$-$4548 $1.9$ $6.66 \times 10^{-4}$ $0.84$ $1.32 \times 10^{6}$ $0.02$ $4.51 \times 10^{5}$ $0.0$ $-2$ HE1045$-$1434 $2.0$ $4.00 \times 10^{-3}$ $0.51$ $1.23 \times 10^{5}$ $0.35$ $2.10 \times 10^{4}$ $24.6$ $-1$ HE1157$-$0518 $1.5$ $6.66 \times 10^{-4}$ $0.59$ $6.75 \times 10^{4}$ $0.32$ $2.83 \times 10^{4}$ $1.7$ $-1$ HE1429$-$0551 $1.3$ $1.00 \times 10^{-3}$ $0.69$ $2.75 \times 10^{5}$ $0.15$ $1.55 \times 10^{5}$ $0.7$ $-1$ HE1443$+$0113 $1.7$ $0$ $0.74$ $7.01 \times 10^{5}$ $0.09$ $2.60 \times 10^{5}$ $0.1$ $-1$ HE1447$+$0102 $1.7$ $6.66 \times 10^{-4}$ $0.54$ $6.52 \times 10^{4}$ $0.35$ $1.82 \times 10^{4}$ $74.5$ $-1$ HE1523$-$1155 $1.5$ $4.00 \times 10^{-3}$ $0.74$ $3.67 \times 10^{5}$ $0.15$ $1.81 \times 10^{5}$ $1.3$ $-1$ HE1528$-$0409 $1.5$ $1.50 \times 10^{-3}$ $0.54$ $4.84 \times 10^{4}$ $0.31$ $1.72 \times 10^{4}$ $12.3$ $-1$ HE2221$-$0453 $1.8$ $4.00 \times 10^{-3}$ $0.76$ $6.84 \times 10^{5}$ $0.12$ $2.22 \times 10^{5}$ $1.3$ $-1$ HE2227$-$4044 $1.5$ $1.00 \times 10^{-3}$ $0.76$ $5.78 \times 10^{3}$ $0.06$ $3.89 \times 10^{2}$ $0.2$ $1$ 0.2 HE2228$-$0706 $1.7$ $2.00 \times 10^{-3}$ $0.54$ $6.52 \times 10^{4}$ $0.35$ $1.85 \times 10^{4}$ $7.8$ $-1$ HE2240$-$0412 $1.5$ $2.00 \times 10^{-3}$ $0.76$ $4.09 \times 10^{3}$ $0.05$ $1.51 \times 10^{2}$ $1.0$ $-1$ SDSS1707$+$58 $1.8$ $4.00 \times 10^{-3}$ $0.51$ $9.06 \times 10^{4}$ $0.34$ $2.04 \times 10^{4}$ $23.2$ $1$ 23.2 ----------------------- -------- -------- -------- -------- -------- -------- -------- -------- -------- ---------------------- ---------------------- ---------------------- ID best min max best min max best min max best min max CEMP-$s/nr$ CS22880$-$074 $1.10$ $0.85$ $1.15$ $1.00$ $0.00$ $4.00$ $0.74$ $0.73$ $0.85$ $4.54 \times 10^{3}$ $6.64 \times 10^{2}$ $1.64 \times 10^{4}$ $1.45$ $1.55$ $1.56 \times 10^{5}$ $3.47 \times 10^{5}$ CS22942$-$019 $1.80$ $1.75$ $1.85$ $0.20$ $0.00$ $0.58$ $0.54$ $0.48$ $0.68$ $6.48 \times 10^{4}$ $8.73 \times 10^{3}$ $1.70 \times 10^{5}$ CS22964$-$161A $1.60$ $0.95$ $1.65$ $2.00$ $0.58$ $4.00$ $0.79$ $0.75$ $0.85$ $2.74 \times 10^{5}$ $2.18 \times 10^{5}$ $4.87 \times 10^{5}$ CS22964$-$161B $1.60$ $0.95$ $1.65$ $2.00$ $0.58$ $4.00$ $0.71$ $0.68$ $0.85$ $2.74 \times 10^{5}$ $2.18 \times 10^{5}$ $4.87 \times 10^{5}$ CS30301$-$015 $2.50$ $1.95$ $2.75$ $4.00$ $0.58$ $4.00$ $0.76$ $0.70$ $0.85$ $5.14 \times 10^{3}$ $2.90 \times 10^{3}$ $6.47 \times 10^{3}$ $1.15$ $1.25$ $1.30 \times 10^{5}$ $2.27 \times 10^{5}$ HD196944 $1.40$ $0.95$ $1.65$ $1.00$ $0.30$ $4.00$ $0.81$ $0.75$ $0.88$ $5.84 \times 10^{3}$ $1.38 \times 10^{3}$ $9.92 \times 10^{3}$ HD198269 $1.00$ $0.95$ $1.65$ $4.00$ $0.80$ $4.00$ $0.79$ $0.73$ $0.85$ $1.64 \times 10^{3}$ $1.31 \times 10^{3}$ $2.05 \times 10^{5}$ $1.95$ $2.15$ HE0202$-$2204 $0.90$ $0.85$ $1.15$ $3.00$ $0.00$ $4.00$ $0.79$ $0.75$ $0.88$ $1.15 \times 10^{3}$ $4.67 \times 10^{2}$ $1.05 \times 10^{4}$ $1.45$ $1.55$ $2.06 \times 10^{5}$ $2.60 \times 10^{5}$ HE1135$+$0139 $0.90$ $0.85$ $1.95$ $1.00$ $0.00$ $4.00$ $0.79$ $0.75$ $0.88$ $3.35 \times 10^{3}$ $4.67 \times 10^{2}$ $7.29 \times 10^{3}$ HE2158$-$0348 $1.50$ $1.15$ $1.65$ $3.00$ $1.75$ $4.00$ $0.69$ $0.65$ $0.75$ $1.18 \times 10^{5}$ $5.79 \times 10^{3}$ $9.17 \times 10^{3}$ $9.34 \times 10^{4}$ $1.48 \times 10^{5}$ CEMP-$s/r$ CS22881$-$036 $1.50$ $1.45$ $1.55$ $0.67$ $0.05$ $2.50$ $0.66$ $0.65$ $0.80$ $1.35 \times 10^{5}$ $6.25 \times 10^{3}$ $9.90 \times 10^{3}$ $1.05$ $1.15$ $1.38 \times 10^{4}$ $1.71 \times 10^{5}$ CS31062$-$012 $1.50$ $1.35$ $1.65$ $2.00$ $0.58$ $3.50$ $0.56$ $0.45$ $0.65$ $6.92 \times 10^{4}$ $9.65 \times 10^{3}$ $1.21 \times 10^{5}$ HD187861 $1.50$ $1.35$ $1.65$ $2.00$ $0.58$ $3.50$ $0.59$ $0.50$ $0.65$ $4.90 \times 10^{4}$ $9.59 \times 10^{3}$ $1.21 \times 10^{5}$ CEMP-$s/ur$ BD$+$04$^{\circ}$2466 $1.10$ $0.85$ $1.15$ $2.00$ $0.00$ $4.00$ $0.76$ $0.73$ $0.85$ $3.19 \times 10^{3}$ $5.70 \times 10^{2}$ $5.27 \times 10^{3}$ $1.45$ $1.65$ $9.90 \times 10^{3}$ $2.37 \times 10^{5}$ HD13826 $1.70$ $1.45$ $1.85$ $0.00$ $0.00$ $0.30$ $0.79$ $0.68$ $0.88$ $3.90 \times 10^{3}$ $4.70 \times 10^{2}$ $1.16 \times 10^{4}$ $0.85$ $0.95$ $1.62 \times 10^{5}$ $2.57 \times 10^{5}$ HD201626 $1.40$ $1.05$ $2.15$ $4.00$ $0.80$ $4.00$ $0.71$ $0.60$ $0.78$ $1.18 \times 10^{5}$ $2.31 \times 10^{3}$ $1.63 \times 10^{5}$ HD5223 $1.10$ $1.05$ $1.15$ $1.50$ $0.30$ $3.50$ $0.76$ $0.68$ $0.83$ $1.79 \times 10^{4}$ $1.31 \times 10^{3}$ $2.07 \times 10^{3}$ $1.45$ $1.55$ $6.63 \times 10^{3}$ $1.05 \times 10^{4}$ $1.38 \times 10^{4}$ $1.88 \times 10^{5}$ HE0231$-$4016 $0.90$ $0.85$ $1.25$ $4.00$ $0.05$ $4.00$ $0.79$ $0.73$ $0.85$ $8.42 \times 10^{2}$ $4.70 \times 10^{2}$ $7.44 \times 10^{3}$ $1.45$ $1.65$ $9.65 \times 10^{3}$ $1.53 \times 10^{4}$ $1.49 \times 10^{5}$ $2.27 \times 10^{5}$ HE0430$-$4404 $0.90$ $0.85$ $1.15$ $4.00$ $0.00$ $4.00$ $0.84$ $0.70$ $0.88$ $5.88 \times 10^{2}$ $4.70 \times 10^{2}$ $5.30 \times 10^{3}$ $1.45$ $1.55$ $1.04 \times 10^{4}$ $1.66 \times 10^{4}$ $1.49 \times 10^{5}$ $2.37 \times 10^{5}$ HE1430$-$1123 $1.50$ $1.35$ $1.65$ $2.00$ $0.05$ $4.00$ $0.76$ $0.68$ $0.85$ $1.83 \times 10^{5}$ $7.26 \times 10^{3}$ $2.05 \times 10^{5}$ $0.85$ $1.25$ $1.24 \times 10^{3}$ $3.17 \times 10^{3}$ HE2150$-$0825 $0.90$ $0.85$ $1.15$ $1.50$ $0.05$ $4.00$ $0.84$ $0.73$ $0.85$ $5.88 \times 10^{2}$ $3.33 \times 10^{2}$ $3.75 \times 10^{3}$ $1.45$ $1.55$ $9.65 \times 10^{3}$ $2.37 \times 10^{5}$ HE2232$-$0603 $0.90$ $0.85$ $1.15$ $1.00$ $0.00$ $1.75$ $0.84$ $0.73$ $0.85$ $5.88 \times 10^{2}$ $3.33 \times 10^{2}$ $3.24 \times 10^{3}$ SDSS2047$+$00 $1.50$ $0.85$ $1.75$ $0.67$ $0.05$ $4.00$ $0.54$ $0.43$ $0.83$ $1.36 \times 10^{5}$ $3.33 \times 10^{2}$ $5.28 \times 10^{2}$ $9.82 \times 10^{2}$ $2.41 \times 10^{5}$ ----------------------- -------- -------- -------- -------- -------- -------- -------- -------- -------- ---------------------- ---------------------- ----------------------

--- abstract: 'Every year millions of men, women and children are forced to leave their homes and seek refuge from wars, human rights violations, persecution, and natural disasters. The number of forcibly displaced people came at a record rate of 44,400 every day throughout 2017, raising the cumulative total to 68.5 million at the year’s end, overtaken the total population of the United Kingdom. Up to 85% of the forcibly displaced find refuge in low- and middle-income countries, calling for increased humanitarian assistance worldwide. To reduce the amount of manual labour required for human-rights-related image analysis, we introduce DisplaceNet, a novel model which infers potential displaced people from images by integrating the control level of the situation and conventional convolutional neural network (CNN) classifier into one framework for image classification. Experimental results show that DisplaceNet achieves up to 4% coveragethe proportion of a data set for which a classifier is able to produce a predictiongain over the sole use of a CNN classifier. Our dataset, codes and trained models will be available online at <https://github.com/GKalliatakis/DisplaceNet>' author: - | Grigorios Kalliatakis, Shoaib Ehsan, Maria Fasli, Klaus McDonald-Maier\ School of Computer Science and Electronic Engineering, University of Essex\ [{gkallia, sehsan, mfasli, kdm}@essex.ac.uk]{} bibliography: - 'egbib.bib' title: 'DisplaceNet: Recognising Displaced People from Images by Exploiting Dominance Level' --- =1 Introduction ============ The displacement of people refers to the forced movement of people from their locality or environment and occupational activities [^1]. It is a form of social change caused by a number of factors such as armed conflict, violence, persecution and human rights violations. Globally, there are now almost 68.5 million forcibly displaced peopleand most are hosted in developing regions, while today 1 out of every 110 people in the world is displaced [@global_trends_report]. In the era of social media and big data, the use of visual evidence to document conflict and human rights abuse has become an important element for human rights organizations and advocates. However, the omnipresence of visual evidence may deluge those accountable for analysing it. Currently, information extraction from human-rights-related imagery requires manual labour by human rights analysts and advocates. Such analysis is time consuming, expensive, and remains emotionally traumatic for analysts to focus on images of horrific events. In this work, we strive to reconcile this gap by automating parts of this process; given a single image we label the image as either *displaced people* or *non displaced people*. Figure \[Fig. 1\] illustrates that naive schemes based solely on object detection or scene recognition are doomed to fail in this binary classification problem. If we can exploit existing smartphone cameras, which are ubiquitous, it may be possible to turn recognition of displaced populations into a powerful and cost-effective computer vision application that could improve humanitarian responses. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![Displaced people recognition poses a challenge at a higher level for the well-studied, deep image representation learning methods. Regularly, emotional states can be the only notifying difference between the encoded visual content of an image that depicts a non-violent situation and the encoded visual content of an image displaying displaced people.[]{data-label="Fig. 1"}](figures/Fig1_a.jpg "fig:"){width="20.00000%" height="0.38\textheight"} ![Displaced people recognition poses a challenge at a higher level for the well-studied, deep image representation learning methods. Regularly, emotional states can be the only notifying difference between the encoded visual content of an image that depicts a non-violent situation and the encoded visual content of an image displaying displaced people.[]{data-label="Fig. 1"}](figures/Fig1_b.jpg "fig:"){width="22.00000%" height="0.43\textheight"} \(a) Child playing \(b) Displaced people \[7pt\] ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ Recently, Kalliatakis *et al.* [@kalliatakis2019exploring] shown that a two-stage fine-tuning of deep Convolutional Neural Networks (CNNs) can address the multi-class classification problem of human rights violations to a certain extent. In this paper, we introduce *DisplaceNet*, a novel method designed with a human-centric approach for solving a sought-after, binary classification problem in the context of human rights image analysis; *displaced people recognition*. Our hypothesis is that the control level of the situation by the person, ranging from *submissive* / *non-control* to *dominant* / *in-control*, is a powerful cue that can help our network make a distinction between displaced people and non-violent instances. First, we develop an end-to-end model for recognising rich information about people’s emotional states by jointly analysing the person and the whole scene. We use the continuous dimensions of the *VAD Emotional State Model* [@mehrabian1995framework], which describes emotions using three numerical dimensions: Valence (V); Arousal (A); and Dominance (D). In the context of this work, we have focused only on dominancemeasures the control level of the situation by the personbecause it is considered as the most relevant for the task of recognising displaced people. Second, following the estimation of emotional states, we introduce a new method for interpreting the overall dominance level of an entire image sample based on the emotional states of all individuals on the scene. As a final step, we propose to assign weights to image samples according to the image-to-overall-dominance relevance to guide prediction of the image classifier. We carried out comprehensive experimentation to evaluate our method for displaced people recognition on a subset of HRA dataset [@kalliatakis2019exploring]. This subset contains all image samples from the *displaced people* category (positive samples) alongside the same number of images taken from the *no violation* category (negative samples). Experimental results show that DisplaceNet can improve the coveragethe proportion of a data set for which a classifier is able to produce a predictionby 4% compared to the sole use of a CNN classifier that is trained end-to-end using the same training data. Related Work ============ **Human rights image analysis.** Image analysis in the context of human rights plays a crucial role in human rights advocacy and accountability efforts. Automatic perception of potential human rights violations enables scientists and investigators to discover content, that may otherwise be concealed by sheer volume of visual data. The automated systems concerned with human rights abuses identification are not producing evidence, but are instead narrowing down the volume of material that must be examined by human analysts who are making legitimate claims, that they then present in justice, accountability, or advocacy settings [@aronson2018computer]. There is a considerable body on literature for video analysis with respect to human rights [@piracés_2018; @aronson2015video]. A different group of methods based on still images alongside the first ever publicly available image dataset for the purpose of human rights violations recognition was introduced in [@visapp17]. Recently, Kalliatakis *et al.* [@kalliatakis2019exploring] introduced a larger, verified-by-experts image dataset for fine-tuning object-centric and scene-centric CNNs. **Object detection.** On of the most improved areas of computer vision in the past few years is object detection, the process of determining the instance of the class to which an object belongs and estimating the location of the object. Object detectors can be split into two main categories: one-stage detectors and two-stage detectors. One of the first modern one-stage object detectors based on deep networks is OverFeat [@OverFeat], which applies a sliding window approach based on multi-scaling for jointly performing classification, detection and localization. More recent works such as YOLO [@redmon2016you; @redmon2016yolo9000] and SSD [@Fu2017DSSDD; @liu2016ssd], have revived interest in one-stage methods, mainly due to their real time efficiency. The leading model in modern object detectors is based on a two-stage approach which was established in [@uijlings2013selective]. R-CNN, a notably successful family of methods, [@girshick2015fast; @girshick2014rich] enhanced the second-stage classifier to a convolutional network, resulting in large accuracy improvements. After that, the speed of R-CNN has also improved over the years by integrating region proposal networks (RPN) with the second-stage classifier into a single convolution network, known as the Faster R-CNN framework [@ren2015faster]. Our method belongs to one-stage detectors. Specifically, we adopt the RetinaNet framework [@lin2018focal] that handles class imbalance by reshaping the standard cross entropy loss to focus training on a sparse set of hard examples and down-weights the loss assigned to well-classified examples. **Emotion recognition.** Most of the research in computer vision to recognise people’s emotional states is focused on facial expression analysis [@fabian2016emotionet; @eleftheriadis2016joint] where a large variety of methods have been developed to recognise the 6 basic emotions defined by Ekman and Friesen [@ekman1971constants]. Lately, CNNs have been used as backbone for the facial expression recognition of Action Units [@fabian2016emotionet]. The second family of methods for emotion recognition use the continuous dimensions of the *VAD Emotional State Model* [@mehrabian1995framework; @mehrabian1974approach] to represent emotions instead of discrete emotion categories. The VAD model uses a 3-dimensional approach to describe and measure the emotional experience of humans: Valence (V) describes affective states from highly negative (unpleasant) to highly positive (pleasant); Arousal (A) measures the intensity of affective states ranging from calm to excited or alert; and Dominance (D) represents the feeling of being controlled or influenced by external stimuli. Method ====== ![Model Architecture. Our model consists of (a) an *object detection* branch, (b) *human-centric* branch, and (c) a *displaced people* branch. The image features and their layers are shared between the human-centric and displaced people branches (blue boxes).[]{data-label="Fig. 2"}](figures/DisplaceNet.pdf){width="48.00000%" height="0.40\textheight"} We now describe our method for recognising displaced people by exploiting the dominance level of the entire image. Our goal is to label challenging everyday photos as either ‘displaced populations’ or ‘no displaced populations’. First, in order to detect the emotional traits of an image, we need to accurately localise the box containing a $human$ and the associated object of interaction (denoted by $b_h$ and $b_o$, respectively), as well as identify the emotional states $e$ of each human using the VAD model. Our proposed solution adopts the RetinaNet [@lin2018focal] object detection framework alongside an additional *human-centric* branch that estimates the continuous dimensions of each detected person and then determines the overall dominance level of the given image. Specifically, given a set of candidate boxes, RetinaNet outputs a set of object boxes and a class label for each box. While the object detector can predict multiple class labels, our model is concerned only with the ‘person’ class. The region of the image comprising the person whose feelings are to be estimated at $b_h$ is used alongside the entire image for simultaneously extracting their most relevant features. These features, are fused and used to perform continuous emotion recognition in VAD space. Our model extends typical image classification by assigning a triplet score $s_{img,d}^{DP}$ to pairs of candidate human boxes $b_h$ and the displaced people category *a*. To do so, we decompose the triplet score into three terms: $$\label{eq:1} s_{img,d}^{DP} = s_h \cdot s_{h,img}^{d} \cdot s_{img}^{DP}$$ We discuss each component next, followed by details for training and inference. The overall architecture of DisplaceNet is shown in Figure \[Fig. 2\]. Model components ---------------- **Object detection.** The object detection branch of DisplaceNet is identical to that of RetinaNet [@lin2018focal] single stage classifier. First, an image is forwarded through ResNet-50 [@he2016deep], then in the subsequent pyramid layers, the more semantically important features are extracted and concatenated with the original features for improved bounding box regression. **Human-centric branch.** The first role of the human-centric branch is to assign an emotion classification score to each human box . Similar to [@kosti2017emotion], we use an end-to-end model with three main modules: two feature extractors and a fusion module. The first module takes the region of the image comprising the person whose emotional traits are to be estimated, $b_h$, while the second module takes as input the entire image and extracts global features. This way the required contextual support is accommodated in our emotion recognition process. Finally, the third module takes as input the extracted image and body features and estimates the continuous dimensions in VAD space. The second role of the human-centric branch is to assign a dominance score $s_{img}^{d}$ which characterises the entire input image. $s_{img}^{d}$ is the encoding of the overall dominance score relative to human box $b_h$ and entire image $img$, that is: $$\label{eq:2} s_{img}^{d} = \frac{1}{n} \sum_{i=1}^{n} s_{h,img}^d$$ Figure \[Fig. 3\] (a),(c) illustrates the three different emotional states over the estimated target objects locations while Figure \[Fig. 3\] (b),(d) shows the overall dominance score proposed here. Note that although all three predicted numerical dimensions are depicted, only dominance is considered to be the most relevant to the task of recognising displaced people since the other two dimensions can be ambiguous for several situations. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Estimating continuous emotions in VAD space vs overall dominance from the combined body and image features. The left column shows the predicted emotional states and their scores from the person region of interest (RoI), while the right column show the same images analysed for overall dominance. The dominance score will be integrated with the standard image classification scores $s_{img}^{DP}$ to identify displaced people. []{data-label="Fig. 3"}](figures/VAD_neg.jpg "fig:"){width="23.00000%" height="0.4\textheight"} ![Estimating continuous emotions in VAD space vs overall dominance from the combined body and image features. The left column shows the predicted emotional states and their scores from the person region of interest (RoI), while the right column show the same images analysed for overall dominance. The dominance score will be integrated with the standard image classification scores $s_{img}^{DP}$ to identify displaced people. []{data-label="Fig. 3"}](figures/DOMINANCE_neg.jpg "fig:"){width="23.00000%" height="0.4\textheight"} (a) (b) \[7pt\] ![Estimating continuous emotions in VAD space vs overall dominance from the combined body and image features. The left column shows the predicted emotional states and their scores from the person region of interest (RoI), while the right column show the same images analysed for overall dominance. The dominance score will be integrated with the standard image classification scores $s_{img}^{DP}$ to identify displaced people. []{data-label="Fig. 3"}](figures/VAD_pos.jpg "fig:"){width="23.00000%" height="0.4\textheight"} ![Estimating continuous emotions in VAD space vs overall dominance from the combined body and image features. The left column shows the predicted emotional states and their scores from the person region of interest (RoI), while the right column show the same images analysed for overall dominance. The dominance score will be integrated with the standard image classification scores $s_{img}^{DP}$ to identify displaced people. []{data-label="Fig. 3"}](figures/DOMINANCE_pos.jpg "fig:"){width="23.00000%" height="0.4\textheight"} (c) (d) \[7pt\] ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- **Displaced people recognition.** The first role of the displaced people branch is to assign a classification score to the input image. Similar to two-phase transfer learning scheme introduced in [@kalliatakis2019exploring], we train an end-to-end model for binary classification of everyday photos as either ‘displaced populations’ or ‘no displaced populations’. In order to improve the discriminative power of our model, the second role of the displaced people branch is to integrate $s_{img}^{d}$ in the recognition pipeline. Specifically, the raw image classification score is readjusted based on the inferred dominance score. Each dominance unit, that is deltas from the neutral state, is expressed as a numeric weight varying between 1 and 10, while the neutral states of dominance are assigned between 4.5 and 5.5 based on the number of examples per each of the scores in the continuous dimensions reported in [@kosti2017emotion]. The adjustment that will be assigned to the raw probability, $s_{img}^{DP}$ is the weight of dominance multiplied by a factor of 0.11 which has been experimentally set. When the input image depicts positive dominance, the adjustment factor is subtracted from the positive human-rights-abuse probability and added to the negative human-rights-abuse probability. Similarly, when the input image depicts negative dominance the adjustment factor is added to the negative human-rights-abuse probability and subtracted from the positive human-rights-abuse probability. This is formally written in Algorithm \[algorithm\]. Finally, when no $b_h$ is detected from the object detection branch, (\[eq:1\]) is reduced into plain image classification as follows: $$\label{eq:6} s_{img,d}^{DP} = s_{img}^{DP}$$ $b_h > 0$ 0.05in $s_{pos}\gets s_{img}^{dp}$ $s_{neg}\gets s_{img}^{ndp}$ 0.05in Return $s_{pos}, s_{neg}$ $diff = weight-5.5$ $adj = diff*0.11$ $s_{pos} = s_{pos}-adj$ $s_{neg} = s_{neg}+adj$ $diff = 4.5-weight$ $adj = diff*0.11$ $s_{pos} = s_{pos}+adj$ $s_{neg} = s_{neg}-adj$ Return $s_{pos}, s_{neg}$ ------------------------------------------------------------------------------------- ---------------------------------------------------------------------------- ---------------------------------------------------------------------------- {width="28.00000%" height="0.3\textheight"} {width="28.00000%" height="0.3\textheight"} {width="31.00000%" height="0.4\textheight"} (a) (b) (c) \[-1pt\] {width="29.00000%" height="0.3\textheight"} {width="35.00000%" height="0.4\textheight"} {width="26.00000%" height="0.3\textheight"} (e) (f) (g) \[1pt\] ------------------------------------------------------------------------------------- ---------------------------------------------------------------------------- ---------------------------------------------------------------------------- Training -------- Due to different datasets, convergence times and loss imbalance, all three branches have been trained separately. For object detection we adopted an existing implementation of the RetinaNet object detector, pre-trained on the COCO dataset [@lin2014microsoft], with a ResNet-50 backbone. For emotion recognition in continuous dimensions, we formulate this task as a regression problem using the Euclidean loss. The two feature extraction modules are designed as truncated versions of various well-known CNNs and initialised using pretrained models on two large-scale image classification datasets, ImageNet [@krizhevsky2012imagenet] and Places [@zhou2018places]. The truncated version of those CNNs removes the fully connected layer and outputs features from the last convolutional layer in order to maintain the localisation of different parts of the images which is significant for the task at hand. Features extracted from these two modules (red and blue boxes in Fig. \[Fig. 2\]B) are then combined by a fusion module. This module first uses a global average pooling layer to reduce the number of features from each network and then a fully connected layer, with an output of a 256-D vector, functions as a dimensionality reduction layer for the concatenated pooled features. Finally, we include a second fully connected layer with 3 neurons representing valence, arousal and dominance. The parameters of the three modules are learned jointly using stochastic gradient descent with momentum of 0.9. The batch size is set to 54 and we use dropout with a ratio of 0.5. We formulate displaced people recognition as a binary classification problem. We train an end-to-end model for classifying everyday images as displaced people positive or displaced people negative, based on the context of the images. We fine-tune various CNN models for the two-class classification task. First, we conduct feature extraction utilising only the convolutional base of the original networks in order to end up with more generic representations as well as retaining spatial information similar to emotion recognition pipeline. The second phase consists of unfreezing some of the top layers of the convolutional base and jointly training a newly added fully connected layer and these top layers. All the CNNs presented here were trained using the Keras Python deep learning framework [@chollet2015keras] over TensorFlow [@abadi2016tensorflow] on Nvidia GPU P100. Dataset and Metrics =================== There are a limited number of image datasets for human rights violations recognition [@kalliatakis2019exploring; @visapp17]. In order to find the main test platform on which we could demonstrate the effectiveness of DisplaceNet and analyse its various components, we construct a new image dataset by maintaining the verified samples intact for the category *displaced populations*. The constructed dataset contains 609 images of displaced people and the same number of non displaced people counterparts for training, as well as 100 images collected from the web for testing and validation. The dataset is made publicly available for future research. We evaluate DisplaceNet with two metrics *accuracy* and *coverage* and compare its performance against the sole use of a CNN classifier. Experiments =========== **Implementation details.** Our emotion recognition implementation is based on the emotion recognition in context (EMOTIC) model [@kosti2017emotion], with the difference that our model estimates only continuous dimensions in VAD space. We train the three main modules on the EMOTIC database, which contains a total number of 18,316 images with 23,788 annotated people, using pre-trained CNN feature extraction modules. We treat this multiclass-multilabel problem as a regression problem by using a weighted Euclidean loss to compensate for the class imbalance of EMOTIC. For the classification part, we fine-tune our models for 50 iterations on the HRA subset with a learning rate of 0.0001 using the stochastic gradient descent (SGD) [@lecun1989backpropagation] optimizer for cross-entropy minimization. These *vanilla* models will be examined against DisplaceNet. Here, vanilla means pure image classification using solely fine-tuning without any alteration. **Baseline.** To enable a fair comparison between vanilla CNNs and DisplaceNet, we use the same backbone combinations for all modules described in Fig. \[Fig. 2\]. We report comparisons in both *accuracy* and *coverage* metrics for fine-tuning up to two convolutional layers in order to be consistent with the implementation of [@kalliatakis2019exploring]. The per-network results are shown in Table \[tab1\]. The implementation of vanilla CNNs is solid with up to 61.5% accuracy. Regarding coverage, vanilla CNNs achieve up to 16.83%. This shows that it is possible to trade coverage with accuracy in the context of human rights image analysis. One can always obtain high accuracy by refusing to process a number of examples, but this reduces the coverage of the system. Nevertheless, vanilla CNNs provide a strong baseline to which we will compare our method. Our method, has a mean coverage of 20.83%. This is an absolute gain of 4 points over the baseline of 16.83%. This is a relative improvement of 23.76%. In relation to accuracy, DisplaceNet has a mean accuracy of 57.33% which is an absolute drop of 4.17 points over the strong baselines of 61.5%. This indicates a relative loss of only 6.7%. We believe that this negligible drop in accuracy is mainly due to the fact that the $test$ set is not solely made up of images with people in their context, it also contains images of generic objects and scenes, where only the sole classifier’s prediction is taken into account. **Qualitative results.** We show our human rights abuse detection results in Fig. \[Fig. 5\]. Each subplot illustrates two predictions alongside their probability scores. The top of the two predictions is given by DisplaceNet, while the bottom one is given by the respective vanilla CNN sole classifier. Our method can successfully classify displaced people by overturning the initial-false prediction of the vanilla CNN. Moreover, DisplaceNet can strengthen the initial-true prediction of the sole classifier. Finally, our method can be incorrect, because of false dominance score inferences. Some of them are caused by a failure of continuous dimensions emotion recognition, which is an interesting open problem for future research. Conclusion ========== We have presented a human-centric approach for recognising displaced people from still images. This two-class labelling problem is not trivial, given the high-level image interpretation required. Understanding a person’s control level of the situation from his frame of reference is closely related with situations where people have been forcibly displaced. Thus, the key to our computational framework is people’s dominance level, which resonates well with our own common sense in judging potential displacement cases. We introduce the overall dominance level of the image which is responsible for weighting the classifiers prediction during inference. We benchmark performance of our DisplaceNet model against sole CNN classifiers. Our experimental results showed that this is an effective strategy, which we believe has good potential beyond human rights related classification. We hope this paper will spark interest and subsequent work along this line of research. All our code and data are publicly available. [^1]: A distinction is often made between conflict-induced and disaster-induced displacement, yet the lines between them may be blurred in practice.

--- --- c i ł ø u Ł Ø ¶ § \#1[| \#1|]{} /\#1[/\#1]{} |\#1 \#1 \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1|]{} \#1[| \#1]{} \#1\#2[\#1\#2]{} \#1[\#1]{} =\#1[ ]{} \#1 **** .  \#1 \#1[ ]{} \#1 [**[.  \#1]{}**]{} \#1[ ]{} =10000 16 cm 0.5cm 0.5cm =-0.5cm -.09cm 1.0cm SU-4228-533\ March 1993\ TOPOLOGY IN PHYSICS - A PERSPECTIVE A.P. Balachandran *Department of Physics, Syracuse University,* *Syracuse, NY 13244-1130* ABSTRACT This article, written in honour of Fritz Rohrlich, briefly surveys the role of topology in physics. When I joined Syracuse University as a junior faculty member in September, 1964, Fritz Rohrlich was already there as a senior theoretician in the quantum field theory group. He was a very well-known physicist by that time, having made fundamental contributions to quantum field theory, and written his splendid book with Jauch. My generation of physicists grew up with Jauch and Rohrlich, and I was also familiar with Fritz’s research. It was therefore with a certain awe and a great deal of respect that I first made his acquaintance. Many years have passed since this first encounter. Our interests too have gradually evolved and changed in this intervening time. Starting from the late seventies or thereabouts, particle physicists have witnessed an increasing intrusion of topological ideas into their discipline. Our group at Syracuse has responded to this development by getting involved in soliton and monopole physics and in investigations on the role of topology in quantum physics. Meanwhile, especially during the last decade, there has been a perceptible shift in the direction of Fritz’s research to foundations and history of quantum physics. Later I will argue that topology affects the nature of wave functions (or more accurately of wave functions in the domains of observables) and has a profound meaning for the fundamentals of quantum theory. It may not therefore be out of place to attempt a partly historical and occasionally technical essay on topology in physics for the purpose of dedication to Fritz. This article has no pretension to historical accuracy or scholarship. As alluded to previously, particle theorists have come to appreciate the importance of topology in the classical and especially in the quantum domain over the years, and that has evoked a certain curiosity about its role in the past, and about the circumstances leading to its prominence since the late seventies. The present article is an outgrowth of this idle curiosity. I have been greatly helped in its preparation by the book on soliton physics written by Russian colleagues$^1$, the essay on Skyrme by Dalitz$^2$ and a speech by Skyrme reconstructed by Aitchison$^3$, and have relied on these sources for information when necessary. Our recent changed perceptions about topology is well brought out by the following incident. Some time in the early part of this year, I received a book from Physics Today entitled [*Knots and Physics*]{}. It is written by the mathematician Louis Kauffman and Physics Today wanted me to review it. Now, twenty five years ago, it would have been remarkable to send a book on knots to a physicist for review. Indeed, a book with a title [*Knots and Physics*]{} would have been considered bizarre by physicists and mathematicians alike. Twenty five years takes us back to 1968. It was a time when the phenomenon of spontaneous symmetry breakdown was only beginning to be widely appreciated, and electroweak theory had not yet been fully articulated. Particle physicists were immersed in studies of symmetry principles, and Ken Wilson was yet to launch lattice QCD. Physicists and mathematicians had cordial, but generally distant relations. Ideas on topology were far from our minds. A knot to me meant no more at that time than what is tied at Hindu weddings. True, there were a handful of physicists, like David Finkelstein and Tony Skyrme, who talked of solitons and fundamental groups. But they were the oddities. We were content with Feynman diagrams and current commutators. But this was not always so. The first idea on solitons had already occurred to Scott Russell in 1842. He was observing the motion of a boat in a narrow channel and discovered that the water formed by its wake formed a remarkably stable structure. He coined the phrase “solitary elevation” while discussing the phenomenon he witnessed. I reproduce his report on what he saw below. [*Report on Waves. By*]{} J. Scott Russell, [*Esq., M.A., F.R.S. Edin.,*]{} [*made to the Meetings in*]{} 1842 [*and*]{} 1843. [*Members of the Committee*]{} $\left\{ \begin{array}{l} \mbox{Sir John Robison*, {\it Sec. R.S. Edin.}}\\ \mbox{J. Scott Russell, {\it F.R.S. Edin.}} \end{array} \right.$ > I believe I shall best introduce this phenomenon by describing the circumstances of my own first acquaintance with it. I was observing the motion of a boat which was rapidly drawn along a narrow channel by a pair of horses, when the boat suddenly stopped–not so the mass of water in the channel which it had put in motion; it accumulated round the prow of the vessel in a state of violent agitation, then suddenly leaving it behind, rolled forward with great velocity, assuming the form of a large solitary elevation, a rounded, smooth and well-defined heap of water which continued its course along the channel apparently without change of form or diminution of speed. I followed it on horseback, and overtook it still rolling on at a rate of some eight or nine miles an hour, preserving its original figure some thirty feet long and a foot to a foot and a half in height. Its height gradually diminished, and after a chase of one or two miles I lost it in the windings of the channel. Such, in the month of August 1834, was my first chance interview with that singular and beautiful phenomenon which I have called the Wave of Translation, a name which it now very generally bears; which I have since found to be an important element in almost every case of fluid resistance, and ascertained to be the type of that great moving elevation of the sea, which, with the regularity of a planet, ascends our rivers and rolls along our shores. It was not Scott Russell alone who came across topological notions in the last century. Sir William Thomson had published his ideas about atoms being vortices in a fluid in 1867$^1$. Thomson later became Lord Kelvin. Kelvin did not like the rigid point-like atoms of chemists. He very much wanted to describe them visually as extended structures. He was much impressed by Helmholtz’s discovery of “vorticity” in fluids. On the basis of experiments with Tait on smoke rings and analytical results on vortices, he developed his vortex atom which maintained that atoms are vortices in a perfect fluid. Some of the vortex atoms of Kelvin are shown in Fig. 1. He already had a good intuition about certain knot invariants and seemed upset that he knew Riemann’s “Lehrsätze aus der Analysis Situs” only through Helmholtz. [**Fig. 1 .**]{} [Kelvin proposed that atoms are vortices in a perfect fluid. The figure shows some of his vortex atoms.]{} But Kelvin’s ideas did not catch on. Fantasies of his sort soon came to be regarded as reactionary desires to preserve a dissolving mechanistic world. The ancien régime was losing its ability to rule in the face of the revolutionary onslaughts of the emerging relativists and quantum theorists. Physicists generally soon ceased to be seriously bothered by knots and topology for many years. But not all physicists. The young Dirac had published his remarkable text book on quantum mechanics in 1930. Soon thereafter$^4$, he came to recognize certain basic features of wave functions with implications which sharply differentiate classical from quantum physics. I can outline his insights from a modern perspective as follows$^5$. The dynamics of a system in classical mechanics can be described by equations of motion on a configuration space $Q$. These equations are generally of second order in time. Thus if the position $q(t_0)$ of the system in $Q$ and its velocity $\dot{q}(t_0)$ are known at some time $t_0$, then the equations of motion uniquely determine the trajectory $q(t)$ for all time $t$. When the classical system is quantized, the state of a system at time $t_0$ is not specified by a position in $Q$ and a velocity. Rather, it is described by a wave function $\psi$ which in elementary quantum mechanics is a (normalized) function on $Q$. The correspondence between quantum states and wave functions however is not one to one since two wave functions which differ by a phase describe the same state. The quantum state of a system is thus an equivalence class $\{e^{i\a}\psi|\a~{\rm real}\}$ of normalized wave functions. The physical reason for this circumstance is that experimental observables correspond to functions like $\psi^*\psi$ which are insensitive to this phase. In discussing the transformation properties of wave functions, it is often convenient to enlarge the domain of definition of wave functions in elementary quantum mechanics in such a way as to naturally describe all the wave functions of an equivalence class. Thus instead of considering wave functions as functions on $Q$, we can regard them as functions on a larger space $\hat{Q}=Q \times S^1\equiv\{(q,e^{i\a})\}$. The space $\hat{Q}$ is obtained by associating circles $S^1$ to each point of $Q$ and is said to be a $U(1)$ bundle over $Q$. Wave functions on $\hat{Q}$ are not completely general functions on $\hat{Q}$, rather they are functions with the property $\psi(q,e^{i(\a+\theta)})=\psi(q,e^{i\a})e^{i\theta}$. \[Here we can also replace $e^{i\theta}$ by $e^{in\theta}$ where $n$ is a fixed integer.\] The behaviour of wave functions under the action $(q,e^{i\a})\rightarrow(q,e^{i\a}e^{i\theta})$ of $U(1)$ is thus fixed. Because of this property, experimental observables like $\psi^*\psi$ are independent of the extra phase and are functions on $Q$ as they should be. The standard elementary treatment which deals with functions on $Q$ is recovered by restricting the wave functions to a surface $\{q,e^{i\a_0}|q\in \hat{Q}\}$ in $\hat{Q}$ where $\a_0$ has a fixed value. Suc h a choice $\a_0$ of $\a$ corresponds to a phase convention in the elementary approach. When the topology of $Q$ is nontrivial, it is often possible to associate circles $S^1$ to each point of $Q$ so that the resultant space $\hat{Q}=\{\hat{q}\}$ is not $Q \times S^1$ although there is still an action of $U(1)$ on $\hat{Q}$. We shall indicate this action by $\hat{q} \rightarrow \hat{q}e^{i\theta}$. It is the analogue of the transformation $(q,e^{i\a}) \rightarrow (q,e^{i\a}e^{i\theta})$ we considered earlier. We shall require this action to be free, which means that $\hat{q}e^{i\theta}=\hat{q}$ if and only if $e^{i\theta}$ is the identity of $U(1)$. When $\hat{Q}\neq Q \times S^1$, the $U(1)$ bundle $\hat{Q}$ over $Q$ is said to be twisted. It is possible to contemplate wave functions which are functions on $\hat{Q}$ even when this bundle is twisted provided they satisfy the constraint $\j(\hat{q}e^{i\theta})=\j(\hat{q})e^{in\theta}$ for some fixed integer $n$. If this constraint is satisfied, experimental observables, being invariant under the $U(1)$ action, are functions on $Q$ as we require. However, when the bundle is twisted, it does not admit globally valid coordinates of the form $(q,e^{i\a})$ so that it is not possible (modulo certain technical qualifications) to make a global phase choice, as we did earlier. In other words, it is not possible to regard wave functions as functions on $Q$ when $\hat{Q}$ is twisted. \[We are assuming for ease of presentation here that wave functions are always smooth functions. That is not of course always the case. The significance of $\hat{Q}$ is that smooth functions on $\hat{Q}$ can provide us with physically acceptable domains for observables. Discussions involving domains of operators tend to be technical. We will not therefore pursue this remark further here.\] It was a great merit of Dirac that already in his seminal paper of 1931 on the role of phases in quantum theory$^4$, he isolated a physical system where $\hat{Q}$ was twisted and these phases had an important role. This was the system of a particle with electric charge $e$ and a particle with magnetic charge $g$. Now we all know that if there is a magnetic monopole (or monopole for short) at the origin, the Maxwell equation $\vec{\nabla}\cdot \vec{B}=4\pi g\d^3(x)$ excludes the existence of a smooth vector potential for the magnetic induction $\vec{B}$ and causes endless trouble in formulating quantum theory. The idea of Dirac was to replace the monopole by a semi-infinite, infinitely thin current loop \[see Fig. 2\]. This semi-infinite loop is often [**Fig. 2 .**]{} [With a magnetic monopole at the origin,$\vec \nabla . \vec B=4\pi g\,\,\,\delta ^3 (x)$, $\vec B=$ magnetic induction. Dirac represented magnetic monopoles by semi-infinite, infinitely thin current loops (Dirac strings). He showed that the loop can not be observed if electric charge $e=\frac {\hbar}{2g}n$, $(n=0,\pm 1,...)$. Thus if a magnetic monopole exists, electric charge is quantized.]{} called the Dirac string. The effect of the string away from itself is that of a monopole, but that is not so exactly on the string. Dirac argued that the effect of this string is undetectable if $e=\frac{\hbar}{2g}n$ (in units where the speed of light $c$ is 1), $n$ being an integer. When that is the case, shifting the string amounts only to changing the phase of the wave function in a way that cannot be observed. Thus, the position of the string, and hence the string itself, ceases to be an observable when $n$ is an integer. In that case, then, Dirac’s approach yields a quantum theory for a charge and a monopole \[whereas it gives only a theory of a charge and a semi-infinite current loop when $n$ is not an integer\]. In this way, Dirac derived the quantization of $eg$ in a charge-monopole theory. This paper of Dirac is fundamental to quantum theory, and much of modern mathematics as well. On the physical side, it predicts charge quantization in units of $\frac{\hbar}{2g}$ and charge quantization is a basic experimental fact. Further, the charge-monopole system has the remarkable feature that it can have half-odd integral angular momenta even if the charge and monopole have integral spin. We are thus led to understand from Dirac’s work and later developments that composites can have half-odd integral spin even if its constituents have integral spin. Dirac’s paper also suggests the ideas which later were discovered, developed and forcefully articulated by Aharonov and Bohm in the context of their celebrated effect. As for mathematics, it must be among the first publications on fibre bundles, the integer $n$ defining what the mathematicians call the Chern class. In Molière’s book, [*Le Bourgeois Gentilhomme*]{}, M. Jourdain somewhere exclaims to his philosophy teacher that he was making prose for more than forty years without knowing it. The following is a reproduction of the relevant passage. > M. Jourdain: Quoi? quand je dis: ‘Nicole, apportez-moi mes pantoufles, et\ > me donnez mon bonnet de nuit’, c’est de la prose ? > > Maître de Philosophie: Oui, Monsieur. > > M. Jourdain: Par ma foi! il y a plus de quarante ans que je dis de la prose\ > sans que j’en susse rien. > > M. Jourdain: What? when I say: ‘Nicole, bring me my slippers, and give me\ > my night-cap’, is that prose ? > > Philosophy Teacher: Yes, Sir. > > M. Jourdain: Good heavens! For more than forty years I have been speaking\ > prose without knowing it. > > [*Le Bourgeois Gentilhomme*]{} (1670), II. iv. It would seem that Dirac was unknowingly making mathematics in the same way that M. Jourdain was unconsciously making prose. All this happened in 1931. But we knew nothing about Dirac’s work when we learned quantum theory in Madras. Indeed, quantum theory is still taught in universities missing out on all the beautiful and fundamental discoveries coming from the youthful Dirac of 1931. Dirac’s paper, I suppose, must have been too difficult for physicists for several decades. What is surely true is that it was largely ignored till the ’70’s. Then events began to unfold in particle and condensed matter theory bringing topological issues to centerstage, and reviving forgotten memories of Dirac. But before describing these events, let us first go back to the late ’50’s and the ’60’s. At that time, Tony Skyrme and David Finkelstein had also discovered novel topological ideas which they were developing independently. They were working alone, or with a colleague or two. Their ideas too remained uncomprehended by the community for many years. Tony Hilton Royle Skyrme was born in England in 1922 in the house of his maternal grandparents$^2$. His maternal great-grandfather Edward Robert knew and admired Kelvin, and was associated with the construction of the Tidal Predictor under the direction of Kelvin and Tait. This machine was for predicting tides worldwide. The admiration went to the extent of naming his son Herbert William Thomson Roberts. This machine was in the house where Tony was born. Tony has said in a speech$^3$ that he was greatly impressed by the ingenuity of its mechanism. Tony grew up in a world beset with increasing turbulence. In 1943, after Cambridge, he joined the British war effort in making the atomic bomb. It was only in 1946 that he began fundamental research. During 1946-61, he was associated with Cambridge, Birmingham and Harwell and was engaged in wide ranging investigations in nuclear physics. It was this work, especially the work on nuclear matter and the fluid drop model, which eventually culminated in his beautiful proposition that nucleons are solitons made of pions. In the speech of Tony I mentioned above, he has described the reasons behind his extraordinary suggestions. He knew of Kelvin from the Tidal Predictor, and was vaguely aware of Kelvin’s vortex atoms. Like Kelvin, he too desired a model of the nucleon which was visualizable and extended. He felt that fermions can emerge from self-interacting Bose fields just as bosons arise as bound states of fermions. Moved by these imprecise and intuitive desires, Tony began his work on nonlinear field theories, the sine-Gordon equation and the chiral model and was led to the proposition that nucleons are twisted topological lumps of pion fields$^6$. In his papers, Tony also had initiated ideas on bosonization, vertex operators, and quantum theories on multiply connected spaces, all years and years ahead of his time, and all topics of central interest today. David Finkelstein had a grasp of topology and differential geometry which was exceptional for physicists in the ’60’s. Like Skyrme, he had understood that solitons can acquire spin 1/2 from the topology of the configuration space. He must have realized what little role relativistic quantum field theory played in the theory of solitons, and struck by the fact that all existing proofs of the spin-statistics theorem relied on relativistic quantum field theory. But solitons can acquire spin half in nonrelativistic models and can not always be described by relativistic quantum fields. He and Rubinstein, I suppose, were led by such thoughts to seek and find an alternative proof of the spin-statistics theorem. Their proof was published in 1968$^7$. It is this proof which is important for chiral solitons. There are grounds to expect that it is the Finkelstein-Rubinstein approach which will be found significant in quantum gravity as well. An absolutely fundamental result, namely the spin-statistics theorem, is getting topologized. Even more striking, it is still not properly understood. The closing ’60’s and the ’70’s herald the dawn of modern times for theorists. In condensed matter physics, attention began to focus on vortices in superconductors and superfluids, and defects in liquid crystals. Soon a classification of defects based on its topological properties emerged. In particle physics, the bootstrap theory of Chew was leading to unexpected developments. According to Chew, there is nuclear democracy, all particles are bound states of each other and none is more elementary than another. These ideas evolved into string theory with its explosive implictions for mathematics and mathematical physics. Physicists in search of the ultimate solution claim that “strings are “TOE”, “TOE” meaning “Theory of Everything.” If that is so, physicists have found a scientific substitute for God. QCD and electroweak theory also began to take shape in the late ’60’s and the early ’70’s. It soon became plausible that they had the ability to account for physics at energies less than about 100 GeV, and perhaps up to much higher energies as well. Attention turned to unification of strong and electroweak theories by the construction of “grand unified” models. It was in this ambience that topology was discovered in particle theory. There was first the striking discovery that grand unified theories predicted monopoles and vortices. Realization soon dawned that the correct approach to their study involved the ideas put forth by Dirac in 1931 and the defect theories of condensed matter. There was then the discovery of instantons in QCD and its implications for vacuum structure, time reversal violation and neutron electric dipole moment. The remarkable observation was also made by ’t Hooft$^8$ that electroweak theory predicted baryon and lepton number violation, although at that time, it was felt that the effect was too small to observe. There was another dimension to this story. These developments involved intense cross fertilization between fields. There were basic contributions to physics by leading mathematicians. It involved as well greater interaction between condensed matter and particle theorists at least because of their shared interest in defect theory. The result of all this activity was that we were bombarded with papers on topology, and by sheer exposure, if not by effort, began to be aware of topological issues. For some of us, this was a period of excitement at learning beautiful new ideas. It was also a period of hope that nonperturbative effects with topological roots can now be investigated with greater ease. Instantons, vortices and monopoles began to play a role in papers on phenomenology and have continued to do so to this day. The startling suggestion was made that monopoles can catalize proton decay at strong interaction rates with life times of the order of perhaps $10^{-19}-10^{-20}$ seconds. The discovery of ’t Hooft about baryon and lepton number violation was also given teeth by the realization that their rates were greatly enhanced at high temperatures. It is nowadays widely speculated that this effect has a significant bearing on the observed baryon number asymmetry in the universe. It was during this time that we discovered Skyrme’s work at Syracuse$^9$. There was, before us, the paper of Pak and Tze$^{10}$ reviewing Skyrme’s research. That too suggested to us that it was worth our while to work on Skyrmions. It was in this way that we came to write our papers on Skyrmions and tell Witten about these ideas. Soon there followed Witten’s remarkable papers$^{11}$. It did not take much time thereafter for the general acceptance of Skyrme’s ideas. Topological notions have flourished with extraordinary vigour in particle physics, and what is at times called physical mathematics, for ten or more years. Much of the recent impetus comes from string and conformal field theories and their relation to complex manifolds. Topology is not likely to go away from classical and quantum physics in the foreseeable future. It is now appreciated that it has at least two important roles in physics. Firstly, it can suggest the existence of stable structures like defects, vortices, monopoles and Skyrmions. Secondly, it has a profound influence on the nature of quantum states. This influence comes about because, as mentioned previously, the phases of wave functions can have serious consequences in quantum theory. This second feature is still poorly understood. Even its existence is not widely known even though sixty two years have passed after Dirac’s paper. There is also a nonabelian generalisation of these phases which we have not discussed in this article. There is also an entirely new role topological physics has recently begun to assume, brought about because fundamental mathematical developments are nowadays significantly influenced by quantum field theory. The contributions to knot theory by Witten, the discovery of certain mirror manifolds in string theory and the recent developments in Riemann surface theory are all dramatic examples of this role of physics in mathematics. This then is one more reason for us to anticipate the vitality and longevity of topological trends in physics. It is time to come to an end. Although there are certain prominent names associated with topological physics and its developments, it is good to be conscious that physics is a social activity in which many humans, not all well-remembered, participate, and that whatever we do necessarily partakes of the collective knowledge and creativity of the physics community. From this perspective, all the advances and insights which have emerged from studying topological aspects of physics are also ultimately the fruits of labor of generations of physicists. I conclude this essay by quoting a poem by Brecht which aptly describes these thoughts. A WORKER READS HISTORY > Who built the seven gates of Thebes ?\ > The books are filled with names of kings.\ > Was it kings who hauled the craggy blocks of stone ?\ > And Babylon, so many times destroyed,\ > Who built the city up each time ? In which of Lima’s houses,\ > That city glittering with gold, lived those who built it ?\ > In the evening when the Chinese wall was finished\ > Where did the masons go ? Imperial Rome\ > Is full of arcs of triumph. Who reared them up ? Over whom\ > Did the Caesars triumph ? Byzantium lives in song,\ > Were all her dwellings palaces ? And even in Atlantis of the legend\ > The night the sea rushed in,\ > The drowning men still bellowed for their slaves. > > Young Alexander conquered India.\ > He alone ?\ > Caesar beat the Gauls.\ > Was there not even a cook in his army ?\ > Philip of Spain wept as his fleet\ > Was sunk and destroyed. Were there no other tears ?\ > Frederick the Great triumphed in the Seven Years War. Who\ > Triumphed with him ? > > Each page a victory, At whose expense the victory ball ?\ > Every ten years a great man,\ > Who paid the piper ? > > So many particulars.\ > So many questions. I am most grateful to Paulo Teotonio for drawing the figures and typing their legends, and to Carl Rosenzweig and Paulo Teotonio for help with the text. This work was supported by the Department of Energy under contract number DE-FG02-85ER40231. **[References]{}** A long list of references is not appropriate in an informal essay of this sort. Only a limited number of publications are therefore cited. 1. V.G. Makhankov, Yu. P. Rybakov and V.I. Sanyuk, “The Skyrme Model, Fundamentals, Methods, Applications” \[Springer-Verlag (in press)\]. This book has been especially useful for information on Kelvin and on his vortex model. 2. R.H. Dalitz, [*Int. J. Mod. Phys. A*]{}[**3**]{} (1988) 2719. 3. T.H.R. Skyrme, [*Int. J. Mod. Phys. A*]{}[**3**]{} (1988) 2745. This speech was reconstructed by I.J.R. Aitchison. References 2 and 3 have been freely used while discussing Skyrme. 4. P.A.M. Dirac, [*Proc. Roy. Soc.*]{} (London) [**A133**]{} (1931) 60. 5. A.P. Balachandran, G. Marmo, B.S. Skagerstam and A. Stern, “Classical Topology and Quantum States” \[World Scientific, 1991\]. 6. T.H.R. Skyrme, [*Proc. Roy. Soc.*]{} (London) [**A247**]{} (1958) 260; [**A260**]{} (1961) 121; [**A262**]{} (1961) 237; [*Nucl. Phys.*]{} [**31**]{} (1962) 556; [*J. Math. Phys.*]{} [**12**]{} (1971) 1735. 7. D. Finkelstein and J. Rubinstein, [*J. Math. Phys.*]{} [**9**]{} (1968) 1762. For recent developments involving the Finkelstein-Rubinstein ideas, see R.D. Tscheuschner, [*Int. J. Theor. Phys.*]{} [**28**]{} (1989) 1269 \[Erratum: [*Int. J. Theor. Phys.*]{} [**29**]{} (1990) 1437\]; A.P. Balachandran, A. Daughton, Z-C. Gu, G. Marmo, A.M. Srivastava and R.D. Sorkin, [*Mod. Phys. Letters*]{} [**A5**]{} (1990) 1575 and [*Int. J. Mod. Phys.*]{} [**A**]{} (in press); A.P. Balachandran, T. Einarsson, T.R. Govindarajan and R. Ramachandran, [*Mod. Phys. Letters*]{} [**A6**]{} (1991) 2801; A.P. Balachandran, W.D. McGlinn, L. O’Raifeartaigh, S. Sen and R.D. Sorkin, [*Int. J. Mod. Phys.*]{} [**A7**]{} (1992) 6887; A.P. Balachandran, W.D. McGlinn, L. O’Raifeartaigh, S. Sen, R.D. Sorkin and A.M. Srivastava, [*Mod. Phys. Letters*]{} [**A7**]{} (1992) 1427. These paper s develop proofs of spin-statistics theorems which do not use relativity or quantum field theory. 8. G. ’t Hooft, [*Phys. Rev. Letters*]{} [**37**]{} (1976) 9; [*Phys. Rev.*]{} [**D14**]{} (1976) 3432 \[Erratum: [**D18**]{} (1978) 2199\]. 9. A.P. Balachandran, V.P. Nair, S.G. Rajeev and A. Stern, [*Phys. Rev. Letters*]{} [**49**]{} (1982) 1124; [*Phys. Rev.*]{} [**D27**]{} (1983) 1153. 10. N.K. Pak and H. Ch. Tze, Ann. Phys. (N.Y.) [**117**]{} (1979) 164. 11. E. Witten, [*Nucl. Phys.*]{} [**B223**]{} (1983) 422, 433.

--- abstract: 'The spectacular metal-to-insulator transition of V$_2$O$_3$ can be progressively suppressed in thin film samples. Evidence for phase separation was observed using microbridges as a mesoscopic probe of transport properties where the same film possesses domains that exhibit a metal-to-insulator transition with clear first order features or remain metallic down to low temperatures. A simple model consisting of two parallel resistors can be used to quantify a phase coexistence scenario explaining the measured macroscopic transport properties. The interaction between film and substrate is the most plausible candidate to explain this extended phase coexistence as shown by a correlation between the transport properties and the structural data.' author: - 'C. Grygiel' - 'A. Pautrat' - 'W. C. Sheets' - 'W. Prellier' - 'B. Mercey' - 'L. Méchin' title: 'Mesoscopic electronic heterogeneities in the transport properties of V$_2$O$_3$ thin films' --- Epitaxial thin films have been the focus of intensive research, owing to their potential application in emergent technologies and fundamental studies of physical phenomena. In particular, strain introduced by a lattice mismatch between thin film and substrate, can be used to alter their properties and stabilize metastable states. An extended temperature regime for phase coexistence has been observed for certain chemical compositions that, in principle, are not allowed by the Gibbs’ phase rule.[@gibbs] One well documented case is that of perovskite manganite R$_{1-x}$A$_x$MnO$_3$ ($R$ = rare earth and $A$ = alkaline earth cations) thin films, which undergo a metal-to-insulator phase transition at a significantly different temperature than those of bulk samples.[@manga] It should also be noted that the disorder in the ionic radius of the $A$-site cation has a profound effect on the average ordering temperature of bulk manganites, which often complicates the determination of how much strain contributes to phase separation in thin film samples.[@moreo1999; @dagotto2001] Nevertheless, a recent comparison of solid-solution alloy and $A$-site ordered superlattice La$_{2/3}$Ca$_{1/3}$MnO$_3$ thin film samples on different substrates revealed that strain is more important than chemical disorder in stabilizing the mixed phase regime near the average ordering temperature.[@pala1] To investigate further the role of the substrate in the phase separation of a thin film sample, it is interesting to work on a sample which is, in its bulk form, a prototype of a discontinuous transition. In this contribution, the transport properties of epitaxial V$_2$O$_3$ thin films are examined. Vanadium sesquioxide V$_2$O$_3$ exhibits a remarkable first-order metal-to-insulator (M-I) transition around 150[K]{}, below which an antiferromagnetic insulating phase exists. Hydrostatic and chemical pressure (V$_{2-x}$M$_{x}$O$_3$ with M=Ti...) or sample non-stoichiometry (V$_{2-y}$O$_3$) significantly lowers this transition temperature.[@shivashankar83; @ueda80; @whan73; @yethiraj90] In particular, application of hydrostatic pressure above a critical threshold of 26[KBars]{} suppresses the M-I phase transition.[@whan69] Under particular growth conditions, substrate-induced strain also suppresses partially the M-I transition,[@lab1; @allimi] and metallic-like behavior, with some indication of a Fermi liquid regime, is observed below 20[K]{}.[@lab1] The intermediate-temperature regime, which is apparently dependent upon film thickness, is more complex, and a non-monotonic variation of the slope ($dR$/$dT$) is observed.[@grygiel2007] Surprisingly, using c-Al$_2$O$_3$ [@grygiel2007] or c-LiTaO$_3$ [@allimi] , the thicker films present a metallic character. Films thinner than 220[Å]{} recover the M-I transition, although it is attenuated strongly when compared to the transition observed in the bulk. To clarify such behavior, we have investigated the homogeneity of the samples at the mesoscopic level (where mesoscopic represents a scale below the large statistical collections at which average properties have been measured previously). Using microbridges as local probes, we show in these thin films a phase separation, which explains by using a simple model of two parallel resistors the evolution of the transport properties at the macroscopic scale. A clear correlation with the structural data is also found. High quality V$_2$O$_3$ films were grown on (0001)-oriented sapphire substrates using the pulsed laser deposition technique. A pulsed KrF excimer laser beam ($\lambda$=248[nm]{}, pulse length 20[ns]{}, repetition rate 3[Hz]{}) was focused on a V$_2$O$_5$ target, depositing films onto the substrate under optimized deposition conditions (650${{}^{\circ }}$C, 0.02[mbar]{} Ar pressure, and a laser fluence of 4[J cm$^{-2}$]{}). The structure of the resulting films was examined by X-Ray Diffraction, which attests to their high quality and epitaxial relation with the substrate. The in-depth details of the growth conditions, and some structural and microstructural properties have been reported previously.[@grygiel2007] The resistance of the samples was measured using a physical property measurement system. Electrical transport measurements were made on unpatterned films using a four probe geometry. Silver contact pads, separated by 1[mm]{}, were deposited by thermal evaporation through a mechanical mask. A 230[Å]{} thick sample, which exhibits a low rms roughness of 0.47[nm]{} (averaged over 3$\times$3$\mu m^{2}$), was selected for patterning. A silver layer was deposited onto the film. Contact pads were first defined using ultraviolet (UV) photolithography and chemical etching in a KI/I$_{\text{2}}$ solution. The V$_{2}$O$_{3}$ thin film was then patterned by UV photolithography and argon ion etched to form 20[$\mu $m]{} and 100[$\mu $m]{} wide bridges. The final microbridge, shown in the figure \[f1\], allows for different measurement geometries. For this experiment, the voltage was measured between $V_1$ and $V_3$, which corresponds to a length of 200[$\mu $m]{}. The external circuitry and the patterned film with silver contact pads were connected using aluminum-silicon wires attached by ultrasonic bonding. Transport measurements were made in 20$\times$200$\mu m^{2}$ (M1) and 100$\times$200$\mu m^{2}$ (M2) bridges. M1 and M2 are located close to each other in the center of the sample. In unpatterned films, we find that the resistivity has an average value of $450 \pm 150 $[$\mu \Omega $.cm]{} at room temperature. In patterned film, we measure 1030$\pm $80[$\mu \Omega $.cm]{} for M1 and 950$\pm $80[$\mu \Omega $.cm]{} for M2. This increase is almost entirely due to an increase of the residual resistivity. From this we conclude that the patterning process did not significantly alter the film properties and, importantly, that each bridge is nearly identical. The resistivity of numerous unpatterned V$_2$O$_3$ thin films were measured as a function of their thickness (approximately 20 samples with thicknesses $t$, 40[Å]{}$\leq t\leq$ 1000[Å]{}). As shown in figure \[rho300\], the film resistivities at room temperature are close to $450 \pm 150 $[$\mu \Omega $.cm]{}, values which are similar to the those of the crystals, and do not evolve with the thickness. The resistivity versus temperature curves measured during cooling for samples of different thicknesses are presented in figure \[courbes-ech\]. While the high temperature behavior is similar for all the samples notable differences appear when $T\leq 150$[K]{}. Insulating behavior is favored for the thinnest samples and metallic for the thickest. A pure metallic behavior was never observed at the macroscopic scale over the 2-300[K]{} range since a memory of the M-I transition is always observed (even if it is tiny), but metallic behavior is recovered at low temperatures.[@grygiel2007] It should be noted that even the more metallic V$_2$O$_3$ behaves as a bad metal and the classical size effect is not observed down to the lowest thickness ($t \approx 42$[Å]{}), which is in agreement with the anomalously low mean free path ($\ell \leq 0.2nm $ in a Drude approximation). This explains the thickness independence of the room temperature resistivity. At lower temperatures, the complex behavior of the resistivity requires further investigations. In particular, the resistivity maximum occuring at intermediate temperature may indicate a competition between metallic and insulating states as observed previously in manganites.[@mayr] The phase separation scenario has been used to explain such characteristics and one may ask if the situation is similar in our films. Accordingly, we have performed local resistivity measurements on a sample possessing an intermediate thickness ($t$=230[Å]{}). This film was selected because it represents the cross-over between the metallic and the insulating unpatterned film behaviours, and displays a clear maximum in $\rho(T)$ at $T\approx 70K$. Local resistivity measurements were done using two microbridges (M1 and M2) patterned on this sample in order to observe if the transport properties are identical at a smaller scale than the macroscopic one. Figure \[f3\] displays the evolution of the resistance ratio $R$/$R_{300K}$ for each microbridge of this patterned film. From the plots, it is clear that each microbridge exhibits significantly different transport properties for $T\leq 150$[K]{}, and that the macroscopic behavior is not simply recovered. The data for the first microbridge (M1) present a metal-to-insulator transition. While the largest increase in resistance is observed close to 150[K]{}, smaller ones are also observed at certain discrete values: 120[K]{}, 110[K]{} and at approximately each 10[K]{} down to 70[K]{} (as shown on figure \[f3\] by arrows). The transition at 150[K]{} appears similar to the first order M-I transition observed in pure bulk V$_2$O$_3$ samples. The resistance increases around three orders across the transition between 150[K]{} and the low temperatures, and this is notably less than what is reported for the transition in crystals where the increase can be up to seven orders of magnitude. Since other transitions occur at different temperatures, each jump in resistance can be attributed to the transition of a part of, and not all of, the bridge. Such behavior can explain, at least partially, why the M-I transition is notably attenuated. Owing to the multiple transitions observed, the domain size of one insulating domain is estimated to be smaller than that of the microbridge (&lt; 20$\times$200$\mu m^{2}$), but not too small to largely affect the resistivity, i.e. in the micrometer scale. Two features demonstrate that this transition remains locally first order, even though it is affected by the disorder. Firstly, after cooling the sample, resistance time series were measured at fixed temperatures (Fig. \[noise\]). During these measurements resistance fluctuations were observed, which are large non-Gaussian $1/f^{\alpha}$ noise, and can be attributed to highly inhomogeneous current paths that form when close to an incipient M-I transition,[@DCR] or two states fluctuators. We focus here on the discrete two-state fluctuators near the transition, shown in the inset of figure \[noise\], in the absence of large non-equilibrium noise. The statistics of the two states are largely independent of the applied magnetic field, which demonstrates that the fluctuation of magnetic domains can not be involved as an origin of this noise (the M-I transition is also a paramagnetic to antiferromagnetic transition, so magnetic effects can not be completely neglected). Based on work in colossal magnetoresistive films,[@merithew] the measurement of the resistance during long intervals (approximately 1 hour for each time series) can be used to extract the main relaxation time $\tau_i$ for each state (high resistance (i=1) and low resistance (i=2) states). Assuming $\tau_i \propto exp(-G_i/K_BT)$ where $G_i$ is the free energy in the state $i$, the Boltzman factor $r=\tau_1/\tau_2$ depends on the free energy difference. The latter term then can be used to calculate the entropy difference ($\Delta S$), which we find to be $\Delta S \approx 60 K_B$ using the thermodynamic identity $\Delta S = K_B \partial(T ln r) /\partial T$. Such behavior is expected when a barrier between two phases exists, and the transition is first order (rather than second order).[@pala] Secondly, when supercooling from a high temperature, the low resistance state is blocked, and after a certain time the resistance switches suddenly to its high resistance state (see Fig. \[noise\]). This indicates that a metastable state can be supercooled, which is also in agreement with a first order nature for the transition. The second curve on the figure \[f3\] corresponds to the measurements made on the second microbridge (M2) on the same sample. For this microbridge there is no trace of the M-I transition and the sample behaves as a metal down to 2[K]{}. At low temperature, the resistance follows the expected $T^2$ temperature variation for electron-electron scattering, as already shown for V$_2$O$_3$ thin films and single crystals subjected to high pressures (26-52[KBars]{}).[@grygiel2007; @whan69] The detailed analysis of this metallic phase, including magnetoresistance, noise properties, and localisation effects will be discussed in another paper. Here, we address if the suppression of the M-I transition in these domains can be accounted for substrate induced pressure. Assuming the film is confined over the surface area of the substrate at the M-I transition temperature leads to an increase of pressure in the film. The maximum effective pressure generated can be estimated [@pala] using experimental data from bulk samples, such as the stress coefficients $C_{11}$ and $C_{12}$,[@stress] and the volume change during the bulk M-I transition (we take 2.3 $\%$ as a representative value, even if some dispersion exists in the litterature).[@whan69] This results in a calculated effective pressure of ($\Delta P$)$\approx 27 KBars$, and because the decrease in the critical temperature owing to elastic distortion arises from the experimental $dP$/$dT_c$,[@whan69] the variation in the critical temperature is ($\Delta T_c$)$ \geq 160 K$. This value is consistent with the suppression of the M-I transition. Since a first order transition is heterogeneous when the volume is kept constant because it induces an heterogeneous pressure, such mechanism will certainly lead to a broad distribution of critical temperatures at a macroscopic scale. As a consequence, for a fixed temperature $T<150 K$, both metallic and insulating domains with different critical temperatures can coexist. These results obtained for the microbridges indicate the presence of mesoscopic electronic heterogeneity in V$_2$O$_3$ films where metallic and insulating regions coexist over a very large temperature range. Since their properties contrast, clear consequences can be expected for transport properties measured on a large scale, i.e. for unpatterned samples. We suggest that these films consist of different metallic and insulating domains, such as those observed using the microbridges, which are formed below 150[K]{}. In heterogeneous medium, percolation theory can be used to calculate the effective resistivity [@perco], using for example random resistors networks. In the case of large heterogeneities, a coarse-grain approach can be sufficient to give a good description of the resistivity, and the effective resistivity can be described using a parallel-resistor model with a metallic (percolative) resistance and an insulating resistance [@mayr]. We do not have a direct measure of the size of the domains, but as discussed above, we propose that the each insulating domain is in the micrometer scale. In the metallic microbridge M2, we do not observe any features (non linearities in voltage-current characteristics, thermal hysteresis) which could indicate mixed states at a lower scale, showing that metallic domains are homogeneous over the size of the microbridge or that it is always shorted out by metallic paths, i.e. above the percolation threshold. In phase separated manganites, a coarse grain approach was strongly guided by the existence of a peak at intermediate temperatures in the effective resistivity, by the large scale (micrometric) of the electronic heterogeneities, and by the recovery to metallic properties at low temperature indicating that current flows trough metallic paths [@mayr; @andres; @taran]. Here we observe the same characteristics. Consequently, to reconstruct the observed macroscopic behaviours, we also propose a two-parallel-resistors model, with each resistance directly given by the measured resistance in the microbridge (M1 for $R_I$ and M2 for $R_M$). The ratio of the metallic phase, denoted as $x$ can be used in the following expression to determine the effective conductivity $\sigma$, $$\label{"the modele"} \sigma = x \sigma _M + (1-x) \sigma _I ,$$ where $\sigma _M$ and $\sigma _I$ are the conductivity of the metallic and insulating phases of the film, respectively, measured experimentally with the microbridges. The variable in this basic model, except for the $x$ value, is the value of the residual resistivity of the metallic phase, which depends slightly (in a non-trivial way) on the fluctuations of the growth parameters. This variable influences the absolute resistivity value at low temperatures where the current path is essentially in the metallic phase, but, overall, does not significantly alter the functional form of the $R(T)$. We have applied this phenomenological model to our unpatterned films. Three simulated curves of resistivity as function of the temperature, each corresponding to a different $x$, are shown in the figure \[courbes-modele\]. The $x$ values, which are temperature independent, were ajusted to match the temperature dependence of simulated curves with those of the experimental curves. The experimental and the calculated curves have very similar appearances, demonstrating that the model (even if simple) is able to explain semi-quantitatively the transport properties. Consequently, the relevant parameter depending on the thickness is the ratio of conducting phase, $x$. For each sample, the corresponding $x$ parameter was deduced. The figure \[x(t)\] shows its thickness-dependence, with a large change at a critical thickness $t_c \approx 200$[Å]{} (as shown on the curve by the hatching zone). For $t\geq t_c$, $x$ is close to $1$, indicating that almost the whole film is metallic. In contrast, for $t<t_c$, $x$ becomes small and the domains with the M-I transition form the majority of sample, with an electrical behaviour similar to the bulk material. The film thickness can therefore be used to control the ratio of conducting phase and the mesoscopic phase separation. The range of thickness where the phase separation is critical is restricted to the domains of $100$[Å]{}$<t<250$[Å]{} (see Fig. \[x(t)\]). These observations show that the metallic phase in our V$_2$O$_3$ films is favored by the large thicknesses. Recalling that metallic V$_2$O$_3$ crystals result from an effective applied pressure, we come to the suprising conclusion that the thicker samples are under larger stress than the ultra-thin films. To explain this unusual evolution of the properties, a structural study was performed. It is known that the ratio of lattice parameters $c/a$ is extremely sensitive to lattice distorsions, especially in the case of hexagonal stacking.[@finger80; @luo2004] Figure \[c/a\]a represents the evolution of $c/a$ as a function of the film thickness, deduced from the analysis of X-Ray diffraction at room temperature.[@grygiel2007] A small variation in this ratio is observed close to a thickness of 200[Å]{} (indicated by a dotted line in figure \[c/a\]a), which is close to the $t_c$ deduced from the transport properties. Films with $t<t_c$ have a reduced ratio $c/a$ ($\sim 2.835$), and compare favorably with the value of $2.828$ reported for a pure $V_2O_3$ crystal (dashed line). Films with $t>t_c$ possess a ratio closer to $2.85$, the value for a crystal subjected to high hydrostatic pressure (solid line).[@finger80] Further indication of a structural cross-over at $t\approx t_c$ is observed in figure \[c/a\]b. The thick films ($t>t_c$) follow a Poisson law with a calculated Poisson coefficient $\nu \approx 0.5$ (incompressibility limit). In contrast films with $t<t_c$ do not follow the Poisson law, points surrounded on the curves (Fig.\[c/a\]), indicating a change in the elastic properties of the films. Overall, both observations indicate that the thinner films have structural properties close to those of the bulk material, and thus undergo the M-I transition associated with a volume variation. We have shown that heterogeneous strains induced by the M-I phase change and substrate-induced volume confinement are likely responsible for the broadening of the critical temperature from 150[K]{} until complete suppression for the conducting phase. However, the role of non-stoichiometry which is known to influence $V_2O_3$ properties must also be considered. For V$_{2-y}$O$_3$ compounds, the ratio $c/a$ remains constant, whatever the $y$ value, and is equal to $2.828$ because both $a$ and $c$ are compressed with respect to stoichiometric $V_2O_3$.[@ueda80] In our films, $c/a$ varies with $t$ and both $a$ and $c$ exhibit an in-plane compressive lattice parameter and an out-of-plane tensile lattice parameter.[@grygiel2007] In general, the evolution of properties as function of $t$ in our films can not be explained by non-stoichiometry, but is consistent with pressure effects.[@ueda80; @whan69] Further studies, including detailed structural studies with Transmission Electron Microscopy, are under way to understand the details of the growth process responsible for the cross-over in the unit cell distortion at $t\approx 200$[Å]{}. In summary, we have used microbridges as local probes to examine the transport properties of V$_2$O$_3$ thin films. Mesoscopic electronic heterogeneities are observed and consist of insulating and metallic zones that coexist over a very large temperature range. This phase separation has a strong influence on the macroscopic transport properties where the evolution of the resistivity as a function of film thickness appears to be governed by the change in quantity of each phase, which is in agreement with the structural data. As in the case of manganite films,[@pala] we believe that such heterogeneous states should be common for films with first order transitions interacting with a substrate. This work is carried out in the frame of the STREP CoMePhS (NMP3-CT-2005-517039) supported by the European Commission and by the CNRS, France. W. C. Sheets was supported additionally by a Chateaubriand postdoctoral fellowship. [99]{} V. M. Kaganer, B. Jenichen, F. Schippan, W. Braun, L. D$\ddot{a}$weritz and K. H. Ploog, Phys. Rev. Lett. **85**, 341 (2000). W. Prellier, P. Lecoeur and B. Mercey, J. Phys. Condens. Matter **13**, R915 (2001). A. Moreo, S. Yunoki and E. Dagotto, Science **283**, 2034 (1999). E. Dagotto, T. Hotta and A. Moreo, Phys. Rep. **344**, 1 (2001). A. Palanisami, M. P. Warusawithana, J. N. Eckstein, M. B. Weissman and N. D. Mathur, Phys. Rev. B **72**, 024454 (2005) S. A. Shivashankar and J. M. Honig, **28**, 5695 (1983). Y. Ueda, K. Kosuge and S. Kachi, J. Solid State Chem. **31**, 171(1980). D. B. McWhan, A. Menth, J. P. Remeika, W. F. Brinkman and T. M. Rice, Phys. Rev. B **7**, 1920 (1973). M. Yethiraj, J. Solid State Chem. **88**, 53 (1990). D. B. McWhan and T. M. Rice, Phys. Rev. Lett. **22**, 887 (1969). S. Autier-Laurent, B. Mercey, D. Chippaux, P. Limelette and Ch. Simon, Phys. Rev. B **74**, 195109 (2006). B. S. Allimi, M. Aindow, and S. P. Alpay, Appl. Phys. Lett. **93**, 112109 (2008). C. Grygiel, Ch. Simon, B. Mercey, W. Prellier, R. Frésard and P. Limelette, **91**, 262103 (2007). M. Mayr, A. Moreo, J. A. Verg$\acute{e}$s, J. Arispe, A. Feiguin, and E. Dagotto, Phys. Rev. Lett **86**, 135 (2000). G. T. Seidler, S. A. Solin and A. C. Marley, Phys. Rev. Lett. **76**, 3049 (1996). R. D. Merithew, M. B. Weissman, F. M. Hess, P. Spradling, E. R. Nowak, J. O’Donnell, J. N. Eckstein, Y. Tokura and Y. Tomioka, Phys. Rev. Lett. **84**, 3442 (2000). A. Palanisami, R. D. Merithew, M.  B. Weissman, Maitri P. Warusawithana, F. M. Hess and J. N. Eckstein, Phys. Rev. B **66**, 092407 (2002). S. Kirkpatrick, Rev. Mod. Phys. 45, 574 (1973). D. N. Nichols, R. J. Sladek and H. R. Harrison, Phys. Rev. B **24**, 3025 (1981). A. de Andrés, M. García-Hernández, and J. L. Martínez, Phys. Rev. B **60**, 7328 (1999). S. Taran, S. Karmakar, S. Chatterjee, B. K. Chaudhuri, C. P. Sun, C. L. Huang, and H. D. Yang, J. Appl. Phys. **99**, 073703 (2006). L. W. Finger and R. M. Hazen, J. Appl. Phys. **51**, 5362 (1980). Q. Luo, Q. Guo and E. G. Wang, Appl. Phys. lett. **84**, 2337 (2004). ![An optical microscopy image of one microbridge patterned (M2, width : 100[$\mu m$]{}) for a 230[Å]{} thick V$_2$O$_3$ film, where the scale represents 100[$\mu m$]{}. The microbridge includes two silver pads for supplying the current ($I^{+}$, $I^{-}$) and eight silver pads with five different lengths depending on the position of the voltage contacts ($V_1$, $V_2$, $V_3$ and $V_4$). Our measurements were made between $V_1$ and $V_3$, a distance corresponding to 200[$\mu m$]{}.[]{data-label="f1"}](fig1.eps){width="9cm"} ![Room temperature resistivity values ($\rho _{300K}$) versus the thickness ($t$) of unpatterned V$_2$O$_3$ films.[]{data-label="rho300"}](fig2.eps){width="9cm"} ![Linear resistivity as function of the temperature measured during cooling for unpatterned V$_2$O$_3$ films of certain thicknesses (noted $t$).[]{data-label="courbes-ech"}](fig3.eps){width="16cm"} ![The resistance, normalized over the 300K value, of a patterned 230[Å]{} thin film of V$_2$O$_3$ measured during cooling on two different microbridges (M1 : 20$\times$200$\mu m^2$, M2 : 100$\times$200$\mu m^2$). The arrows represent for M1 different transition temperatures observed. The inset is a close-up near the M-I transition and down to low temperatures.[]{data-label="f3"}](fig4.eps){width="9cm"} ![Resistance time series after supercooling for the insulating microbridge M1 of 230[Å]{} thick V$_2$O$_3$ film at 158.8[K]{}. The inset shows the resistance fluctuations of two states respectively at 160.4[K]{} and 160[K]{}.[]{data-label="noise"}](fig5.eps){width="9cm"} ![a) Normalized resistivity ratio $\rho(T)/\rho(300K)$ as function of temperature measured during cooling for three thicknesses. b) Simulated resistivity ratio ($\rho $/$\rho _{300K}$) versus temperature for the model of parallel resistances applied to certain metallic ratios $x$.[]{data-label="courbes-modele"}](fig6.eps){width="16cm"} ![Ratio $x$ of the metallic phase, extracted from the application of the parallel-resistance model, versus thickness of our V$_2$O$_3$ films. The hatching zone corresponds to the cross-over thickness, which is close to 200[Å]{}.[]{data-label="x(t)"}](fig7.eps){width="9cm"} ![a) : $c/a$ ratio of the lattice parameters versus thickness of V$_2$O$_3$ films at room temperature. The hatching zone corresponds to the cross-over thickness, close to 200[Å]{}. The doted line corresponds to the bulk value, whereas the solid line refers to the value for crystals subjected to a hydrostatic pressure of 47[kbar]{}.[@finger80] b) : Out-of-plane lattice parameter ($c$) versus in-plane lattice parameter ($a$) for the different films. The line indicates films following the Poisson law. On both curves, the surrounded points correspond to the same sample thickness ($t< 200$[Å]{}, see the text for details).[]{data-label="c/a"}](fig8.eps){width="9cm"}

--- abstract: 'Angular Momentum Loss is important for understanding astrophysical phenomena such as stellar rotation, magnetic activity, close binaries, and cataclysmic variables. Magnetic breaking is the dominant mechanism in the spin down of young late-type stars. We have studied angular momentum loss as a function of stellar magnetic activity. We argue that the complexity of the field and its latitudinal distribution are crucial for angular momentum loss rates. In this work we discuss how angular momentum is modulated by magnetic cycles, and how stellar spin down is not just a simple function of large scale magnetic field strength.' author: - 'Cecilia Garraffo, Jeremy J. Drake, and Ofer Cohen' title: Magnetic Modulation of Stellar Angular Momentum Loss --- INTRODUCTION ============ Stars spin-down through their magnetized winds that carry away mass and angular momentum (magnetic breaking). These winds corotate with them up to a certain distance, called the Alfvén radius ($R_A$), where the speed of the wind reaches the Alfvénic speed ($ u_A=B/\sqrt{4\pi \rho}$, being $B$ the magnetic field strength and $\rho$ the density of the plasma). The strength of the wind is expected to scale up with field strength and, therefore, stellar spin-down is a function of magnetic activity. @1967ApJ...148..217W provided the first analytical expression for magnetic breaking, $\dot{J}=\frac{2}{3} \Omega \dot{M}R_A^2 $, where $\dot{J}$ is the angular momentum loss rate, $\Omega$ the angular velocity, $\dot{M}$ the mass loss rate, and where spherical symmetry has been assumed (a split magnetic monopole was used and $R_A$ is constant). Recently, improved observational techniques like Zeeman Doppler Imaging have provided us with new information about the magnetic topology of almost a hundred stars . The fact that most of them show much more complex structure than that of a dipole has drawn a great amount of attention towards the role of magnetic topology in angular momentum loss. Since then, some authors have discussed the importance of topology for the winds structure and efficiency of magnetic braking [@2009ApJ...699.1501C; @2013ApJ...764...32G; @2014MNRAS.439.2122L; @2011ApJ...741...54C]. Both analytical and numerical models have been built aiming, with increasing success, to realistically study the interplay between spin-down rates and magnetic structure at the surface of stars [@1987MNRAS.226...57M; @1988ApJ...333..236K; @1995ApJ...446..435C; @1958ApJ...128..677P; @2012ApJ...754L..26M; @2011ApJ...741...54C; @2009ApJ...699.1501C; @2014ApJ...783...55C]. In this work we study the relevance of magnetic active regions and their latitudinal locations on stellar spin-down rates. Using a three dimensional magnetohydrodynamic (MHD) code we perform simulations to explore the effect of stellar spots on mass and angular momentum loss rates. NUMERICAL SIMULATION ==================== We use the generic [*BATS-R-US*]{} code [@1999JCoPh.154..284P; @2012JCoPh.231..870T] that solves the set of MHD equations for the conservation of mass, momentum, magnetic induction, and energy on a spherical, logarithmic in the $\hat{r}$ coordinate grid. We take advantage of the code’s Adaptive Mesh Refinement capabilities to refine the grid around regions where the magnetic field changes its sign. This way we better resolve current sheets that form during the simulation. We carry out simulations for a set of magnetic maps, starting with two pure dipoles (with field strengths of 10 and 20 Gauss at the poles) and superimposing the magnetic active regions observed for the sun near solar maximum (Carrington Rotation 1958). We place these two rings of spots at three different latitudes, using the same shift in latitude for both hemispheres, and with three different scalings for the spots strength. From the resulting solutions we compute the mass and angular momentum loss rates and study the spatial distribution of the contributing quantities (density and wind speed) at the Alfvén and stellar surfaces, as well as the shape and size of the Alfvén Surface itself. RESULTS AND DISCUSSION ====================== The general behaviour is ruled by the latitude of the magnetic active regions and is well described by two limiting cases: magnetograms with low-latitude and high-latitude spots (see figure \[fig:1\]). We group our results in these two categories that represent the general trend of the whole sample. In order to illustrate the qualitative effect of the magnetic active regions in each of these two regimes, in Fig \[fig:2\] we compare the mass (both at the Alfvén and at the stellar surfaces) and angular momentum losses as a function of latitude for a magnetogram with low-latitude spots and the same with the spots shifted towards high-latitudes. The transition latitude from one regime to the other is determined by the limiting latitude between open and closed field lines regions on the stellar surface of the pure dipolar solution. We find that, for all cases, there is a significant reduction of mass and angular momentum loss when magnetic active regions are located at high-latitudes compared to both the pure dipolar solution and the one with low-latitude spots. In general, the introduction of magnetic active regions within the “dead-zone” from which wind does not escape for the dipolar case (low-latitudes) does not change systematically the mass or angular momentum loss rates. However, any reallistic distribution of active regions will produce a small residual dipolar component as a result of the spatial separation of the spots. For rings of spots at low-latitudes, this competing dipole will be almost perpendicular to the rotation axis and original dipole, and will result in a tilt of the Alfvén surface. Consequently, the low-latitude active regions lead to a change in orientation of the Alfvén surface. In contrast, the introduction of active regions at high-latitudes leads to a significant reduction in both mass and angular momentum loss rates. The reason for this is that spots located in the open field lines regions of a star can efficiently couple to them and close them, therefore reducing the amount of wind carrying plasma away (see Fig \[fig:3\] for a qualitative plot). It is also worth noticing from Fig \[fig:2\] that most of the mass being lost is coming from mid-latitudes at the stellar surface and it funnels towards the equator when approaching the Alfvén surface. This makes sense because closed field lines cannot contribute to the mass loss mechanism. As a result, most of the torque experienced by the star will be located in a narrow band at mid latitudes, fundamentally at the transition latitude where the open field lines begin to appear and that that divides the low and high-latitude regimes. In summary, we find a bi-modal regime regulated by the latitude of spots by which magnetic active regions efficiently reduce mass and angular momentum loss rates [*only*]{} when located outside of the dead zone (the closed field line region on the stellar surface). The mechanism behind it is the closing of otherwise open field that leads to a reduction in the plasma being carried away (see Fig \[fig:3\]). As a result, magnetic cycles of stars whose spots cross the limiting latitude of the dead-zone, and therefore turn on and off this mechanism, will experience a large modulation of mass and angular momentum loss, by a typical factor of the order of a few (see Fig \[fig:2\]). We thank Jean-René Galarneau for providing Figure \[fig:3\]. CG was supported by [*Smithsonian Institution Consortium for Unlocking the Mysteries of the Universe*]{} grant” Lessons from Mars: Are Habitable Atmospheres on Planets around M Dwarfs Viable?” during the course of this research. JJD was supported by NASA contract NAS8-03060 to the [*Chandra X-ray Center*]{} (CXC) and thanks the CXC director, Belinda Wilkes, and the CXC science team for advice and support. OC was supported by the [*Smithsonian Institution Consortium for Unlocking the Mysteries of the Universe*]{} grant ”Lessons from Mars: Are Habitable Atmospheres on Planets around M Dwarfs Viable?” and by the [*Smithsonian Institute Competitive Grants Program for Science*]{} (CGPS) grant ”Can Exoplanets Around Red Dwarfs Maintain Habitable Atmospheres?.” The simulations were performed on the NASA HEC Pleiades system under award SMD-13-4526. [16]{} Chaboyer, B., Demarque, P., Guenther, D. B., & Pinsonneault, M. H. 1995, , 446, 435 Cohen, O., Drake, J. J., Kashyap, V. L., & Gombosi, T. I. 2009, , 699, 1501 Cohen, O., & Drake, J. J. 2014, , 783, 55 Cranmer, S. R., & Saar, S. H. 2011, , 741, 54 Donati, J.-F., & Brown, S. F. 1997, , 326, 1135 Donati, J.-F., & Landstreet, J. D. 2009, , 47, 333 Garraffo, C., Cohen, O., Drake, J. J., & Downs, C. 2013, , 764, 32 Kawaler, S. D. 1988, , 333, 236 Lang, P., Jardine, M., Morin, J., et al. 2014, , 439, 2122 Matt, S. P., MacGregor, K. B., Pinsonneault, M. H., & Greene, T. P. 2012, , 754, L26 Mestel, L., & Spruit, H. C. 1987, , 226, 57 Morin, J., Donati, J.-F., Petit, P., et al. 2010, , 407, 2269 Parker, E. N. 1958, , 128, 677 Powell, K. G., Roe, P. L., Linde, T. J., Gombosi, T. I., & De Zeeuw, D. L. 1999, Journal of Computational Physics, 154, 284 T[ó]{}th, G., van der Holst, B., Sokolov, I. V., et al. 2012, Journal of Computational Physics, 231, 870 Weber, E. J., & Davis, L., Jr. 1967, , 148, 217 ![Typical stellar magnetograms used in our simulations for magnetic active regions at low (top) and high (bottom) latitudes[]{data-label="fig:1"}](fig1){width="\textwidth"} ![Mass loss and angular momentum loss as a function of latitude for low-latitude (red) and high-latitude (green) magnetic active regions at the Alfvén Surface (solid line) and at the stellar surface (dashed line).[]{data-label="fig:2"}](fig2){width="\textwidth"} ![Qualitative plot of the wind structure for a dipole (left), with the limiting latitude between open and closed field line regions shown in white. The wind structure for the same dipole with low-latitude (center) and high-latitude (right) magnetic active regions. The addition of spots at fairly low latitudes—within the “dead zone” where field is already closed—makes little difference to the wind morphology or mass and angular momentum loss rates, and only results in a tilt of the magnetic axis. Adding spots at higher latitudes closes field lines and quenches mass and angular momentum loss.[]{data-label="fig:3"}](fig3){width="\textwidth"}

--- abstract: 'We have operated a quantum point contact (QPC) charge detector in a radio frequency (RF) mode that allows fast charge detection in a bandwidth of tens of megahertz. We find that the charge sensitivity of the RF-QPC is limited not by the noise of a secondary amplifier, but by non-equilibrium noise [$S_{I}$]{} of the QPC itself. We have performed frequency-resolved measurements of the noise within a [$10\:{\ensuremath{\mathrm{MHz}}}$]{} bandwidth around our carrier wave. When averaged over our bandwidth, we find that [$S_{I}$]{} is in good agreement with the theory of photon-assisted shot noise. Our measurements also reveal strong frequency dependence of the noise, asymmetry with respect to the carrier wave, the appearance of sharp local maxima that are correlated with mechanical degrees of freedom in the sample, and noise suppression indicative of many-body physics near the 0.7 structure.' author: - Madhu Thalakulam - 'W. W. Xue' - Feng Pan - 'Z. Ji' - 'J. Stettenheim' - Loren Pfeiffer - 'K. W. West' - 'A. J. Rimberg' title: 'Shot-Noise-Limited Operation of a Fast Quantum-Point-Contact Charge Sensor' --- All measurements, including electrical amplification, are subject to quantum mechanical limits [@Caves:1982; @Devoret:2000]: a standard measurement of a quantum system must add noise with a strictly determined minimal size. To reach this quantum limit, the output noise of a measurement system must be dominated by the intrinsic noise of an initial quantum amplifier and not that of a subsequent classical one [@Clerk:2004b; @Korotkov:2003]. This requires that shot noise arising from the flow of current through the quantum amplifier dominates the measurement system noise. Here, we report shot-noise limited operation of a quantum-point-contact (QPC) charge sensor in a radio-frequency (RF) mode analogous to that used for single electron transistors [@Schoelkopf:1998]. QPCs, one of the simplest nanoscale systems, are surprisingly complex. Recently, study of QPC has been focused on two areas in particular. First, there is a strong interaction between electronic and mechanical degrees of freedom in GaAs-based QPCs, allowing both detection of mechanical resonances using a QPC as a detector [@Cleland:2002] and synchronized transport of electrons through QPCs in the tunneling regime [@Shilton:1996]. Second, there is both experimental [@Expt07; @Cronenwett:2002] and theoretical [@Theory07] evidence of the formation of a many-body magnetic impurity state in QPCs that manifests itself as an anomalous plateau in the QPC conductance at ${\ensuremath{G_{\text{QPC}}}}\approx 0.7{\ensuremath{G_{0}}}$ where ${\ensuremath{G_{0}}}=2e^{2}/h$. In this Letter, we use frequency-resolved measurements of shot noise [@Shotnoise; @Shotnoise07] in a heretofore unexplored limit to characterize our RF-QPCs. We find the shot noise in the vicinity of the carrier wave frequency [$f_{0}$]{} shows surprising frequency dependence and reflects both the physics of the 0.7 structure and the interplay between vibrational and electronic degrees of freedom. Coupling of electronic and mechanical degrees of freedom and the presence of a local moment in a QPC do not appear to have been considered previously with regard to its potential as a quantum limited charge detector. Our measurements of the intrinsic noise and charge sensitivity of an RF-QPC charge detector lie at the intersection of these three areas of investigation. ![\[fig1\] (Color online) (a) Schematic diagram of the measurement circuit. The electron micrograph shows the QPC geometry; only one constriction is used in the measurements. The RF carrier wave is applied via a directional coupler, which also directs the reflected wave to a cryogenic HEMT amplifier with noise temperature [$2.3\:{\ensuremath{\mathrm{K}}}$]{} followed by a GaAs FET amplifier at room temperature. There was [$48\:{\ensuremath{\mathrm{dB}}}$]{} of attenuation in the input RF lines. All dc lines passed through cascaded $\pi$-type, $RC$ and microwave filters. A circulator between the tank circuit and the HEMT amplifier isolated the sample from noise sources on the output line. (b) [$G_{\text{QPC}}$]{}  versus gate voltage [$V_{g}$]{} at zero magnetic field and $T={\ensuremath{25\:{\ensuremath{\mathrm{mK}}}}}$. Inset: [$G_{\text{QPC}}$]{}  after exposure of the sample to light, showing multiple conductance plateaus. (c) Nonlinear conductance ${\ensuremath{G_{\text{QPC}}}}({\ensuremath{V_{\text{dc}}}})$. Measurements were performed for a series of values of [$V_{g}$]{} with spacing $\Delta{\ensuremath{V_{g}}}={\ensuremath{1\:{\ensuremath{\mathrm{mV}}}}}$ and plotted without offset. The vertical dashed lines indicate the estimated rms rf voltage applied to the QPC for subsequent noise measurements. ](Fig1.eps){width="7.5cm"} Our QPCs were formed via the split gate technique in a [GaAs/AlGaAs]{}heterostructure containing a 2DEG with sheet density $n_{s}={\ensuremath{1.3{\ensuremath{\times 10^{11}}}\:{\ensuremath{\mathrm{cm^{-2}}}}}}$ and mobility $\mu={\ensuremath{7.4{\ensuremath{\times 10^{6}}}\:{\ensuremath{\mathrm{cm^{2}V^{-1}s^{-1}}}}}}$ located [$100\:{\ensuremath{\mathrm{nm}}}$]{} beneath the heterostructure surface. We fabricated two samples, A and B. Except where noted all data shown is from sample A; results for sample B were similar. Measurements were performed in a dilution refrigerator with a base temperature of $T={\ensuremath{25\:{\ensuremath{\mathrm{mK}}}}}$ and effective electron temperature $T_{e}\approx{\ensuremath{80\:{\ensuremath{\mathrm{mK}}}}}$ . The QPCs were imbedded in an $LC$ tank circuit consisting of a Nb spiral chip inductor with $L={\ensuremath{140 (125)\:{\ensuremath{\mathrm{nH}}}}}$ for sample A (B), parasitic capacitance ${\ensuremath{C_{p}}}={\ensuremath{0.28 (0.25)\:{\ensuremath{\mathrm{pF}}}}}$ and resonant frequency ${\ensuremath{f_{0}}}=1/2 \pi\sqrt{LC}={\ensuremath{800 (900)\:{\ensuremath{\mathrm{MHz}}}}}$. A bias-tee in our rf circuitry \[Fig. \[fig1\](a)\] allowed application of an RF ([$V_{\text{rf}}$]{}) signal for microwave reflectometry measurement of the QPC charge sensitivity and noise [@Schoelkopf:1998; @Lu:2003] and near-dc voltages ([$v_{\text{ac}}$]{} and [$V_{\text{dc}}$]{}) for lockin measurements of the QPC conductance [$G_{\text{QPC}}$]{}. Conductance data for our QPCs (${\ensuremath{v_{\text{ac}}}}={\ensuremath{20\:{\ensuremath{\mathrm{\mu V}}}}}$ rms at [$13\:{\ensuremath{\mathrm{Hz}}}$]{}) show well-defined plateaus in [$G_{\text{QPC}}$]{} versus the voltage [$V_{g}$]{} applied to the split gates \[Fig. \[fig1\](b)\]. Application of a dc voltage [$V_{\text{dc}}$]{} allowed measurement of nonlinear differential conductance ${\ensuremath{G_{\text{QPC}}}}({\ensuremath{V_{\text{dc}}}})$ versus both [$V_{g}$]{} and [$V_{\text{dc}}$]{}. For $T<{\ensuremath{500\:{\ensuremath{\mathrm{mK}}}}}$, we observed a peak in [$G_{\text{QPC}}$]{} around ${\ensuremath{V_{\text{dc}}}}=0$ for QPC conductance in the range $0<{\ensuremath{G_{\text{QPC}}}}<{\ensuremath{G_{0}}}$ \[Fig. \[fig1\](c)\]. This zero-bias anomaly (ZBA) has been studied previously [@Cronenwett:2002] and interpreted as an indication of the onset of Kondo physics in the QPC [@Theory07], as has an additional plateau at finite bias (${\ensuremath{V_{\text{dc}}}}\approx{\ensuremath{700\:{\ensuremath{\mathrm{\mu V}}}}}$) for which ${\ensuremath{G_{\text{QPC}}}}\approx 0.8{\ensuremath{G_{0}}}$ [@Cronenwett:2002]. These measurements of ${\ensuremath{G_{\text{QPC}}}}({\ensuremath{V_{\text{dc}}}})$ provide clear evidence that the physics associated with the 0.7 structure is present in our QPCs and are indicative of their high quality. To operate our QPC as a charge detector, we tuned [$V_{g}$]{} to maximize $d{\ensuremath{G_{\text{QPC}}}}/d{\ensuremath{V_{g}}}$ (typically ${\ensuremath{G_{\text{QPC}}}}\approx 0.5{\ensuremath{G_{0}}}$) at ${\ensuremath{V_{\text{dc}}}}=0$ and applied an rf carrier wave ${\ensuremath{V_{\text{rf}}}}\cos{\ensuremath{\omega_{0}}}t$ where ${\ensuremath{\omega_{0}}}=2\pi{\ensuremath{f_{0}}}$ to the tank circuit. Some portion $\Gamma{\ensuremath{V_{\text{rf}}}}\cos{\ensuremath{\omega_{0}}}t$ of the wave is reflected (the reflection coefficient $\Gamma$ of the tank circuit depends on [$G_{\text{QPC}}$]{}) and is measured at the output of our amplifier chain [@Schoelkopf:1998; @Lu:2003]. The RF-QPC bandwidth is determined by the width $\Delta f ={\ensuremath{f_{0}}}/Q\approx{\ensuremath{60\:{\ensuremath{\mathrm{MHz}}}}}$ of the tank circuit resonance, allowing very fast charge detection. For a QPC coupled to a quantum dot, an electron tunneling event typically changes [$G_{\text{QPC}}$]{} by 1–3% [@Vandersypen:2004]. To mimic this effect, we apply a small ac voltage [$v_{m}$]{} \[Fig. \[fig1\](a)\] at [$97\:{\ensuremath{\mathrm{kHz}}}$]{} to one QPC gate so that $\Delta{\ensuremath{G_{\text{QPC}}}}/{\ensuremath{G_{\text{QPC}}}}\approx 2.7\%$. The RF-QPC output shows side peaks at ${\ensuremath{f_{0}}}\pm{\ensuremath{97\:{\ensuremath{\mathrm{kHz}}}}}$ indicative of amplitude modulation riding on a broad noise background \[right inset, Fig. \[fig2\](a)\]. We estimate the charge sensitivity of the QPC to be ${\ensuremath{\delta q}}\approx{\ensuremath{5{\ensuremath{\times 10^{-4}}}\:{\ensuremath{e/\sqrt{\mathrm{Hz}}}}}}$ referred to a hypothetical quantum dot. ![\[fig2\] (Color online) (a) Output power spectrum [$P_{n}$]{} of an RF-QPC (sample A) for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-98\:{\ensuremath{\mathrm{dBm}}}}}$ (dashed green) and [$-88\:{\ensuremath{\mathrm{dBm}}}$]{} (solid blue) and HEMT noise floor [$P_{n}^{A}$]{}. Both [$P_{\text{in}}$]{} and [$P_{n}$]{} are referred to the input of the HEMT amplifier. Right inset: output of the RF-QPC subject to conductance modulation and HEMT noise floor [$P_{n}^{A}$]{}. Measurement is for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-57\:{\ensuremath{\mathrm{dBm}}}}}$ and ${\ensuremath{G_{\text{QPC}}}}\approx 0.5{\ensuremath{G_{0}}}$. Left inset: sample A geometry. (b) [$P_{n}$]{} for sample B. Inset: sample B geometry. (c) PS.  (squares) and [$\mathcal{P}_{n}^{E}$]{}  (circles) versus [$P_{\text{in}}$]{}. Dashed lines are guides to the eye that scale as [$P_{\text{in}}$]{}  and ${\ensuremath{P_{\text{in}}}}^{1/2}$. (d) Signal-to-noise ratio for the RF-QPC on a linear scale versus conductance modulation.](Fig2.eps){width="7.5cm"} Two aspects of the noise limiting the QPC sensitivity are striking: first, it is larger than the noise [$P_{n}^{A}$]{} of the HEMT amplifier, which usually limits the performance of RF-SETs; second, it is frequency-dependent rather than white. To investigate, we measure the spectrum of reflected noise power [$P_{n}$]{} in a [$10\:{\ensuremath{\mathrm{MHz}}}$]{} bandwidth around [$f_{0}$]{} for different values of the input power [$P_{\text{in}}$]{} and with no conductance modulation \[Fig. \[fig2\](a)\]. In addition to broadband noise that decreases away from [$f_{0}$]{}, there are large peaks in [$P_{n}$]{}at ${\ensuremath{f_{0}}}\pm{\ensuremath{580\:{\ensuremath{\mathrm{kHz}}}}}$. For ${\ensuremath{P_{\text{in}}}}={\ensuremath{-98\:{\ensuremath{\mathrm{dBm}}}}}$ the broadband noise is clearly visible and the peaks are relatively small; for larger input power ${\ensuremath{P_{\text{in}}}}={\ensuremath{-88\:{\ensuremath{\mathrm{dBm}}}}}$ the broadband noise decreases while the peaks at ${\ensuremath{f_{0}}}\pm{\ensuremath{580\:{\ensuremath{\mathrm{kHz}}}}}$ become more pronounced. [$P_{n}$]{} for sample B shows similar peaks but a more complex spectrum. Since the measured noise [$P_{n}$]{} depends on [$P_{\text{in}}$]{}  (and on [$G_{\text{QPC}}$]{}, see below), it is associated with the sample. There are two broad categories into which such noise might fall: modulation noise, for which the current through the QPC is amplitude modulated; and shot noise [@Kogan:1996]. Modulation noise scales with input power as ${\ensuremath{P_{n}}}\propto{\ensuremath{P_{\text{in}}}}$ whatever its origin, whether motion of trapped charges in the substrate, electromagnetic noise coupled to the QPC gates, mixing due to the QPC nonlinearity, or some other source. Shot noise, in contrast, scales as ${\ensuremath{P_{\text{in}}}}^{1/2}$. In our experiment shot noise arises from the partition noise of electron-hole pairs created by the RF voltage [$v_{\text{rf}}^{\text{QPC}}$]{} across the QPC [@Reydellet:2003]; for an ideal matching network this (rms) voltage is given by ${\ensuremath{v_{\text{rf}}^{\text{QPC}}}}=2Q{\ensuremath{V_{\text{rf}}}}=2Q\sqrt{{\ensuremath{P_{\text{in}}}}{\ensuremath{Z_{0}}}}$. Such “photon assisted” shot noise (PASN) has been examined theoretically [@PASNTheory] and measured both in normal metals [@Schoelkopf:1998a] and QPCs [@Reydellet:2003]. Previous work has studied PASN at a frequency $\omega$ much less than the drive frequency [$\omega_{0}$]{}. Here, we measured PASN for $\omega\approx{\ensuremath{\omega_{0}}}$. Assuming energy-independent transmission coefficients [$T_{n}$]{} it can be shown that the spectral density of photon-assisted shot noise is given by ${\ensuremath{S_{I}}}(\omega ,{\ensuremath{\omega_{0}}}) = \frac{{4e^2 }}{h}\sum\limits_n {{\ensuremath{T_{n}}}(1 - {\ensuremath{T_{n}}})} \sum\limits_{l = - \infty }^\infty {(\hbar \omega + l\hbar {\ensuremath{\omega_{0}}})\,J_l^2 (\alpha )\coth \left[ {\frac{{\hbar \omega + l\hbar {\ensuremath{\omega_{0}}}}}{{2k_B T}}} \right]}$ where $\alpha=\sqrt{2}e{\ensuremath{v_{\text{rf}}^{\text{QPC}}}}/\hbar{\ensuremath{\omega_{0}}}$. For low temperature and $\alpha\gg 1$ the infinite sum can be evaluated easily and scales as $\alpha\propto{\ensuremath{P_{\text{in}}}}^{1/2}$. In addition to shot noise, [$P_{n}$]{} includes contributions [$P_{n}^{T}$]{}  from thermal noise and [$P_{n}^{A}$]{} from the HEMT amplifier that we account for by extracting the excess noise ${\ensuremath{P_{n}^{E}}}={\ensuremath{P_{n}}}-{\ensuremath{P_{n}^{T}}}-{\ensuremath{P_{n}^{A}}}$ from our raw data. The prediction for ${\ensuremath{S_{I}}}(\omega,{\ensuremath{\omega_{0}}})$ above allows us to determine the origin of the excess noise [$P_{n}^{E}$]{} by measuring its dependence on [$P_{\text{in}}$]{} and ${\ensuremath{G_{\text{QPC}}}}={\ensuremath{G_{0}}}\sum_{n}{\ensuremath{T_{n}}}$. We varied [$P_{\text{in}}$]{}  over a six decade range, and measured both the power PS.  in a charge modulation signal and the integrated excess noise ${\ensuremath{\mathcal{P}_{n}^{E}}}=\int{\ensuremath{P_{n}^{E}}}df$ in a 4.8 MHz bandwidth above [$f_{0}$]{} (with no charge modulation). We find ${\ensuremath{P_{s}}}\propto{\ensuremath{P_{\text{in}}}}$ over a range of three decades in [$P_{\text{in}}$]{} before the RF-QPC response begins to saturate \[Fig. \[fig2\](c)\]. The linearity of the RF-QPC in this range is excellent: the SNR for the modulation signal rises linearly with increasing $\Delta{\ensuremath{G_{\text{QPC}}}}/{\ensuremath{G_{\text{QPC}}}}$ up to $\Delta{\ensuremath{G_{\text{QPC}}}}/{\ensuremath{G_{\text{QPC}}}}=15$% \[Fig. \[fig2\](d)\]. In contrast [$\mathcal{P}_{n}^{E}$]{}  scales as ${\ensuremath{P_{\text{in}}}}^{1/2}$ over a nearly five decade range, eliminating modulation noise as the source of [$P_{n}^{E}$]{}. We also measured [$\mathcal{P}_{n}^{E}$]{}  versus [$G_{\text{QPC}}$]{} over the range $0<{\ensuremath{G_{\text{QPC}}}}<2{\ensuremath{G_{0}}}$ for two different values of [$P_{\text{in}}$]{}, corresponding to [$v_{\text{rf}}^{\text{QPC}}$]{} indicated by the dashed lines in Fig. \[fig1\](c). [$P_{n}^{E}$]{} vanishes for ${\ensuremath{G_{\text{QPC}}}}\approx 0$, is maximal for ${\ensuremath{G_{\text{QPC}}}}\approx 0.5{\ensuremath{G_{0}}}$, and vanishes again for ${\ensuremath{G_{\text{QPC}}}}=1.0{\ensuremath{G_{0}}}$ \[Fig. \[fig3\](a)–(c)\]. A more detailed set of measurements \[Fig. \[fig3\](d)\] confirms that the magnitude of [$\mathcal{P}_{n}^{E}$]{} is well described by the shot noise ${\ensuremath{S_{I}}}(\omega,{\ensuremath{\omega_{0}}})$ integrated over the same bandwidth and converted to voltage noise [@Korotkov:1999] by the tank circuit \[Fig. \[fig3\](d), dashed lines\]. Interestingly, [$\mathcal{P}_{n}^{E}$]{} is noticeably suppressed for [$G_{\text{QPC}}$]{} in the vicinity of $0.7{\ensuremath{G_{0}}}$, in agreement with recent measurements of dc shot noise in QPCs [@Shotnoise07]. These observations, combined with the scaling as ${\ensuremath{P_{\text{in}}}}^{1/2}$ described earlier, conclusively identify shot noise as the source of the excess noise [$P_{n}^{E}$]{}. ![\[fig3\] (Color online) (a)–(c) [$P_{n}$]{} for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-88\:{\ensuremath{\mathrm{dBm}}}}}$ and ${\ensuremath{G_{\text{QPC}}}}\approx 0$ (a), $0.5{\ensuremath{G_{0}}}$ (b), and [$G_{0}$]{}  (c). (d) [$\mathcal{P}_{n}^{E}$]{} versus [$G_{\text{QPC}}$]{}  for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-88\:{\ensuremath{\mathrm{dBm}}}}}$ (top) and[$-93\:{\ensuremath{\mathrm{dBm}}}$]{} (bottom), corresponding to ${\ensuremath{V_{\text{rf}}}}^{\text{QPC}}={\ensuremath{230\:{\ensuremath{\mathrm{\mu V}}}}}$ and [$130\:{\ensuremath{\mathrm{\mu V}}}$]{} respectively, as indicated relative to the QPC $IV$ characteristics in Fig. \[fig1\](c) by the vertical dashed lines. There is no rise in the noise floor for ${\ensuremath{G_{\text{QPC}}}}\approx{\ensuremath{G_{0}}}$ versus ${\ensuremath{G_{\text{QPC}}}}\approx 0$ \[compare (c) and (a)\], indicating that there is no significant sample heating for these input powers. To compare with theory, we integrated the predicted PASN power over the same bandwidth as for [$\mathcal{P}_{n}^{E}$]{} and converted to noise power at the HEMT amplifier (dashed lines). Current noise ${\ensuremath{S_{I}}}(\omega,{\ensuremath{\omega_{0}}})$ in the QPC is transformed by an ideal $LC$ matching network into noise power $(2L/{\ensuremath{C_{p}}}{\ensuremath{Z_{0}}}){\ensuremath{S_{I}}}(\omega,{\ensuremath{\omega_{0}}})$ at the input to the HEMT amplifier, where ${\ensuremath{Z_{0}}}={\ensuremath{50\:{\ensuremath{\mathrm{\Omega}}}}}$ is the impedance of the coaxial cable connecting it to the tank circuit. The results were shifted downward by [$3.9\:{\ensuremath{\mathrm{dB}}}$]{} but no other fitting parameter was used. The reduction of the measured noise relative to theory is likely due to losses in the matching network. ](Fig3.eps){width="7.5cm"} In contrast to the calculated ${\ensuremath{S_{I}}}(\omega,{\ensuremath{\omega_{0}}})$, [$P_{n}^{E}$]{} depends strongly on $\omega$. It is not uncommon, however, for noise to show spectral features corresponding to physical excitations of a system [@Balatsky:2006]. We hypothesize that a surface acoustic wave (SAW) with a half-wavelength equal to a typical sample dimension $\ell$ is excited in the piezoelectric GaAs substrate by the rf drive and take the SAW frequency to be $f_{\ell}=\beta {\ensuremath{v_{s}}}/2\ell$ where ${\ensuremath{v_{s}}}={\ensuremath{3010\:{\ensuremath{\mathrm{m/s}}}}}$ is the speed of sound in GaAs and $\beta=1.05$ is a scaling parameter. We expect the SAW to produce features in [$P_{n}$]{} at ${\ensuremath{f_{0}}}\pm f_{\ell}$. For sample A there is only one relevant length scale $\ell_{a}$ \[left inset, Fig. \[fig2\](a)\] for which $f_{a}\approx{\ensuremath{580\:{\ensuremath{\mathrm{kHz}}}}}$. Agreement of $f_{a}$ with the offset of the noise peaks in Fig. \[fig2\](a) from [$f_{0}$]{} is remarkable. For sample B there are three relevant length scales \[inset, Fig. \[fig2\](b)\], $\ell_{b1}$ ([$2.8\:{\ensuremath{\mathrm{mm}}}$]{}), $\ell_{b2}$ ([$1.6\:{\ensuremath{\mathrm{mm}}}$]{}) and $\ell_{b3}$ ([$0.8\:{\ensuremath{\mathrm{mm}}}$]{}) with corresponding frequencies $f_{b1}$ ([$560\:{\ensuremath{\mathrm{kHz}}}$]{}), $f_{b2}$ ([$990\:{\ensuremath{\mathrm{kHz}}}$]{}) and $f_{b3}$ ([$1.98\:{\ensuremath{\mathrm{MHz}}}$]{}). For each $f_{bi}$ there is a broad peak in [$P_{n}$]{} at ${\ensuremath{f_{0}}}\pm f_{bi}$ that scales as ${\ensuremath{P_{\text{in}}}}^{1/2}$, providing strong evidence of coupling between shot noise and mechanical degrees of freedom in our RF-QPCs. The peak marked by the asterisk scales as [$P_{\text{in}}$]{} identifying it as modulation noise. For [$P_{\text{in}}$]{} used in Fig. \[fig3\] [$v_{\text{rf}}^{\text{QPC}}$]{} was sufficiently large to drive the QPC away from the ZBA. However, we were able to measure [$P_{n}$]{} near ${\ensuremath{G_{\text{QPC}}}}\approx 0.5{\ensuremath{G_{0}}}$ for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-103\:{\ensuremath{\mathrm{dBm}}}}}$ for which [$v_{\text{rf}}^{\text{QPC}}$]{} lies entirely within the ZBA at a series of temperatures \[Fig. \[fig4\]\]. Interestingly, the broadband noise is noticeably asymmetric with respect to [$f_{0}$]{} \[Fig. \[fig4\](b)\] for $T<{\ensuremath{500\:{\ensuremath{\mathrm{mK}}}}}$. As the temperature is raised, the broadband noise both weakens and becomes more symmetric, so that for $T>{\ensuremath{500\:{\ensuremath{\mathrm{mK}}}}}$ it has nearly vanished. In contrast, the peaks at ${\ensuremath{f_{0}}}\pm{\ensuremath{580\:{\ensuremath{\mathrm{kHz}}}}}$ are clearly visible for $T={\ensuremath{1\:{\ensuremath{\mathrm{K}}}}}$, suggesting different physical origins for the two phenomena. Note that the ZBA has a temperature dependence similar to that of the broadband noise, weakening rapidly for temperatures above [$115\:{\ensuremath{\mathrm{mK}}}$]{} and nearly vanishing for $T>{\ensuremath{550\:{\ensuremath{\mathrm{mK}}}}}$. The asymmetry in [$P_{n}$]{}  also vanishes when [$v_{\text{rf}}^{\text{QPC}}$]{} is far out of the ZBA: in Fig. \[fig2\](a) some asymmetry is visible in [$P_{n}$]{}  for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-98\:{\ensuremath{\mathrm{dBm}}}}}$ but not for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-88\:{\ensuremath{\mathrm{dBm}}}}}$. Similar dependence on $T$ and [$P_{\text{in}}$]{} for the broadband noise and the ZBA suggest they may be related; further experiments are needed for a conclusive demonstration. In conclusion, we have operated an RF-QPC at the shot noise limit. The noise both shows coupling to mechanical degrees of freedom in the sample and reflects the many-body physics of the 0.7 structure. Our results suggest that such phenomena may have important implications for the ultimate charge sensitivity of the QPC and how nearly it can approach the quantum limit. Our results have immediate implications for study of spin-based quantum information processing in quantum dots [@Petta:2005; @Koppens:2006]. The techniques employed here may also be applicable to studies of noise in other semiconductor devices such as quantum dots in the Kondo regime. ![\[fig4\] (Color online) (a) ${\ensuremath{G_{\text{QPC}}}}({\ensuremath{V_{\text{dc}}}})$ near $0.5{\ensuremath{G_{0}}}$ versus [$V_{\text{dc}}$]{} for $T=25$ to [$550\:{\ensuremath{\mathrm{m K}}}$]{}. The size of the rms rf bias ${\ensuremath{V_{\text{rf}}}}^{\text{QPC}}={\ensuremath{40\:{\ensuremath{\mathrm{\mu V}}}}}$ for the noise measurements in (b) is indicated by the yellow shaded region around ${\ensuremath{V_{\text{dc}}}}=0$. Inset: differential conductance versus gate voltage for $T=25$ (solid) and [$550\:{\ensuremath{\mathrm{mK}}}$]{} (dashed). (b) Noise power [$P_{n}$]{}  frequency for $T = 225$ to [$1000\:{\ensuremath{\mathrm{mK}}}$]{}, near ${\ensuremath{G_{\text{QPC}}}}\approx 0.5{\ensuremath{G_{0}}}$ and for ${\ensuremath{P_{\text{in}}}}={\ensuremath{-103\:{\ensuremath{\mathrm{dBm}}}}}$. Successive curves are offset by [$5\:{\ensuremath{\mathrm{dB}}}$]{} for clarity. Peaks at ${\ensuremath{f_{0}}}=\pm{\ensuremath{250\:{\ensuremath{\mathrm{kHz}}}}}$ scale as [$P_{\text{in}}$]{} and are due to modulation noise. ](Fig4.eps){width="7.5cm"} This work was supported by the NSF under Grant No. DMR-0454914, by the ARO under Agreement No. W911NF-06-1-0312 and by the NSA, LPS and ARO under Agreement No. W911NF-04-1-0389. We thank M. Blencowe for many helpful conversations. [22]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , ****, (). , ****, (). , ****, (). , *et al.*, ****, (). , , , , ****, (). , *et al.*, ****, (). , *et al.*, ****, (); , *et al.*, , ****, (); , *et al.*, ****, (). , *et al.*, ****, (). , , , ****, (); , , , ****, (); , ****, (); , , , ****, (). , ****, (); , ****, (); , , , , ****, (); , *et al.*, ****, (). , *et al.*, ****, (); , *et al.*, ****, (). , *et al.*, ****, (). , *et al.*, ****, (). , ** (, , ). , *et al.*, ****, (). , ****, ().; , ****, (). , , , , ****, (). , ****, (). , , , , ****, (). , *et al.*, ****, (). , *et al.*, ****, ().

--- abstract: 'In this article we study the structure and stability of compact astrophysical objects which are ruled by the dark energy equation of state (EoS). The existence of dark energy is important for explaining the current accelerated expansion of the universe. Exact solutions to Einstein field equations (EFE) have been found by considering particularized metric potential, Finch and Skea ansatz [@Finch]. The obtained solutions are relevant to the explanation of compact fluid sphere. Further, we have observed at the junction interface, the interior solution is matched with the Schwarzschild’s exterior vacuum solution. Based on that, we have noticed the obtained solutions are well in agreement with the observed maximum mass bound of $\approx$ 2$M_{\bigodot}$, namely, PSR J1416-2230, Vela X-1, 4U 1608-52, Her X-1 and PSR J1903+327, whose predictable masses and radii are not compatible with the standard neutron star models. Also, the stability of the stellar configuration has been discussed briefly, by considering the energy conditions, surface redshift, compactness, mass-radius relation in terms of the state parameter ($\omega$). Finally, we demonstrate that the features so obtained are physically acceptable and consistent with the observed/reported data [@Gondek-Rosinska; @Glendenning]. Thus, the present dark energy equation of state appears talented regarding the presence of several exotic astrophysical matters.' author: - Ayan Banerjee - 'M. K. Jasim' - Anirudh Pradhan title: Analytical model of dark energy stars --- Introduction ============ The present accelerated expansion of the universe has been established by various self-governing high-precision observational data such as the galaxy rotation curves, Supernovae type Ia [@Grant; @Perlmutter] and cosmic microwave background radiation [@Bennett; @Hinshaw]. Several models have been proposed to account for this observed late-time accelerated expansion of the Universe. One of the assumptions of the current paradigm of cosmology is the existence of so-called dark energy, where general relativity is assumed to be correct. The dark energy component is characterized by a negative pressure. A large volume of phenomenological and theoretical models have been proposed to describe the actual situation (see [@Copeland] and references therein). In this regard, the simplest and most efficient way to describe the present situation is by introducing a cosmological constant $\Lambda$ within general relativity (GR). Overall, DE is thought to contribute 70% of the worldwide energy budget in the universe. See Ref [@Li:2011sd] for a recent review on DE models. However, despite its success, $\Lambda$CDM model, which consists of a cosmological constant ($\Lambda$) plus Cold Dark Matter (CDM), suffers the theoretical problems for instance the *fine-tuning* and *cosmic happenstance* mysteries. So, it is important to investigate other available theoretical schemes such as employing an EoS p = $\omega \rho$, where $p$ is the pressure and $\rho$ is the energy density, respectively. Without entering further details, the requirement for cosmic expansion is $\omega <-1/3$ , as shown through Friedmann equation $\ddot{a}/a$ = -$4 \pi G(p + \rho/3)$, whereas $\omega=-1$ reduces to the particular case of cosmological constant. On the other hand, the parameter range $-1< \omega<-1/3$ is called quintessence models and the dark energy is reducing accordingly with a scale feature $a(t)$ as $\rho_Q \equiv a^{-3(1+\omega)}$ [@Turner]. In [@Alam:2003sc], authors have shown that depending on the evolution of density perturbations one can distinguish a cosmological constant ($\omega=-1$) from quintessence models with $\omega \geqslant - 0.9$ at the 3$\sigma$ level. It is therefore the equation of state parameter violates the null energy condition, i.e. $\omega_{de}< -1$. However, the EoS of a phantom scalar field is always less than $-1$. Our interest is to search for local astrophysical objects within this scenario. In this line of thought, gravastar (gravitational vacuum star) model received special attention - as an alternative to a black hole solution has been proposed by Mazur and Mottola [@Mazur]. This model consist of five layers including two thin–shells. The de-Sitter geometry in the interior governed by an EoS p = -$\rho$, matches to an exterior Schwarzschild solution. It is argued that in between the interior and exterior geometry, there is a finite thickness shell with an equation of state p = +$\rho$. The shell is comprised of stiff fluid matter. See Ref [@Visser] for a recent development on gravastar model where five-layer model can be simplified to three-layer with the phase transition layer was replaced by a single spherical $\delta$-shell. In Ref [@Bilic:2005sn] an argument claiming for alternative model of gravastar where the de-Sitter regime was replaced by an interior solution governed by a Chaplygin gas equation of state, interpreted as a Born-Infeld phantom gravastar. We refer to our reader for a recent discussion see e.g., [@Wiltshire]. Apart from the aforementioned gravastar solution, one may consider extended objects supported by dark energy EoS (interior region). The motivation comes from the fact that dark energy exerts a repulsive force on its surrounding, and this repulsive force may prevent the star from collapsing. This model found great success and denoted as a *dark energy star* [@GChapline]. After this work, Lobo [@lobo1] generalized of the gravastar picture by considering the dark energy EoS, $\omega= p/ \rho <-1/3$, which is matched to the exterior Schwarzschild geometry at a junction interface. According to the proposed model large stability regions do exist, but sufficiently close to where the event horizon is expected to form. The results of this analysis had led to another possibility that massive star doesn’t simply collapse to form a black hole. Furthermore, Chaplygin dark star was proposed in [@Bertolami:2005pz]. A related simplified model with time-dependent spacetime was proposed in Ref. [@DeBenedictis:2008qm] with an evolving parameter $\omega$ crossing the phantom divide, $\omega$ = -1. Various kinds of dark stars were found to be stable under small perturbations (See in detail in Refs. [@Lobo:2006ue]). In [@Yazadjiev:2011sm], Yazadjiev found a class of exact interior solutions describing mixed relativistic stars. According to the model dark energy was provided by scalar fields with negative kinetic energy. Whereas the dark energy imprints in gravitational wave spectrum of mixed neutron-dark-energy stars (containing both ordinary matter and dark energy) have been found in the literature, and we refer the reader to Ref [@Yazadjiev:2011sd]. Lately, an interest in G-lump - a vacuum self gravitation particle-like structure without horizons was introduced by Dymnikova [@Dymnikova:2001fb]. Motivated by the undergoing a phase of accelerated expansion of our universe which was confirmed by extremely luminous stellar explosions, known as type Ia supernovae (SNeIa), provokes us to rethink the commonly accepted scenario. Another related issue is to assess the properties of neutron stars and there formation in our modern scenarios, which is changed after the discovery of PSR J1614-2230 [@Demorest] as 1.97 $\pm$ 0.04 $M_{\odot}$. This discovery puts a severe constraint on the EoS for nuclear matter. In particular, one needs to consider the appearance of exotic particle at densities $\sim$ 5 - 8 $\times 10^{14}$ g/cm$^3$ [@Nagae]. Nevertheless, ultra-luminous X-ray sources (ULXs) are X-ray sources with luminosity above the Eddington limit cannot be explained by the conventional idea of normal stellar mass black hole. Therefore, the basic nature of ultraluminous X-ray pulsars are [@pcc; @mlr], as of now, remains unsolved [@Bachetti:2015pwa]. However, different models have been suggested for the ULX-pulsars including magnetic field of different strength, but none of them are conventional one. With this viewpoint, we explore dark energy as a possible source which could constraints on the mass-radius relation for compact objects whose estimated masses and radii are not compatible with our known sources, such as X-ray burster 4U 1820-30, X-ray sources PSR J1614 - 2230 and X-ray pulsar Vela X - 1, Cen X - 3. On the other hand, the effect of anisotropy on the evolution of self gravitating objects has been studied extensively by several authors using both analytical and numerical methods. The anisotropy means that there are two components of pressure, one is radial direction and the other is along the transverse directions (see Ref. [@Bowers; @Ruderman] and the references within). Interestingly, the interior solution for anisotropy fluid distribution can arise from various reasons: the presence of type 3A superfluid phase, a mixture of different types of fluids or the presence of a magnetic field, etc. (see Ref. [@Kippenhahm; @Sokolov]). From recent investigations of the stellar structure, Herrera and Barreto [@Herrera] considered polytropes for anisotropic matter both in the Newtonian and the GR regimes [@Herrera1; @Herrera2]. In [@Feroze; @Maharaj2], charged anisotropic solutions with a quadratic equation of state was obtained. In particular static anisotropic fluid spheres with uniform energy density was considered in [@Maharaj1]. Recently, Mak and Harko [@Mak], showed that anisotropy may play a vital role in the stability of a dense stars with strange matter. Other studies are also available (check references for instance [@Maurya:2018kxg]). Through this suggestions one could look for anisotropic dark energy model at small scales, and hence hope for yet another approach to find other possible solution satisfying the physical criteria and necessarily contains. This paper is organized as follows. In Sec. \[sec2\], the EFEs have been developed assuming a specific choice of interior space-time (Finch and Skea type), which is static and spherically symmetric. Then, the dark energy EoS have been used to present the structural equations of dark energy star and verify our solutions with compact star Vela X-1. In Sec. \[sec3\], the matching of interior space-time to the exterior Schwarzschild vacuum solution at a junction interface have been studied. In Sec. \[sec4\], we have explored physical properties of compact object for known mass - radius and then determine the mass-radius relation and compactness of star for different compact objects. Finally, a brief discussion is given in Sec. \[sec5\]. Basic concepts of Einstein’s field equations (EFE) {#sec2} ================================================== The interior metric of a static and spherically symmetric star solution is defined as $$\begin{aligned} \label{eq1} ds^{2}&=& -exp\left[{-2\int_r^{\infty}g(\tilde{r})d\tilde{r}}\right]dt^{2} +\frac{dr^{2}} {1-\frac{2m}{r}} \nonumber \\ &\;& +r^{2}(d\theta^{2}+\sin^{2}\theta d\phi^{2}),\end{aligned}$$ where the functions $g(r)$ and $m(r)$ are random functions of the radial coordinate $r$. The factor $g(r)$ is locally measured for gravitational acceleration. As a consequence, positive and negative $g(r)$ means inwardly gravitational attraction, and an outward gravitational repulsion, respectively. The matter contained in the spherical object is described by the anisotropic fluid, and the corresponding energy-momentum tensor will be then $$\label{eq2} T_{\mu\nu}=(\rho+p_t)U_{\mu}U_{\mu}+p_t g_{\mu\nu}+(p_r-p_t)\chi_{\mu}\chi_{\nu},$$ where $U_{\mu}$ represents the 4-velocity of the fluid and $\chi_{\mu}$ is the unit 4-vector along the radial direction. As a matter source in (\[eq2\]), we can also justify by summoning the diagonal components of stress-energy tensor $T^{\mu} _\nu$ = diag \[−$\rho$, $p_r$, $p_t$, $p_t$\]. The GR field equations for the metric (\[eq1\]) with the following energy-momentum tensor (\[eq2\]) can be written as $$\begin{aligned} \label{eq3} m'(r) &=& 4\pi r^{2}\rho, \\ \label{eq4} g(r) &=& \frac{m+4\pi r^{3}p_r}{r(r-2m)}, \\ p_r'&=& -\frac{(\rho+p_r)(m+4\pi r^{3}p_r)}{r(r-2m)}+\frac{2}{r}(p_t-p_r) \label{eq5},\end{aligned}$$ where ‘prime’ denotes the differentiation with respect to $r$. Along these lines, $\rho(r)$ is the energy density, whereas $p_r(r)$ and $p_t(r)$ are the radial and transverse pressures, respectively. Eq. (\[eq5\]), corresponds to the Bianchi’s identity suggesting that (the covariant conservation of the stress-energy tensor) $\Delta_\mu T^{\mu\nu}=0$. By considering relativistic Tolman-Oppenheimer-Volkoff (TOV) equation, one can also achieve the conservation equation. Since, the system of Eqs. (\[eq3\]-\[eq5\]) having three equations with five unknowns, namely, $\rho$(r), $p_r$(r), $p_t$(r), $g(r)$ and $m(r)$. Thus, the unknown functions have to solve for achieving all the plausible features of a compact star. To find the solution of these differential equations, we assume a metric potential *ansatz*, namely, Finch and Skea type [@Finch]. The metric potential function $e^{\lambda}$ is given by the equality: $$\label{eq6} e^{\lambda}=\left(1-\frac{2m}{r}\right)^{-1} =1+\frac{r^{2}}{R^{2}},$$ where $R$ is related with curvature parameter of the configuration which satisfy all physical criteria for a stellar source. Now, solving (\[eq6\]) one can derive the mass function $$\label{eq7} m(r)=\frac{r^{3}}{2(R^{2}+r^{2})}.$$ In next, we adopt the dark energy EoS, $p_r$ = $\omega \rho$ with $\omega < 0$, which is one of the more promising direction to elucidate the current accelerated expansion of the universe. Recently, Ghezzi [@Ghezzi] found a compact object coupled to inhomogeneous anisotropic dark energy. Moreover, in essence of Ref. [@Bertolami], Bertolami and Paramos studied spherically symmetric dark energy structure using a polytropic EoS of negative index. Dark stars were further extended that describe the collapse of a spherical object from an initial state of positive pressure to a final state with negative pressure inside a finite radius core [@Beltracchi:2018ait]. ![The “gravity profile", $g(r)$, is plotted within the interval for $-1/3<\omega<0 $, and then for $-1<\omega<-1/3 $ respectively on the top and bottom panels. Analysis shows that $g(r)$ takes positive value i.e. indicating an inwardly gravitational attraction for $-1/3<\omega<0 $. To be a solution of a dark star we necessarily exclude this region. []{data-label="f7"}](fig1.pdf "fig:"){width="8.cm"} ![The “gravity profile", $g(r)$, is plotted within the interval for $-1/3<\omega<0 $, and then for $-1<\omega<-1/3 $ respectively on the top and bottom panels. Analysis shows that $g(r)$ takes positive value i.e. indicating an inwardly gravitational attraction for $-1/3<\omega<0 $. To be a solution of a dark star we necessarily exclude this region. []{data-label="f7"}](fig2.pdf "fig:"){width="8cm"} Now, taking into account the Eqs. (\[eq3\]) and (\[eq4\]), and using the above EoS we obtain the function $g(r)$ as follows: $$\label{eq8} g(r)=\frac{r}{2R^{2}}\left\{1+\frac{\omega(3R^{2}+r^{2})}{R^{2}+r^{2}}\right\},$$ then the Eqs. (\[eq3\])-(\[eq5\]) reduce to a simple system of algebraic equations $$\begin{aligned} \rho &=& \frac{1}{8\pi}\frac{3R^{2}+r^{2}}{(R^{2}+r^{2})^{2}},\label{eq9}\\ p_r &=& \frac{\omega}{8\pi}\frac{3R^{2}+r^{2}}{(R^{2}+r^{2})^{2}}, \label{eq10}\\ p_t &=&\frac{1+\omega}{32\pi}\frac{3R^{2}+r^{2}}{(R^{2}+r^{2})^{2}}\frac{r^{2}}{R^{2}} \left\{1+\omega\frac{3R^{2}+r^{2}}{R^{2}+r^{2}}\right\} \nonumber \\ &\;& -\frac{\omega}{8\pi}\frac{R^{2}(r^{2}-3R^{2})}{(R^{2}+r^{2})^3}, \label{eq11}\end{aligned}$$ where $\Delta$ = $p_t - p_r$ stands for measure of the pressure anisotropy of the fluid comprising the dark energy star, can be expressed in the following form $$\label{eq12} \Delta=\frac{1+\omega}{32\pi}\frac{3R^{2}+r^{2}}{(R^{2}+r^{2})^{2}}\frac{r^{2}}{R^{2}} \left\{1+\omega\frac{3R^{2}+r^{2}}{R^{2}+r^{2}}\right\}-\frac{\omega r}{4\pi}\frac{5R^{2}+r^{2}}{(R^{2}+r^{2})^{3}}.$$ It is known that $\Delta$ represents a force due to the anisotropic nature of the stellar model. The force is being directed outward when $\Delta$ $>$ 0 i.e. $p_t> p_r$ and inward when $\Delta$ $<$ 0 i.e. $p_t < p_r$. Corresponding to $\Delta$ = 0 is a certain case of an isotropic pressure. At the stellar centre, $r = 0$ we have $\Delta$ = 0, which is expected. Indeed, the central density is a non-zero constant, $\rho_0$= 3/8$\pi R^2$. This reflects that there is no singularity inside the star. Our intention is to study and analyze the dark energy star that can be considered as a suitable way to explain those massive stellar systems like white dwarfs, massive pulsars and magnetars etc. The main assumption that leads to Eq. (\[eq8\]) is that $g(r) >$ 0, for $\omega > -(r^2+R^2)/(r^2+ 3R^2)$ indicating an inwardly gravitational attraction. The condition is verified from Fig. **1**, that $g(r)$ is positive in the interval $-1/3<\omega< 0$, while $g(r) < 0$ in the interval $ -1<\omega<-1/3$. The above discussions show that to be a gravastar like solution, the local acceleration due to gravity of the interior solution be repulsive. In spite of these constraints, we consider only the region where $g(r) < 0$ for further precision (see Refs. [@lobo1] for more details). Specifically, we investigate the maximum mass of a dark star using the dark energy equations of state $\omega <-1/3$, and study the effect of the state parameter $\omega $ on the other physical properties such as the density, mass-radius and gravitational redshift. In addition, we compare the obtained results of this theory with some strange/compact star candidates like PSR J1416-2230, Vela X-1, 4U 1608-52, Her X-1 and PSR J1903+327, respectively. However, the X-ray pulsar Vela X-1, whose estimated mass and radius are 1.77$\pm$ 0.08 $M_{\bigodot}$ and $R = 10.852$ Km has been allowed for testing the physical acceptability of the developed model. ![Variation of the density and anisotropy $\Delta$ have been plotted as a function of the radial coordinate $r$. For density profile we consider $\omega = - 0.35$ and radii are on the order of 10 km. For anisotropy factor, we verify that $\Delta >0$ in the range $-1<\omega<-1/3 $.[]{data-label="f7"}](fig3.pdf "fig:"){width="8cm"} ![Variation of the density and anisotropy $\Delta$ have been plotted as a function of the radial coordinate $r$. For density profile we consider $\omega = - 0.35$ and radii are on the order of 10 km. For anisotropy factor, we verify that $\Delta >0$ in the range $-1<\omega<-1/3 $.[]{data-label="f7"}](fig4.pdf "fig:"){width="8cm"} Junction Condition {#sec3} =================== At this stage, the interior solution is matched to an exterior Schwarzschild vacuum solution with $p$ = $\rho$ = 0 at a junction interface $\Sigma$, with junction radius, $a$. The Schwarzschild exterior solution is given by $$\label{eq13} ds^{2}=-\left(1-\frac{2M}{r}\right)dt^{2}+\frac{dr^{2}} {1-\frac{2M}{r}}+r^{2}(d\theta^{2}+\sin^{2}\theta d\phi^{2}),$$ which possesses an event horizon at $r_h = 2M$. We have chosen the value of $a > r_h$, to avoid the presence of horizons i.e. the junction radius lies outside $2M$. Let us turn now our attention in computing the surface stress-energy tensor $S_{ij}$, which can be expressed in terms of the jump of the extrinsic curvature by means of the Lanczos equation [@Israel] (see Ref. [@Lobo:2005zu] for more details) and defined by $$\label{eq14} S^{i}_{j} = -\frac{1}{8\pi}\left(\kappa^i_j-\delta^i_j\kappa^m_m\right),$$ where $k_{ij}$ represents discontinuity in the extrinsic curvature $K_{ij}$, across the junction interface, which is defined by $k_{ij} = K^{+}_{ij}-K^{-}_{ij}$. Therefore, the second fundamental form has been used to yields the extrinsic curvature in its final form, which is $$\label{eq15} K^{\pm}_{ij} = -\eta_{\nu}\left(\frac{\partial^2 x^{\nu}}{\partial\xi^i \partial\xi^j}+\Gamma^{\nu\pm}_{\alpha\beta}\frac{\partial x^{\alpha}}{\partial\xi^{i}}\frac{\partial x^{\beta}}{\partial\xi^{j}}\right),$$ where $\eta_{\nu}$ represents the unit normal at the junction, the symbol $`\pm'$ represents the interior and exterior spacetime, and $\xi^i$ represents the intrinsic coordinates. Now, by using the metrics (\[eq1\]) and (\[eq13\]), the non-trivial components of the extrinsic curvature are given by $$\label{eq16} K_{\tau}^{\tau~+}=\frac{\frac{M}{a^{2}}+\ddot{a}}{\sqrt{1-\frac{2M}{a}+\dot{a}^{2}}},$$ $$\label{eq17} K_{\tau}^{\tau~-}=\frac{\frac{a}{2}\left\{\frac{1}{R^{2}+a^{2}}+\omega\frac{3R^{2}+a^{2}}{(R^{2}+a^{2})^{2}} \right\}+\ddot{a}-\frac{(1+\omega)a}{2R^{2}}\frac{3R^{2}+a^{2}}{R^{2}+a^{2}}\dot{a}^{2}}{\sqrt{\frac{R^{2}}{R^{2}+a^{2}}+\dot{a}^{2}}},$$ and $$\begin{aligned} \label{eq18} K_{\theta}^{\theta~+}&=& \frac{1}{a}\sqrt{1-\frac{2M}{a}+\dot{a}^{2}}, \\ \label{eq19} K_{\theta}^{\theta~-} &=& \frac{1}{a}\sqrt{\frac{R^{2}}{R^{2}+a^{2}}+\dot{a}^{2}}.\end{aligned}$$ We have adopted the usual notation in which the dot and prime represent $d/d\tau$ and $d/dr$, respectively. Since our metrics are diagonal, $S^{i}_{j}$ is also diagonalized and written as $S^{i}_{j}= \text{diag} ~(-\sigma,\mathcal{P},\mathcal{P})$. Thus, the Lanczos equations give the energy density and pressure on the shell: $$\label{eq20} \sigma=-\frac{1}{4\pi a}\left[\sqrt{1-\frac{2M}{a}+\dot{a}^{2}}-\sqrt{\frac{R^{2}}{R^{2}+a^{2}}+\dot{a}^{2}}\right],$$ $$\begin{aligned} \label{eq21} \mathcal{P}&=&\frac{1}{8\pi a}\left[\frac{1-\frac{M}{a}+\dot{a}^{2}+a\ddot{a}}{\sqrt{1-\frac{2M}{a}+\dot{a}^{2}}} -\frac{1+\frac{\omega a^{2}}{2}\frac{3R^{2}+a^{2}}{(R^{2}+a^{2})^{2}}}{\sqrt{\frac{R^{2}}{R^{2}+a^{2}}+\dot{a}^{2}}} \right.\nonumber\\ && \left. -\frac{\dot{a}^{2}+a\ddot{a}-\frac{a^{2}}{2(R^{2}+a^{2})} +(1+\omega)\frac{a^{2}}{2R^{2}}\frac{3R^{2}+a^{2}}{R^{2}+a^{2}}\dot{a}^{2}} {\sqrt{\frac{R^{2}}{R^{2}+a^{2}}+\dot{a}^{2}}}\right],\end{aligned}$$ where $\sigma$ and $\mathcal{P}$ are the surface energy density and surface pressure at the junction interference. Here, we interpret the quantities $\sigma$ and $\mathcal{P}$ arising from the thin-shell formalism for the spherically symmetric case. In this regard, the equation of motion for the surface stress-energy tensor is given by $S^{i}_{j|i}$= $\left[T_{\mu\nu}e^{\mu}_{(j)}n^{\nu}\right]^{+}_{-}$, where $[X]^{+}_{-}$ denotes the discontinuity across the surface interface, i.e. $[X]^{+}_{-}=[X]^{+}|_{\Sigma}-[X]^{-}|_{\Sigma}$. According to the Ref. [@Lobo:2005zu], the momentum flux term $F_{\mu}$= $T_{\mu\nu}U^{\nu}$ in the right hand side corresponds to the net discontinuity across the shell. Consequently, the energy-conservation equation on the shell $\nabla^{i}S^{i}_{i}$ can be written as $S^{i}_{\tau|i}$ =$\dot{\sigma}+2\dot{a}(\sigma+\mathcal{P})/a$ and the energy flux is given by $$\left[T_{\mu\nu}e^{\mu}_{(j)}n^{\nu}\right]^{+}_{-}= - \frac{\left(\rho+p_r\right)\dot{a}\sqrt{1-2m/a+\dot{a^2}}}{1-2m/a}.$$ One may deduced $\rho$ and $p_r$ from Eqs. (\[eq9\]-\[eq10\]), respectively, evaluated at the junction radius, $a$. Now, using the above relationship the conservation identity is much more appealing form $$\sigma'= -\frac{2}{a}(\sigma+\mathcal{P})+ \Theta,$$ where $\Theta$ is defined as $$\label{th24} \Theta = -\frac{1}{4\pi a} \frac{m'(1+ \omega)}{(a-2m)}\sqrt{1-2m/a+\dot{a^2}}.$$ After some algebraic manipulations, from Eq. (\[eq7\]), we then obtain $$\label{th25} \Theta = -\frac{(w+1) \left(r^4+3 r^2 R^2\right)}{8 \pi a \left(r^2+R^2\right) \left(a \left(r^2+R^2\right)-r^3\right)}\sqrt{1-2m/a+\dot{a^2}}.$$ For $\omega = -1$, the Eq (\[th25\]) reduces to zero, which relate to the solution obtained in Ref [@Visser:2003ge]. We would also like to emphasize that for closing the system of equations we need an equation of state (EoS) $\mathcal{P}=\mathcal{P}(\sigma)$. Such an EoS would embrace the microphysics of the matter inside the shell. For instant, gravastar model has an interior de-Sitter spacetime surrounded by a thin shell of ultra-stiff matter with an equation of state $p = \rho$, which is again matched to an exterior Schwarzschild vacuum solution. [TABLE 1: Numerical values of physical parameters for the different compact stars for $\omega$]{} \[Table11-1\] ------------------------------------------------ ------------------- ---------------------- ------------------------ ------------------------ ---------- ---------------- Compact Observed Mass Predicted $\rho_{c}$ $\rho_{s}$ Surface $\frac{2M}{R}$ Stars $(M_{\bigodot})$ Radius (Km) $(gm/cm^{3})$ $(gm/cm^{3})$ Redshift PSR J1416-2230 (Demorest *et al.* [@Demorest]) 1.97 $\pm$ 0.04 11.083 $\pm$ 0.037 7.965 $\times 10^{14}$ 5.695 $\times 10^{14}$ 0.391 0.445 Vela X-1  (Rawls *et al.* [@R]) 1.77 $\pm$ 0.08   10.852 $\pm$ 0.108 7.231 $\times 10^{14}$ 5.371 $\times 10^{14}$ 0.365 0.579 4U 1608-52  (Gu$\dot{v}$er et al. [@Guver]) 1.74 $\pm$ 0.14   10.811 $\pm$ 0.197 7.695 $\times 10^{14}$ 5.975 $\times 10^{14}$ 0.342 0.436 Her X-1 (Abubekerov et al. [@Abubekerov]) 0.85 $\pm$ 0.15   8.836 $\pm$ 0.481 8.014 $\times 10^{14}$ 6.453 $\times 10^{14}$ 0.317 0.327 PSR J1903+327 (Freire *et al.* [@F]) 1.667 $\pm$ 0.021   10.703 $\pm$ 0.032 7.695 $\times 10^{14}$ 5.832 $\times 10^{14}$ 0.215 0.320 ------------------------------------------------ ------------------- ---------------------- ------------------------ ------------------------ ---------- ---------------- Structural properties of compact objects {#sec4} ========================================= In order to examine more details about the stellar structure, we perform some analytical calculations and study the corresponding constraint for interior fluid sphere. Then, the effect of dark energy has been discussed to describe the mass-radius relation and explore massive stellar objects like massive pulsars, super-Chandrasekhar stars and magnetars, namely, PSR J1416-2230, Vela X-1 4U 1608-52, Her X-1 and PSR J1903+327 as given in Gangopadhyay *et al* [@Gangopadhyay]. To be physically acceptable, the model should be free from any geometrical singularities i.e. energy density and pressure are regularity and finite at the center $r = 0$. Owing to the several observations, the obtained solutions have been studied and analyzed it’s physical acceptability in terms of the star Vela  X-1 (mass $1.77 \pm 0.08$ $M_{\bigodot}$ and radius $R = 10.852$ Km) by assuming it as the representative of compact star candidates. On substituting these values into Eqs. (\[eq9\]-\[eq10\]), we plot the dependence of energy density $\rho$ vs $r$ in Fig **2**. It may be pointed out here that energy density is finite inside the stellar interior and monotonic decreasing function of $r$. Since we have chosen dark energy EoS then it is obvious that radial pressure $p_r$ is negative inside the stellar interior. In Fig. **2** (right panel), we plot the anisotropic factor inside the stellar interior for $\omega<-1/3$. We notic from the figure that $\Delta>0$ for $\omega = -0.35$ i.e. in the phantom region anisotropic force directed outward. Energy Condition ---------------- We are interested here to investigate the energy conditions in terms of the components of the energy-momentum tensor. To begin with, we consider different energy conditions, namely, null energy condition (NEC), weak energy condition (WEC), strong energy condition (SEC) and dominant energy condition (DEC) for the compact star candidate Vela  X-1 (as the representative of other stellar model)- read as follows: $$\label{eq26} (i) ~~~\text{\textbf{NEC}}: \rho(r)+p_r(r) \geq 0,\\$$ $$(ii) ~~~\text{\textbf{WEC}}: \rho(r)+p_r(r) \geq 0~~ \text{and} ~~\rho(r) \geq 0,$$ $$(iii) ~~\text{\textbf{SEC}}: \rho(r)+p_r(r) \geq 0~~ \text{and} ~~\rho+p_r(r)+p_t(r) \geq 0,$$ $$\label{eq29} (iv) ~~~\text{\textbf{DEC}}: \rho(r)> \lvert p_r(r)\lvert ~~ \text{and} ~~\rho (r) > \lvert p_t(r)\rvert.$$ Utilizing the inequalities, the nature of energy conditions for the astral structure Vela X-1 has been studied. The behavior of these conditions are shown graphically. In Fig. **3**, we plot the left hand side of the above inequalities as a function of $r$, which shows that all energy conditions of our model are satisfied for $\omega= -0.35$. [TABLE II: Physical parameters of the strange star candidate Vela X-1 due to the different values of state parameter $\omega$ with mass 1.77 $\pm$ 0.08 (Rawls *et al.* [@R])]{} ----------- ------------- ------------------------ ------------------------ ---------- ---------------- -- Values of Predicted $\rho_{c}$ $\rho_{s}$ Surface $\frac{2M}{R}$ $\omega$ Radius (Km) $(gm/cm^{3})$ $(gm/cm^{3})$ Redshift - 0.37 10.736 7.221 $\times 10^{14}$ 5.351 $\times 10^{14}$ 0.357 0.546 - 0.41 10.652 7.131 $\times 10^{14}$ 5.142 $\times 10^{14}$ 0.348 0.529 - 0.47 10.611 7.057 $\times 10^{14}$ 4.957 $\times 10^{14}$ 0.337 0.486 ----------- ------------- ------------------------ ------------------------ ---------- ---------------- -- ![The plot depicts the energy conditions for the inequalities Eqs. (\[eq26\]-\[eq29\]) as a function of the radial coordinate $r$. The curves for null energy condition (NEC), weak energy condition (WEC), strong energy condition (SEC), dominant energy condition (DEC) are shown in this figures, for the compact star Vela X-1.](fig5.pdf){width="8.6cm"} {width="8cm"} {width="8cm"} Mass-radius relation and Surface gravitational red shift --------------------------------------------------------- Here, we extend our analysis towards the effective mass-radius relation and surface gravitational red shift. For this purpose, the mass function within the radius of a compact star is given by $$\label{eq30} m(r)=\int_0^{r}4\pi \rho r^{2}dr=\frac{r^{3}}{2(r^{2}+R^{2})}.$$ We observe that the mass function $m(r)\rightarrow 0$ as $r\rightarrow 0$. Since the spherically symmetric perfect fluid configurations have the allowable mass-radius ratio fall within the limit of $ 2M/R < 8/9$ (in the unit $c = G = 1$) [@Buchdahl]. Thus, putting the values considered for ‘Vela X-1’ in Eq. (\[eq30\]), we obtain the profile for mass function as shown in Fig. **4**. In order to check the condition we have plotted total mass normalized in solar mass, i.e. $M/M{_\odot}$ with the radius $R$ in Fig. **4**, for the specific value of $\omega$ = - 0.35. Indeed, it is shown that $m(r)$ is monotonic increasing function of radial co-ordinate and $m(r)>0$ for $0 <r <a$. To obtain the mass-radius relation for different compact stars, which are in general very closely equal to the observed values are tabulated in Table-**I**. Moreover, central density, surface density as well as surface redshift are tabulated Table-**I** in describing the other compact configurations. From the observation one can see that the mass-radius relation for the different compact stars do not cross the proposed range of Buchdahl-Bondi inequality [@Buchdahl]. {width="8.6cm"} Following the basic definition of the compactness of the star is obtained as $$\label{eq31} u(r)=\frac{m(r)}{r}=\frac{r^{2}}{2(R^{2}+r^{2})}.$$ We show the behaviour of the compactness $u(r)$ of the star in Fig. **5**. More precisely, the figure indicating that $u(r)$ is a monotonic increasing function of $r$, and the redshift function $z_s $ of the compact star is given by $$\label{eq32} 1+z_s=(1-2u)^{-\frac{1}{2}}.$$ Thus, the surface redshift function $z_s$ can be define as $$\label{eq33} z_s=\sqrt{\left(\frac{R^{2}}{R^{2}+r^{2}}\right)}-1.$$ In Fig. **6**, the surface redshift is shown for the compact star Vela X-1, which is determined from the compactness parameter. We also perform the maximum surface redshift ($z_s$) for different compact stellar configuration which is shown in Table **I** and Table **II**, respectively for different values of state parameter $\omega$. It is also of interest to observe that maximum value of the surface redshift, $z_s < 1$. In this sense, one may extract that obtained results are compatible with the result obtained in [@hamity]. Results and Discussion {#sec5} ====================== The driving forces behind scientific progress are contradictions between entrenched theories and new observations. Typically, dark energy models are based on scalar fields minimally coupled to gravity, is believed to responsible for the present accelerated expansion of the Universe. In fact, DE has opened up new possibilities in theoretical research for cosmology as well as astrophysical objects. In this article we present a formalism, based on dark energy EoS, describing a new class of static spherically symmetric stellar model satisfying all the physical realistic conditions. The interior space-time metric is matched with the Schwarzschild exterior vacuum solution at the boundary. The constraints of the field equations are determined by the following conditions (i) the interior metric is describe by Finch-Skea type and (ii) the subsequent analysis is based on the dark energy EoS $p_r$ = $\omega \rho$. The motivation for implementing this model satisfies all the physical requirements like energy density, radial and transverse pressures which are finite and regular at the centre. Therefore, it can potentially describe as a compact object which is neither a neutron stars nor a quark star. It is evident from Figs. **1** that the ‘gravitarional profile’ $g(r)>$ 0, indicating an inwardly gravitational attraction within the range of $-1/3<\omega<0 $, and $g(r)<0$ for $\omega<-1/3$, which indicate the gravitational repulsion. For a dark star solution it is necessary that the local acceleration due to gravity of the interior solution be repulsive in nature, so that the region where $g(r) >$ 0 is necessarily excluded. For value of $w = -0.35$, we consider Vela X-1 of mass 1.77 $M_{\bigodot}$ as the representative of compact star candidates. The results reported in Figs. **1**-**6**. In particular, we pointed out that the central density is finite and maximum at the stellar interior (Fig. **2**), whereas the variation of anisotropy shown in Fig. **2** (right panel) is positive throughout the system for $-1<\omega <-1/3$. Hence, the direction of the anisotropic force is outward for our system. Next, we matched our interior solution to the exterior Schwarzschild solution in presence of thin shell. We set $G = c = 1$, while solving Einstein equations and plotting the graphs. Based on physical requirements and plugging the values of $G$ and $c$ into the relevant equations, we calculate the numerical values of mass-radius relation for different compact stars. Physically the mass of the star is strongly dependent on its central density, and we know that high central density stars have lower gravitational masses. The obtained solutions are compared with the observed evidence for the existence of compact stars which is consistent with our model and the of mass-radius for different stars lies in the proposed range by Buchdahl [@Buchdahl] (The results reported in Table **I** and  **II**). To examine the nature of physical quantities, we consider a particular star, namely, Vela X-1 to investigate the physical properties. In order to investigate internal structure of the dark star in more details, we plot Figs. **3** - **6**, for energy conditions, mass-radius relations and surface redshift, respectively. In our model we have found the surface redshift ($z_s$) for the different compact stars are of finite values and vanishes outside of the star (see Table **I** and **II** for more details), which typically fall within the proposed range in [@hamity], are physically acceptable. Finally, at very high mass measurement of about 2$M_{\bigodot}$ requires a really stiff equation of state in neutron stars, and DE may be used as a possible candidate to study observable compact astrophysical objects. Acknowledgement {#acknowledgement .unnumbered} =============== MKJ acknowledge continuous support and encouragement from the administration of University of Nizwa. A. P also thanks the IUCAA, Pune for the support under visiting associateship programme where a part of this work has been done. We would also like to thank the referee for his or her valuable comments. [99]{} M. R. Finch and J.E.F. Skea, Class. Quantum. Grav., [**6**]{}, 467 (1989). D. Gondek-Rosinska *et al*, Astron. and Astrophys., [**363**]{}, 1005 (2000). N. K. Glendenning, Nuclear Physics, Particle Physics and General Relativity (Springer-Verlag, New York, Inc. 1997). A. Grant *et al*, Astrophys. J. [**560**]{}, 49 (2001). S. Perlmutter, M. S. Turner and M. White, Phys. Rev. Lett. [**83**]{}, 670 (1999). C. L. Bennett *et al*, Astrophys. J. Suppl. [**148**]{}, 1 (2003). G. Hinshaw *et al.* \[WMAP Collaboration\], Astrophys. J. Suppl. [**148**]{}, 135 (2003). E.J. Copeland, M. Sami and S. Tsujikawa, Int.J.Mod.Phys.D [**15**]{}, 1753 (2006); T. Padmanabhan, Gen. Relativ. Gravit. [**40**]{}, 529 (2008); R. Durrer and R. Maartens, Gen. Relativ. Gravit. [**40**]{}, 301 (2008). M. Li, X. D. Li, S. Wang and Y. Wang, Commun. Theor. Phys.  [**56**]{}, 525 (2011); Y. F. Cai, E. N. Saridakis, M. R. Setare and J. Q. Xia, Phys. Rept.  [**493**]{}, 1 (2010). M.S. Turner, \[arXiv:astro-ph/0108103\]. U. Alam, V. Sahni, T. D. Saini and A. A. Starobinsky, Mon. Not. Roy. Astron. Soc.  [**344**]{}, 1057 (2003). P.O. Mazur and E. Mottola, \[arXiv:gr-qc/0109035\] (2001). M. Visser and D. L. Wiltshire, Class. Quant. Grav. [**21**]{}, 1135 (2004). N. Bilic, G. B. Tupper and R. D. Viollier, JCAP [**0602**]{}, 013 (2006). C. B. M. H. Chirenti and L. Rezzolla, Class. Quant. Grav.  [**24**]{}, 4191 (2007); A. Banerjee, J.R. Villanueva, P. Channuie and K. Jusufi, Chin.Phys. C [**42**]{} 115101 (2018); S. Ghosh, S. Ray, F. Rahaman and B.K. Guha, Annals Phys.[**394**]{} 230 (2018); P. Bhar and F. Rahaman, Eur. Phys. J. C [**75**]{}, 41 (2015). G. Chapline, eConf C, [**041213**]{} 0205 (2004). F. Lobo, Class. Quantum. Grav. [**23**]{}, 1525 (2006). O. Bertolami and J. Paramos, Phys. Rev. D [**72**]{}, 123512 (2005). A. DeBenedictis, R. Garattini and F. S. N. Lobo, Phys. Rev. D [**78**]{}, 104003 (2008). F. S. N. Lobo, Phys. Rev. D [**75**]{}, 024023 (2007); D. Horvat and A. Marunovic, Class. Quant. Grav.  [**30**]{}, 145006 (2013); P. Bhar, T. Manna, F. Rahaman and A. Banerjee, Can. J. Phys.  [**96**]{}, 594 (2018); F. Rahaman, R. Maulick, A. K. Yadav, S. Ray and R. Sharma, Gen. Rel. Grav.  [**44**]{}, 107 (2012); M. Kalam, F. Rahaman, S. Molla and S. M. Hossein, Astrophys. Space Sci.  [**349**]{}, 865 (2014). S. S. Yazadjiev, Phys. Rev. D [**83**]{}, 127501 (2011). S. S. Yazadjiev and D. D. Doneva, JCAP [**1203**]{}, 037 (2012). I. Dymnikova, Class. Quant. Grav.  [**19**]{}, 725 (2002); I. Dymnikova, Int. J. Mod. Phys. D [**12**]{}, 1015 (2003). P. Demorest *et al*, [*Nature* ]{}, [**467**]{}, 1081 (2010). T. Nagae, Prog. Theor. Phys. Suppl.[**185**]{}, 299 (2010). P. C. C. Freire *et al*, Mon. Not. Roy. Astron. Soc., [**412**]{}, 2763 (2011). M. L. Rawls *et al*, ApJ, [**730**]{}, 25 (2011). M. Bachetti, Astron. Nachr.  [**337**]{}, 349 (2017); H. Feng and R. Soria, New Astron. Rev. [**55**]{}, 166 (2011). R.L. Bowers and E. P. T. Liang, Class. Astrophys. J. [**188**]{}, 657 (1974). R. Ruderman, Class. Ann. Rev. Astron. Astrophys [**10**]{}, 427 (1972). R. Kippenhahm and A. Weigert, Stellar Structure and Evolution (Springer, Berlin, 1990). A. I. Sokolov, Zh. Eksp. Teor. Fiz. [**79**]{}, 1137 (1980). L. Herrera and W. Barreto, Phys. Rev. D [**87**]{}, 087303(2013). L. Herrera andW. Barreto, Phys. Rev. D [**88**]{}, 084022 (2013). L. Herrera, A. Di Prisco, W. Barreto, and J. Ospino, Gen. Relativ. Gravit. [**46**]{}, 1827 (2014). T. Feroze and A. A. Siddiqui, Gen. Relativ. Gravit. [**43**]{}, 1025 (2011). S. D. Maharaj and P. Mafa Takisa, Gen. Relativ. Gravit. [**44**]{}, 1419 (2012). S. D. Maharaj and R. Maartens, Gen. Relativ. Gravit. [**21**]{}, 899 (1989). M. K. Mak and T. Harko, Chin. J. Astron. Astrophys. [**2**]{}, 248 (2002). S. K. Maurya, A. Banerjee, M. K. Jasim, J. Kumar, A. K. Prasad and A. Pradhan, Phys. Rev. D [**99**]{}, 044029 (2019); S. K. Maurya, A. Banerjee and P. Channuie, Chin. Phys. C [**42**]{}, 055101 (2018); S. K. Maurya and S. D. Maharaj, Eur. Phys. J. A [**54**]{}, 68 (2018); D. Deb, M. Khlopov, F. Rahaman, S. Ray and B. K. Guha, Eur. Phys. J. C [**78**]{}, 465 (2018); M. Kalam, F. Rahaman, S. Molla and S. M. Hossein, Astrophys. Space Sci.  [**349**]{}, 865 (2014); M. Kalam, F. Rahaman, S. M. Hossein and S. Ray, Eur. Phys. J. C [**73**]{}, 2409 (2013); B. C. Paul, P. K. Chattopadhyay and S. Karmakar, Astrophys. Space Sci.  [**356**]{}, 327 (2015); R. Sharma and S. Mukherjee: Mod. Phys. Lett. A. [**17**]{}, 2535 (2002). C. R. Ghezzi, Astrophys.Space Sci. [**333**]{}, 437 (2011). O. Bertolami and J. Paramos, Phys. Rev. D [**72**]{}, 1235 (2005). P. Beltracchi and P. Gondolo, Phys. Rev. D [**99**]{}, 044037 (2019). W. Israel, Nuovo Cimento [**44B**]{}, 1 (1966); and corrections in ibid. [**48B**]{}, 463 (1966). F. S. N. Lobo and P. Crawford, Class. Quant. Grav.  [**22**]{}, 4869 (2005). M. Visser and D. L. Wiltshire, Class. Quant. Grav.  [**21**]{}, 1135 (2004). H. A. Buchdahl, Phys. Rev. [**116**]{}, 1027 (1959). D. E. Barraco and V. H. Hamity, Phys.Rev.D [**65**]{}, 124028 (2002). M. L. Rawls *et al*, Astrophys. J. [**730**]{} 25 (2011). T. Gu$\dot{{v}}$er, F. $\dot{{O}}$zel, A. Cabrera-Lavers and P. Wroblewski, Astrophys. J. [**712**]{}, 964 (2010). M. K. Abubekerov, E. A. Antokhina, A. M. Cherepashchuk and V. V. Shimanskii Astron. Rep. [**52**]{}, 379 (2008). P. C. C. Freire *et al*, Mon. Not. Roy. Astron. Soc. [**412**]{}, 2763 (2011). T. Gangopadhyay *et al*, Mon. Not. R. Astron. Soc. [**431**]{}, 3216 (2013).

--- abstract: | An $r$-simple $k$-path is a [path]{} in the graph of length $k$ that passes through each vertex at most $r$ times. The [$r$-SIMPLE $k$-PATH]{} problem, given a graph $G$ as input, asks whether there exists an $r$-simple $k$-path in $G$. We first show that this problem is NP-Complete. We then show that there is a graph $G$ that contains an $r$-simple $k$-path and no simple path of length greater than $4\log k/\log r$. So this, in a sense, motivates this problem especially when one’s goal is to find a short path that visits many vertices in the graph while bounding the number of visits at each vertex. We then give a randomized algorithm that runs in time $${\mathrm{poly}}(n)\cdot 2^{O( k\cdot \log r/r)}$$ that solves the [$r$-SIMPLE $k$-PATH]{} on a graph with $n$ vertices with one-sided error. We also show that a randomized algorithm with running time ${\mathrm{poly}}(n)\cdot 2^{(c/2)k/ r}$ with $c<1$ gives a randomized algorithm with running time ${\mathrm{poly}}(n)\cdot 2^{cn}$ for the Hamiltonian path problem in a directed graph - an outstanding open problem. So in a sense our algorithm is optimal up to an $O(\log r)$ factor. author: - | Hasan Abasi Nader H. Bshouty Ariel Gabizon[^1] Elad Haramaty[^2]\ Department of Computer Science\ Technion, Haifa title: '**[On $r$-Simple $k$-Path]{}** ' --- Introduction ============ Let $G$ be a directed graph on $n$ vertices. A path $\rho$ is called [*simple*]{} if all the vertices in the path are distinct. The [SIMPLE $k$-PATH]{} problem, given $G$ as input, asks whether there exists a simple path in $G$ of length $k$. This is a generalization of the well known [HAMILTONIAN-PATH]{} problem that asks whether there is a simple path passing through *all* vertices, i.e., a simple path of length $n$ in $G$. As [HAMILTONIAN-PATH]{} is NP-complete, we do not expect to find polynomial time algorithms for [SIMPLE $k$-PATH]{} for general $k$. Moreover, we do not even expect to find good approximation algorithms for the corresponding optimization problem: the *longest path problem*, where we ask what is the length of the longest simple path in $G$. This is because Björklund et al. [@BHK04] showed that the longest path problem cannot be approximated in polynomial time to within a multiplicative factor of $n^{1-\epsilon}$, for any constant $\epsilon>0$, unless P=NP. This motivated finding algorithms for [SIMPLE $k$-PATH]{} with running time whose dependence on $k$ is as small as possible. The first result in this venue by Monien [@M85] achieved a running time of $k!\cdot {\mathrm{poly}}(n)$. Since then, there has been extensive research on constructing algorithms for [SIMPLE $k$-PATH]{} running in time $f(k)\cdot{\mathrm{poly}}(n)$, for a function $f(k)$ as small as possible [@B93; @AYZ95; @KMRR06; @CLSZ07; @K08]. The current state of the art is $2^k\cdot poly(n)$ by Williams [@W09] for directed graphs and $O(1.657^k)\cdot {\mathrm{poly}}(n)$ by Bj[ö]{}rklund [@BHKK10] for undirected graphs. Our results ----------- In this paper we look at a further generalization of [SIMPLE $k$-PATH]{} which we call [$r$-SIMPLE $k$-PATH]{}. In this problem instead of insisting on $\rho$ being a simple path, we allow $\rho$ to visit any vertex a fixed number of times. We now formally define the problem [$r$-SIMPLE $k$-PATH]{}. \[dfn:r\_simp\] Fix integers $r\leq k$. Let $G$ be a directed graph. - We say a path $\rho$ in $G$ is *$r$-simple*, if each vertex of $G$ appears in $\rho$ at most $r$ times. Obviously, $\rho$ is a simple path if and only if it is a $1$-simple path. - The *[$r$-SIMPLE $k$-PATH]{}* problem, given $G$ as input, asks whether there exists an $r$-simple path in $G$ of length $k$. At first, one may wonder whether for some fixed $r>1$, [$r$-SIMPLE $k$-PATH]{} always has a polynomial time algorithm. We show this is unlikely by showing that for any $r$, for some $k$ [$r$-SIMPLE $k$-PATH]{} is NP-complete. See Theorem \[thm:simpathNPC\] in Section \[sec:prelim\_results\] for a formal statement and proof of this. Thus, as in the case of [SIMPLE $k$-PATH]{}, one may ask what is the best dependency of the running time on $r$ and $k$ that can be obtained in an algorithm for [$r$-SIMPLE $k$-PATH]{}. Our main result is \[thm:main\] Fix any integers $r,k$ with $2\leq r\leq k$. There is a randomized algorithm running in time $${\mathrm{poly}}(n)\cdot O\left(r^{ \frac{2k}{r} +O(1)}\right) = {\mathrm{poly}}(n)\cdot 2^{O( k\cdot \log r/r)}$$ solving [$r$-SIMPLE $k$-PATH]{} on a graph with $n$ vertices with one-sided error. One may ask how far from optimal is the dependency on $k$ and $r$ in Theorem \[thm:main\]. Theorem \[thm:simpathNPC\] implies that a running time of ${\mathrm{poly}}(n)\cdot 2^{o(k/r)}$ would give an algorithm with running time $2^{o(n)}$ for [HAMILTONIAN-PATH]{}. Moreover, even a running time of ${\mathrm{poly}}(n)\cdot 2^{c\cdot k/r}$, for a small enough constant $c<1/2$, would imply a better algorithm for [HAMILTONIAN-PATH]{} than those of [@W09; @BHKK10] which are the best currently known. So, in a sense our algorithm is optimal up to an $O(\log r)$ factor. We find closing this $O(\log r)$ gap, e.g. by a better reduction to [HAMILTONIAN-PATH]{}, or a better algorithm for [$r$-SIMPLE $k$-PATH]{}, to be an interesting open problem. Finding a path with many distinct vertices ------------------------------------------ We give more motivation for the [$r$-SIMPLE $k$-PATH]{} problem. Suppose we are in a situation where we wish to find a relatively short path passing through many distinct vertices. Note that an $r$-simple path of length $k$ must path through at least $k/r$ distinct vertices. Thus, in case, for example, a $2$-simple path of length $k$ exists, our algorithm for [$2$-SIMPLE $k$-PATH]{} can be used to find a path of length $k$ with at least $k/2$ distinct vertices in time ${\mathrm{poly}}(n)\cdot 2^{k/2}$. One may ask how this would compare to the number of distinct vertices returned by the algorithms for [SIMPLE $k$-PATH]{}. We show there can be a large gap. Specifically, for any given $k$ , we show there is a graph $G$ where all simple paths are of length less than $4\cdot \log k$, but $G$ contains a $2$-simple path of length $k$. See Theorem \[Gap\] for a precise statement. Overview of the proof of Theorem \[thm:main\] ============================================= We give an informal sketch of Theorem \[thm:main\]. We are given a directed graph $G$ on $n$ vertices, and integers $r\leq k$. We wish to decide if $G$ contains an $r$-simple path of length $k$. There are two main stages in our algorithm. The first is to reduce the task to another one concerning multivariate polynomials. This part, described below, is very similar to [@AB13]. #### Reduction to a question about polynomials We want to associate our graph $G$ with a certain multivariate polynomial $p_G$. We associate with the $i$’th vertex a variable $x_i$. The monomials of the polynomial will correspond to the paths of length $k$ in $G$. So we have $$p_G({{\bf x}})=\sum_{i_1\to i_2\to \cdots \to i_{k} \in G} x_{i_1}\cdots x_{i_{k}},$$ where $i_1\to i_2\to \cdots \to i_{k} \in G$ means that $i_1, i_2, \cdots , i_{k}$ is a directed path in $G$. An important issue is *over what field $\mathbb{F}$ is $p_G$ defined?* A central part of the algorithm is indeed choosing the appropriate field to work over. Another issue is how efficiently $p_G$ can be evaluated? (Note that it potentially contains $n^k$ different monomials.) Williams shows in [@W09] that using the adjacency matrix of $G$ it can be computed in ${\mathrm{poly}}(n)$-time. See Section \[sec:path-to-poly\]. For now, think of $p_G$ as defined over $\mathbb{Q}$, i.e., having integer coefficients. It is easy to see that $G$ contains an $r$-simple path of length $k$ if and only if $p_G$ contains a monomial such that the individual degrees of all variables are at most $r$. Let us call such a monomial an *$r$-monomial*. Thus our task is reduced to checking whether a homogenous polynomial of degree $k$ contains an $r$-monomial. #### Checking whether $p_G$ contains an $r$-monomial Let us assume in this overview for simplicity that $p=r+1$ is prime. Let us view $p_G$ as a polynomial over ${\mathbb{F}_p}$. One problem with doing this is that if we have $p$ directed paths of length $k$ passing through the same vertices in different order, this translates in $p_G$ to $p$ copies of the same monomial summing up to $0$. To avoid this we need to look at a variant of $p_G$ that contains auxiliary variables that prevent this cancelation. For details on this issue see [@AB13] and Section \[sec:path-to-poly\]. For this overview let us assume this does not happen. Recall that we have the equality $a^p=a$ for any $a\in {\mathbb{F}_p}$. Let us look at a monomial that *is not* an $r$-monomial, say $x_1^{r+1}\cdot x_2=x_1^p\cdot x_2$. The equality mentioned implies this monomial is equivalent as a function from ${\mathbb{F}_p}^n$ to ${\mathbb{F}_p}$ to the monomial $x_1\cdot x_2$. By the same argument, any monomial that is *not* an $r$-monomial will be ‘equivalent’ to one of smaller degree. More generally, $p_G$ that is homogenous of degree $k$ over $\mathbb{Q}$ will be equivalent to a polynomial of degree smaller than $k$ as a function from ${\mathbb{F}_p}^n$ to ${\mathbb{F}_p}$ if and only if it *does not* contain an $r$-monomial. Thus, we have reduced our task to the problem of *low-degree testing*. In this context, this problem is as follows: Given black-box access to a function $f:{\mathbb{F}_p}^n\to {\mathbb{F}_p}$ of degree at most $k$, determine whether it has degree exactly $k$ or less than $k$, using few queries to the function. Here, for a function $f:{\mathbb{F}_p}^n\to{\mathbb{F}_p}$, by its degree we mean the total degree of the lowest-degree polynomial $p\in {\mathbb{F}_p}[x_1,\ldots,x_n]$ representing it. Haramati, Shpilka and Sudan [@HSS] gave an optimal solution (in terms of the number of queries) to this problem for prime fields. A little work is required to determine the exact running time of the test of [@HSS] (in addition to the bound on the number of queries to $f$). See Section \[sec:ldtest\] for details. For details on dealing with the case that $r+1$ is not prime, see Section \[sec:rmon\]. Definitions and Preliminary Results {#sec:prelim_results} =================================== In this section we give some definitions and preliminary results that will be used throughout this paper. Let $G(V,E)$ be a directed graph where $V$ is the set of vertices and $E\subseteq V\times V$ the set of edges. We denote by $n=|V|$ the number of vertices in the graph and by $m=|E|$ the number of edges in the graph. A $k$-[*path*]{} or a [*path of length $k$*]{} is a sequence $\rho=v_1,\ldots,v_{k}$ such that $(v_i,v_{i+1})$ is an edge in $G$ for all $i=1,\ldots,k-1$. A [*path*]{} is a $k$-path for some integer $k>0$. A path $\rho$ is called [*simple*]{} if all the vertices in the path are distinct. We say that a path $\rho$ in $G$ is *$r$-simple*, if each vertex of $G$ appears in $\rho$ at most $r$ times. Obviously, a simple path is a $1$-simple path. Given as input a directed graph $G$ on $n$ vertices, the [$r$-SIMPLE $k$-PATH]{} problem asks for a given $G$ whether it contains an $r$-simple path of length $k$. When $r=1$ then the problem is called [SIMPLE $k$-PATH]{}. The [$r$-SIMPLE PATH]{} problem asks for a given $G$ and integer $k$ whether $G$ contains an $r$-simple $k$-path of length $k$. The problem SIMPLE PATH is [$1$-SIMPLE PATH]{}. In this paper we will study the above problems. The following result gives a reduction from [$r$-SIMPLE $k$-PATH]{} to [SIMPLE $k$-PATH]{}. If [$r$-SIMPLE $k$-PATH]{} can be solved in time $T(r,k,n,m)$ then [$sr$-SIMPLE $k$-PATH]{} can be solved in time $T(r,k,sn,s^2m)$. In particular, If [SIMPLE $k$-PATH]{} can be solved in time $T(k,n,m)$ then [$r$-SIMPLE $k$-PATH]{} can be solved in time $T(k,rn,r^2m)$. Let $G$ be a directed graph. Define the graph $G'=G\odot I_s$ where each vertex $v$ in $G$ is replaced with an independent set $I_s$ of size $s$ in $G'$ with the vertices $v^{(1)},\ldots,v^{(s)}$. Each edge $(u,v)$ in $G$ is replaced by the edges $(u^{(i)},v^{(j)})$, $1\le i,j\le s$. It is easy to see that there is a $rs$-simple $k$-path in $G$ if and only if there is a $r$-simple $k$-path in $G'$. We now show that [$r$-SIMPLE PATH]{} is NP-complete. \[thm:simpathNPC\] For any $r$ the decision problem [$r$-SIMPLE PATH]{} is NP-complete. We will reduce deciding [HAMILTONIAN-PATH]{} on a graph of $n$ vertices, to deciding [$r$-SIMPLE $(2rn-n+2)$-PATH]{} on a graph of $2\cdot n$ vertices.\ Given an input graph $G=(V,E)$ to [HAMILTONIAN-PATH]{}, we define a new graph $G'=(V',E')$ as follows. We let $V' = V \bigcup \bar{V}$, where $\bar{V}=\{ \bar{v}_1,\bar{v}_2,...,\bar{v}_n\}$ and $E'=E\bigcup \bar{E}$ where $$\bar{E}=\{(\bar{v}_i,v_i),(v_i,\bar{v}_i) | i\in[n] \}.$$ For $j\in [n]$, it will be convenient to denote by $\rho_j$, the path of length $2r-1$ that begins at $v_j$, goes back and forth from $v_j$ to $\bar{v}_j$ and ends in $v_j$, i.e., $\rho_j \triangleq (v_j,\bar{v}_j,\ldots,v_j,\bar{v}_j,v_j)$. We make 2 observations.\ 1. If a vertex $\bar{v}_j\in \bar{V}$ appears $r$ times in an $r$-simple path $\rho$ then it must be the start or end vertex of $\rho$: To see this, note that by construction of $G'$, if $\bar{v}_j$ is *not* the start or end vertex of $\rho$, visiting it $r$ times requires visiting $v_j$ $r+1$ times. 2. Suppose $\rho$ is an $r$-simple path that begins and ends in a vertex of $V$. If $\rho$ visits a vertex $\bar{v}_j\in \bar{V}$ $r-1$ times, then it must contain $\rho_j$ as a subpath: To see this, note that as $\rho$ does not start in $\bar{v}_j$, any visit to $\bar{v}_j$ must have a visit to $v_j$ before and after. The only way this would sum up to at most $r$ visits in $v_j$ is if these visits where continuous. In other words, only if $\rho$ contains $\rho_j$. We want to show that $G$ contains a Hamiltonian path if and only if $G'$ contains an $r$-simple path of length $2rn-n+2$. Assume first that $G$ contains a Hamiltonian path $v_{i_1}\cdot v_{i_2}\cdots v_{i_n}$. Define the path $\rho = \bar{v}_{i_1}\cdot\rho_{i_1}\cdot\rho_{i_2}\cdots\rho_{i_n}\cdot \bar{v}_{i_n}$. It is of length $$n\cdot(2r-1)+2 = 2rn - n +2,$$ and it is $r$-simple.\ Now assume that we have an $r$-simple path $\rho$ in $G'$ of length $2nr-n+2$. We first claim that $\rho$ must start and end with a vertex from $\bar{V}$: Otherwise, using the first observation above, $\rho$ contains at most $n+1$ vertices appearing $r$ times, and thus has length at most $$(n+1)\cdot r + (n-1)\cdot (r-1) = 2rn - n+1 .$$ Let $\rho'$ be the path $\rho$ with the first and last vertex deleted. So $\rho'$ has length $2rn - n$ and begins and ends in a vertex of $V$. Note that by the first observation $\rho'$ visits all vertices of $\bar{V}$ at most $r-1$ times. We now claim that for every $j\in [n]$, $\rho'$ must contain $\rho_j$ as a subpath. Otherwise, by the second observation, $\rho'$ visits *some* vertex of $\bar{V}$ *less* than $r-1$ times. In this case it has length less than $n\cdot r + n\cdot(r-1) = 2nr -n$. A contradiction. Thus $\rho'$ contains every $\rho_j$ as a subpath. It cannot contain anything else ‘between’ the $\rho_j$’s, as then it would visit some vertex of $V$ *more* than $r$ times. So $$\rho'=\rho_{i_1}\cdots\rho_{i_n},$$ for some ordering $i_1,\ldots,i_n$ of $[n]$. It follows that $v_{i_1}\cdots v_{i_n}$ is a Hamiltonian path in $G$. The above result implies \[LoB\] If [$r$-SIMPLE $k$-PATH]{} can be solved in $T(r,k,n,m)$ time then [HAMILTONIAN-PATH]{} can be solved in $T(r,2rn-n+2,2n,m+2n)$. In particular, if there is an algorithm for [$r$-SIMPLE $k$-PATH]{} that runs in time $poly(n)\cdot 2^{(c/2)(k/r)}$ then there is an algorithm for [HAMILTONIAN-PATH]{} that runs in time $poly(n)\cdot 2^{cn}$. Gap === In this section we show that the gap between the longest simple path and the longest $r$-simple path can be exponentially large even for $r=2$. We first give the following lower bound for the gap If $G$ contains an $r$-simple path of length $k$ then $G$ contains a simple path of length $\lceil \frac{\log k}{\log r}\rceil$. Let $t=\lfloor \frac{\log {k} -1}{\log r}\rfloor$. Let $\rho$ be an $r$-simple path whose first vertex is $v_0$. We will use $\rho$ to construct a simple path $\bar{\rho}$ of length $\lfloor\frac{logk}{logr}\rfloor$. We denote $\rho_0=\rho$. As $v_0$ appears at most $r$ times in $\rho_0$, there must be a subpath $\rho_1$ of $\rho_0$ of length at least $(k-r)/r$ where $v_0$ does not appear. Let $v_1$ be the first vertex of $\rho_1$. Similarly, for $1< i \leq t$, we define the subpath $\rho_i$ of $\rho_{i-1}$ to be a subpath of length at least $$(k-r-\ldots -r^i)/r^i \geq (k-r^{i+1})/r^i,$$ where $v_1,\ldots,v_{i-1}$ do not appear, and define $v_i$ to be the first vertex of $\rho_i$. Note that we can always assume there is an edge from $v_{i-1}$ to $v_i$ as we can start $\rho_i$ just after an appearance of $v_{i-1}$ in $\rho_{i-1}$. Note that for $1\leq i \leq t$, such a $v_i$ as defined indeed exists as $(k-r^{i+1})/r^i\geq 1$ when $$k\geq 2\cdot r^{i+1}\leftrightarrow i+1 \leq (\log k -1)/\log r$$ Thus, $v_0\cdots v_{t-1}$ is a simple path of the desired length. Before we give the upper bound we give the following definition. A [*full $r$-tree*]{} is a tree where each vertex has $r$ children and all the leaves of the tree are in the same level. The root is on level $1$. \[Gap\] There is a graph $G$ that contains an $r$-simple path of length $k$ and no simple path of length greater than $4\log k/\log r$. We first give the proof for $r\ge 3$. Consider a full $(r-1)$-tree of depth $\lceil \log n/\log(r-1)\rceil$. Remove vertices from the lowest level (leaves) so the number of vertices in the graph is $n$. Obviously there is an $r$-simple path of length $k\ge n$. Any simple tour in this tree must change level at each step and if it changes from level $\ell$ to level $\ell+1$ it cannot go back in the following step to level $\ell$. So the longest possible simple path is $2\lceil \log n/\log(r-1)\rceil-2\le 3.17 (\log k/\log r)$. For $r=2$ we take a full binary tree ($2$-tree) and add an edge between every two children of the same vertex. The $2$-simple path starts from the root $v$, recursively makes a tour in the left tree of $v$ then moves to the root of the right tree of $v$ (via the edge that we added) then recursively makes a tour in the right tree of $v$ and then visit $v$ again. Obviously this is a $2$-simple path of length $k>n$. A simple tour in this graph can stay in the same level only twice, can move to a higher level or can move to a lower level. Again here if it moves from level $\ell$ to $\ell+1$ it cannot go back in the following step to level $\ell$. Therefore the longest simple path is of length at most $4\log n \leq 4\log k$. From $r$-Simple $k$-Path to Multivariate Polynomial {#sec:path-to-poly} =================================================== The purpose of this section is to reduce the question of whether a graph $G$ contains an $r$-simple $k$-path, to that of whether a certain multivariate polynomial *contains an $r$-monomial*, as defined below. Fix a field ${\mathbb{F}}$. Fix a monomial $M= M({{\boldsymbol z}})=z_1^{i_1}\cdots z_t^{i_t}$. - We say $M$ is an *$r$-monomial* if no variable appears with degree larger than $r$ in $M$. That is, for all $1\leq j \leq t$, $i_j\leq r$. - Let $f({{\boldsymbol z}})$ be a multivariate polynomial over ${\mathbb{F}}$. We say $f$ *contains an $r$-monomial*, if there is an $r$-monomial $M({{\boldsymbol z}})$ appearing with a nonzero coefficient $c\in {\mathbb{F}}$ in $f$. We now describe this reduction. Let $G(V,E)$ be a directed graph where $V=\{1,2,\ldots,n\}$. Let $A$ be the adjacency matrix and $B$ be the $n\times n$ matrix such that $B_{i,j}=x_i\cdot A_{i,j}$ where $x_i$, $i=1,\ldots,n$ are indeterminates. Let $\bf1$ be the row $n$-vector of $1$s and ${{\boldsymbol x}}=(x_1,\ldots,x_n)^T$. Consider the polynomial $p_G({{\boldsymbol x}})={{\boldsymbol 1}}\cdot B^{k-1}\cdot {{\boldsymbol x}}$. It is easy to see $$p_G({{\boldsymbol x}})=\sum_{i_1\to i_2\to \cdots \to i_{k} \in G} x_{i_1}\cdots x_{i_{k}}$$ where $i_1\to i_2\to \cdots \to i_{k} \in G$ means that $i_1, i_2, \cdots , i_{k}$ is a directed path in $G$. Obviously, for field of characteristic zero there is an $r$-simple $k$-path if and only if $p_G({{\boldsymbol x}})$ contains an $r$-monomial. For other fields the later statement is not true. For example, in undirected graph, $k=2$, and $r=1$ if $(1,2)\in E$ and the field is of characteristic $2$ then the monomial $x_1x_2$ occurs twice and will vanish in $p_G({{\boldsymbol x}})$. We solve the problem as follows. Let $B^{(m)}$ be an $n\times n$ matrices, $m=2,\ldots,k$, such that $B^{(m)}_{i,j}=x_i\cdot y_{m,i}\cdot A_{i,j}$ where $x_i$ and $y_{m,i}$ are indeterminates. Let, ${{\boldsymbol y}}=({{\boldsymbol y}}_1,\ldots,{{\boldsymbol y}}_{k})$ and ${{\boldsymbol y}}_m =(y_{m,1},\ldots,y_{m,n})$. Let ${{\boldsymbol x}}\centerdot{{\boldsymbol y}}=(x_1y_{1,1},\ldots,x_ny_{1,n})$. Consider the polynomial $P_G({{\boldsymbol x}},{{\boldsymbol y}})={{\boldsymbol 1}}B^{(k)}B^{(k-1)}\cdots B^{(2)}({{\boldsymbol x}}\centerdot{{\boldsymbol y}})$. It is easy to see that $$P_G({{\boldsymbol x}},{{\boldsymbol y}})=\sum_{i_1\to i_2\to \cdots \to i_{k} \in G} x_{i_1}\cdots x_{i_{k}} y_{1,i_1}\cdots y_{k,i_{k}}$$ Obviously, no two paths have the same monomial in $P_G$. Note that as $P_G$ contains only ${\{0,1\}}$ coefficients, we can define it over any field ${\mathbb{F}}$. It will actually be convenient to view it as a polynomial $P_G({{\boldsymbol x}})$ whose coefficients are in the field of rational functions ${\mathbb{F}}({{\boldsymbol y}})$. Therefore, for any field, there is an $r$-simple $k$-path if and only if $P_G({{\boldsymbol x}},{{\boldsymbol y}})$ contains an $r$-monomial in ${{\boldsymbol x}}$. We record this fact in the lemma below. \[lem:path\_to\_pG\] Fix any field ${\mathbb{F}}$. The graph $G$ contains an $r$-simple $k$-path if and only if the polynomial $P_G$, defined over ${\mathbb{F}}({{\boldsymbol y}})$, contains an $r$-monomial $M({{\boldsymbol x}})$. Low Degree Tester {#sec:ldtest} ================= In this section we present a tester that determines whether a function $f:{\mathbb{F}_p}^n\to {\mathbb{F}_p}$ of degree *at most* $d$ has, in fact, degree *less than* $d$. The important point is that the tester will be able to do this using few black-box queries to $f$. The results of this section essentially follow from the work of Haramaty, Sudan and Shpilka [@HSS]. First, let us say precisely what we mean by the *degree* of a function $f:{\mathbb{F}_p}^n\to {\mathbb{F}_p}$. We define this to be the degree of the lowest degree polynomial $f'\in{\mathbb{F}_p}[{{\bf x}}]$ that agrees with $f$ as a function from ${\mathbb{F}_p}^n$ to ${\mathbb{F}_p}$. It is known from the theory of finite fields that there is a unique such $f'$, and that the individual degrees of all variables in $f'$ are smaller than $p$. Moreover, given any polynomial $g\in{\mathbb{F}_p}[{{\boldsymbol x}}]$ agreeing with $f$ as a function from ${\mathbb{F}_p}^n$ to ${\mathbb{F}_p}$, $f'$ can be derived from $g$ by replacing, for any $1\leq i \leq n$, occurrences of $x_i^t$ with $x_i^t \mod x_i^p-x_i$ (i.e., $x_i^{((t-1)\mod (p-1))+1}$ when $t\not=0$). We do not prove these basic facts formally here. They essentially follow from the fact that $a^p=a$ for $a\in {\mathbb{F}_p}$. This motivates the following definition. \[dfn:degp\] Fix positive integers $n,d$ and a prime $p$. Let $f\in{\mathbb{F}_p}[{{\boldsymbol x}}] = {\mathbb{F}_p}[x_1,\ldots,x_n]$. We define $\deg_p(f)$ to be the degree of the polynomial $f$ when replacing, for $1\leq i \leq n$, $x_i^t$ by $(x_i^t\mod x_i^p-x_i)$. More formally, $\deg_p(f)\triangleq \deg(f')$ where $$f'(x_1,\ldots,x_n) \triangleq f(x_1,\ldots,x_n)\; \mathrm{mod} \;x_1^p-x_1,\;\ldots,\;\mathrm{mod}\; x_n^p-x_n.$$ Moreover, for a function $g:V \to {\mathbb{F}}_p$ where $V \subseteq {\mathbb{F}}^n_p$ is a subspace of dimension $k$, we define $\deg_p(g)=\min_f deg_p(f)$ where $f\in {\mathbb{F}}_p [x_1,...,x_n]$ and $f|_V=g$. Here $g$ can be regarded as a function in ${\mathbb{F}}_p[x_1,\ldots,x_k]$. We note that this notion of degree is affine invariant, i.e does change after affine transformations. In addition it has the property that for any affine subspace $V$, $\deg_p(f|_V) \leq \deg_p(f)$. We now present the main result of this section. \[lem: low degree testing\] There is a randomized algorithm $A$ running in time ${\mathrm{poly}}(n)\cdot p^{\left\lceil \frac{d}{p-1}\right\rceil +1}$ that determines with constant one-sided error whether a function $f$ of degree at most $d$ has degree less than $d$. More precisely, given black-box access to a function $f:{\mathbb{F}_p}^n\to {\mathbb{F}_p}$ with $\deg_p(f)\leq d$, - If $\deg_p(f)= d$, $A$ accepts with probability at least $99/100$. - If $\deg_p(f)<d$, $A$ rejects with probability one. The result essentially follows from the work of Haramaty, Shpilka and Sudan [@HSS]. A technicality is to analyze the precise running time, and not just the query complexity as in [@HSS] Before proving Lemma \[lem: low degree testing\], we state some required preliminary lemmas. \[lem: not-so-low degree testing\] Suppose we have black-box access to a function $f:{\mathbb{F}_p}^t\to {\mathbb{F}_p}$. Then we can determine in deterministic time $O(p^t)$ whether $deg_p(f)\geq (p-1)\cdot t$. Consider the algorithm that yields a positive answer if and only if $\sum_{{{\bf a}}\in{\mathbb{F}}_{p}^{t}}f({{\bf a}})=0$ . It is clear that the running time is indeed $O(p^{t})$. Let us now show correctness. As in Definition \[dfn:degp\], define $$f'(x_1,\ldots,x_n) \triangleq f(x_1,\ldots,x_n)\; \mathrm{mod} \;x_1^p-x_1,\;\ldots,\;\mathrm{mod}\; x_n^p-x_n,$$ so that $\deg_p(f) = \deg(f')$. We show that 1. The only monomial of degree $\geq t(p-1)$ in $f'$ is $M_{\max}\triangleq \prod_{i=1}^{t}x_{i}^{p-1}$ and 2. the coefficient of $M_{\max}$ in $f'$, is $(-1)^{t}\cdot \sum_{{{\bf a}}\in{\mathbb{F}_p}^t}f({{\bf a}})$.  From these two items, it is clear that indeed $\deg_p(f)=\deg(f') \geq t\cdot (p-1)$ if and only if $ \sum_{{{\bf a}}\in{\mathbb{F}_p}^t}f({{\bf a}})\neq 0$. The first item is obvious, as the individual degrees in $f'$ are at most $p-1$. For the second item, let us calculate the coefficient of $M_{\max}$ in $f'$. For every ${{\bf a}}\in {\mathbb{F}_p}^t$, consider the function $g_{{{\bf a}}}:{\mathbb{F}_p}^t\to {\mathbb{F}_p}$ that is one on ${{\bf a}}$ and zero elsewhere. One can verify that $g_{{{\bf a}}}({{\bf x}})=\prod_{i=1}^{t}\frac{\prod_{\alpha\in{\mathbb{F}_p}\backslash\{{{\bf a}}_i\}}({{\bf x}}_{i}-\alpha)}{\prod_{\beta\in{\mathbb{F}_p}\backslash\{0\}}\beta}$. Clearly, the coefficient of $M_{\max}$ in $g_{{{\bf a}}}$ is $(\prod_{\beta\in{\mathbb{F}_p}\backslash\{0\}}\beta)^{-t}=(-1)^t$. Note that in $g_{{{\bf a}}}$, all individual degrees are smaller than $p$. Hence, $f'=\sum_{{{\bf a}}\in{\mathbb{F}_p}^{t}}f({{\bf a}})\cdot g_{{{\bf a}}}$ and the coefficient of $M_{\max}$ in $f'$ is $(-1)^{t}\cdot \sum_{{{\bf a}}\in {\mathbb{F}_p}^t} f({{\bf a}})$. The algorithm for Lemma \[lem: low degree testing\] checks the degree of the function only on a small subspace. The key for its correctness is to show that when you restrict the function to a subspace (even for $n-1$ dimensional subspace) the degree does not decrease with high probability. The Lemma appeared in [@HSS]. We give a proof sketch here for completeness \[lem:HSS-step\] Let ${\mathbb{F}}_{p}$ be a field of prime size $p$ and $f:{\mathbb{F}_p}^n\to{\mathbb{F}_p}$ be a function with $\deg_p(f) = t(p-1)$. The number of hyperplanes $H$ such that $\deg_p(f|_{H})<t(p-1)$ is at most $p^{t+1}$ We will assume w.l.o.g that $f$ has the monomial $\prod_{i=1}^{t}x_{i}^{p-1}$. One can show that for any degree $t(p-1)$ polynomial $f$ there is linear transformation $A$ such that $f(Ax)$ has the monomial $\prod_{i=1}^{t}x_{i}^{p-1}$. So it will be enough to prove the lemma for the suitable transformation of $f$. We will assume for simplicity that all the hyperplanes are of the form of $H_{\alpha}=\{x\mid x_{1}=\sum_{i=2}^{n}\alpha_{i}x_{i}+\alpha_{0}\}$ for some $\alpha_{2},\ldots,\alpha_{n}$. Indeed, there are few more hyperplanes that does not depend on the first coordinate, but they don’t contribute much to the upper bound. To prove the lemma we will show that for any of the $p^{t}$ possible values for $\alpha_{2},\ldots,\alpha_{t},\alpha_{0}$ there are $<p$ possibilities for $\alpha_{t+1},\ldots,\alpha_{n}$ such that $\deg(f|_{H_{\alpha}})<t(p-1)$. Fix $\alpha_{2},\ldots,\alpha_{t},\alpha_{0}$. For simplicity we assume they are all zero, but the same bound goes for any $\alpha_{2},\ldots,\alpha_{t},\alpha_{0}$ (one can reduce the general case to the zero case by some affine transformation). Now consider all the monomials $M$ in $f$ with the following properties: (1) $M$ divides $\prod_{i=2}^{t}x_{i}^{p-1}$ and (2) $\deg(M)=t(p-1)$. We can write the sum of all those monomials as $ $$\prod_{i=2}^{t}x_{i}^{p-1}g(x_{1},x_{t+1},\ldots,x_{n})$. By definition, $g$ is homogenous polynomial of degree $p-1$. Because $\prod_{i=1}^{t}x_{i}^{p-1}$ is a monomial of $f$, $x_{1}^{p-1}$ is a monomial of $g$. Because the hyperplanes does not depend on the variables $x_{2},\ldots,x_{t}$ (recall, we assumed $\alpha_2 = \cdots = \alpha_t = 0$) the degree of $f$ can decrease on $H_{\alpha}$ only if the degree of $g$ decrees on $H_{\alpha}$. Because $g$ is homogenous of degree $p-1$ and we consider only linear hyperplanes of the form $x_{1}=L(x_{t+1},\ldots,x_{n})$, then $g|_{x_{1}=L}$ is still homogenous of degree $p-1$, so if the degree $\deg(g|_{x_{1}=L})<p-1$ then $g|_{x_{1}=L}\equiv0$. Now consider $g$ as an univariate polynomial in $x_{1}$ over the field of rational functions in $x_{t+1},\ldots,x_{n}$. In this view our question is: how many field elements $L\in{\mathbb{F}}_{p}(x_{t+1},\ldots,x_{n})$ are there such that $g(L)=0$. From the fundamental theorem of the algebra the answer is $p-1$ and we are done.  From Lemma \[lem:HSS-step\] we get the following corollary. \[cor: low deg tester: completeness\] Let $n>t$ and $f:{\mathbb{F}}_p^n \to {\mathbb{F}}_p$ be a polynomial such that $\deg_p(f)=t(p-1)$. Then ${{\bf Pr}}_{V}\left[\deg_p(f|_{V})=t(p-1)\right]\geq\frac{1}{p+1}\prod_{k=1}^{n-t-1}\left(1-p^{-k}\right)=\Omega(\frac{1}{p})$, where $V$ is a random $t$-dimensional affine subspace. We proceed by induction on $n$. Consider first the base case, where $n=t+1$. In this case the number of $t$-dimensional affine subspaces $V \subseteq {\mathbb{F}}^{t+1}_p$ is $\frac{p^{t+2} -1}{p-1} > p^{t+1} + p^{t}$. By Lemma \[lem:HSS-step\] on at most $p^{t+1}$ of them $\deg(f|_V) <t(p-1)$ so the probability that $\deg(f|_V) = t(p-1)$ is $\frac{1}{p+1}$ as required. Now assume the claim is true for $n-1$, and consider the following way of choosing a random $t$-dimensional affine subspace $V$. First choose a random hyperplane $H\subseteq {\mathbb{F}}_p^n$ and then choose a random $t$-dimensional affine subspace $V\subseteq H$. There are more than $p^n$ hyperplanes $H\subseteq {\mathbb{F}}_p^n$, so by Lemma \[lem:HSS-step\] the probability that $\deg_p(f|_H)=t(p-1)$ is at least $1-p^{t+1-n}$. Moreover, in the event that $\deg_p(f|_H)=t(p-1)$, we can apply the induction hypothesis to $f|_H$. Hence, $$\begin{aligned} {{\bf Pr}}\left[\deg_p(f|_{V})=t(p-1)\right] & = & {{\bf Pr}}\left[\deg_p((f|_{H})|_{V})=t(p-1)\mid\deg_p(f|_{H})=t(p-1)\right]\cdot{{\bf Pr}}\left[\deg_p(f|_{H})=t(p-1)\right]\\ & \leq & \frac{1}{p+1}\prod_{k=1}^{(n-1)-t-1}\left(1-p^{-k}\right)\cdot(1-p^{t+1-n})=\frac{1}{p+1}\prod_{k=1}^{n-t-1}\left(1-p^{-k}\right)\end{aligned}$$ We are now ready to prove Lemma \[lem: low degree testing\]. Let $t= \left\lceil \frac{d}{p-1}\right\rceil$. We assume without lost of generality that $d = t(p-1)$: Otherwise, let $a=t(p-1)-d$ and consider the function $f'(x_0,x_1,....,x_n) \triangleq x_0^a\cdot f(x_1,...,x_n)$. It is easily checked that $\deg_p(f')\leq t(p-1)$. Also $\deg_p(f')=t(p-1)$ if and only if $\deg_p(f) =d $. We will present an algorithm for the problem with one sided error probability $1-\Omega\left(\frac{1}{p}\right)$ that runs in time ${\mathrm{poly}}(n)\cdot O(p^t)$. By repeating it $O(p)$ times, we can get down to error probability $1/100$ in running time ${\mathrm{poly}}(n)\cdot O(p^{t+1})$ as required. Consider the following algorithm. Choose a random $t$-dimensional affine subspace $V$. Accept if and only if the $\deg_p(f|_V)<t(p-1)$. Assume first that $\deg_p(f)<t(p-1)$. Then for any affine subspace $V$, $\deg_p(f|_V) \leq \deg_p(f)<t(p-1)$. On the other hand, if $\deg_p(f|_V)= t(p-1)$, Corollary \[cor: low deg tester: completeness\] implies we will accept with probability at least $\Omega(\frac{1}{p})$. We conclude by analyzing the running time. Choosing $V$ can be done in ${\mathrm{poly}}(n)$-time. For checking whether $\deg_p(f|_V) =t(p-1)$, Lemma \[lem: not-so-low degree testing\] gives running $O(p^t)$ assuming black-box access to $f|_V$. Given black-box access to $f$, we can compute $f|_V({{\bf a}})$ for ${{\bf a}}\in {\mathbb{F}_p}^t$ in ${\mathrm{poly}}(n)$-time. The claimed running time of ${\mathrm{poly}}(n)\cdot O(p^t)$ follows. Testing if $P_G$ contains an $r$-monomial {#sec:rmon} ========================================= In this section we present a method for testing whether the polynomial $P_G$, described in Section \[sec:path-to-poly\], contains an $r$-monomial. This is done using the low-degree tester from the previous section. As stated in Lemma \[lem:path\_to\_pG\], this is precisely equivalent to whether $G$ contains an $r$-simple $k$-path. Recall we viewed $P_G$ as a polynomial over a field of rational functions ${\mathbb{F}_p}({{\bf y}})$. To obtain efficient algorithms, we first reduce the question to checking whether a different polynomial defined over ${\mathbb{F}_p}$ rather than ${\mathbb{F}_p}({{\bf y}})$ contains an $r$-monomial. It is important in the next Lemma that we are able to do this reduction for *any* $p$, in particular a ‘small’ one. \[lem:only\_x\_var\] Fix any integers $r,k$, with $r\leq k$. Let $p$ be any prime and $t=\lceil \log_p {10k} \rceil $. Let $G$ be a directed graph on $n$ vertices. Given an adjacency matrix $A_G$ for $G$, we can return in ${\mathrm{poly}}(n)$-time ${\mathrm{poly}}(n)$-size circuits computing polynomials $f^1_G,\ldots,f^t_G:{\mathbb{F}_p}^n\to {\mathbb{F}_p}$ on inputs in ${\mathbb{F}_p}^n$ such that - For $1\leq i \leq t$, $f^i_G$ is (either the zero polynomial or) homogenous of degree $k$. - If $G$ contains an $r$-simple $k$-path then with probability at least $9/10$, for some $1\leq i\leq t$, $f^i_G$ contains an $r$-monomial. - If $G$ does not contain an $r$-simple $k$-path, for all $1\leq i \leq t$, $f^i_G$ does not contain an $r$-monomial. Note that the discussion in Section \[sec:path-to-poly\] implies we can compute $P_G$ in ${\mathrm{poly}}(n)$-time over inputs in ${\mathbb{F}}_{p}^{2n}$. We choose random ${{\bf b}}\in {\mathbb{F}}_{p^t}^n$ and let $$f_G({{\bf x}})\triangleq P_G({{\bf x}},{{\bf b}}).$$ Suppose $P_G$, as a polynomial over ${\mathbb{F}}({{\bf y}})$, contains an $r$-monomial $M'({{\bf x}})$. The coefficient $c_{M'}({{\boldsymbol y}})$ of $M'$ in $P_G$ is a nonzero polynomial of degree $k$. So, by the Schwartz-Zippel Lemma, $c_{M'}({{\bf b}})=0$ with probability at most $k/p^t\leq 1/10$. In the event that $c_{M'}({{\bf b}})\neq 0$, $f_G({{\bf x}})$ is a homogenous polynomial of degree $k$ in ${\mathbb{F}}_{p^t}[{{\bf x}}]$ containing an $r$-monomial. Let us assume from now on, we chose a ${{\bf b}}$ such that indeed $a_{M'}\triangleq c_{M'}({{\bf b}})\neq 0$. We now discuss how to end up with polynomials having coefficients in ${\mathbb{F}_p}$ rather than ${\mathbb{F}}_{p^t}$. Let $T_1,\ldots, T_t:{\mathbb{F}}_{p^t}\to {\mathbb{F}_p}$ be independent ${\mathbb{F}_p}$-linear maps. Suppose $f_G = \sum_M a_M\cdot M({{\bf x}})$. For $1\leq i \leq t$, define a polynomial $f_G^i\in {\mathbb{F}_p}[{{\bf x}}]$ by $$f_G^i ({{\boldsymbol x}}) \triangleq \sum_M T_i(a_M)\cdot M({{\boldsymbol x}}).$$ Note that for all $1\leq i \leq t$, $f_G^i$ is the zero polynomial or homogenous of degree $k$. As $a_{M'}\neq 0$, for some $i$, $T_i(a_{M'})\neq 0$. For this $i$, $f_G^i$ is homogenous of degree $k$ and contains an $r$-monomial, specifically, the $r$-monomial $a_{M'}\cdot M'({{\bf x}})$. We claim that for all $1\leq i \leq t$, $f_G^i$ can be computed by a $poly(n)$-size circuit on inputs ${{\bf a}}\in {\mathbb{F}_p}^n$. This is because $f_G$ and $T_i$ are efficiently computable, and because for ${{\bf a}}\in {\mathbb{F}_p}^n$, $$T_i(f_G({{\bf a}})) = T_i\left(\sum_M a_M\cdot M({{\bf a}})\right) = \sum_M T_i(a_M)\cdot M({{\bf a}}) = f_G^i({{\bf a}}),$$ where the second equality is due to the ${\mathbb{F}_p}$-linearity of $T_i$. The above lemma implies \[cor:enough\_over\_Fp\] Fix any prime $p$. Suppose that given black-box access to a polynomial $g\in {\mathbb{F}_p}[{{\bf x}}]$ that is homogenous of degree $k$, we can determine in time ${\mathrm{poly}}(n)\cdot S$ if it contains an $r$-monomial. Then we can also determine in time ${\mathrm{poly}}(n)\cdot S$ whether $P_G$ as a polynomial over ${\mathbb{F}_p}({{\bf y}})$ contains an $r$ monomial. [Our reduction to low-degree testing is based on the following simple observation that, for the right $p$ and for homogenous polynomials, containing an $r$-monomial is equivalent to having a certain $\deg_p$-degree.]{} \[lem:degiffr-mon\] Suppose $g\in{\mathbb{F}_p}[{{\bf x}}]$ is a homogenous polynomial of degree $k$. Suppose $r=p-1$. Then $\deg_p(g)= k $ if and only if $g$ contains an $r$-monomial. If $g$ contains an $r$-monomial $M$ then, as $r<p$, $\deg_p(M)=k$, which implies that $\deg_p(g)= k$. If $g$ does not contain an $r$-monomial, then for every monomial $M$ in $g$ there is an $i\in [n]$ such that the degree of $x_i$ in $M$ is at least $r+1=p$. So replacing $x_i^p$ by $x_i$ will reduce the degree of $M$ and therefore $\deg_p(M)<k$. Since this happens for all monomials of $g$, $\deg_p(g)< k$. We introduce another element on notation that will be convenient in the rest of this section. Fix integers $n,d$ and prime $p$. Let $f\in {\mathbb{F}_p}[{{\bf x}}]$ be an $n$-variate polynomial of degree at most $d$. We define $LDT(f,n,d,p)$ to be $1$ if $\deg_p(f) = d$, and $0$ otherwise. Before proceeding, we note that the results of Section \[sec:ldtest\] imply that given $n,d,p$ and black-box access to $f$, $LDT(f,n,d,p)$ can be computed in time ${\mathrm{poly}}(n)\cdot O(p^{\left\lceil {d}/{(p-1)}\right\rceil +1})$. In particular, if given ${{\bf a}}\in {\mathbb{F}_p}^n$, we can compute $f({{\bf a}})$ in ${\mathrm{poly}}(n)$-time, then we can compute $LDT(f,n,d,p)$ in time ${\mathrm{poly}}(n)\cdot O(p^{\left\lceil {d}/({p-1})\right\rceil +1})$.\ The following lemma is an easy corollary of Lemma \[lem:degiffr-mon\]. Fix integers $r,k$ with $r\leq k$. Suppose $p=r+1$ is prime. Let $g\in {\mathbb{F}_p}[{{\bf x}}]$ be homogenous of degree $k$ and computable in ${\mathrm{poly}}(n)$-time. There is a randomized algorithm running in time $$poly(n)\cdot O((r+1)^{\left\lceil \frac{k}{r}\right\rceil +1})$$ determining whether $g$ contains an $r$-monomial. The algorithm simply returns $LDT(g,n,d=k,p=r+1)$. The running time follows from the discussion above. The correctness follows from Lemma \[lem:degiffr-mon\]. We wish to have a similar result when $r+1$ is not a prime. \[lem:r,k,p\] Fix integers $r,k$ with $r\leq k$. Let $p$ be the smallest prime such that $\frac{p-1}{r} \in \mathbb{Z}$. Let $g\in {\mathbb{F}_p}[{{\bf x}}]$ be homogenous of degree $k$ and computable by a ${\mathrm{poly}}(n)$-size circuit. There is a randomized algorithm running in time $poly(n)\cdot O(p^{\left\lceil \frac{k}{r}\right\rceil +1})$ determining whether $g$ contains an $r$-monomial. Denote $l\triangleq \frac{p-1}{r}$ and define $$h(x_1,x_2,\ldots ,x_n):= g(x^l_1,x^l_2,\ldots ,x^l_n).$$ The algorithm returns $LDT(h,n,d=k\cdot l,p)$. Note that $h$ is homogenous of degree $k\cdot l$. Note also that $h$ contains an $r\cdot l$-monomial if and only if $g$ contains an $r$-monomial. As $r\cdot l + 1=p$ correctness now follows from Lemma \[lem:degiffr-mon\]. The best known bound for the smallest prime number $p$ that satisfies $r|p-1$ is $r^{5.5}$ due to Heath-Brown [@R96]. This gives a randomized algorithm running in time $$poly(n)\cdot O(r^{\frac{5.5k}{r} +O(1)}).$$ Schinzel, Sierpinski, and Kanold have conjectured the value to be 2 [@R96]. In the following Theorem we give a better bound. We first give the following \[lem:r,k,p,l\] Fix integers $r,k$ with $r\leq k$. Let $p$ be the smallest prime such that there is an $ l \in \mathbb{Z}$ for which $r\cdot l \leq p-1 $ and $(r+1)\cdot l > p-1$. Let $g\in {\mathbb{F}_p}[{{\bf x}}]$ be homogenous of degree $k$ and computable by a ${\mathrm{poly}}(n)$-size circuit. There is a randomized algorithm running in time $$poly(n)\cdot O\left(p^{\left\lceil \frac{l\cdot k}{p-1}\right\rceil +1}\right)$$ determining whether $g$ contains an $r$-monomial. As in the proof of Lemma \[lem:r,k,p\], we define $h(x_1,x_2,...,x_n)\triangleq g(x^l_1,x^l_2,...,x^l_n)$. The algorithm returns $LDT(h,n,d=k\cdot l,p)$. As in the proof of Lemma $\ref{lem:r,k,p}$, $h$ is homogenous of degree $k\cdot l$ and contains an $(r\cdot l)$-monomial if and only if $g$ contains an $r$-monomial. Furthermore, as $r\cdot l \leq p-1$ and $(r+1)\cdot l \geq p$, $h$ contains a $(p-1)$-monomial if and only if $g$ contains an $r$-monomial. Correctness now follows from Lemma \[lem:degiffr-mon\]. The main result of this section contains two results. The first is unconditional. The second is true if Cramer’s conjecture is true. Cramer’s conjecture states that the gap between two consecutive primes $p_{n+1}-p_n=O(\log^2 p_n), \cite{C36}$. (Unconditional Result) Fix any integers $r,k$ with $2\leq r\leq k$. Let $g\in {\mathbb{F}_p}[{{\bf x}}]$ be homogenous of degree $k$ and computable by a ${\mathrm{poly}}(n)$-size circuit. There is a randomized algorithm running in time $${\mathrm{poly}}(n)\cdot O\left(r^{\frac{2k}{r} +O(1)}\right)$$ determining whether $g$ contains an $r$-monomial. (Conditional Result) If Cramer’s Conjecture is true then the time complexity of the algorithm is $${\mathrm{poly}}(n)\cdot O\left( r^{\frac{k}{r}+o\left(\frac{k}{r}\right)}\right).$$ We will find $p$ and $l$ as required in Lemma \[lem:r,k,p,l\]. Fix a prime $p$ such that $r^2 + r +1 <p < 2r^2 + 2r\leq 3r^2 $ . (This can be done as for any positive integer $t>3$, there is always a prime between $t$ and $2t$.) Define $l\triangleq \lfloor \frac{p-1}{r} \rfloor $. We have $$\begin{aligned} r\cdot l = r\cdot \lfloor \frac{p-1}{r}\rfloor \leq p-1 \end{aligned}$$ $$\begin{aligned} (r+1)\cdot l &\geq& (r+1)\cdot (\frac{p-1}{r} -1)\\ &=&(p-1)+\frac{p-1}{r}-r-1 >(p-1) \end{aligned}$$ The first claim now follows from Lemma \[lem:r,k,p,l\] and Corollary \[cor:enough\_over\_Fp\]. If Cramer’s conjecture is true then there is a constant $c$ such that for every integer $x$ there is a prime number in $[x,x+c\log^2(x)]$ . Then there is a prime number $p$ in the interval $[2cr\log^2 r, 2cr\log^2 r+c\log^2(2cr\log^2 r)]$ and we can choose $l=2c\log^2 r$. Then the time complexity will be $$poly(n)\cdot O\left( r^{\frac{k}{r}+o\left(\frac{k}{r}\right)}\right).$$ The following table summarizes the result for $r\le 11$. See Lemma \[lem:r,k,p,l\]. $r$ Result Field and $l$ ------ -------------- ----------------------------- $1$ $2^k$ [@W09] ${\mathbb{F}}_2$ , $l=1$ $2$ $1.73^k$ ${\mathbb{F}}_3$ , $l=1$ $3$ $1.912^k$ ${\mathbb{F}}_7$ , $l=2$ $4$ $1.495^k$ ${\mathbb{F}}_5$ , $l=1$ $5$ $1.615^k$ ${\mathbb{F}}_{11}$ , $l=2$ $6$ $1.383^k$ ${\mathbb{F}}_{7}$ , $l=1$ $7$ $1.533^k$ ${\mathbb{F}}_{23}$ , $l=3$ $8$ $1.424^k$ ${\mathbb{F}}_{17}$ , $l=2$ $9$ $1.387^k$ ${\mathbb{F}}_{19}$ , $l=2$ $10$ $1.27^k$ ${\mathbb{F}}_{11}$ , $l=1$ $11$ $1.329^k$ ${\mathbb{F}}_{23}$ , $l=2$ [X]{} H. Abasi and N. H. Bshouty. A simple algorithm for undirected hamiltonicity. Electronic Colloquium on Computational Complexity (ECCC), 20:12, 2013. Alon, N., Yuster, R., and Zwick, U. 1995. Color-Coding. J. ACM 42, 4 (Jul. 1995), 844-–856. R. Bellman. Dynamic programming treatment of the travelling salesman problem, J. Assoc. Comput. Mach. 9, pp. 61–-63, 1962 R. Bellman. Combinatorial processes and dynamic programming, Combinatorial Analysis (R. Bellman, M. Hall. Eds.), Proceedings of Symposia in Applied Mathematics 10, American Mathematical Society, pp. 217–-249, 1960. H. L. Bodlaender, On linear time minor tests with depth-first search, J. Algorithm. 14(1):1–-23, 1993. A. Björklund, T. Husfeldt, S. Khanna. Approximating Longest Directed Paths and Cycles. ICALP 2004: 222–233 A. Björklund, T. Husfeldt, P. Kaski, M. Koivisto. Narrow sieves for parameterized paths and packings. CoRR abs/1007.1161 (2010). R. C. Baker, G. Harman and J. Pintz. The Difference between Consecutive Primes, II. Proc. London Math. Soc. (3) 83 pp. 532–562. (2001). H. Cramer. On the order of magnitude of the difference between consecutive prime numbers, Acta Arithmetica 2, pp. 23-–46, (1936) J. Chen, S. Lu, S.-H. Sze, and F. Zhang, Improved algorithms for path, matching, and packing problems, in Proc. 18th Annual ACM–SIAM Symposium on Discrete Algorithms, SODA 2007 (Philadelphia, PA, USA, 2007), pp. 298-–307. H. N. Gabow, S. Nie. Finding Long Paths, Cycles and Circuits. ISAAC 2008, pp. 752–763. E. Haramaty, A. Shpilka, M. Sudan. Optimal Testing of Multivariate Polynomials over Small Prime Fields. SIAM J. Comput. 42(2): 536–562 (2013) I. Koutis, Faster algebraic algorithms for path and packing problems, in Proc. 35th International Colloquium on Automata, Languages and Programming, ICALP (Reykjavik, Iceland, July 7–11, 2008), Springer LNCS 5125, pp. 575–-586, 2008. D. R. Karger, R. Motwani, G. D. S. Ramkumar. On Approximating the Longest Path in a Graph. Algorithmica 18(1): 82–98 (1997). J. Kneis, D. Mölle, S. Richter, and P. Rossmanith, Divide-and-color, in Proc. 32nd International Workshop on Graph-Theoretic Concepts in Computer Science, WG (Bergen, Norway, June 22–24, 2006), Springer LNCS 4271, pp. 58-–67, 2006. B. Monien, How to find long paths efficiently, Annals of Discrete Mathematics 25 (1985), 239–-254. P. Ribenboim, The New Book of Prime Number Records, Springer, New York, 1996. R. Williams, Finding paths of length $k$ in $O^*(2^k)$, Inform. Process Lett. 109(6):301–-338, 2009. [^1]: The research leading to these results has received funding from the European Community’s Seventh Framework Programme (FP7/2007-2013) under grant agreement number 257575. [^2]: This research was partially supported by the Israel Science Foundation (grant number 339/10)

--- abstract: 'Let $S$ be an integral domain with field of fractions $F$ and let $A$ be an $F$-algebra. An $S$-subalgebra $R$ of $A$ is called $S$-nice if $R$ is lying over $S$ and the localization of $R$ with respect to $S \setminus \{ 0 \}$ is $A$. Let $\mathbb S$ be the set of all $S$-nice subalgebras of $A$. We define a notion of open sets on $\mathbb S$ which makes this set a $T_0$ Alexandroff space. This enables us to study the algebraic structure of $\mathbb S$ from the point of view of topology. We prove that an irreducible subset of $\mathbb S$ has a supremum with respect to the specialization order. We present equivalent conditions for an open set of $\mathbb S$ to be irreducible, and characterize the irreducible components of $\mathbb S$.' author: - Shai Sarussi title: Alexandroff Topology of Algebras over an Integral Domain --- \[section\] \[thm\][Corollary]{} \[thm\][Lemma]{} \[thm\][Proposition]{} \[thm\][Definition]{} \[thm\][Remark]{}\[thm\][Example]{} \[thm\][Conjecture]{} Introduction and some preliminary results ========================================= As the title suggests, in this paper we discuss algebras over integral domains from the point of view of Alexandroff topology, which will shortly be defined. In \[Sa1\] and \[Sa2\] we studied algebras over valuation domains, concentrating on quasi-valuations that extend a valuation on a field. In \[Sa3\] we prove several existence theorems of integral domains that may be applied to the study of quasi-valuations. More specifically, let $S$ be an integral domain which is not a field, let $F$ be its field of fractions, and let $A \neq F$ be an $F$-algebra. In \[Sa3\] we study $S$-subalgebras of $A$, that are lying over $S$ and whose localizations with respect to $S \setminus \{ 0 \}$ is $A$. We call them $S$-nice subalgebras of $A$. Namely, an $S$-subalgebra $R$ of $A$ is called an $S$-nice subalgebra of $A$ if $R \cap F=S$ and $FR=A$; we shall use this notation throughout this paper. We denote by $\mathbb S$ the set of all $S$-nice subalgebras of $A$. We recall now some definitions and results from \[Sa3\]. The following concept is used quite frequently: let $B$ be a basis of $A$ over $F$. We say that $B$ is $S$-stable if there exists a basis $C$ of $A$ over $F$ such that for all $c \in C$ and $b \in B$, one has $cb \in \sum_{y \in B} Sy$. We note in \[Sa3, Remark 3.4\] that if a basis $B$ is closed under multiplication then $B$ is $S$-stable. Thus, for example, every free (noncommutative) $F$-algebra with an arbitrary set of generators has an $S$-stable basis; in particular, every polynomial algebra with an arbitrary set of indeterminates over $F$ has an $S$-stable basis. We also show in \[Sa3, Proposition 3.12\] that if $A$ is finite dimensional over $F$, then every basis of $A$ over $F$ is $S$-stable. The first existence theorem is as follows. \[existence of S-nice\] (cf. \[Sa3, Theorem 3.14\]) If there exists an $S$-stable basis of $A$ over $F$, then there exists an $S$-nice subalgebra of $A$. In particular, if $A$ is finite dimensional over $F$ then there exists an $S$-nice subalgebra of $A$. The following result is a going-down lemma for $S$-nice subalgebras. \[going down\] (cf. \[Sa3, Lemma 3.20\]) Let $S_1 \subseteq S_2$ be integral domains with field of fractions $F$ such that $S_2 \neq F$. Assume that there exists an $S_1$-stable basis of $A$ over $F$. Let $R$ be an $S_2$-nice subalgebra of $A$. Then there exists an $S_1$-nice subalgebra of $A$, which is contained in $R$. We conclude that a minimal $S$-nice subalgebra of $A $ does not exist. More precisely, we prove \[not minimal\] (cf. \[Sa3, Proposition 3.21\]) Assume that there exists an $S$-stable basis of $A$ over $F$. Let $R$ be an $S$-nice subalgebra of $A$. Then there exists an infinite decreasing chain of $S$-nice subalgebras of $A $ starting from $R$. In particular, a minimal $S$-nice subalgebra of $A $ does not exist. Let $\mathcal C$ be a chain of prime ideals of $S$ and let $R$ be a faithful $S$-algebra. Let $\mathcal D$ be a chain of prime ideals of $R$. We say that $\mathcal D$ covers $\mathcal C$ if for every $P \in \mathcal C$ there exists $Q \in \mathcal D$ lying over $P$; namely, $Q \cap S=P$. \[every chain is covered commutative case\] (cf. \[Sa3, Theorem 3.24\]) Assume that there exists an $S$-stable basis of $A$ over $F$. Let $\mathcal C$ be a chain of prime ideals of $S$. If $A$ is commutative then there exists an $S$-nice subalgebra $R$ of $A$ such that there exists a chain of prime ideals $\mathcal F$ of $R$ covering $\mathcal C$. In fact, there exists an infinite descending chain of such $S$-nice subalgebras of $A$. Note that by Theorem \[existence of S-nice\] and Proposition \[not minimal\] if $A$ contains an $S$-stable basis then $\mathbb S$ is not empty and is, in fact, infinite. In particular, by the above-mentioned note before Theorem \[existence of S-nice\], if $A$ is finite dimensional over $F$ then $\mathbb S$ is infinite. In the second section of this paper, which is its main part, we do not assume that $A$ contains an $S$-stable basis; we merely assume that $\mathbb S$ is not empty. We do, however, present an example in which the existence of an $S$-stable basis is assumed. In the third section of this paper we assume that $A$ contains an $S$-stable basis We present now some of the common definitions we use from order theory. Let $P$ be a set. A relation $\leq$ on $P$ that is reflexive and transitive is called a quasi order or a preorder; if the relation is also antisymmetric then it is called a partial order and $P$ is called a partially ordered set, or a poset. Let $L \subseteq P$. We say that $a \in P$ is a lower (resp. upper) bound of $L$ if $a \leq x$ (resp. $a \geq x$) for all $x \in L$. If the set of lower (resp. upper) bounds of $L$ has a unique greatest (resp. smallest) element, this element is called the greatest lower (resp. least upper) bound of $L$, and is denoted by $\text{inf}L$ (resp. $\text{sup}L$). We say that $L$ is a lower set if $L=\{ y \in P \mid y \leq x \text{ for some } x \in L\}$. We say that $L$ is an upper set if $L=\{ y \in P \mid y \geq x \text{ for some } x \in L\}$. A subset $\emptyset \neq D \subseteq P$ is called directed if for all $a,b \in D$ there exists $c \in D$ such that $a \leq c$ and $b \leq c$. We say that $P$ is a dcpo (directed complete partial order) if every directed subset of $P$ has a supremum. A subset $I$ of $P$ is called an ideal of $P$ if $I$ is a lower set and directed. $P$ is called an inf semilattice (resp. sup semilattice) if for all $a,b \in P$, $\text{inf}\{a,b\}$ (resp. $\text{sup}\{a,b\}$) exists in $P$. If $P$ is both an inf semilattice and a sup semilattice, we say that $P$ is a lattice. A subset $L$ of $P$ is called a sublattice of $P$ if $L$ is a lattice with respect to the partial order of $P$. Similarly, one defines inf subsemilattice and sup subsemilattice. We briefly discuss now the notion of an Alexandroff topological space. A topological space whose set of open sets is closed under arbitrary intersections is called an Alexandroff space, after P. alexandroff who first introduced such topological spaces in his paper \[Al\] from 1937. Equivalently, A topological space is called an Alexandroff space if every element has a minimal open set containing it. A finite topological space is the most important particular case of an Alexandroff space. In fact, Alexandroff spaces share many properties with finite topological spaces; in particular, Alexandroff spaces have all the properties of finite spaces relevant for the theory of digital topology (see \[He\] and \[Kr\]). Thus, in the eighties the interest in Alexandroff spaces arose as a consequence of the very important role of finite spaces in digital topology. In 1999 F. G. Arenas studied the topological properties of Alexandroff spaces (see \[Ar\]). Let $(T,\tau)$ be a topological space. For $X \subseteq T$ we denote by $clX$ the closure of $X$. It is well known and easy to prove, that if one defines $x \leq_{\tau} y$ whenever $x \in cl \{ y \}$, then $ \leq_{\tau} $ is a quasi order; i.e., a reflexive and transitive relation. $ \leq_{\tau}$ is called the specialization order. Recall that $(T,\tau)$ is called $T_0$ if for every two distinct elements in $T$, there exists an open set containing one of them but not the other. It is known that if $(T,\tau)$ is $T_0$ then $ \leq_{\tau} $ is a partial order. On the other hand, for any quasi order $ \leq $ on a set $T$, one can define the topology whose open sets are the upper sets of $T$ with respect to $ \leq $, denote it by $A_\leq$. So, there are two functors, the specialization order from the class of all topological spaces to the class of all quasi ordered sets, sending $(T,\tau)$ to $(T,\leq_\tau)$; and the functor in the opposite direction sending $(T,\leq)$ to $(T,A_\leq)$. If one restrict the class of all topological spaces to the the class of all alexandroff topological spaces, then one has an isomorphism between the categories. In this paper the symbol $\subset$ means proper inclusion and the symbol $\subseteq$ means inclusion or equality. The Alexandroff Topology ======================== Inspired by the Zarisky topology on the prime spectrum of a ring, for every $M \subseteq A$ we denote by $V(M)$ the set of all $S$-nice subalgebras of $A$ containing $M$. It is easy to see that $ V(\{0\})=\mathbb S$, $V(F)=\emptyset$, and for every $M_1,M_2 \subseteq A$, we have $$V(M_1 \cup M_2)=V(M_1) \cap V(M_2).$$ Thus, the set B=$\{ V(M) \}_{M \subseteq A}$ is a basis for a topology on $\mathbb S$. Namely, every open set in $\mathbb S$ is a union of elements of $B$. Moreover, for every set $\{M_i\}_{i \in I} $ of subsets of $ A$, we have $$V(\cup _{i \in I}M_i)= \cap_{i \in I} V(M_i).$$ Now, an intersection of union of elements of $B$ can be presented as a union of intersection of elements of $B$. Indeed, let $I$ be a set and let $\{ J_i\}_{i \in I}$ be a set of sets such that for all $i \in I$ and $j_i \in J_i$ there exists a set $ X_{i,j_i}$; then $$\bigcap_{i \in I} (\bigcup_{j_i \in J_i} X_{i,j_i})=\bigcup_{t \in \prod_{i \in I} J_i} (\bigcap_{i \in I} X_{i,t_i}),$$ where $t_i$ denotes the $i$-th component of $t$. We apply the above equation to elements of $B$ and deduce that every intersection of open sets of $\mathbb S$ is open. Thus, $\mathbb S$ is an Alexandroff topological space with respect to the topology defined above. Let $T$ be a topological space and let $x,y \in T$; then $x \in cl\{y\} $ iff for every open set $U$ containing $x$, we have $y \in U$. If $T$ is alexandroff, this is equivalent to $y \in U_x$, where $U_x$ denotes the minimal open set containing $x$. Now, let $R \in \mathbb S$. It is easy to see that the minimal open set containing $R$ is $V(R)$. Indeed, let $U$ be an open set containing $R$; then $U$ is of the form $\cup_{i \in I} V(M_i)$ where $M_i \subseteq A$ for all $i \in I$. Thus, $R \in V(M_i)$ for some $i \in I$; hence $M_i \subseteq R$ and $V(R) \subseteq V(M_i)$. On the other hand, $V(R) $ is an open set containing $R$. Thus, the specialization order on $\mathbb S$ is the order of inclusion; i.e., for $R_1, R_2 \in \mathbb S$, $R_1 \leq R_2 $ iff $R_1 \subseteq R_2 $. Moreover, as in any alexandroff topological space, $U \subseteq \mathbb S$ is open iff $U$ is an upper set with respect to the specialization order; dually, $C \subseteq \mathbb S$ is closed iff $C$ is a lower set. We will frequently use the following four basic lemmas. \[basic lemma\] 1. Let $R_1 $ and $R_2 $ be two elements of $ \mathbb S$ and let $R$ be an $S$-algebra satisfying $R_1 \subseteq R \subseteq R_2$; then $R$ is an $S$-nice subalgebra of $A$. 2. Let $\{R_i\}$ be a finite subset of $\mathbb S$; then $ \cap_{1 \leq i \leq n} R_i \in \mathbb S$. Straightforward. For subsets $M \subseteq A$ and $T \subseteq F$ we define $$T \cdot M = \{\sum_{1 \leq i \leq n}t_i m_i \mid t_i \in T, m_i \in M\}.$$ \[union satisfies basic properties\] Let $K=\{R_i\}_{i \in I}$ be a nonempty subset of $\mathbb S$ and denote $R_0=\cup_{i \in I} R_i$. Then the following three properties are valid: \(a) $R_0 \cap F=S$; \(b) $F \cdot R_0=A$; and \(c) If $R_0$ is closed under addition then $S \cdot R_0 \subseteq R_0$. In particular, if $R_0$ is a ring then it is an $S$-nice subalgebra of $A$. \(a) Clearly, $S \subseteq R_0 \cap F$. Let $\alpha \in R_0 \cap F$; then $\alpha \in R_1$ for some $R_1 \in K$. Since $R_1 \cap F=S$, we have $\alpha \in S$. (b) Since $K$ is not empty there exists $R_1 \in K$ such that $FR_1=A$; thus, $A= FR_1 \subseteq FR_0 \subseteq A$. (c) Let $\{r_j\}_{1 \leq j \leq n} \subseteq R_0$ and let $\{s_j\}_{1 \leq j \leq n} \subseteq S$; then for all $1 \leq j \leq n$, $r_j \in R_{i_{j}}$ for appropriate $R_{i_{j}} \in K$. Thus, for all $1 \leq j \leq n$, $s_j r_j \in R_{i_j} \subseteq R_0$; hence, $S \cdot R_0 \subseteq R_0$. Note that $R_0$ is not necessarily an $S$-algebra since it is not necessarily a ring. \[sup of a chain exists\] Let $C= \{ R_{i} \}_{i \in I}$ be a nonempty chain in $\mathbb S$. Then the supremum of $C$ exists in $\mathbb S$. Let $R_0=\cup_{i \in I}R_{i}$; since $C$ is a chain, $R_0$ is ring. By Lemma \[union satisfies basic properties\], $R_0 \in \mathbb S$. It is clear that $R_0$ is the smallest $S$-nice subalgebra of $A$ containing every element of $C$. \[a maximal element exists in an open=upper set\] Let $U$ be a nonempty open set of $\mathbb S$. Then there exists a maximal element in $U$, which is also a maximal element in $\mathbb S$. Consider $U$ with the partial order of containment. Let $C$ be a nonempty chain in $U$. By Lemma \[sup of a chain exists\], $\text{sup}C$ exists in $\mathbb S$. Now, $\text{sup}C \supseteq R_1$ for some $R_1 \in U$; thus, by the definition of an open set in $\mathbb S$, $\text{sup}C \in U$. Therefore, by Zorn’s Lemma, there exists a maximal element in $U$. \[in a maximal chain there exists a maximal element\] In view of Lemma \[a maximal element exists in an open=upper set\], one can be more precise. In fact, for every maximal chain $C$ in a nonempty open set $U$ of $\mathbb S$ (there exists such a chain by Zorn’s Lemma), $\text{sup}C=\cup_{R \in C}R$ is a maximal element of $C$, $U$ and $\mathbb S$. Moreover, for the same reason, for any $R \in U$ there exists a maximal element $R' \in U$ containing $R$; indeed, take any maximal chain in $U$ containing $R$, and use the reasoning above. Let $U \subseteq \mathbb S$ be an open set. We consider the following properties: \(a) $U$ is of the form $V(R_0)$ for some $R_0 \in \mathbb S$. \(b) $U$ is closed under arbitrary nonempty intersection. \(c) $V(\cap_{R \in U} R) = U$. \(d) $V(\cap_{R \in U} R) \subseteq U$. \(e) $U$ is closed under finite nonempty intersection; i.e., $U$ is an inf subsemilattice of $\mathbb S$. \(f) $U$ contains no more than one minimal element. \[a implies b implies c for open U\] Notation as above; the following implications hold: $(a) \Leftrightarrow (b) \Rightarrow (c) \Leftrightarrow (d) \Rightarrow (e) \Rightarrow (f)$. $(a) \Rightarrow (b)$. By definition, every element of $U$ contains $R_0$. Let $R'=\cap_{R \in W} R$ where $\emptyset \neq W \subseteq U$. Then $R_0 \subseteq R' \subseteq R''$ for some $R'' \in W$; hence by Lemma \[basic lemma\], $R' \in \mathbb S$ and thus, by the definition of $U$, $R' \in U$. $(b) \Rightarrow (a)$ and $(b) \Rightarrow (c)$. By assumption $\cap_{R \in U} R \in U \subseteq \mathbb S$, which is clearly the smallest element of $U$; since every $R \in U$ contains $\cap_{R \in U} R$ and $U$ is an open set, we get $U=V(\cap_{R \in U} R)$. $(c) \Rightarrow (d)$ is trivial. $(d) \Rightarrow (c)$. As above, every $R \in U$ contains $\cap_{R \in U} R$; thus, $V(\cap_{R \in U} R) \supseteq U$. $(d) \Rightarrow (e)$. Let $R_1,R_2 \in U$; by Lemma \[basic lemma\], $R_1 \cap R_2 \in \mathbb S$; clearly, $R_1 \cap R_2 \supseteq \cap_{R \in U} R$, and thus $R_1 \cap R_2 \in U$. $(e) \Rightarrow (f)$. Assume to the contrary that there exists two different minimal elements $R_1 ,R_2$ in $U$. Then, by assumption $R_1 \cap R_2 \in U$, but it is clearly strictly contained in both $R_1$ and $R_2$, a contradiction. We present now examples which demonstrate that the left to right implications in the previous proposition cannot be reversed. \[left implications may not hold\] To show that the implication $(c) \Rightarrow (b)$ may not hold, we can consider any case in which $\mathbb S$ is taken as the open set and $\cap_{R \in \mathbb S} R \notin \mathbb S$; it is clear that $V(\cap_{R \in \mathbb S} R) = \mathbb S$. As an explicit example, let $S= \mathbb{Z}$, $A=M_2(\mathbb{Q})$ and $U=\mathbb S$. More generally, in \[Sa3, discussion after Proposition 3.21\] we showed that whenever $A$ contains an $S$-stable basis, we have $\cap_{R \in \mathbb S} R \notin \mathbb S$. To show that the implication $(e) \Rightarrow (d)$ may not hold, one can consider an infinite chain of $S$-nice subalgebras of $A$ such that their intersection is an $S$-nice subalgebra of $A$ that is strictly contained in each of them. As an explicit example, let $S= O_v$ be a valuation domain with value group $\mathbb{R}$, and let $A=M_2(F)$, where $F$ is the field of fractions of $O_v$. Take $0<r_0 \in \mathbb{R}$ and for all $0< r \leq r_0$ let $I_r=\{x \in O_v \mid v(x) \geq r\}$. For all $0< r \leq r_0$ denote $$R_r= \left( \begin{array}{cccccccc} O_v& I_r \\ J_1 & O_v \\ \end{array} \right),$$ where $J_1$ is any nonzero ideal of $O_v$. Let $U=\cup_{0< r <r_0} V(R_r)$. Since $\{ I_r \}_{0<r<r_0}$ is a chain, $U$ is closed under finite nonempty intersection; however, $$\cap_{R \in U} R=\cap_{0< r <r_0} R_r = R_{r_{0}} \notin U.$$ Finally we demonstrate that $(f)$ does not necessarily imply $(e)$. With the notation presented above, let $J_2$ be an ideal of $O_v$ strictly containing $J_1$; and let $$R'= \left( \begin{array}{cccccccc} O_v& I_{r_0} \\ J_2 & O_v \\ \end{array} \right).$$ Let $U=\cup_{0< r <r_0} V(R_r) \cup V(R')$. Then the unique minimal element of $U$ is $R'$, but $U$ in not closed under finite nonempty intersection since for all $0< r <r_0$, we have $R' \supset R' \cap R_r $; thus, $R' \cap R_r \notin U$ \[exists a lower bound iff\] Let $H$ be a nonempty subset of $\mathbb S$; then 1\. There exists a lower bound for $H$ iff the infimum of $ H$ exists. 2\. There exists an upper bound for $H$ iff the supremum of $ H$ exists. Clearly, the right to left implication in both statements is trivial. We prove the left to right implication of the first statement. Let $R_0 \in \mathbb S$ be a lower bound of $H$, let $R_1 = \cap_{R \in H} R$ and let $R_2$ be any element of $H$ (note that $H$ is not empty). Then by Lemma \[basic lemma\], $R_1 \in \mathbb S$ and it is clearly the infimum of $H$. We prove now the left to right implication of the second statement. Let $K$ denote the set of all upper bounds of $H$; by assumption, $K$ is not empty. Since $H$ is not empty, $K$ has a lower bound; indeed, any element of $H$ is a lower bound for $K$. Thus, by the first statement, the infimum of $K$, $R_3 = \cap_{R \in K} R$, exists in $\mathbb S$. Now, let $R \in H$, then for every $R' \in K$, we have $R \subseteq R'$; thus, $R \subseteq R_3$. Therefore, $R_3 \in K$ and $R_3$ is the supremum of $H$. We note that the assumption that $H \neq \emptyset$ is crucial. Indeed, the empty set clearly has an upper bound but the supremum of $\emptyset$ does not exist, since $\mathbb S$ may not contain a smallest element, as shown in Example \[left implications may not hold\]. Also, the empty set clearly has a lower bound but the infimum of $\emptyset$ exists iff $\mathbb S$ is irreducible, as we shall see in Theorem \[one maximal in U equivallent conditions\]. Dually to Proposition \[a implies b implies c for open U\], let $C \subseteq \mathbb S$ be a closed set. We consider the following properties: \(a) $C=cl \{R \}$ for some $R \in \mathbb S$. \(b) Every nonempty subset of $C$ has a supremum, which belongs to $C$. \(c) $\text{sup} C \in \mathbb{S}$ and $C=cl \{\text{sup} C\}$. \(d) $\text{sup} C \in \mathbb{S}$ and $cl \{\text{sup} C\} \subseteq C$. \(e) $C$ is a sup subsemilattice of $\mathbb S$. \(f) $C$ contains no more than one maximal element. The proof of the following proposition is quite similar to the proof of Proposition \[a implies b implies c for open U\]. The implications $(a) \Rightarrow (b)$, $(c) \Rightarrow (b)$, and $(d) \Rightarrow (e)$ rely on Proposition \[exists a lower bound iff\]; we shall not prove it here. \[a implies b implies c for closed C\] Notation as above; the following implications hold: $(a) \Leftrightarrow (b) \Leftrightarrow (c) \Leftrightarrow (d) \Rightarrow (e) \Rightarrow (f)$. Note the small difference in the implication $(c) \Rightarrow (b)$, which is valid in Proposition \[a implies b implies c for closed C\] but not in Proposition \[a implies b implies c for open U\]. The reason for this difference is the fact that $V(M)$ is defined for all $M \subseteq A$ whereas $cl \{ R\}$ is defined only for $R \in \mathbb{S}$. While in condition (c) of Proposition \[a implies b implies c for open U\], $ \cap_{R \in U} R$ is not necessarily in $\mathbb{S}$, in conditions (c) and (d) of Proposition \[a implies b implies c for closed C\] we require that $\text{sup}C$ would be in $\mathbb{S}$. We also note that condition (e) of Proposition \[a implies b implies c for closed C\] implies that $C$ is an ideal of $\mathbb{S}$, in the sense of order theory defined in the introduction. Indeed, by assumption $C$ is closed and thus it is a lower set. By the assumption in (e), $C$ is a sup subsemilattice of $\mathbb S$; in particular, $C$ is directed. In fact, in this case $C$ is actually a sublattice of $\mathbb S$. To show this, let $R_1,R_2 \in C$, the infimum of $R_1$ and $R_2$ is their intersection which is in $\mathbb{S}$; by assumption $C$ is a closed set and thus it is a lower set, so $\text{inf}\{R_1,R_2\} \in C$; and by the assumption in condition (e), $C$ is a sup subsemilattice of $\mathbb S$. The following theorem is important to our study. \[irreducible has sup\] Let $I $ be an irreducible subset of $ \mathbb S$. Then $\cup_{R \in I}R$ is an $S$-nice subalgebra of $A$; in particular, $\text{sup}I=\cup_{R \in D}R$. Denote $R_0=\cup_{R \in I}R$. Let $a,b \in R_0$; we prove that there exists $R_1 \in I$ such that $a,b \in R_1$. Assume to the contrary that there exists no such $R_1$. Let $G_1 $ denote the set of all elements in $\mathbb S$ not containing $a$, and let $G_2 $ denote the set of all elements in $\mathbb S$ not containing $b$. It is clear that $G_1$ and $G_2$ are closed in $\mathbb S$. However, by our assumption $I \subseteq G_1 \cup G_2$, while $I \nsubseteq G_1$ and $I \nsubseteq G_2$, a contradiction. Thus, $R_0$ is a ring. By Lemma \[union satisfies basic properties\], $R_0 \in \mathbb S$. \[one maximal in U equivallent conditions\] Let $U \subseteq \mathbb S$ be a nonempty open set. The following statements are equivalent: \(a) $U$ has a greatest element. \(b) There exists $R_0 \in \mathbb S$ such that $clU=cl\{R_0\}$. \(c) $U$ has a unique maximal element. \(d) $U$ is irreducible. \(e) $U$ is a sup subsemilattice of $\mathbb S$. (f) Every nonempty subset $H \subseteq U$ has a supremum. We prove $(a) \Rightarrow (b) \Rightarrow (c) \Rightarrow (d) \Rightarrow (e) \Rightarrow (f) \Rightarrow (a)$. To show that $(a) \Rightarrow (b)$ we denote by $R_0$ the greatest element of $U$. It is clear that every closed set containing $U$ also contains $R_0$; on the other hand, since $R_0$ is the greatest element of $U$, by the definition of the topology on $\mathbb S$, every closed set containing $R_0$ also contains $U$. $(b) \Rightarrow (c)$. We prove that $R_0$ is the unique maximal element of $U$. It is clear that $R_0$ is a maximal element of $U$, since otherwise $U \setminus cl\{R_0\} \neq \emptyset$ and thus $clU \setminus cl\{R_0\} \neq \emptyset$. Similarly, Assuming there exists another maximal element $R_1 \in U$, we get $R_1 \in U \setminus cl\{R_0\}$. $(c) \Rightarrow (d)$. Let $R_0$ denote the unique maximal element of $U$. Assume to the contrary that $U \subseteq G_1 \cup G_2$ where $G_1 , G_2$ are closed in $\mathbb S$ while $U \nsubseteq G_1$ and $U \nsubseteq G_2$. Let $R_1 \in U \setminus G_1$ and $R_2 \in U \setminus G_2$. By Remark \[in a maximal chain there exists a maximal element\], there exist maximal elements $R_1 ', R_2 ' \in U$ containing $R_1$ and $R_2$, respectively. Since $R_0$ is the unique maximal element of $U$, we have $R_1 '= R_2 '=R_0$. Thus, $R_0 \notin G_1$ and $R_0 \notin G_2$, a contradiction. We prove $(d) \Rightarrow (e)$. Let $R_1, R_2 \in U$; by assumption $U$ is irreducible and thus by Theorem \[irreducible has sup\], $R_0=\cup_{R \in U} R$ is an $S$-nice subalgebra of $A$. Clearly, $R_0 $ contains both $R_1$ and $R_2$. Thus, by Proposition \[exists a lower bound iff\], $\text{sup}\{R_1 ,R_2 \}$ exists. Hence, by the definition of an open set, $\text{sup}\{R_1 ,R_2 \} \in U$. We prove now $(e) \Rightarrow (f)$. Let $\emptyset \neq H \subseteq U$. By Remark \[in a maximal chain there exists a maximal element\], for every $R \in H$ there exists a maximal element $T_R \in U$ containing $R$. By assumption, every two elements of $U$ have a supremum, thus these $T_R$ must all be equal. So, $H$ is bounded from above and therefore, by Proposition \[exists a lower bound iff\], the supremum of $H$ exists. Finally, we show $(f) \Rightarrow (a)$. By assumption $U$ has a supremum, and since $U$ is an open set, its supremum is its greatest element. In view of Theorem \[one maximal in U equivallent conditions\], we characterize now the irreducible components of $\mathbb S$. \[irreducible components\] $I \subseteq \mathbb S$ is an irreducible component of $\mathbb S$ iff $I=cl \{ R \}$ for some maximal $R \in \mathbb S$. Let $R$ be a maximal element of $\mathbb S$. It is clear that $cl \{ R \}$ is irreducible. Assume to the contrary that there exists an irreducible set $G \supset cl \{ R \}$. Let $R' \in G \setminus cl \{ R \}$. Then $R \nsubseteq R'$ and $R' \nsubseteq R$. By Theorem \[irreducible has sup\], $\cup_{T \in G}T \in \mathbb S$. However, $\cup_{T \in G}T$ strictly contains $R$, a contradiction. On the other hand, let $I \subseteq \mathbb S$ be an irreducible component of $\mathbb S$. Let $R_0=\cup_{T \in I}T$; by Theorem \[irreducible has sup\], $R_0\in \mathbb S$. Thus, $I \subseteq cl \{ R_0\}$. Since $cl \{ R_0\}$ is irreducible and $I $ is an irreducible component of $\mathbb S$, one has $I=cl \{ R_0\}$. Now, $R_0$ is maximal in $\mathbb S$, since otherwise there exists $R_0 \subset R_1 \in \mathbb S$, but then $cl \{ R_0 \} \subset cl \{ R_1 \}$, a contradiction. Recall from \[GHKLMS, Def. O-5.6.\] that a topological space $T$ is called sober if for every irreducible closed subset $C$ of $T$, there exists a unique $t \in T$ such that $C$ is the closure of $t$; i.e., $C$ has a unique generic point. Also recall that a poset $P$ is called dcpo (directed complete partial order) if every directed subset of $P$ has a supremum. It is known (see, for example \[GHKLMS, Ex. O-5.15.\]) that every sober space is a dcpo, under the specialization order. Note that by Lemma \[a maximal element exists in an open=upper set\] there exists a maximal $S$-nice subalgebra of $A$. In \[Sa3, discussion after Corollary 3.21\] we noted that in case $S$ is a valuation domain of $F$ and $A$ is a field, then the maximal $S$-nice subalgebras of $A$ are precisely the valuation domains (whose valuations extend $v$) of $A$. So, by Proposition \[irreducible components\], the closures of these valuation domains are precisely the irreducible components of $\mathbb S$. We also showed in \[Sa3, Example 3.26\] that even in the case of a central simple $F$-algebra, one can have an infinite ascending chain of $S$-nice subalgebras of $A$ (even when $S$ is a valuation domain). For the reader’s convenience we present here the example. (cf. \[Sa3, Example 3.26\]) \[example of an infinite accending chain of s-nice\] Let $C$ be a non-Noetherian integral domain with field of fractions $F$. Let $\{ 0 \} \neq I_{1} \subset I_{2} \subset I_{3} \subset ...$ be an infinite ascending chain of ideals of $C$ and let $A=M_{n}(F)$. Then $$R_1= \left( \begin{array}{cccccccc} C& C & ... & C & I_{1} \\ C & C & ... & C & I_{1} \\ . & . & ... & . & . \\ . & . & ... & . & . \\ . & . & ... & . & . \\ C & C & ... & C & I_{1} \\ C & C & ... & C & C \\ \end{array} \right) \subset R_2= \left( \begin{array}{cccccccc} C& C & ... & C & I_{2} \\ C & C & ... & C & I_{2} \\ . & . & ... & . & . \\ . & . & ... & . & . \\ . & . & ... & . & . \\ C & C & ... & C & I_{2} \\ C & C & ... & C & C \\ \end{array} \right) \subset ...$$ is an infinite accending chain of $C$-nice subalgebras of $A$. So, let $$I=\cup_{n \in \mathbb N} cl \{ R_n\};$$ then $I$ is closed and irreducible with no generic point. In particular, we have an example in which $\mathbb S$ is not sober. Nevertheless, in a subsequent paper we will show that $\mathbb S$ is indeed a dcpo and has some interesting properties from the point of view of domain theory. Note that by Lemma \[basic lemma\], $\mathbb S$ is an inf semilattice, where the infimum is actually an intersection of sets. In particular, taking $U=\mathbb S$ in Theorem \[one maximal in U equivallent conditions\], the conditions of the theorem are also equivalent to the condition that $\mathbb S$ is a lattice. In fact we can say even more. By the definition of an irreducible space, every finite intersection of nonempty open sets is nonempty. In our case, whenever $\mathbb S$ satisfies the equivalent conditions of Theorem \[one maximal in U equivallent conditions\], the intersection of **all** nonempty open sets of $\mathbb S$ is the singleton $\{ R_0 \}$, where $R_0$ denotes the greatest element of $\mathbb S$. Moreover, this property is also equivalent to the equivalent conditions presented in Theorem \[one maximal in U equivallent conditions\]. Recall (cf. \[En, Theorem 16.4\]) that a valuation $v$ on a field $K$ is called henselian if $v$ extends uniquely to every algebraic field extension of $K$; in this case, one also says that the corresponding valuation domain is henselian. Thus, in view of the previous remark and the discussion before Example \[example of an infinite accending chain of s-nice\], we have, If $S$ is an henselian valuation domain, $F$ is its field of fractions and $A$ is an algebraic field extension of $F$, then $\mathbb S$ is a lattice (viewed from the point of view of order theory), and an irreducible topological space (viewed as an Alexandroff topological space). Prime ideals ============ In this short section we discuss the prime spectra of $S$-nice subalgebras of $A$ and the subsets of $\text{Spec}(S)$ covered by them. We assume that $A$ is commutative (in this case whenever $R_1 \subset R_2$ are $S$-subalgebras of $A$ and $Q$ is a prime ideal of $R_2$, then $Q \cap R_1$ is a prime ideal of $R_1$). We also assume that $A$ contains an $S$-stable basis, in order to be able to use the going down lemma for $S$-nice subalgebras of $A$ (Lemma \[going down\]) and Theorem \[every chain is covered commutative case\]. For a ring $T$, we denote by $\text{Spec}(T)$ the prime spectrum of $T$; i.e., the set of all prime ideals of $T$. Recall the following definition from \[Sa3\]: if for every $P \in \text{Spec}(S)$ there exists $Q \in \text{Spec}(R)$ lying over $P$, we say that $R$ satisfies “Lying Over" (LO, in short) over $S$. We denote by $\text{Spec}_R (S)$ the set of all prime ideals of $S$ having a prime ideal of $R$ lying over them; namely, $\text{Spec}_R (S)= \{ P \in \text{Spec}(S) \mid$ there exists $Q \in \text{Spec}(R)$ lying over $ P\}$. Note that, by definition $\text{Spec}_R (S) \subset \text{Spec} (S)$ iff $R$ does not satisfy LO over $S$. As usual, we use the term “almost all" to mean that a property is satisfied to all but finitely many elements. In the following lemma we show that whenever $R$ does not satisfy LO over $S$, there exists an $S$-nice subalgebra of $A$ whose prime spectrum lies over a larger set of primes of $S$ than the prime spectrum of $R$. \[not LO then there exists a better one\] Let $R_1 $ be an $S$-nice subalgebra of $A$ such that $\text{Spec}_{R_1} (S) \neq \text{Spec} (S)$, and let $P \in \text{Spec} (S) \setminus \text{Spec} _{R_1} (S)$. Then there exists $R_1 \supset R \in \mathbb S$ such that $\text{Spec}_{R_1} (S) \subset \text{Spec}_{R} (S)$ and $P \in \text{Spec}_{R} (S)$. Consider the chain $\{ P\}$. By Theorem \[every chain is covered commutative case\], there exists $R_2 \in \mathbb S$ having a prime ideal lying over $P$. By Lemma \[basic lemma\], $R=R_1 \cap R_2$ is an $S$-nice subalgebra of $A$. Now, since $A$ is commutative, every prime ideal of $R_i$ ($i=1,2$) intersect to a prime ideal of $R$. Thus, $\text{Spec}_{R_1} (S) \subset \text{Spec}_{R} (S)$ with $P \in \text{Spec}_{R} (S)$. \[LO iff for almost all\] There exists $R \in \mathbb S$ satisfying LO over $S$ iff there exists $R_1 \in \mathbb S$ whose prime spectrum is lying over almost all prime ideals of $S$; namely, $\text{Spec}(S) \setminus \text{Spec}_{R_1} (S)$ is finite. $(\Rightarrow)$ is trivial; just take $R_1=R$. $(\Leftarrow)$ Let $R_1$ be an $S$-nice subalgebra of $A$ whose prime spectrum is lying over almost all prime ideals of $S$. If $R_1$ satisfies LO over $S$ then we are done. Otherwise, by Lemma \[not LO then there exists a better one\] there exists $R_1 \supset R_2 \in \mathbb S$ with $\text{Spec}_{R_1} (S) \subset \text{Spec}_{R_2} (S)$. After finitely many such steps we get $R_n \in \mathbb S$ that satisfies LO over $S$. Note that if there exists $R \in \mathbb S$ that satisfies LO over $S$, then for every nonempty closed set $C \subseteq \mathbb S$ there exists $R_1 \in C$ that satisfies LO over $S$; indeed, take any $R_2 \in C$ and denote $R_1 = R \cap R_2$. In view of the previous remark, in the following proposition we present equivalent conditions for the non-existence of an $S$-nice subalgebra of $A$ that satisfies LO over $S$. The following conditions are equivalent: \(a) There exists no $R \in \mathbb S$ that satisfies LO over $S$. \(b) There is no maximal subset $Y \subseteq \text{Spec}(S)$ (with respect to inclusion) such that there exists $R \in \mathbb S$ with $\text{Spec}_{R} (S)=Y$. \(c) For every $R \in \mathbb S$ there exists an infinite descending chain $\{R_i \}_{i \in I}$ of $S$-nice subalgebras of $A$ such that $R_i \subseteq R$ for all $i \in I$, and $\text{Spec}_{R_i} (S) \subset \text{Spec}_{R_j} (S)$, whenever $R_j \subset R_i$. \(d) Every nonempty closed subset of $\mathbb S$ contains an infinite descending chain $\{R_i \}_{i \in I}$ of $S$-nice subalgebras of $A$ such that $\text{Spec}_{R_i} (S) \subset \text{Spec}_{R_j} (S)$, whenever $R_j \subset R_i$. We prove $(a) \Rightarrow (b) \Rightarrow (c) \Rightarrow (d) \Rightarrow (a)$. To show $(a) \Rightarrow (b)$, assume to the contrary that there exists $R_1 \in \mathbb S$ such that $Y=\text{Spec}_{R_1} (S)$ is a maximal subset of $\text{Spec}(S)$. By assumption $\text{Spec}_{R_1} (S) \neq \text{Spec} (S)$. By Lemma \[not LO then there exists a better one\] there exists $R_2 \in \mathbb S$ such that $\text{Spec}_{R_1} (S) \subset \text{Spec}_{R_2} (S)$, a contradiction to the maximality of $Y$. We prove now $(b) \Rightarrow (c)$. Let $R_1 \in \mathbb S$. By assumption $\text{Spec}_{R_1} (S) \neq \text{Spec} (S)$, since otherwise $\text{Spec}_{R_1} (S)$ is a maximal subset of $\text{Spec} (S)$. By Lemma \[not LO then there exists a better one\] there exists $R_1 \supset R_2 \in \mathbb S$ such that $\text{Spec}_{R_1} (S) \subset \text{Spec}_{R_2} (S)$. By assumption $\text{Spec}_{R_2} (S) \neq \text{Spec} (S)$; so again by Lemma \[not LO then there exists a better one\], there exists $R_2 \supset R_3 \in \mathbb S$ such that $\text{Spec}_{R_2} (S) \subset \text{Spec}_{R_3} (S)$. We continue this way to get an infinite descending chain $\{R_k \}_{k \in \mathbb N}$ of $S$-nice subalgebras of $A$ such that $\text{Spec}_{R_i} (S) \subset \text{Spec}_{R_j} (S)$, whenever $i<j$. To prove that $(c) \Rightarrow (d)$, let $C$ be a nonempty closed subset of $\mathbb S$ and let $R_1 \in C$. The result follows by applying the assumption on $R_1$ and recalling the definition of the topology on $\mathbb S$. Finally, we show $(d) \Rightarrow (a)$. Assume to the contrary that there exists $R_1 \in \mathbb S$ satisfying LO over $S$. Then the closed subset $cl \{R_1 \}$ contains only elements that satisfy LO over $S$, a contradiction. [99]{} , P. Alexandroff, *Diskrete Räume*, Mat. Sb. (N.S.) 2 (1937), 501-518. , F. G. Arenas, *Alexandroff spaces*, Acta Math. Univ. Comenianae 1 (1999), 17–25. , O. Endler, *Valuation Theory*. Springer-Verlag, New York, 1972. , G. Gierz, K.H. Hofmann, K. Keimel, J.D. Lawson, M. Mislove, D.S. Scott, *Continuous Lattices and Domains*, Encyclopedia of Mathematics, Vol. 93, Cambridge University Press, Cambridge, MA, 2003. , G.T. Herman, *On topology as applied to image analysis*, Comput. Vision, Graphics Image Process, 52, (1990), 409-415. , E.H. Kronheimer, *The topology of digital images*, Top. and its Appl., 46 (1992), 279-303. , S. Sarussi, *Quasi-valuations extending a valuation*, J. Algebra 372 (2012), 318-364. , S. Sarussi, *Quasi-valuations and algebras over valuation domains*, Comm. Algebra, (2019), DOI: 10.1080/00927872.2018.1522322. , S. Sarussi, *Extensions of integral domains and quasi-valuations*, preprint available at https://arxiv.org/pdf/1805.02273.

--- abstract: | Given a reconfigurable system $X$, such as a robot moving on a grid or a set of particles traversing a graph without colliding, the possible positions of $X$ naturally form a cubical complex ${\mathcal{S}}(X)$. When ${\mathcal{S}}(X)$ is a CAT(0) space, we can explicitly construct the shortest path between any two points, for any of the four most natural metrics: distance, time, number of moves, and number of steps of simultaneous moves. Using Ardila, Owen, and Sullivant’s result that CAT(0) cubical complexes are in correspondence with posets with inconsistent pairs (PIPs), we can prove that a state complex ${\mathcal{S}}(X)$ is CAT(0) by identifying the corresponding PIP. We illustrate this very general strategy with one known and one new example: Abrams and Ghrist’s positive robotic arm on a square grid, and the robotic arm in a strip. We then use the PIP as a combinatorial “remote control" to move these robots efficiently from one position to another. author: - --- There are numerous contexts in mathematics, robotics, and many other fields where a discrete system moves according to local, reversible moves. For example, one might consider a robotic arm moving around a grid, a number of particles moving around a graph, or a phylogenetic tree undergoing local mutations. Abrams, Ghrist, and Peterson [@AG; @GP] introduced the formalism of *reconfigurable systems* to model a very wide variety of such contexts. Perhaps the most natural and important question that arises is the *motion-planning* or *shape-planning* question: how does one efficiently get a reconfigurable system $X$ from one position to another one? A first approach is to study the *transition graph* $G(X)$ of the system, whose vertices are the states of the system and whose edges correspond to the allowable moves between them. Abrams, Ghrist, and Peterson observed that $G(X)$ is the 1-skeleton of the *state complex* ${\mathcal{S}}(X)$: a cubical complex whose vertices are the states of $X$, whose edges correspond to allowable moves, and whose cubes correspond to collections of moves which can be performed simultaneously. In fact, ${\mathcal{S}}(X)$ can be regarded as the space of all possible positions of $X$, including the positions in between states. The geometry and topology of the state complex ${\mathcal{S}}(X)$ can help us solve the motion-planning problem for the system $X$. More concretely, ${\mathcal{S}}(X)$ is locally non-positively curved for *any* configuration system. [@AG; @GP] This implies that each homotopy class of paths from $p$ to $q$ contains a unique shortest path. Furthermore, the state complex of *some* reconfigurable systems is globally non-positively curved, or *CAT(0)*. This stronger property implies that for any two points $p$ and $q$ there is a unique shortest path between them. Ardila, Owen, and Sullivant [@AOS] gave an explicit algorithm to find this path. (In fact, there are at least four natural ways to measure the efficiency of a path between $p$ and $q$: by its Euclidean length inside ${\mathcal{S}}(X)$, by the amount of time that we take to traverse it, by the number of moves we perform, or by the number of steps of simultaneous moves that we perform. We show in this paper that, if ${\mathcal{S}}(X)$ is CAT(0), then we can find an optimal path under any of these four metrics.) It is therefore extremely useful to find out when a state complex ${\mathcal{S}}(X)$ is CAT(0). The first ground-breaking result in this direction is due to Gromov [@Gr], who gave a topological-combinatorial criterion for this geometric property. He proved that a cubical complex is CAT(0) if and only if it is simply connected (a topological condition) and the link of every vertex is a flag simplicial complex (a combinatorial condition). Roller [@Ro] and Sageev [@Sa], and Ardila, Owen, and Sullivant [@AOS] then gave two completely combinatorial descriptions of CAT(0) cubical complexes. These two descriptions are slightly different but equivalent, and they have different advantages depending on the context. Ardila, Owen, and Sullivant gave a bijection between (pointed) CAT(0) cubical complexes and certain simple combinatorial objects, which they called *posets with inconsistent pairs* (PIPs). These objects had been studied earlier in the computer science literature, where they were called *coherent event structures*. [@WN; @S] We restate this characterization of CAT(0) cubical complexes in Theorem \[th:poset\], and use it in a crucial way throughout the paper. In this paper, we put into practice the paradigm introduced by Ardila, Owen, and Sullivant to prove that cubical complexes are CAT(0). In principle, this method is completely general, and we illustrate it with one known and one new example of robotic arms: Abrams and Ghrist’s positive robotic arm in a quadrant, and the robotic arm in a strip of length 1. We use Theorem \[th:poset\] to prove that the state complexes of these robotic arms are CAT(0). It follows that we can find the optimal way to move these robots from one position to another one, under any of the four metrics. The key to navigating ${\mathcal{S}}(X)$ efficiently is to use its PIP as a combinatorial “remote control" that governs our movement inside ${\mathcal{S}}(X)$. While the state complex ${\mathcal{S}}(X)$ is often too large to be computed explicitly, the corresponding PIP is much smaller and more manageable. The paper is organized as follows. In Section \[sec:prelim\] we review the basic definitions of reconfigurable systems, cubical complexes, CAT(0) spaces, and the combinatorial characterization of CAT(0) cubical complexes in terms of *PIP*s (posets with inconsistent pairs). In Section \[sec:robots\] we introduce our two main objects of study: the positive robotic arm on a quadrant, and the robotic arm on a strip. In Sections \[sec:quadrant\] and \[sec:strip\], respectively, we prove that these two robotic arms give rise to CAT(0) cubical complexes by identifying their corresponding PIPs. In contrast, we show in Section \[sec:notCAT(0)\] an example of a state complex that is not CAT(0). In Section \[sec:optimization\] we discuss the four natural ways to measure a path between two positions $p$ and $q$ of a system $X$. When the state complex ${\mathcal{S}}(X)$ is CAT(0), we show how to find a shortest path between $p$ and $q$ under any of these metrics. {#sec:prelim} {#section-1} Abrams, Ghrist, and Peterson [@AG; @GP] introduced the general framework of *reconfigurable systems* to study discrete systems which vary according to local, reversible moves. This model applies to a wide variety of contexts, such as a robot moving on a grid without self-intersecting [@AG], a set of particles traversing a graph without colliding [@GP], or a variable element of a right angled Coxeter group $(W,S)$ which is subsequently getting multiplied by generators in $S$. [@Sc] We now recall the basic definitions pertaining to reconfigurable systems [@AG; @GP] and illustrate them with an example. Let $\mathcal{G}=\left(V,E\right)$ be a graph and $\mathcal{A}$ be a set of labels. A *state* is a labeling of the vertices of $\mathcal{G}$ by elements of $\mathcal{A}$. Roughly speaking, a *reconfigurable system* is given by a collection of states, together with a given set of moves called *generators* that one can perform to get from one state to another. More precisely: A *generator* $\varphi$ for a local reconfigurable system consists of: - A subset $SUP(\varphi)\subset V$ called the *support* of $\varphi$. - A subset $TR(\varphi)\subset SUP(\varphi)$ called the *trace* of $\varphi$. - An unordered pair of *local states* $u_{0}^{loc}$ and $u_{1}^{loc}$, which are labelings of $SUP(\varphi)$ by elements of $\mathcal{A}$ that agree outside of $TR(\varphi)$: $$u_{0}^{loc}|_{SUP(\varphi) - TR(\varphi)} = u_{1}^{loc}|_{SUP(\varphi) - TR(\varphi)}.$$ A generator $\varphi$ is *admissible* at a state $u$ if $u|_{SUP(\varphi)} \in \{u_{0}^{loc}, u_{1}^{loc}\}$. In that case, the generator $\varphi$ acts on the state $u$ to give the state $$\varphi[u]:= \left\{ \begin{array}{ccll}u & \textrm{on} & G-SUP(\varphi) & \\ u^{loc}_{1-i} & \textrm{on} & SUP(\varphi), & \textrm{where } u|_{SUP(\varphi)} = u_{i}^{loc}. \end{array}\right.$$ Physically, states $u$ represent different positions or configurations of our object of interest. Each generator $\varphi$ is a rule that allows us to move between states: the move $\varphi$ changes $u$ locally at $SUP(\varphi)$, switching it from $u_{0}^{loc}$ to $u_{1}^{loc}$ or vice versa. Naturally, to be able to apply $\varphi$, the state $u$ must look like one of these two labelings locally. A *reconfigurable system* on a graph consists of a collection $\Phi$ of generators, together with a collection of states which is closed under all possible admissible actions of the generators. ![A generator for a metamorphic robot in the hexagonal lattice. \[fig:hexrobot\]](constrainedhex.pdf) Consider a robot made up of identical hexagonal unit cells in the hexagonal lattice, which has the ability to pivot cells on the boundary whenever they are unobstructed, as shown in Figure \[fig:hexrobot\]a.-b. We can represent the hexagonal lattice by its dual graph $G$, where each unit hexagon is a vertex and we join neighboring vertices. The result is a triangular lattice. Then a state of the robot is just a labeling of the vertices of the triangular lattice $G$ with the set ${{\cal A}}=\{0,1\}$, where the $1$s represent the cells of the robot. Figure \[fig:hexrobot\]a.-d. illustrate one generator $\varphi$: a. and b. show the two local states of $SUP(\varphi)$, c. shows $TR(\varphi)$ shaded (the cells where the pivoting takes place) inside $SUP(\varphi)$ (where we now include the cell necessary to pivot, and the three cells that could obstruct the pivoting), and d. shows the effect of applying $\varphi$ to one particular state. A set $\{\varphi_1, \ldots, \varphi_n\}$ of generators *commute* if for any $i \neq j$ we have that $TR(\varphi_i)\cap SUP(\varphi_j)=\emptyset$. In words, generators commute when they are “physically independent", so they can be applied simultaneously to any state where they are all admissible. ![Four states and the two commuting generators connecting them.[]{data-label="fig:hexrobotstates"}](commgens-newnew.pdf) Figure \[fig:hexrobotstates\] shows four states which are connected by two commutative moves. Note that we could reach state $d$ from $a$ by simultaneously performing these two moves. Cubical complexes are a useful tool to model this in general. A *cubical complex* $X$ is a polyhedral complex obtained by gluing cubes of various dimensions, in such a way that the intersection of any two cubes is a face of both. Such a space $X$ has a natural piecewise Euclidean metric space as follows: Each cell is given the Euclidean metric of a cube of length $1$, and the distance between two points $p$ and $q$ in $X$ is the infimum among the lengths of all piecewise linear paths from $p$ to $q$ in $X$. Any reconfigurable system gives rise to a cubical complex: \[def:state complex\]The *state complex* $\mathcal{S}({\mathcal{R}})$ of a reconfigurable system $\mathcal{R}$ is a cubical complex whose vertices correspond to the states of $\mathcal{R}$. We draw an edge between two states if they differ by an application of a single generator. The $k$-cubes correspond to $k$-tuples of commutative moves: Given $k$ such moves which are applicable at a state $u$, we can obtain $2^k$ different states from $u$ by performing a subset of these $k$ moves; these are the vertices of a $k$-cube in $\mathcal{S}({\mathcal{R}})$. Figure \[fig:hexcomplex\] shows the state complex of the robot of 5 cells which starts horizontal in the lower right corner of a hexagonal tunnel of width 3, and is constrained to stay inside that tunnel. Notice that, due to the definition of the moves in Figure \[fig:hexrobot\], the robot is not able to pivot to the top row or out of the lower right corner. ![The state complex of a hexagonal metamorphic robot in a tunnel. \[fig:hexcomplex\]](hexcomplex.pdf) The following observation is part of Definition \[def:state complex\]; it is simple but important. [@GP] Let ${\mathcal{R}}$ be a reconfigurable system. - The 0-skeleton of the state complex ${\mathcal{S}}({{\mathcal{R}}})$ is the set of states ${\mathcal{R}}$. - The 1-skeleton of the state complex ${\mathcal{S}}({{\mathcal{R}}})$ is the transition graph of ${\mathcal{R}}$. Given a reconfigurable system ${\mathcal{R}}$ and a state $u$, there is a natural partial order on the states of ${\mathcal{R}}$ as follows. \[def:posetR\] Let ${\mathcal{R}}$ be a reconfigurable system and let $u$ be any “home" state. Define a poset ${\mathcal{R}}_u$ on the set of states by declaring that $p \leq q$ if there is a shortest edge-path from the home state $u$ to $q$ going through $p$. More precisely, $p \leq q$ if there is a sequence $u= p_0, p_1, \ldots, p_k=q$ of minimal length containing $p$, where each state $p_i$ can be obtained from $p_{i-1}$ by a single move. {#sec:combo} We now define CAT(0) spaces, the spaces of global non-positive curvature that we are interested in. For more information, see [@BBI; @BH]. Let $X$ be a metric space where there is a unique geodesic (shortest) path between any two points. Consider a triangle $T$ in $X$ of side lengths $a,b,c$, and build a comparison triangle $T'$ with the same lengths in the Euclidean plane. Consider a chord of length $d$ in $T$ which connects two points on the boundary of $T$; there is a corresponding comparison chord in $T'$, say of length $d'$. If $d \leq d'$ for any chord in $T$, we say that $T$ is a *thin triangle* in $X$. ![A chord in a triangle in $X$, and the corresponding chord in the comparison triangle in the plane. The triangle in $X$ is *thin* if $d \leq d'$ for all such chords. \[fig:thintriangle\]](thintriangle.pdf) A *CAT(0) space* is a metric space having a unique geodesic between any two points, such that every triangle is thin. A related concept is that of a *locally CAT(0)* or *non-positively curved* metric space $X$. This is a space where all sufficiently small triangles are thin. Testing whether a general metric space is CAT(0) is quite subtle. However, Gromov [@Gr] proved that this is easier if the space is a cubical complex. In a cubical complex, the link of any vertex is a simplicial complex. We say that a simplicial complex $\Delta$ is *flag* if it has no empty simplices; *i.e.*, if any $d+1$ vertices which are pairwise connected by edges of $\Delta$ form a $d$-simplex in $\Delta$. A cubical complex is CAT(0) if and only if it is simply connected and the link of any vertex is a flag simplicial complex. Ardila, Owen, and Sullivant [@AOS] gave a combinatorial description of CAT(0) cube complexes, which we now recall. This description may be seen as a global and purely combinatorial alternative to Gromov’s theorem. If $X$ is a CAT(0) cubical complex and $v$ is any vertex of $X$, we call $(X, v)$ a *rooted CAT(0) cubical complex*. The right side of Figure \[fig:bijection\] shows an example; the cube and the three square flaps are all part of the complex. The vertex labels will become relevant later. ![A poset with inconsistent pairs and the corresponding rooted CAT(0) cubical complex.\[fig:bijection\] ](cat0poset.pdf "fig:"){width="1.5in"} ![A poset with inconsistent pairs and the corresponding rooted CAT(0) cubical complex.\[fig:bijection\] ](cat0complex.pdf "fig:"){width="3in"} Recall that a poset $P$ is *locally finite* if every interval $[i, j] = \{k \in P \, : \, i \leq k \leq j\}$ is finite, and it has *finite width* if every antichain (set of pairwise incomparable elements) is finite. A *poset with inconsistent pairs (PIP)* is a locally finite poset $P$ of finite width, together with a collection of *inconsistent pairs* $\{p,q\}$, such that: 1. If $p$ and $q$ are inconsistent, then there is no $r$ such that $r \geq p$ and $r \geq q$. 2. If $p$ and $q$ are inconsistent and $p' \geq p$ and $q' \geq q$, then $p'$ and $q'$ are inconsistent. Posets with inconsistent pairs are equivalent to *coherent event structures*, defined earlier in the computer science literature; see for example [@WN; @S]. The *Hasse diagram* of a poset with inconsistent pairs (PIP) is obtained by drawing the poset, and connecting each minimal inconsistent pair with a dotted line. An inconsistent pair $\{p,q\}$ is *minimal* if there is no other inconsistent pair $\{p',q'\}$ with $p' \leq p$ and $q' \leq q$. Naturally, the minimal inconsistent pairs determine all other inconsistent pairs. For example, the left side of Figure \[fig:bijection\] shows the Hasse diagram of a PIP whose inconsistent pairs are $\{3,6\}$ and $\{5,6\}$. Recall that $I \subseteq P$ is an *order ideal* if $a \leq b$ and $b \in I$ imply $a \in I$. A *consistent order ideal* is one which contains no inconsistent pairs. \[def:cubePIP\] If $P$ is a poset with inconsistent pairs, we construct the *rooted cube complex of $P$*, which we denote $X(P)$. The vertices of $X(P)$ are identified with the consistent order ideals of $P$. There will be a cube $C(I,M)$ for each pair $(I, M)$ of a consistent order ideal $I$ and a subset $M \subseteq I_{max}$, where $I_{max}$ is the set of maximal elements of $I$. This cube has dimension $|M|$, and its vertices are obtained by removing from $I$ the $2^{|M|}$ possible subsets of $M$. The cubes are naturally glued along their faces according to their labels. The root is the vertex corresponding to the empty order ideal. Figure \[fig:bijection\] shows a PIP $P$ and the corresponding complex $X(P)$, which is rooted at $v$. For example, the compatible order ideal $I=\{1,2,3,4\}$ and the subset $M = \{1,4\} \subseteq I_{max}$ gives rise to the square with vertices labelled $1234, 123, 234, 23$. \[th:poset\] [@AOS] The map $P \mapsto X(P)$ is a bijection between posets with inconsistent pairs and rooted CAT(0) cube complexes. {#section-2} The influential paper of Billera, Holmes, and Vogtmann [@BHV] was one of the first to highlight the relevance of the CAT(0) property in applications. A fundamental question in phylogenetics is to guess the most likely evolutionary tree of $n$ present day species, say by measuring how different their DNA sequences are. To approach this question, Billera, Holmes, and Vogtmann proposed a construction of the *space of phylogenetic trees* $T_n$, a space whose points correspond to all the possible evolutionary trees for $n$ species. Many mathematical aspects of this phylogenetic question translate to understanding and efficiently navigating the space $T_n$. We now know that the space $T_n$ has close connections to important objects in algebraic geometry [@SS], tropical geometry [@AK; @Ar; @SS], topology [@Bo; @Vo], and combinatorics [@TZ]. Most relevantly to this paper, the space $T_n$ was shown in [@BHV] to be a CAT(0) cubical complex. This led to important consequences, such as the existence of unique geodesics and of “average trees" in $T_n$. Furthermore, after numerous partial results by many authors, Owen and Provan [@OP] recently gave the first polynomial time algorithm to compute geodesics in $T_n$. The work of Billera, Holmes, and Vogtmann was generalized in the following two directions: [@AOS]\[th:algorithm\] There is an algorithm to compute the geodesic between any two points in a CAT(0) cubical complex. [@AG; @GP] \[th:NPC\] The state complex of a reconfigurable system is a **locally** CAT(0) cubical complex; that is, all small enough triangles are thin. When the state complex of a reconfigurable system is **globally** CAT(0), we can use the algorithm in Theorem \[th:algorithm\] to navigate it. That will allow us to get our system from one position to another one in the optimal way. This highlights the importance of the following question: \[q:CAT(0)?\] Is the state complex of a given reconfigurable system a CAT(0) space? Theorem \[th:poset\] offers a new technique to provide an affirmative answer to Question \[q:CAT(0)?\]: Rooted CAT(0) cubical complexes are in bijection with PIPs; so to prove that a cubical complex is CAT(0), we “simply" have to choose a root for it, and find the corresponding PIP! The purpose of this paper is to put this paradigm to use for the first time, providing two concrete instances where it succeeds. We introduce the two relevant robots in Section \[sec:robots\], and provide combinatorial proofs that their state complexes are CAT(0) in Sections \[sec:quadrant\] and \[sec:strip\]. We remark that this technique is completely general. In principle, it works for **any** reconfigurable system whose state complex $X$ is CAT(0). In practice, it may not always be easy to identify the corresponding PIP. However, we hope to convince the reader that this can be done cleanly in many interesting special cases. {#sec:robots} In this section we introduce two robotic arms which have CAT(0) cubical complexes. The first robot was introducted by Abrams and Ghrist [@AG], who proved that its state complex is CAT(0), using combinatorial and topological methods. We give a purely combinatorial proof in Section \[sec:quadrant\]. To our knowledge, the second robot is new, and we use the same method in Section \[sec:strip\] to prove that its state complex is also CAT(0). {#subsec:quadrant} The following model, which we call $QR_n$, was first introduced in [@AG]. Consider a robotic arm consisting of $n$ links of unit length, attached sequentially. The robot lives inside an $n \times n$ grid, and its base is affixed to the lower left corner of the grid. Every one of its links must face north or east, starting from the base. The left panel of Figure \[fig:Qarm\] shows a position of the arm. ![The robotic arm in position $3568$ for $n=9$, and the corresponding particles on a line (to be introduced later).\[fig:Qarm\]](arm3568.pdf) The robot is free to move using the two local moves illustrated in Figure \[fig:QRmoves\]: - *NE-Switching corners*: Two consecutive links facing north and east can be switched to face east and north, and vice versa. - *NE-Flipping the end*: If the last link of the robot is facing east, it can be switched to face north, and vice versa. We call this the “positive" robotic arm because its joints can only face north or east. ![The two moves of $QR_n$.\[fig:QRmoves\]](QRmoves.pdf) It is clear that $QR_n$ has $2^n$ possible positions, corresponding to the paths of length $n$ which start at the southwest corner and always step east or north. We call these simply *NE-paths*. We will label each state of the robot using the set of its vertical steps: if a position of the robot has $k$ links facing north at positions $a_1, \ldots, a_k$ (counting from the base), then we label it $\{a_1, \ldots, a_k\}$ or simply $a_1\ldots a_k$. Notice that two states of different lengths can have the same label. We assume implicitly that the length of the robot is specified ahead of time. The system $QR_n$ is a reconfigurable system. We can specify a state of the robotic arm by labelling the edges of the grid with $0$s and $1$s, where a $1$ indicates that an edge is occupied by the robot. The moves can be reinterpreted as shown in the bottom half of Figure \[fig:QRmoves\]. The result follows from the definitions. ### {#sec:Qparticles} Consider a board consisting of $n$ slots on a line, and a system of indistinguishable particles hopping around the board. Any particle can hop to the slot immediately to its left or right whenever that slot is empty. Particles may enter the board by hopping onto the rightmost slot, and they may leave the board by hopping out of the rightmost slot. We say that two reconfigurable systems are *equivalent* if they have isomorphic state complexes. The system $QR_n$ is equivalent to the system of hopping particles on a board of length $n$. There is an obvious bijection between the states of these two systems: If the robot is in position $\{a_1, \ldots, a_k\}$, we associate to it the state where there are $k$ particles at slots $a_1, \ldots, a_k$. Switching a NE-corner of the robot corresponds to moving a particle left or right, and flipping the end of the robot corresponds to a particle entering or leaving the system from the right. This correspondence is illustrated in Figure \[fig:Qarm\]. {#subsec:strip} Now consider a robotic arm $SR_n$ which also consists of $n$ links of unit length, attached sequentially. Now the robot lives inside a $1 \times n$ grid, and its base is still affixed to the lower left corner of the grid, but the links do not necessarily have to face north and east. The left panel of Figure \[fig:SRparticles\] shows a position of the arm. ![The robotic arm in position $1479$ for $n=9$, and the corresponding particles on a line. \[fig:SRparticles\]](SRparticles.pdf) The robot starts out fully horizontal, and is free to move using the local moves illustrated in Figure \[fig:SRmoves\]: - *Switching corners*: Two consecutive links facing different directions can interchange their directions. - *Flipping the end*: The end of the robot can rotate $90^{\circ}$ as long as it does not intersect the rest of the robot. ![The four kinds of moves of $SR_n$.\[fig:SRmoves\]](SRmoves.pdf) The $SR_n$ robot is restricted to a smaller board, but it is not positive, so it has a wider range of moves. It can switch between north-east (NE) and east-north (EN) corners, as well as between south-east (SE) and east-south (ES) corners. The end of the robot can flip from facing east to facing either south or north. (No link ever faces west, due to the small height of the grid.) The system $SR_n$ is a reconfigurable system. Again it is clear how to convert the moves of $SR_n$, shown in Figure \[fig:SRmoves\], to the language of reconfigurable systems. The number of possible positions of the robotic arm $SR_n$ is equal to the $(n+2)$-th Fibonacci number $F_{n+2} = \frac1{\sqrt 5}\left(\left(\frac{1+\sqrt 5}2 \right)^{n+2} - \left( \frac{1-\sqrt 5}2 \right)^{n+2} \right)$. Let $a_n$ be the number of positions of $SR_n$. The positions whose first step faces east are in bijection with the positions of $SR_{n-1}$. On the other hand, if the first step of the robot faces north, then the second step must face east, and the rest of the robot – after flipping it upside down – is a position of $SR_{n-2}$. Therefore $a_n=a_{n-1}+a_{n-2}$. Since $a_1=2$ and $a_2=3$, the result follows. For this reason, we call a position of $SR_n$ a *Fibonacci path*. We also say that a subset of $[n]$ is *spread out* if it does not contain any two consecutive integers. Again we label each state of the robotic arm using the set of its vertical steps: if a position of the robot has $k$ vertical links at positions $a_1, \ldots, a_k$ (counting from the base), then we label it $\{a_1, \ldots, a_k\}$ or simply $a_1\ldots a_k$. For example, the label of the position of Figure \[fig:SRparticles\] is $\{1,4,7,9\}$, or simply $1479$. It is clear how to recover the position of the arm from its label, so this gives a bijection between the states of $SR_n$ and the spread out subsets of $[n]$. ### {#sec:Sparticles} In analogy with what we did for $QR_n$, consider a board consisting of $n$ slots on a line, and a system of indistinguishable *repellent particles* hopping around the board. The repellent particles must stay at distance at least $2$ from each other; so a particle can hop to the slot immediately to its left or right whenever that slot and its other neighbor are empty. Particles may enter the board by hopping onto the rightmost slot, and they may leave the board by hopping out of the rightmost slot. The system $SR_n$ is equivalent to the system of hopping repellent particles on a board of length $n$. Again, there is an obvious bijection between the states of these two systems: If the arm is in position $\{a_1, \ldots, a_k\}$, we associate to it the state where there are $k$ particles at slots $a_1, \ldots, a_k$. Switching a corner of the robot still corresponds to moving a particle left or right, and flipping the end of the robot still corresponds to a particle entering or leaving the system from the right. This correspondence is illustrated in Figure \[fig:SRparticles\]. ### {#section-3} We will find it useful to construct a second robot, closer to $QR_n$ in spirit, which is equivalent to $SR_n$. Consider the *pyramidal grid of size $n$*, which we denote ${\mathrm{Pyr}}_n$, consisting of strips of sizes $1 \times n, \, 1 \times (n-2), \, 1 \times (n-4), \ldots $ stacked on top of each other in decreasing order, so that each level is centered on top of the previous level. Now consider the positive robotic arm in the grid ${\mathrm{Pyr}}_n$ whose base is still affixed to the lower left corner of the grid, and whose links must point north or east. The robot moves using the following restricted versions of the two local moves of $QR_n$. - *NE-Switching corners*: Two consecutive links facing north and east **which are preceded and followed by east links** can be switched to face east and north, and vice versa - *NE-Flipping the end*: If the last link of the robot is facing east and **it is preceded by an east link**, then it can be switched to face north, and vice versa. Now the possible positions of the robot correspond to the *F-paths*: the paths in ${\mathrm{Pyr}}_n$ which start at the southwest corner and take steps north and east, without ever taking two consecutive steps north. The lower right corner of Figure \[fig:unfurl\] shows an F-path of length $9$. \[prop:model2\] The system $SR_n$ is equivalent to the robotic arm in the pyramidal grid ${\mathrm{Pyr}}_n$. Given a Fibonacci path of length $n$, we create a unique F-path by *unfolding*, as follows. First we place the Fibonacci path in the bottom row of ${\mathrm{Pyr}}_n$, attached at the bottom left-corner. Now we follow the path from left to right. Whenever we encounter a vertical link, starting from the second one, we “unfold" the arm into the next level of ${\mathrm{Pyr}}_n$ by taking the rest of the arm and flipping it vertically. We repeat this process until the path has no edges facing south; the result is a unique F-path. This construction is illustrated in Figure \[fig:unfurl\]. It is clear how to recover the Fibonacci path from the F-path. ![Unfolding the Fibonacci path $1479$ to obtain an F-path.\[fig:unfurl\]](unfurl1479-rounded.pdf) One then easily checks that this bijection is compatible with the moves of the systems: switching the corners and flipping the end of $SR_n$ correspond to NE-switching the corners and flipping the end of the arm in ${\mathrm{Pyr}}_n$, respectively. {#sec:quadrant} Having introduced all the background information, we now present a combinatorial proof that the state complex ${\mathcal{S}}(QR_n)$ of the robotic arm in a quadrant $QR_n$ is CAT(0). Abrams and Ghrist give a combinatorial-topological proof in [@AG]. In view of Theorem \[th:poset\], our strategy is as follows. We root the complex ${\mathcal{S}}(QR_n)$ at a natural vertex $v$, namely, the one corresponding to the fully horizontal robot. If ${\mathcal{S}}(QR_n)$ really is CAT(0), then Theorem \[th:poset\] puts it in correspondence with a PIP (poset with inconsistent pairs) $QP_n$. By drawing the first few examples, it is not difficult to guess what the correct PIP should be in general. (For the two examples in this paper, it turns out that the PIP does not have any inconsistent pairs.) We then prove that, under the bijection of Theorem \[th:poset\], the PIP $QP_n$ is mapped to the (rooted) state complex of $QR_n$. Therefore this complex must be CAT(0). Define the PIP $QP_n$ to be the set of lattice points inside the triangle $y \geq 0, y \leq x,$ and $x \leq n-1$, with componentwise order (so $(x,y) \leq (x',y')$ if $x \leq x'$ and $y \leq y'$) and no inconsistent pairs. The poset $QP_n$ has the triangular shape shown in Figure \[fig:QPn\] for $n=6$. ![The poset $QP_6$. \[fig:QPn\]](QP6.pdf) \[prop:bijQ\] There is a bijection between the states of the robot $QR_n$ and the order ideals of the poset $QP_n$. Consider the $n(n+1)/2$ cells on the lower left half of the $n \times n$ board, below the main diagonal. Partially order them so that each cell is less than the cell to its left and the cell above it (if they exist). Clearly this order is isomorphic to $QP_n$; if we rotate the board $45^\circ$ clockwise, we will get the cells to line up with the Hasse diagram of this poset. Any NE-path is contained in this lower left half of the board, and divides it into two parts. The part below the path clearly forms an order ideal of $QP_n$. Conversely, any order ideal comes from an NE-path in this way. ![The bijection between states of $QR_n$ and order ideals of $QP_n$. \[fig:Qbijection\]](bijQP9.pdf) Recall that a *lattice* is a poset where any two elements $x$ and $y$ have a least upper bound $x \vee y$ and a greatest lower bound $x \wedge y$. A lattice is *distributive* if any three elements $x,y,z$ satisfy $x \vee (y \wedge z) = (x \vee y) \wedge (x \vee z)$ (which implies the dual relation.) An element is *join-irreducible* if it is not the join of two smaller elements. Birkhoff’s theorem states that given a poset $P$, the poset of order ideals of $P$ ordered by containment is a distributive lattice. Conversely, every distributive lattice $L$ arises uniquely in this way from a poset $P$. In fact, $P$ can be recovered as the induced poset of join-irreducibles of $L$. For more information, see [@EC1]. In light of Birkhoff’s theorem, the following is immediate from Proposition \[prop:bijQ\]. If we declare the “home" state of $QR_n$ to be the fully horizontal state, then the poset of states of $QR_n$ is a distributive lattice. Let the *word* *of a subset* $A=\{a_1 < a_2 < \cdots < a_k\} \subseteq [n]$ be the length $n$ word $w(A) = (a_1, a_2, \ldots a_k, (n+1), (n+1), \ldots, (n+1))$. \[prop:Qorder\] The lattice of states of $QR_n$ is isomorphic to the poset on the subsets of $[n]$, where $A \leq B$ if $w(A) \geq w(B)$ coordinatewise. This is clear from the hopping particles model for $QR_n$ of Section \[sec:Qparticles\]. Each state corresponds to the subset of $[n]$ indicating the locations of the particles. We move up the poset by moving a particle to the left or having a particle enter from the right; and that precisely corresponds to decreasing a coordinate of $w(A)$ by one. Birkhoff’s theorem tells us that the join-irreducible elements of the poset of states $QR_n$ should form a poset isomorphic to $QP_n$. A state of the hopping particles model is join-irreducible when only one particle can jump to the right. These are the states where the particles occupy positions $\{i, i+1, \ldots, j\}$ for some $i \leq j$. We can verify directly that these $n(n+1)/2$ states indeed from an isomorphic copy of the poset $QP_n$. Figure \[fig:Qjoinirreds\] illustrates this in the example $n=4$. We have rotated the lattice $90^\circ$ clockwise to save space. ![The join-irreducibles of the poset of $QR_n$ form a copy of $QP_n$.\[fig:Qjoinirreds\]](Qjoinirreds.pdf) Having established these results about the 1-skeleton of the state complex, we now extend them to the higher-dimensional cubes. A *partial NE-path* is a path of consecutive links which may be north edges $N$, east edges $E$, unit squares $\square$, or partial unit squares $\llcorner$, such that - Each unit square is attached to the rest of the path by its southwest and northeast corners. - There is at most one partial unit square $\llcorner$, which must be the last link, and must be attached to the rest of the path by its southwest corner. The *length* of a partial NE-path is $e+2f+g$, where $e$ is the number of edges, $f$ is the number of squares, and $g$ is the number of half-squares (which is $0$ or $1$). The partial NE-paths form a poset by containment, whose minimal elements are the NE-paths. To illustrate this definition, Figure \[fig:NEbijection2\] shows a partial NE-path which contains the NE-path of Figure \[fig:Qbijection\]. Recall that $X(QP_n)$ is the rooted cube complex corresponding to the PIP $QP_n$ under the bijection of Theorem \[th:poset\]. We use the notation of Definition \[def:cubePIP\]. \[lemma:QP\] The partial NE-paths of length $n$ having $k$ squares or half-squares are in order-preserving bijection with the $k$-dimensional cubes of $X(QP_n)$. Consider a partial NE-path $p$. Its upper boundary is an NE-path corresponding to an order ideal $I \subseteq QP_n$, and its $k$ squares (possibly including a half-square at the end) correspond to $k$ maximal elements $m_1, \ldots, m_k$ of $I$. Let $p$ correspond to the cube $C(I, \{m_1, \ldots, m_k\})$ of $X(QP_n)$. This correspondence is illustrated in Figure \[fig:NEbijection2\]. This is clearly an order-preserving bijection. Now recall that ${\mathcal{S}}(QR_n)$ is the state complex of the reconfigurable system $QR_n$. \[lemma:QR\] The partial NE-paths of length $n$ having $k$ squares or half-squares are in order-preserving bijection with the cubes of the state complex ${\mathcal{S}}(QR_n)$. ![The bijection between partial NE-paths and pairs $(I,M)$ consisting of an order ideal $I$ and a subset $M$ of maximal elements of $I$ in $QP_n$. \[fig:NEbijection2\]](NEbijection9.pdf) A $k$-cube $C$ of ${\mathcal{S}}(QR_n)$ is given by a state $u$ and a collection $\{\varphi_1, \ldots, \varphi_k\}$ of $k$ commutative moves that can be applied to $u$. The state $u$ is given by an NE-path, and each one of the $k$ moves $\varphi_1, \ldots, \varphi_k$ corresponds to a corner of the NE-path that could be switched. The two positions of this corner before and after the move $\varphi_i$ form a square (or half-square). Since the moves are commutative, two of these squares cannot share an edge. Adding these $k$ squares to the NE-path $u$ gives rise to a partial NE-path corresponding to the $k$-cube $C$. Conversely, consider a partial NE-path with $k$ squares. There are $2^k$ NE-paths contained in it, obtained by “resolving" each square into an NE or an EN corner, or “resolving" the half-square into an N or an E step. The resulting $2^k$ NE-paths form a cube of ${\mathcal{S}}(QR_n)$. This bijection is clearly order-preserving. \[th:QR\] The state complex of the robotic arm $QR_n$ in an $n \times n$ grid is a CAT(0) cubical complex. Lemmas \[lemma:QP\] and \[lemma:QR\] show that the state complex ${\mathcal{S}}(QR_n)$ is isomorphic to the cube complex $X(QP_n)$, which is CAT(0) by Theorem \[th:poset\]. Let $q_{n,d}$ be the number of $d$-cubes in the state complex of the robot in a quadrant $QR_n$. Then $$\sum_{n, d \geq 0} q_{n,d}\, x^ny^d = \frac{1+xy}{1-2x-x^2y}$$ The number $q_{n,d}$ counts the partial NE-paths of length $n$ having $d$ squares or half-squares. Say that such a path $P$ has *weight* $wt(P)=x^ny^d$. Each partial NE-path corresponds to a sequence of symbols $N,E,\square$, or [$\llcorner$]{}, where the symbol $\llcorner$ can only occur at the end of the sequence. Each $N$ contributes a weight of $x$, each $E$ contributes a weight of $x$, each $\square$ contributes a weight of $x^2y$, and each [$\llcorner$]{} contributes a weight of $xy$. Therefore $$\begin{aligned} \sum q_{n,d}\, x^ny^d &=& \sum_{\textrm{NE-paths } P} {\mathrm{wt}}(P) \\ &=& \sum_{\textrm{NE-paths } P \textrm{ with no } \llcorner} {\mathrm{wt}}(P) + \sum_{\textrm{NE-paths } P \textrm{ with } \llcorner} {\mathrm{wt}}(P) \\ &=& \frac{1}{1-{\mathrm{wt}}(N) - {\mathrm{wt}}(E) - {\mathrm{wt}}(\square)} + \frac{{\mathrm{wt}}({\large \llcorner})}{1-{\mathrm{wt}}(N) - {\mathrm{wt}}(E) - {\mathrm{wt}}(\square)} \\ &=& \frac{1+xy}{1-2x-x^2y}.\end{aligned}$$ Here we are using the observation that $$\frac{1}{1-{\mathrm{wt}}(N) - {\mathrm{wt}}(E) - {\mathrm{wt}}(\square)} = \sum_{\stackrel{\textrm{ sequences } a_1 \ldots a_k \textrm{ with }}{a_i \in \{N,E,\square\}}} {\mathrm{wt}}(a_1)\cdots {\mathrm{wt}}(a_k)$$ is the generating function for the weights of the NE-paths which do not include the symbol $\llcorner$. The result follows. Notice that the PIP of this robot is much smaller than its state complex. While the state complex $QR_n$ has exponentially many faces, the size of the PIP $QP_n$ is quadratic in $n$. {#sec:strip} Now we carry out the same approach for the robotic arm in a strip $SR_n$. Define the PIP $SP_n$ to be the set of lattice points inside the triangle $y \geq 0, y \leq 2x,$ and $x \leq n-1$, with componentwise order (so $(x,y) \leq (x',y')$ if $x \leq x'$ and $y \leq y'$) and no inconsistent pairs. ![The poset $SP_n$. \[fig:SPn\]](SP9.pdf) The poset $SP_n$ has the triangular shape shown in Figure \[fig:SPn\] for $n=9$. It is the union of a “stack" of chains of lengths $n, n-2, n-4, \ldots$. \[prop:bijS\] There is a bijection between the states of the robot $SR_n$ and the order ideals of the poset $SP_n$. We wish to imitate the proof of Proposition \[prop:bijQ\]. It is not obvious how to do it for the robot in a strip, and this is the reason that we introduced the equivalent model of the robot in a pyramid in Proposition \[prop:model2\]. Consider the $n + (n-2) + (n-4) + \cdots$ cells of the pyramidal board ${\mathrm{Pyr}}_n$. Partially order them so that each cell is less than the cell to its left and the cell northeast of it (if they exist), as shown in Figure \[fig:Sbijection\]. This order is isomorphic to $SP_n$. Any F-path divides the pyramid into two parts. Since an F-path does not contain two consecutive steps north, the part of ${\mathrm{Pyr}}_n$ below it forms an order ideal of $QP_n$. Conversely, any order ideal comes from an F-path in this way. ![The bijection between states of $SR_n$ and order ideals of $SP_n$. \[fig:Sbijection\]](bijSR9SP9.pdf) If we declare the “home" state of $SR_n$ to be the fully horizontal state, then the poset of states of $SR_n$ is a distributive lattice. Recall that a set of integers is *spread out* if it contains no two consecutive integers. The lattice of states of $SR_n$ is isomorphic to the poset on the spread out subsets of $[n]$, where $A \leq B$ if $w(A) \geq w(B)$ coordinatewise. This is clear from the repellent hopping particles model for $SR_n$ of Section \[sec:Sparticles\], as in Proposition \[prop:Qorder\]. The join-irreducible elements of the poset of $SR_n$ correspond to the states of the hopping particles model where only one particle can jump to the right. These are the states where the particles occupy positions $\{i, i+2, i+4,, \ldots, j\}$ for some $i \leq j$ of the same parity. These states form a copy of $QP_n$ inside the poset of $QR_n$, as illustrated in Figure \[fig:Sjoinirreds\] for $n=6$. This lattice is also rotated $90^\circ$ clockwise. ![The join-irreducibles of the poset of $SR_n$ form a copy of $SP_n$.\[fig:Sjoinirreds\]](bijSR6.pdf) A *partial F-path* is a partial NE-path such that the link following any vertical edge or square, if there is one, must be a horizontal edge. Recall that $X(SP_n)$ is the rooted cube complex corresponding to the PIP $SP_n$ under the bijection of Theorem \[th:poset\]. We then have the following results. The proofs are essentially the same as those of Lemmas \[lemma:QP\] and \[lemma:QR\] and Theorem \[th:QR\]. \[lemma:SP\] The partial F-paths of length $n$ having $d$ squares or half-squares are in bijection with the $d$-dimensional cubes of the state complex of $X(SP_n)$. \[lemma:SR\] The partial F-paths of length $n$ having $d$ squares or half-squares are in bijection with the $d$-dimensional cubes of the state complex of $SR_n$. \[th:SR\] The state complex of the robotic arm $SR_n$ of length $n$ in a strip of width 1 is a CAT(0) cubical complex. Let $s_{n,d}$ be the number of $d$-cubes in the state complex of the robot in a strip $SR_n$. Then $$\sum_{n, d \geq 0} s_{n,d}\, x^ny^d = \frac{1+x+xy+x^2y}{1-x-x^2-x^3y}.$$ The number $s_{n,d}$ counts the sequences of $N,E$, $\square$, and [$\llcorner$]{} having weight $x^ny^d$ which contain none of the subwords $NN, N\square, \square N$, and $\square\square$ (so that an $N$ or an $\square$ can only be followed by an $E$), and which may only contain $\llcorner$ at the end of the sequence. Each such word may be regarded as a sequence of the “clusters" $NE, \square E,$ and $E$, possibly followed by a single $N$, $\square$, or [$\llcorner$]{}. Therefore $$\begin{aligned} \sum_{n, d \geq 0} s_{n,d}\, x^ny^d &=& \sum_{\textrm{F-paths } P} {\mathrm{wt}}(P) \\ &=& \frac{1+{\mathrm{wt}}(N) + {\mathrm{wt}}(\square)+{\mathrm{wt}}({\large \llcorner})}{1-{\mathrm{wt}}(NE) - {\mathrm{wt}}(\square E) - {\mathrm{wt}}(E)} \\ &=& \frac{1+x+xy+x^2y}{1-x^2 - (x^2y)x-x},\end{aligned}$$ as desired. The PIP of this robot is also quadratic in size, while the state complex is exponential in size. {#sec:notCAT(0)} It is worthwhile to exhibit an example of a similar robot whose state complex is not CAT(0). There are many such examples; *e.g.*, see [@AG]. Here we consider an “unpinned" version of the robotic arm in Section \[sec:strip\], which we can think of as a robotic snake. Now we are allowed to flip the end and switch any corners as long as we do not make the robot self-intersect. In general, the cubical complex of such an unpinned planar robotic snake is not CAT($0$). Let $U^{\ell}_{m,n}$ denote the reconfigurable system corresponding to the unpinned robotic snake of length $\ell$ in an $m\times n$ grid, and ${\mathcal{S}}(U^{\ell}_{m,n})$ its corresponding state complex. The state complex ${\mathcal{S}}(U^{1}_{m,n})$ is not CAT($0$) for any $n,m\in\mathbb{Z}$. In fact, one checks easily that ${\mathcal{S}}(U^{1}_{m,n})$ consists of $mn$ empty squares arranged diagonally on an $m \times n$ rectangular grid, glued corner to corner as shown in Figure \[fig:notCAT(0)\]. This phenomenon can be observed for any given $\ell$ and sufficiently large $m$ and $n$. ![The state complexes $SC^{1}_{1,6}$ and $SC^{1}_{3,5}$.\[fig:notCAT(0)\]](SC16-SC35.pdf) {#sec:optimization} Consider a robot, or some other reconfigurable system ${\mathcal{R}}$, whose state complex ${\mathcal{S}}({\mathcal{R}})$ is CAT(0). As in the two examples above, there may be a natural choice of a “home state" $u$, such that the PIP $P_u$ corresponding to the rooted complex $({\mathcal{S}}({\mathcal{R}}),u)$ has a particularly simple description. Now suppose that we want to take the robot from state $a$ to state $b$ in an optimal way. Equivalently, we wish to get from vertex $a$ to vertex $b$ of the state complex ${\mathcal{S}}({\mathcal{R}})$. {#subsec:reroot} To find the optimal path from $a$ to $b$, the first step will be to reroot the complex at $a$, and find the PIP $P_a$ corresponding to the rooted CAT(0) cubical complex $({\mathcal{S}}({\mathcal{R}}),a)$. Fortunately, this is very easy to do. If $p$ and $q$ are an inconsistent pair in a PIP, write $p \nleftrightarrow q$. \[prop:reroot\] Let $u$ and $a$ be vertices of the CAT(0) cube complex $X$ and let $P_u$ and $P_a$ be the PIPs corresponding to the rooted complexes $(X,u)$ and $(X,a)$ respectively. Let $I$ be the consistent order ideal of $P_u$ corresponding to $a$, and let $J=P_u-I$. The PIP $P_a$ has an element $p'$ corresponding to each element $p \in P_u$, and it can be described in terms of $P_u$ as follows: - If $j_1<j_2$ in $P_u$, then $j_1'<j_2'$ in $P_a$. - If $i_1<i_2$ in $P_u$ then $i_1'<i_2'$ in $P_a$. - If $i<j$ in $P_u$ then $i' \nleftrightarrow j'$ in $P_a$. - If $j_1 \nleftrightarrow j_2$ in $P_u$, then $j_1' \nleftrightarrow j_2'$ in $P_a$. - If $i \nleftrightarrow j$ in $P_u$ then $i' < j'$ in $P_a$. Here the $i$s and the $j$s represent arbitrary elements of $I$ and $J$, respectively.[^1] ![The PIPs $P_u$ and $P_a$ before and after rerooting the CAT(0) cube complex.\[fig:reroot\]](reroot.pdf) \[cor:reroot\] The Hasse diagram of $P_a$ is obtained from that of $P_u$ by - turning $I$ upside down, and - converting all solid edges from $I$ to $J$ into dotted edges, and vice versa. The effect of rerooting on the PIP is illustrated in Figure \[fig:reroot\]. Notice that, even if $P_u$ has no inconsistent pairs, the PIP $P_a$ probably will have inconsistent pairs. (One can easily show that a CAT(0) cube complex has at most two roots whose PIPs have no inconsistent pairs, corresponding to the top and bottom elements of a distributive lattice.) Let us establish a bijection between the vertices of $X(P_u)$ and the vertices of $X(P_a)$. Given a consistent order ideal $B$ of $P_u$, write $B=B_I \cup B_J$ where $B_I = B \cap I$ and $B_J = B \cap J$. We claim that $B' = (I'-B_I') \cup B_J'$ is a consistent order ideal of $P_a$. (We write $S'=\{s' \, : \, s \in S\} \subseteq P_a$ for each $S \subseteq P_u$.) First assume, for the sake of contradiction, that $B'$ is not an order ideal of $P_a$. Since $B_I$ and $B_J$ are ideals of $I$ and $J$ respectively, $I'-B_I'$ and $B_J'$ are ideals of $I'$ and $J'$ respectively. Therefore we must have $j' > i'$ for some $j' \in B_J'$ and $i' \notin (I'-B_I')$; that is, $i' \in B_I'$. By the definition of $P_a$, we must then have $i \nleftrightarrow j$ in $P_u$ where $i,j \in B$, contradicting the consistency of $B$. Now assume that the order ideal $B'$ contains an inconsistent pair. Now, $I'$ contains no inconsistent pairs by the definition of $P_a$. If $B_J'$ contained an inconsistent pair, then $B_J$ would have to contain one as well. So the inconsistent pair must consist of $i' \in I' - B_I'$ and $j' \in B_J'$. But then $i < j$ for $i \notin B_I$ and $j \in B_J$, contradicting the fact that $B$ is an order ideal. This concludes the proof. The mapping $B \mapsto B'$ above establishes a bijection between the vertices of $X(P_u)$ and the vertices of $X(P_a)$. It is easy to see that adjacencies and cubes in $X(P_u)$ correspond to adjacencies and cubes in $X(P_u)$ as well. Therefore the PIP $P_a$ must be precisely the one we obtain by rerooting $X(P_u)=X(P_a)$ at $a$, as desired. Having described the PIP $P_a$ explicitly in terms of $P_u$, it is straightforward to verify the description of its Hasse diagram. Now that we have rerooted the complex, our goal is to get from the root $a$ to the vertex $b$ in the optimal way. This will get our robot from position $a$ to position $b$ efficiently. There are at least four notions of “optimality" for which we can do this. {#section-4} Suppose we want to find the shortest path from $a$ to $b$ in the Euclidean metric of the cubical complex ${\mathcal{S}}({\mathcal{R}})$. This can be accomplished using Ardila, Owen, and Sullivant’s algorithm [@AOS] to compute the shortest path from $a$ to $b$. As explained there, a prerequisite for this is to write down the PIP $P_a$, which we have done in Proposition \[prop:reroot\]. This metric is very useful in some applications, particularly when navigating the space of phylogenetic trees [@BHV; @OP]. However, this metric does not seem ideally suited to the robotic applications we have in mind here. For instance, suppose that the optimal path from $a$ to $b$ crosses a diagonal of a $d$-dimensional cube. To a robot, this means performing $d$ moves simultaneously. The cost of doing this in the Euclidean metric is $\sqrt{d}$, the length of the diagonal. In practice, it is conceivable that there may be a higher cost to performing $d$ moves simultaneously, but we do not know of a context where that cost would be $\sqrt{d}$. It is probably more natural to consider the following three variants. {#section-5} Suppose we are only allowed to perform one move at a time. Geometrically, we are looking for a shortest edge-path from $a$ to $b$. Let $B$ be the consistent order ideal of $P_a$ corresponding to vertex $v$ in the rooted complex $({\mathcal{S}}({\mathcal{R}}),a)$. We can regard $B$ as a subposet of $P_a$. The shortest edge-paths from $a$ to $b$ are in one-to-one correspondence with the linear extensions of the poset $B$. Their length is $|B|$. Combinatorially, a shortest path from $a$ to $b$ is simply obtained by building up the order ideal $B$ one element at a time, starting with the empty set, and adding a minimal missing element of $B$ at each step. The previous argument also shows how to construct the minimal shortest paths. {#section-6} Now suppose that we can move the robot in steps, where at each step we can perform several moves at a time with no penalty. Geometrically, we are looking for a shortest cube path from $a$ to $b$, where at each step we cross a cube from the current vertex to the one across the diagonal. Again, let $B$ be the consistent order ideal $B$ of $P_a$. Suppose that two opposite vertices $i$ and $j$ of a cube $C$ correspond to the order ideals $I$ and $J$ of $P_a$, respectively. Then in the notation of Definition \[def:cubePIP\], the cube must be $C=C(I \cup J, I \triangle J)$, where $I \triangle J := (I-J) \cup (J-I)$ is the symmetric difference of $I$ and $J$. In particular, $I \triangle J$ is an antichain. It follows that combinatorially, a cube path $a=v_0, v_1, \ldots, v_k=b$ corresponds to a sequence of order ideals $\mathbf{I}: \emptyset=I_0, I_1, \ldots, I_k=B$ such that the symmetric difference $I_j \triangle I_{j+1}$ is an antichain for all $j$. Let the *depth* $d(B)$ of $B$ be the size of the longest chain(s) in $B$. \[def:normal\] Let the *normal cube path* from $a$ to $b$ be the cube path given by the sequence of order ideals $\mathbf{M} : \emptyset = M_0 \subset M_1 \subset \cdots \subset M_{d(B)}=B$, where each ideal is obtained from the previous one by adding to it all the minimal elements that have not yet been added. In other words, $M_{k+1}:= M_{k} \cup (B-M_{k})_{min}$ for $k \geq 0$. Alternatively, $M_k$ consists of the elements of $B$ whose shortest path to a minimal element has lengh $k-1$. From this it follows that $M_{d(B)}=B$ and $M_{d(B)-1} \neq B$. The previous definition is due to Niblo and Reeves [@NR] in a different language; the correspondence with PIPs makes these paths more explicit. It also allows us to give a simple proof of the following result from Reeves’s Ph.D. thesis [@Re]: The shortest cube paths from $a$ to $b$ have size $d(B)$. In particular, the normal cube path from $a$ to $b$ is minimal. To prove that $\mathbf{M}$ is indeed minimal, we claim that any other cube path $\mathbf{I}$ satisfies $I_j \subseteq M_j \subsetneq B$ for all $j < d(B)$. We prove it by induction; the case $j=0$ is trivial. We assume that $I_j \subseteq M_j$, and prove that $I_{j+1} \subseteq M_{j+1}$. Consider an element $i \notin M_{j+1}$. Then we have $i \notin M_j$ (and therefore $i \notin I_j$) and $i \notin (B-M_j)_{min}$, which then implies that $i>i'$ for some $i' \in B-M_j$. This gives $i' \in B-I_j$, which implies that $i \notin (B-I_j)_{min}$. If we had $i \in I_{j+1}$, we would also have $i' \in I_{j+1}$, which would imply that the chain $i<i'$ is a subset of the antichain $I_j \triangle I_{j+1}$, a contradiction. Therefore $i \notin I_{j+1}$ as desired. If a robot has a CAT(0) state complex, the above results give us a way to move the robot from state $a$ to state $b$. First we compute the PIP $P_u$ corresponding to a “natural" choice of a root $u$. This is the only non-trivial step, and its difficulty depends on the description of the reconfiguration system that we are given. Having found $P_u$, the rest is easy. We use Proposition \[prop:reroot\] to reroot the complex to $a$ and find the corresponding PIP $P_a$. We then find the order ideal $B \subseteq P_a$, and compute the normal cube path of Definition \[def:normal\]. The shortest cube path from $a$ to $b$ is not necessarily unique. For instance, the “reverse" normal cube path from $b$ to $a$ also has length $d(B)$, and is generally different. In fact, Abrams and Ghrist [@AG Algorithm 8.1] gave a fast algorithm that starts with any given edge path from $a$ to $b$, and transforms it into a shortest cube path from $a$ to $b$. Their algorithm has the advantage that it does not require one to compute the whole state complex, which is often exponential in size. As we have explained, our approach offers an alternative way to overcome this difficulty. In the examples we have considered, and in most natural reconfiguration systems in the plane, a state complex of exponential size has a PIP of quadratic size. {#section-7} Perhaps the most realistic model is to allow ourselves to move the robot continuously in time, where we can perform several moves simultaneously, as long as these moves are physically independent. We can even perform only part of a move, and perform the rest of the move later. Each move still takes one unit of time, and there is no time penalty for multitasking. Geometrically, we are endowing each cube with the $\ell_\infty$ metric: For $\mathbf{x}, \mathbf{y}$ in a unit $d$-cube, we let $||\mathbf{x} - \mathbf{y}|| := \max(x_1-y_1, \ldots, x_d-y_d)$. Now we are looking for a shortest path from $a$ to $b$ with respect to this $\ell_\infty$ metric. The following result, stated without proof in [@AG], shows that the added flexibility of performing partial moves does not actually help us move our robots more quickly. The fastest paths from $a$ to $b$ take $d(B)$ units of time. In particular, the normal cube path from $a$ to $b$ is a fastest path. Consider an optimal path from $a$ to $b$. When going from $a$ to $b$, we need to perform all the moves in $B$. Consider a longest chain $p_1 < \cdots < p_{d(B)}$ of $P_a$. For each $i$, we must spend a total of at least one unit of time performing the move $p_i$. Also, we can never perform (any part of) the moves $p_i$ and $p_j$ simultaneously, since they cannot be in the same antichain. Therefore we need at least $d(B)$ units of time to carry out these $d(B)$ moves. It remains to remark that the normal cube path indeed takes $d(B)$ units of time. {#section-8} The authors would like to thank Megan Owen, Rick Scott, and Seth Sullivant for numerous enlightening discussions on this subject. They also thank the anonymous referees for their valuable suggestions. [99]{} A. Abrams and R. Ghrist. State complexes for metamorphic robots. *The International Journal of Robotics Research* [**23**]{} (2004) 811-826. F. Ardila and C. Klivans. The Bergman complex of a matroid and phylogenetic trees. *Journal of Combinatorial Theory, Series B* [**96**]{} (2006) 38-49. F. Ardila. Subdominant matroid ultrametrics *Annals of Combinat.* [**8**]{} (2004) 379-389. F. Ardila, M. Owen, and S. Sullivant. Geodesics in CAT(0) cubical complexes. *Advances in Applied Mathematics* [**48**]{} (2012) 142-163. L. Billera, S. Holmes, and K. Vogtmann. Geometry of the space of phylogenetic trees. *Adv. in Appl. Math.* , [**27**]{} 733 - 767, 2001. J. M. Boardman. Homotopy structures and the language of trees. *In Proceedings of the Symposium Pure Math.,* [**22**]{} (1971) 37-58. M. Bridson and A. Haefligher. Metric spaces of non-positive curvature. Springer-Verlag, Berlin, Heidelberg, 1999. D. Burago, Y. Burago, and S. Ivanov. A course in metric geometry. AMS Graduate Studies in Mathematics, Vol 33, 2001. G. Chirikjian. Kinematics of a metamorphic robotic system. *In Proceedings of the IEEE International Conference on Robotics and Automation* (1994) 449-455 R. Ghrist and V. Peterson. The geometry and topology of reconfiguration. *Adv. Apl. Mathematics* [**38**]{} (2007) 302-323. M. Gromov. Hyperbolic groups. In Essays in group theory, volume 8 of *Math. Sci. Res. Inst. Publ.* 75�263. Springer, New York, 1987. G.A. Niblo and L.D. Reeves. The geometry of cube complexes and the complexity of their fundamental groups. *Topology,* [**37(3)**]{} (1998) 621-633. M. Owen and J.S. Provan. A fast algorithm for computing geodesic distances in tree space. *IEEE/ ACM Transactions on Computational Biology and Bioinformatics,* [**8**]{} (2011) 2-13. L. D. Reeves. Biautomatic structures and combinatorics for cube complexes. *Ph.D. thesis, University of Melbourne,* 1995. M. A. Roller. Poc sets, median algebras and group actions. an extended study of Dunwoody’s construction and Sageev’s theorem. Unpublished preprint, 1998. M. Sageev. Ends of group pairs and non-positively curved cube complexes. *Proc. London Math. Soc. (3)*, [**71**]{} (3) 585�617, 1995. L. Santocanale. A nice labelling for tree-like event structures of degree 3. Inf. Comput., 208:652-665, 2010. R. Scott. Rationality and reciprocity for the greedy normal form of a Coxeter group, *Trans. Amer. Math. Soc.* [**363**]{} (2011),385–415. D. Speyer and B. Sturmfels. The tropical Grassmannian. *Advances in Geometry* [**4**]{} (2004) 389-411. R. P. Stanley. *Enumerative Combinatorics, vol. 1* Cambridge University Press, Cambridge, 1999. H. Trappmann and G. M. Ziegler. Shellability of complexes of trees. *Journal of Combinatorial Theory, Series A,* [**82**]{} (1998) 168-178. K. Vogtmann. Local structure of some $OUT(F_n)$-complexes. *In Proceedings of the Edinburgh Math. Soc. (2),* [**33**]{} (1990) 367-379. G. Winskel and M. Nielsen, Models for concurrency, in: *Handbook of logic in computer science*, Vol. 4, Vol. 4 of Handb. Log. Comput. Sci., Oxford Univ. Press, New York, 1995, pp. 1-148. [^1]: Notice that we never have $i>j$ or $i_1 \nleftrightarrow i_2$ in $P_u$.

--- abstract: 'In free completely symmetric tensor gauge field theories on Minkowski space-time, all gauge invariant functions and Killing tensor fields are computed, both on-shell and off-shell. These problems are addressed in the metric-like formalisms.' --- IHES/P/05/15\ [hep-th/0505068]{} Xavier Bekaert$^\clubsuit$[^1] and Nicolas Boulanger$^{\eighthnote}$[^2]\ \ PACS codes: 11.30.Ly, 02.20.Tw, 11.10.Kk Keywords: Higher-spins, Killing tensors, BRST Introduction ============ The old Fronsdal’s programme – that consists in introducing consistent interactions among massless higher-spin fields [@Fronsdal:1978rb] – still remains an important open mathematical problem of classical field theory. Numerous preliminary results toward this goal have been obtained (see e.g. the review paper [@BCIV] and references therein) which reveal surprising properties of higher-spin gauge fields. A better (“geometrical") understanding of the gauge structure underlying these theories will presumably be necessary for actually completing Fronsdal’s programme. As a very preliminary step in this direction, some properties of the Abelian gauge transformations of the free theories are investigated closely in the present paper. More precisely, we focus on the following problems: the determination of all gauge invariant functions and all Killing tensor fields of free completely symmetric tensor gauge field theories in flat background, respectively on-shell and off-shell, under the sole assumption of locality. These problems are solved in both constrained and unconstrained metric-like approaches. The plan of the paper is as follows. Free higher-spin gauge theories are reviewed in Section \[statement\] where the problems addressed are formulated in a more precise way. As explained in Section \[reformulation\], the problems are solved via a cohomological reformulation in the Becchi-Rouet-Stora-Tyutin (BRST) formalism and their general solutions are contained in the corollary of Theorem \[Hgamma\] and in Theorem \[Killing\] which are given in the section \[results\] that concludes the paper. In order to spell out the notation and make the paper as self-contained as possible, Appendix \[irreps\] reviews some textbook material on irreducible representations of the general linear and orthogonal Lie algebras. Theorems \[Hgamma\] and \[Killing\] are respectively proved in Appendices \[proof1\] and \[proof2\]. Statement of the problems {#statement} ========================= Einstein’s gravity theory is a non-Abelian massless spin-$2$ field theory, the two main formulations of which are the “metric" and the “frame" approaches. In a very close analogy, there exist two main approaches to higher-spin ([*i.e.*]{} spin $s>2$) field theories that are by-now referred to as “metric-like" [@Fronsdal:1978rb; @deWit:1979pe] and “frame-like" [@BCIV; @Vframe]. The present paper essentially deals with the former approach where the massless field is represented by a completely symmetric field $\varphi$ of rank[^3] $s>0$, the gauge transformation of which reads $$\delta_\varepsilon\varphi_{\m_1\ldots\,\,\m_s}=s\,\partial_{(\m_1}\varepsilon_{\m_2\ldots\m_s)}\,,\label{gtransfo}$$ where the curved bracket denotes complete symmetrization with strength one[^4] and the Greek indices run over $n$ values ($n\geq 3$). The gauge parameter $\varepsilon$ is a completely symmetric tensor field of rank $s-1$. For spin $s=2$, the gauge field $\varphi_{\mu\nu}$ represents the graviton while the gauge transformations (\[gtransfo\]) correspond to linearized diffeomorphisms. Originally, some algebraic constraints were imposed on the metric-like gauge field and gauge parameters [@Fronsdal:1978rb; @deWit:1979pe]. More precisely, given $\eta_{\m\n}$ the metric tensor of the flat background[^5], the gauge parameter $\varepsilon$ was taken to be “traceless" ($s\geq 3$) and the gauge field $\varphi$ to be “double-traceless" ($s\geq 4$). [**Notation:**]{} The trace of any tensor $T_{\m_1\ldots \m_r}$ is written as $T^\prime_{\mu_1...\mu_{r-2}}$($\equiv\eta^{\n_1\n_2}T_{\mu_1...\mu_{r-2}\n_1\n_2}$). Thus the previous “trace" constraints are written as $$\varepsilon^{\prime}_{\mu_1...\mu_{s-3}}=0\,,\quad\varphi^{\prime\prime}_{\mu_1...\mu_{s-4}}=0\,.\label{traceconstr}$$ The traceless part of any tensor $T_{\m_1\ldots \m_r}$ is denoted by $\widehat{T}_{\m_1\ldots \m_r}$ (that is ${\widehat{T}}_{\m_1\ldots \m_{r-2}}^{\prime}=0$). Analogously, the “double traceless" part of $T$ is denoted by $\widetilde{T}_{\m_1\ldots \m_r}$ ($\widetilde{T}^{\prime\prime}_{\m_1\ldots \m_{r-4}}=0$). Hence, the constraints (\[traceconstr\]) can also be written $$\varepsilon_{\mu_1...\mu_{s-1}}=\widehat{\varepsilon}_{\mu_1...\mu_{s-1}}\,,\quad\varphi_{\mu_1...\mu_s}=\widetilde{\varphi}_{\mu_1...\mu_s}\,.$$ Recently [@Francia:2002aa], Francia and Sagnotti realized that the algebraic constraints (\[traceconstr\]) could be consistently relaxed in the sense that the former theory appears as a gauge fixing of a consistent non-local theory without any trace constraint. The former and latter formulations will be referred to as “constrained" and “unconstrained" approaches (see [@Bouatta:2004kk] for pedagogical reviews on both of them). Though the absence of the constraints (\[traceconstr\]) simplifies our algebraic analysis, the non-locality property of the unconstrained approach is still rather elusive, therefore both cases will be treated for the sake of completeness. Gauge invariant functions ------------------------- There is a general belief that a local function of the completely symmetric gauge field $\varphi_{\m_1\ldots\,\,\m_s}$ is gauge invariant under unconstrained gauge transformations if and only if it depends on $\varphi$ only via the de Wit–Freedman curvature tensor field [@deWit:1979pe]$${\cal R}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s} :=\partial_{\m_1}\ldots\partial_{\m_s}\varphi_{\n_1\ldots\,\,\n_s}+\ldots$$ and its partial derivatives. The dots stand for the terms necessary for the tensor $\cal R$ to belong to the irreducible ${\mathfrak{gl}}(n)$-module associated with the rectangular two-row Young diagram of length $s$ depicted (60,15)(-5,2) (0,7.5)(13.5,0)[1]{}(50,7)(53,7)[$s$]{} (0,0)(13.5,0)[1]{}(50,7)(53,0)[$s$]{} , so that $\cal R$ satisfies the algebraic identities$${\cal R}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s}={\cal R}_{(\m_1\ldots\,\,\m_s)\,,\,\n_1\ldots\,\,\n_s}={\cal R}_{\m_1\ldots\,\,\m_s\,,\,(\n_1\ldots\,\,\n_s)}\,,$$ $${\cal R}_{(\m_1\ldots\,\,\m_s\,,\,\n_1)\n_2\ldots\,\,\n_s}=0\,.$$ For constrained fields and parameters, the most general local gauge invariant function of $\varphi$ is believed to depend only on the de Wit–Freedman curvature tensor $\cal R$ and on the Fronsdal tensor $$\label{Fronsdaltensor}{\cal F}_{\m_1\ldots\,\,\m_s}\,:=\,\Box\,\varphi_{\mu_1...\mu_s}\,-\,s\,\partial^\n\partial_{(\m_1}\varphi_{\mu_2...\mu_s)\,\n} \,+\,{s(s-1)\over 2 }\,\partial_{(\m_1}\partial_{\m_2}\varphi^\prime_{\mu_3...\mu_s)\n_1\n_2}$$ together with their derivatives. We would like to put the previous expectations on firm mathematical grounds. Part of the present paper (Section \[Th1\]) is therefore mainly devoted to a mathematically rigorous analysis of the following problem: the determination of all possible gauge invariant local functions depending on the gauge field $\varphi$ and a finite number of its partial derivatives both in the unconstrained and constrained approaches. Actually, the answer depends on whether the system is studied on-shell or not since, for instance, $${\cal F}_{\m_1\ldots\,\,\m_s}\,\approx 0\label{Fronsdalequ}$$ in the constrained approach. The symbol $\approx$ means that the equality holds on-shell. By combining the Damour-Deser identity [@DD] (which relates the trace of $\cal R$ to $s-2$ exterior derivatives of $\cal F$) with the generalized Poincaré lemma [@DuboisV], we argued in [@BB2] that the field equation $${\cal R}^\prime_{\m_1\ldots\,\,\m_s\,,\,\,\n_1\ldots\,\,\n_{s-2}}\,\approx 0\label{Weylequ}$$ in the unconstrained approach is dynamically equivalent to (\[Fronsdalequ\]) in the constrained approach. More precisely, $$\begin{aligned} {\cal F}_{\m_1\ldots\,\,\m_s}(\widetilde{\varphi})&=& 0\,,\quad \delta\widetilde{\varphi}_{\m_1\ldots\,\,\m_s}=s\,\partial_{(\m_1}\widehat{\varepsilon}_{\m_2\ldots\m_s)}\nn\\ &&\Longleftrightarrow \quad{\cal R}^\prime_{\m_1\ldots\,\,\m_s\,,\,\,\n_1\ldots\,\,\n_{s-2}}(\varphi)= 0\,,\quad \delta\varphi_{\m_1\ldots\,\,\m_s}=s\,\partial_{(\m_1}\varepsilon_{\m_2\ldots\m_s)}\,.\label{dynequiv}\end{aligned}$$ The spin-$s$ Weyl-like tensor $\cal W$ is defined as the traceless part of the spin-$s$ tensor $\cal R$, $${\cal W}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s}:=\widehat{{\cal R}}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s}\,,$$ hence it belongs to the irreducible $\mathfrak{o}(n-1,1)$-module associated with the rectangular two-row Young diagram of length $s$. In both approaches, the on-shell de Wit – Freedman tensor is equal to the Weyl-like tensor $${\cal R}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s}\approx{{\cal W}}_{\m_1\ldots\,\,\m_s\,,\,\n_1\ldots\,\,\n_s}\,.$$ A prerequisite of Vasiliev’s unfolded approach is the determination of all on-shell nontrivial derivatives of the fieldstrengths. The spin-$s$ de Wit – Freedman tensor obeys some differential Bianchi identity and, on-shell, it is traceless, divergenceless and satisfies the massless Klein-Gordon equation [@BB2]. This implies that the $k$th partial derivatives of the spin-$s$ de Wit – Freedman tensor belong off(on)-shell to the irreducible $\mathfrak{gl}(n)$($\mathfrak{o}(n-1,1)$)-module labeled by the two-row Young diagram (115,15)(0,2) (10,7.5)(13.5,0)[1]{}(65,7)(80,7.5)[$s+k$]{} (10,0)(13.5,0)[1]{}(35,7)(50,0)[$s$]{} . They span the so-called “twisted adjoint module" and play a fundamental role in the unfolded approach. To address the previously-stated problem of determining the gauge-invariant local functions of the (un)constrained symmetric gauge field $\varphi$, we reformulate it as a cohomological problem in the BRST formalism (see [*e.g.*]{} [@Henneauxbook] for a comprehensive review). More precisely, we use the fact that the set of such gauge invariant functions is given by the local BRST cohomology group $H^0(s)$ in vanishing ghost number in the sector without antifields. This cohomology group is actually identical to the cohomology group $H^0(\g)$ of the longitudinal exterior differential $\g$ in the sector without antifields. In the present paper, the complete local cohomology of $\g$ is determined. This result is useful because the knowledge of $H^*(\g)$ is the first ingredient in the computation of the local BRST cohomology group $H^{n,0}(s|\,d)$ in top form degree and in vanishing ghost number for higher-spin gauge theories, a subject which will be addressed later in the context of the classification of local consistent vertices for higher-spin gauge theories in flat space-time [@BBetal]. These results will be compared to two other cohomological analyses of somewhat related problems: the generalized Poincaré lemma of [@DuboisV] (extended to arbitrary irreducible tensors under the general linear group in [@BB1]) and the problem of writing down non-trivial equations in the unfolded formalism [@Shaynk] (see also Sections 9 and 10 of [@BCIV] for a review). Killing tensors {#Killingsect} --------------- Another problem of physical interest for a better understanding of the higher-spin symmetries is the determination of all Killing tensor fields on Minkowski space-time, that is, the symmetric tensor fields satisfying the following (off-shell) Killing-like equation [@Eisen] $$\partial_{(\m_1}\varepsilon_{\m_2\ldots\m_s)}=0\,.\label{Killingequ}$$ Actually this problem is easy to solve for tensor fields $\varepsilon_{\m_1\ldots\m_{s-1}}(x)$ which are formal power series in $x$. Killing tensor fields on spaces of constant curvature have been extensively studied by mathematicians (see [*e.g.*]{} [@Killingconstcurvv; @Killingtensorsolution]). A more involved problem is the determination of all local “on-shell Killing tensor fields". Non-trivial on-shell Killing tensor fields on flat space-time are solutions of the conditions $$\partial_{(\m_1}\varepsilon_{\m_2\ldots\m_s)}\approx 0\,,\quad\quad \varepsilon_{\m_1\ldots\m_{s-1}}\not\approx 0\,, \label{weaK}$$ which define a cohomological problem. Again, in order to address this problem we formulate it in the context of the local BRST formalism, by using the fact that these local non-trivial on-shell Killing tensor fields are in one-to-one correspondence with cocycles of the local Koszul-Tate cohomology group $H^n_2(\delta|d)$ in top form degree and antifield number two [@Barnich:2000zw]. Though interesting in its own sake, the knowledge of $H^n_2(\delta|d)$ is the second ingredient in the computation of the local BRST cohomology group $H^{n,0}(s|d)$ [@BBetal]. Generally speaking, the global symmetries of a solution of some field equation correspond to the space of gauge parameters leaving the gauge fields invariant under gauge transformations evaluated at the solution. Furthermore, for the flat vacuum solution they are expected to correspond to the full symmetry algebra of the theory. More specifically, the Killing tensors (\[Killingequ\]) of the infinite tower of higher-spin fields should be related to a higher-spin algebra (if any) in flat space-time. Indeed, we prove in Corollary \[coro3\] (Subsection \[KTate\]) that they are in one-to-one correspondence with the elements of a Minkowski higher-spin algebra that can be obtained as a quotient of the universal enveloping algebra $\cu(\mathfrak{iso}(n-1,1))$ of the Poincaré algebra or as an Inönü-Wigner contraction of the anti de Sitter / conformal higher-spin algebras of Eastwood and Vasiliev [@Eastwood; @Valgebra] in the flat limit $\Lambda\rightarrow 0$. These higher-spin gauge symmetry algebras might eventually find their origin in the general procedure of “gauging" some global higher-symmetry algebras of free theories, as we argue now. Reformulating an observation of [@Berends], the action of a complex bosonic field $\phi_{\m_1\ldots\m_t}$ with arbitrary spin $t$ is invariant under the global infinitesimal transformations $$\d_\lambda\phi_{\m_1\ldots\m_t}\,=\, i^{s-2}\,\lambda_{\n_1\ldots\n_{s-1}}\partial^{\n_1}\ldots\partial^{\n_{s-1}} \phi_{\m_1\ldots\m_t}\,,\label{univtranslationalg}$$ where $\lambda_{\n_1\ldots\n_{s-1}}$ are completely symmetric *constant real* tensors. For $s=1$, one recovers the usual infinitesimal $\mathfrak{u}(1)$ phase transformation and, for $s=2$, one obtains the usual infinitesimal action of the translation group ${\mathbb R}^d$. If one tried to gauge the global symmetry transformations of the form (\[univtranslationalg\]) by replacing the constant parameters $\lambda_{\n_1\ldots\n_{s-1}}$ by arbitrary completely symmetric tensor fields $\varepsilon_{\n_1\ldots\n_{s-1}}$, then the usual prescription would require the introduction of some connection in order to define proper covariant derivative. It is suggestive to interpret the gauge fields $\varphi_{\m_1\ldots\m_s}$ as entering into the definition of the connection so that the Abelian gauge transformations (\[gtransfo\]) should somehow correspond to the linearization of the non-Abelian gauge transformations. Moreover, the trace condition on the gauge parameter may find a natural explanation according to this line of thinking. For a scalar field ($t=0$), the infinitesmal transformation (\[univtranslationalg\]) reads $$\d_\lambda\phi\,\approx\, i^{s-2}\,\widehat{\lambda}_{\n_1\ldots\n_{s-1}}\partial^{\n_1}\ldots\partial^{\n_{s-1}}\phi\,, \label{univtranslationalgonsh}$$ that is, the parameters may be assumed to be traceless without loss of generality, since $(\Box-m^2)\phi\approx 0$. This is in agreement with the fact that trace conditions are present in the so-called “on-shell" higher-spin algebra and not in the “off-shell" one (see Section 5 of [@BCIV]). Let $\{P^\m\}$ be a basis of the (Abelian) translation algebra ${\mathbb R}^n$ that is represented by the Hermitian operators $\textsc{P}^\m=i\partial^\m$ acting on $\phi_{\m_1\ldots\m_t}$. The Hermitian kinetic operator $\textsc{K}(\textsc{P})$ for the field $\phi_{\m_1\ldots\m_t}$ is an element of the polynomial algebra ${\mathbb R}[\textsc{P}^\m]$. The translation operators $-i\lambda_\m\textsc{P}^\m$ obviously generate symmetries of the quadratic action $\int \langle\phi |\textsc{K}|\phi\rangle$ since they are anti-Hermitian and commute with the kinetic operator, but the same is true for any product of such translation operators, therefore the universal enveloping algebra of the translations generates an infinite-dimensional algebra of global symmetries of the free theory. The universal enveloping algebra of an Abelian Lie algebra is the corresponding polynomial algebra, in our case ${\cal U}({\mathbb R}^n)\cong{\mathbb R}[P^\m]$. This algebra is unitarily represented by the anti-Hermitian operators $-i\lambda_{\n_1\ldots\n_{s-1}}\textsc{P}^{\n_1}\ldots \textsc{P}^{\n_{s-1}}$ acting on the field $\phi_{\m_1\ldots\m_t}$ as in (\[univtranslationalg\]). (The symmetry algebra ${\mathbb R}[\textsc{P}^\m]$ should be quotiented by the ideal of elements proportional to $\textsc{K}(\textsc{P})$ in order to obtain the algebra of non-trivial symmetries of the action. For $t=0$ and $\textsc{K}=\textsc{P}^2+m^2$, the action of the latter algebra on the scalar $\phi$ is written explicitly in (\[univtranslationalgonsh\]).) The Abelian Lie algebra[^6] ${\tt hs}({\mathbb R}^n)$ of the higher-symmetries (\[univtranslationalg\]) is obtained from the associative algebra ${\cal U}({\mathbb R}^n)$ with the Lie bracket given by the commutator. The symmetry group ${\tt HS}({\mathbb R}^n)$ is obtained by exponentiation of ${\tt hs}({\mathbb R}^n)$ with generic element $U$ in the unitary representation: $$U=\exp(-i\lambda_{\n_1\ldots\n_{s-1}}\textsc{P}^{\n_1}\ldots \textsc{P}^{\n_{s-1}})\,.$$ This leads to the finite form $$\langle x|\phi\rangle\,\longrightarrow\, \langle x|\,U|\phi\rangle=\int d^n y\, U(x-y)\,\langle y|\phi\rangle=\exp(-i\l)\langle x|\phi\rangle+\langle x+\l\,|\phi\rangle+\ldots\label{finitetransfo}$$ of the infinitesimal transformations (\[univtranslationalg\]) where $$U(z)\,=\,(2\pi)^{-n}\int d^np\,\,\exp\, [-i\,(\,z^\mu p_\mu+\,\lambda_{\n_1\ldots\n_{s-1}}\,p^{\n_1}\ldots p^{\n_{s-1}})]$$ and the dots in (\[finitetransfo\]) denote the transformations corresponding to $s\geq 3$. These new terms are readily seen to be non-local (in the sense that they cannot be written in the form $\langle x^\prime|\phi\rangle$ where $x^\prime=f(x)$) because $U(z)$ is a delta of Dirac only when $s=1,2$. The previous reasoning entirely applies to the more complicated Poincaré algebra $\mathfrak{iso}(n-1,1)$ case and leads to some Minkowski higher-spin algebra. One should point out that the previous arguments provide a mean to evade the conclusions of the S-matrix no-go theorems against higher space-time symmetries [@cm], even in flat space-time. (For a more general discussion of higher-symmetries along the lines of this subsection, see [@B].) Cohomological reformulation of the problems {#reformulation} =========================================== Jet space --------- To reformulate a field theoretical problem – a functional problem – into a finite-dimensional algebraic problem – much more easy to cope with – one usually treats the fields and their partial derivatives as independent coordinates of a so-called “jet space". (We follow closely the terminology of the report [@Barnich:2000zw].) [**Notation:**]{} Frequently, we will omit indices for fields whose rank has been defined previously. The letter $\Phi$ will collectively denote the set of field variables (some of which may be Grassmannian). The $p$th partial derivatives of the variables $\Phi$ will be denoted by $\partial^p \Phi$, that is $\partial^p\Phi\,\sim\,\partial_{\m_1}\ldots\partial_{\m_p}\Phi\,.$ An (off-shell) [*local function of the field variables*]{} $\Phi$ is a function $f=f(x,[\Phi])$ of the space-time coordinates, polynomial in the field variables $\Phi$ and a finite number of their derivatives. The notation $[\Phi]$ means dependence on the variables $\Phi$, $\partial\Phi$, $\partial^2\Phi$, ..., $\partial^k\Phi$ for some finite but otherwise arbitrary integer $k$. The [*jet space of order*]{} $k$ is defined as the direct product $J^k(E)={\cal M}\times V^k$, where $\cal M$ is the $n$-dimensional space-time manifold, $V^k$ the space with coordinates given by $\Phi$, $\partial\Phi$, $\partial^2\Phi$, ..., $\partial^k\Phi$ and $E\equiv J^0(E)$ the jet space of order zero with coordinates $x$ and $\Phi$. A local function is thus a function on a jet space of some finite order (which will always be omitted in the sequel for the sake of readability), that is, an element of the space of sections on the trivial bundle $J^*(E)$. We will denote the supercommutative algebra of (off-shell) local functions of the field variables $\Phi$ by $\Upsilon_0(\Phi)$. Let $\Psi\in\Upsilon_0(\Phi)$ be the left-hand-side of the equations of motion $\Psi(\,[\Phi]\,)\approx 0$. The collection of equations $[\Psi]= 0$ defines a submanifold of the jet space $J^*(E)$ called the [*stationary surface*]{} and denoted by $\Sigma$. It turns out to be convenient to impose the following [*regularity conditions*]{}: (i) the local functions $[\Psi]$ can be split into independent and dependent ones and (ii) the independent functions can be locally taken as the first coordinates of a new, regular, coordinate system on the jet space $J^*(E)$ in the vicinity of the stationary surface $\Sigma$. The algebra of local functions proportional to the variables $[\Psi]$ forms an ideal, which we denote by ${\cal I}$, that leads to the equivalence relation $$f,g\in \Upsilon_0(\Phi)\,;\quad\quad f\approx g\quad\Longleftrightarrow \quad f-g\in{\cal I}\,.$$ One defines the algebra $\Upsilon_\Psi(\Phi)$ of [*on-shell local functions*]{} as the quotient of the algebra $\Upsilon_0(\Phi)$ by the ideal ${\cal I}$. The regularity conditions imply that the space of on-shell local functions is isomorphic to the space $\Gamma(\Sigma)$ of local functions on the stationary surface $\Sigma$. We will not consider topologically non-trivial space-time and/or field manifolds. Therefore a [*field history*]{} is defined as a map $h:{\cal{M}}\rightarrow E:x\mapsto (x,\Phi(x))$. (In the more general case, one should now introduce sections instead of functions, as well as jet bundles.) Any field history naturally induces a map from $\cal M$ to any given jet space $J^k(E)$ by evaluating the partial derivatives at each point of $\cal M$. The evaluation of a local function at a field history yields a [*local space-time function*]{}. Fields and antifields --------------------- It is straightforward to apply the antifield BRST formalism[^7] to free “irreducible" gauge theories, such as Fronsdal’s theory describing rank-$s$ completely symmetric gauge fields. Therefore we directly give the main results. Let us start with the unconstrained gauge theory. The gauge transformations (\[gtransfo\]) are irreducible, hence it is sufficient to introduce a fermionic ghost field $\xi_{\m_1\ldots\m_{s-1}}$ for each bosonic gauge parameter $\varepsilon_{\m_1\ldots\m_{s-1}}$. The “field” content is thereby enlarged to $$\Phi=\{\,\varphi\,,\,\xi\}\,,\quad\quad\epsilon(\varphi)=0\,,\quad\epsilon(\xi)=1\,,$$ where $\epsilon(Z)$ denotes the Grassmann parity of the field $Z$. The corresponding algebra $\Upsilon_0(\Phi)$ of local functions of the fields is $\mathbb N$-graded by the pureghost number. The corresponding diagonal operator $puregh$ is an even derivation defined by the following grading of the generators $$puregh(\varphi)=0,\quad puregh(\xi)=1.$$ To each field $Z\in\Phi$ we associate an antifield $Z^*$ of opposite parity. The set of associated antifields is then $$\Phi^*=\{\,\varphi^{*}\,,\,\xi^*\}\,,\quad\quad\epsilon(\varphi^*)=1\,,\quad\epsilon(\xi^*)=0\,.$$ The pure ghost number of any antifield vanishes. The algebra $\Upsilon(\Phi,\Phi^*)$ of local functions of the fields and antifields is bigraded: first by the pureghost number, second by the antighost number defined from $$antigh(\varphi)=0,\quad antigh(\xi)=0\quad antigh(\varphi^*)=1,\quad antigh(\xi^*)=2.$$ The (total) ghost number equals the difference between the pure ghost number and the antighost number$$gh(Z)=puregh(Z)-antigh(Z)\,,\quad\forall Z\in \Phi\cup\Phi^*\,.$$ The previous paragraph similarly applies to the constrained gauge theory, the only difference being that some trace constraints must be imposed. The field content of the constrained theory is $$\overline{\Phi}=\{\,\widetilde{\varphi}\,,\,\widehat{\xi}\,\}\,, \quad\quad\epsilon(\widetilde{\varphi})=0\,,\quad\epsilon(\widehat{\xi}\,)=1\,.$$ The set of associated antifields is then $$\overline{\Phi}^*=\{\,\widetilde{\varphi}^{*}\,,\,\widehat{\xi}^*\}\,,\quad\quad\epsilon(\widetilde{\varphi}^*)=1\,,\quad\epsilon(\widehat{\xi}^*)=0\,.$$ Following the comments made in the introduction, the stationary surface is the submanifold defined by ${{\cal R}^\prime}(\,[\varphi]\,)\approx 0$ in the unconstrained approach and by ${{\cal F}}(\,[\widetilde{\varphi}]\,)\approx 0$ in the constrained approach. Thus the algebras of on-shell local functions are respectively $\Upsilon_{{\cal R}^\prime}(\varphi)$ in the unconstrained approach and $\Upsilon_{{\cal F}}(\widetilde{\varphi})$ in the constrained approach. BRST differential ----------------- The [*BRST differential*]{} $s$ acts on the algebra $\Upsilon_0(\Phi,\Phi^*)$ and its ghost number is equal to one. Its action obviously induces an action on the subalgebra $\Upsilon_0(\overline{\Phi},\overline{\Phi}^*)$. It decomposes according to the antighost number as $$s=\d + \gamma \,,\quad antigh(\d)=-1\,,\quad antigh(\g)=0$$ and provides the algebra of gauge invariant functions on the stationary surface through its cohomology at ghost number zero $H^0(s)$. The [*Koszul-Tate differential*]{} $\delta$ acts trivially on the fields, $\d\Phi=0$. In the antifield sector, $\d\varphi^*$ is equal to the equations of motion and $\d \xi^*$ is proportional to the “Noether identities". The [*exterior derivative along the gauge orbits*]{} is the differential $\gamma$ defined by replacing the gauge parameters by the corresponding ghosts in the gauge transformation (\[gtransfo\]) $$\label{gammaformula} \g\varphi_{\m_1\ldots\,\,\m_s}=s\,\partial_{(\m_1}\xi_{\m_2\ldots\m_s)}\,,\quad\g \xi_{\m_1\ldots\m_{s-1}}=0\,,$$ On the one hand, the Koszul-Tate differential $\d$ implements the restriction on the stationary surface $\Sigma$ (since it provides a resolution of the algebra of functions on the stationary surface). On the other hand, the longitudinal exterior differential $\g$ picks out the gauge-invariant functions via its cohomology at pureghost number zero. The cohomological groups $H^0(s,\Upsilon_0(\Phi,\Phi^*))\cong H(\g,\Upsilon_{{\cal R}^\prime}(\varphi))$ and $H^0(s,\Upsilon_0(\overline{\Phi},\overline{\Phi}^*))\cong H(\g,\Upsilon_{{\cal F}}(\widetilde{\varphi}))$ can be read off Corollary \[corol\] given in the next section. The results and their physical interpretations {#results} ============================================== Longitudinal exterior cohomology {#Th1} -------------------------------- Let us define the Grassmann algebra $\Xi$ generated by the fermionic variables $d^{k}\xi$, where $k$ is a positive integer not larger than $s-1$ and where each exterior derivative $d$ in the power $d^{k}$ acts on a different space-time index of $\xi$. It is easy to see that the tensors $d^{k}\xi$ belong to the irreducible $\mathfrak{gl}(n)$-module labeled by the Young diagram (80,15)(-5,2)(58,6.5)[s–]{} (0,6.5)(10.5,0)[1]{}(50,7) (0,-1)(10.5,0)[1]{}(40,7)(52,-2)[k]{} . The set of variables $\xi$, $d\xi$, $d^2\xi$, ..., $d^{s-1}\xi$ is collectively denoted by $\langle \xi\rangle$. The traceless parts of the set of constrained variables $\widehat{\xi}$, $d\widehat{\xi}$, $d^2\widehat{\xi}$, ..., $d^{s-1}\widehat{\xi}$ is collectively denoted by $\widehat{\xi}\,$ and the corresponding Grassmann algebra by $\widehat{\Xi}$. \[Hgamma\] For the ggy, the cohomology of the longitudinal exterior differential $\gamma$ is the superalgebra freely generated by - : the space-time coordinates $x^\m$, - : the curvature tensor $\cal R$ with its partial derivatives, - : the ghost $\xi$ and its $s-1$ exterior derivatives, - : the antifields $\Phi^*$ with their partial derivatives, that is, $$H\Big(\,\gamma\,,\Upsilon_0(\Phi,\Phi^*)\Big)\cong \Upsilon_0({\cal R},\Phi^*)\otimes \Xi\,.$$ More explicitly, $$\gamma F\Big(x,[\Phi],[\Phi^*]\Big)=0\quad\Longleftrightarrow \quad F=G\Big(x,[{\cal R}],\langle \xi\rangle,[\Phi^*]\Big)\,+\,\gamma (\ldots)\,,$$ $$G\Big(x,[{\cal R}],\langle \xi\rangle,[\Phi^*]\Big)=\gamma (\ldots) \quad\Longleftrightarrow \quad G=0\,.$$ For the ggy, the cohomology of the longitudinal exterior differential $\gamma$ is the superalgebra freely generated by - : the space-time coordinates $x^\m$, - : the traceless components $\widehat{\,[{\cal R}]\,}$ of (the partial derivatives of) the curvature tensor, - : the Fronsdal tensor $\cal F$ and its partial derivatives, - : the traceless components of the ghost $\widehat{\xi}$ and its $s-1$ exterior derivatives, - : the antifields $\overline{\Phi}^*$ with their partial derivatives, that is[^8] $$H\Big(\,\gamma\,,\Upsilon_0(\overline{\Phi},\overline{\Phi}^*)\Big)\cong \Upsilon_0({\cal R},{\cal F},\overline{\Phi}^*)\otimes \widehat{ \Xi}\,.$$ More explicitly, $$\gamma F\Big(x,[\overline{\Phi}],[\overline{\Phi}^*]\Big)=0\quad\Longleftrightarrow \quad F=G\Big(x,\widehat{\,[{\cal R}]\,},[{\cal F }],{\mbox{\textlquill}}\,\widehat{\xi}\,{\mbox{\textrquill}},[\overline{\Phi}^*]\Big)\,+\,\gamma (\ldots)\,,$$ $$G\Big(x,\widehat{\,[{\cal R}]\,},[{\cal F }],{\mbox{\textlquill}}\,\widehat{\xi}\,{\mbox{\textrquill}},[\overline{\Phi}^*]\Big)=\gamma (\ldots) \quad\Longleftrightarrow \quad G=0\,.$$ Setting the ghosts and antifields to zero, one can make contact with the standard physical field content. Moreover, by restricting the gauge-invariant functions to the stationary surfaces defined by the relations (\[Fronsdalequ\]) or (\[Weylequ\]), one may even put them on-shell in the corresponding approaches. The results are summarized in the following corollary of Theorem \[Hgamma\]: \[corol\] $\bullet$ A local function of the unconstrained gauge field $\varphi$ is gauge invariant under unconstrained gauge transformations if and only if it depends on $\varphi$ via the de Wit – Freedman curvature tensor field $\cal R$ and its partial derivatives. Therefore, on-shell it is a function of the traceless component of the partial derivatives of the Weyl-like tensor $\cal W$ only. $$\delta_\varepsilon f\Big([\varphi]\Big)=0\,,\quad\forall \varepsilon\quad\Longleftrightarrow \quad f=f\Big([{\cal R}]\Big) \quad\Longleftrightarrow \quad f\approx f(\widehat{\,[{\cal W}]\,})\,.$$ $\bullet$ A local function of the unconstrained gauge field $\varphi$ that is invariant under constrained gauge transformations must depend on the Fronsdal tensor $\cal F$ with all its partial derivatives, the traceless component of the partial derivatives of $\cal R$ and on the double trace $\varphi^{\prime\prime}$ of the gauge field. $$\delta_{\widehat{\,\varepsilon}} f\Big([\varphi]\Big)=0\,,\quad\forall \widehat{\varepsilon}\quad\Longleftrightarrow \quad f=f\Big(\widehat{\,[{\cal R}]\,},[{\cal F}],[\varphi^{\prime\prime}]\Big)\,.$$ $\bullet$ For constrained gauge fields and parameters, the most general local gauge-invariant function must depend on $\widetilde{\varphi}$ only through the Fronsdal tensor $\cal F$ with all its partial derivatives and the traceless component of the partial derivatives of $\cal R$. Therefore, on-shell it is a function of the traceless component of the partial derivatives of the Weyl-like tensor $\cal W$ solely. $$\delta_{\widehat{\,\varepsilon}} \,g\Big([\widetilde{\varphi}]\Big)=0\,,\quad\forall \widehat{\varepsilon} \quad\Longleftrightarrow \quad g=g\Big(\widehat{\,[{\cal R}]\,},[{\cal F}]\Big) \quad\Longleftrightarrow \quad g\approx f(\widehat{\,[{\cal W}]\,})\,,$$where $f(\widehat{\,[{\cal R}]\,}):=g(\widehat{\,[{\cal R}]\,},[{\cal F}]=0)$. Some remarks are in order: - : As one can see, the space of on-shell gauge invariant local functions are identical in both approaches. This confirms the fact that both approaches are dynamically equivalent, as written in (\[dynequiv\]). Moreover, this is in complete agreement with Vasiliev’s unfolded formulation where the variables $\widehat{\,[{\cal W}]\,}$ span the on-shell twisted adjoint module for the free theory of a rank-$s$ symmetric tensor gauge field in flat space-time (see Sections 8 and 12 of [@BCIV]). - : Corollary \[corol\] may also be used when writing down local free ([*i.e.*]{} linear) field equations that are gauge-invariant and translation-invariant. Indeed, Corollary \[corol\] implies that the most general possibility is an equation that fixes to zero a local function linear in $[{\cal R}]$ in the unconstrained case and linear in $[{\cal R}]$ and $[{\cal F}]$ in the constrained approach. These possibilities reproduce the “Weyl cohomology" and “Einstein cohomology" obtained through the very different method of [@Shaynk] (see also Section 10 of [@BCIV]). - : Notice that these algebraic results also hold for local spacetime functions obtained by the evaluation of the local function $f$ at an arbitrary field history. Such a spacetime version of Corollary \[corol\] is somehow the converse of deriving the gauge transformations (\[gtransfo\]) as the most general transformations leaving invariant the curvature tensor ${\cal R}(x)$ and, in the constrained approach, the Fronsdal operator ${\cal F}(x)$. (This may be done as a straigthforward application of the results contained in [@DuboisV].) Local Koszul-Tate cohomology {#KTate} ---------------------------- The space $\cal K$ of unconstrained (off-shell) Killing tensor fields on flat space-time is the vector space of all completely symmetric tensor fields $\varepsilon_{\m_1\ldots\m_r}(x)$ satisfying (\[Killingequ\]). The space ${\cal K}=\bigoplus\limits_{r=0}^\infty{\cal K}_r$ is $\mathbb N$-graded by the rank $r$ and can be endowed with a structure of $\mathbb N$-graded commutative algebra via the natural symmetric product $$(\varepsilon_1 \vee\varepsilon_2)_{\m_1\ldots\,\m_{r_1+r_2}}:= (\varepsilon_1)_{(\,\m_1\ldots\,\m_{r_1}}(\varepsilon_2)_{\m_{r_1+1}\ldots\,\m_{r_1+r_2})}\,.\label{symprod}$$ One can explicitly check that the rank-$(r_1+r_2)$ symmetric tensor field $\varepsilon_1 \vee\varepsilon_2$ obeys the off-shell Killing-like equation. This ensures that the product $\vee$ is internal in the associative algebra $\cal K$. \[lem\] For an unconstrained free spin-$s$ ($s\geq 1$) gauge field theory, the most general off-shell Killing tensor field $\varepsilon_{\m_1\ldots\m_{s-1}}(x)$ that is a formal power series in $x$ is a polynomial of degree $s-1$ where the coefficient of the term of homogeneity degree $t$ is an irreducible tensor under $\mathfrak{gl}(n)$ characterized by the Young diagram (85,15)(0,2) (10,7.5)(13.5,0)[1]{}(45,7)(60,7.5)[$s-1$]{} (10,0)(13.5,0)[1]{}(25,7)(41,-1)[$t$]{} . More explicitly, &\_[(\_1]{}\_[\_2…\_[s]{})]{}(x)=0&\ & &\ &\_[\_1…\_[s-1]{}]{}(x)=\_[t=0]{}\^[s-1]{} ł\_[\_1…\_[s-1]{},\_1…\_t]{}x\^[\_1]{}…x\^[\_t]{},ł\_[(\_1…\_[s-1]{},\_1)\_2…\_t]{}=0.&\[Killingsol\] This lemma has been derived previously by several authors in different versions [@Killingtensorsolution; @BGST]. In Appendix \[shproof\], we provide a short and new proof of this useful property. Our main results about Killing tensors (either off-shell or on-shell) are contained in the following theorem: \[Killing\] (i) : Let $\{{P}^\m,{M}^{\n\r}\}$ be a basis of the Lie algebra $\mathfrak{iso}(n-1,1):={\mathbb R}^n\niplus \mathfrak{so}(n-1,1)$ of the isometry group of the flat space-time ${\mathbb R}^{n-1,1}$, where $\{{P}^\m\}$ is a basis of the translation algebra ${\mathbb R}^n$ and $\{{M}^{\m\n}\}$ is a basis of the Lorentz algebra $\mathfrak{so}(n-1,1)$. The commutative algebra $\cal K$ of unconstrained off-shell Killing tensor fields is isomorphic to the quotient of the symmetric algebra $\bigvee(\mathfrak{iso}(n-1,1))$ of the vector space $\mathfrak{iso}(n-1,1)$ by the relations$$R\,\equiv\, \{\,{P}_{[\m}{M}_{\n\r]}\,,\, {M}_{[\m\n}{M}_{\r]\s}\,\}\,.\label{rel}$$ More precisely, any equivalence class of the algebra $${\cal K}\,\cong\,\bigvee\big(\mathfrak{iso}(n-1,1)\big)\,/\,R$$ can be represented by a Weyl-ordered polynomial $$S({M},{P})=\sum\limits_{s\geq 1}\sum\limits_{t=0}^{s-1} \z_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}\,(\,{M}^{\n_1\m_1}\cdots{M}^{\n_t\m_t} {P}^{\m_{t+1}}\cdots{P}^{\m_{s-1}}+\mbox{perms}\,\,)\,,\label{representatives}$$ where the coefficients $\z$ are tensors having the same symmetry properties as the tensors $\l$ in Lemma \[lem\] and “perms" stands for the sum of terms obtained from ${M}^{\n_1\m_1}\cdots{M}^{\n_t\m_t}$ ${P}^{\m_{t+1}}\cdots{P}^{\m_{s-1}}$ by performing all nontrivial permutations of the elements ${P}^\m$ and ${M}^{\m\n}$. (ii) : All non-trivial on-shell Killing tensor fields $\widehat{\,\varepsilon\,}_{\m_1\ldots\m_{s-1}}(x,[\varphi])$ (that are formal power series in $x$) of the constrained theory of spin-$s$ free gauge field can be represented by the traceless component of the off-shell Killing tensor fields given in Lemma \[lem\]. By looking at the details of Theorems 6.5 and 6.7 of [@Barnich:2000zw] (and their proofs), it is straightforward to see that the following corollary is equivalent to the point (ii) of Theorem \[Killing\]: \[Hdeltamod\]In the constrained approach, the top forms $\widehat{\varepsilon}_{\m_1\ldots\m_{s-1}}(x)\,\widehat{\xi}^{*\,\m_1\ldots\m_{s-1}}d^nx$, where $\widehat{\varepsilon}_{\m_1\ldots\m_{s-1}}(x)$ runs over all non-trivial on-shell Killing tensor fields, span the local Koszul-Tate cohomology $H^n_2\Big(\d\,|\,d\,,\,\Upsilon_0(\overline{\Phi},\overline{\Phi}^*)\Big)$ in top form degree $n$ and in antifield number $2$. Let ${\mathbb R}[\textsc{X},\textsc{P}]$ be the real polynomial algebra in the variables $\textsc{X}^{\m}$ and $\textsc{P}^{\n}$. Let us introduce the antisymetric bilinears $\textsc{M}^{\m\n}:=\textsc{X}^{\m}\textsc{P}^{\n}-\textsc{X}^{\n}\textsc{P}^{\m}$ that provide a realization of the relations (\[rel\]). The representative (\[representatives\]) in the corresponding realization of $\bigvee(\mathfrak{iso}(n-1,1))/R$ in ${\mathbb R}[\textsc{X},\textsc{P}]$ is equal to $$\Lambda(\textsc{X},\textsc{P})= \sum\limits_{s\geq 1}\sum\limits_{t=0}^{s-1} \l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}\textsc{X}^{\n_1}\ldots \textsc{X}^{\n_t} \textsc{P}^{\m_1}\ldots\textsc{P}^{\m_{s-1}}\,, \label{representativesll}$$ where the tensor $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ is a linear function of the tensor $\zeta_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$. They satisfy the same symmetry properties as the tensors $\l$ in Lemma \[lem\]. The isomorphism between the algebra $\cal K$ of unconstrained off-shell Killing tensor fields (\[Killingsol\]) and the quotient $\bigvee(\mathfrak{iso}(n-1,1))/R$ is obvious when the representatives of the quotient are realized as in (\[representativesll\]). In that case, the symmetrized tensor product (\[symprod\]) in $\cal K$ is mapped to the pointwise product of polynomials in ${\mathbb R}[\textsc{X},\textsc{P}]$ induced by the following embeddings $${\cal K}_{s-1}\,\hookrightarrow\, {\mathbb R}[\textsc{X},\textsc{P}] \,:\, \varepsilon_{\m_1\ldots\,\m_{s-1}}(x) \mapsto \varepsilon_{\m_1\ldots\,\m_{s-1}}(\textsc{X})\,\textsc{P}^{\m_{1}} \ldots\textsc{P}^{\m_{s-1}}\,.\label{isomorfism}$$ The Poincaré-Birkhoff-Witt theorem shows that there exists a canonical isomorphism *of vector spaces* between the universal enveloping algebra ${\cal U}(\mathfrak{g})$ and the symmetric algebra $\bigvee(\mathfrak{g})$. Given a unitary representation of the Poincaré group $ISO(n-1,1)$ such that the generators $\textsc{P}^\m$ and $\textsc{M}^{\n\r}$ are Hermitian operators acting on some Hilbert space $\{|\phi\rangle\}$, the corresponding Hermitian operators of the form (\[representatives\]) define a unitary representation of the universal enveloping algebra ${\cal U}(\mathfrak{iso}(n-1,1))$. By definition, any Lorentz scalar $\textsc{K}(\textsc{P}^2)$ built out of the quadratic Casimir operator $\textsc{P}^2$ commute with the generators $\textsc{P}^\m$ and $\textsc{M}^{\n\r}$ of $\mathfrak{iso}(n-1,1)$. Therefore, the Hermitian operators of the form (\[representatives\]) generate global symmetries of the quadratic action $\langle\phi|\,\textsc{K}|\phi\rangle$ with kinetic operator $\textsc{K}:=\textsc{K}(\textsc{P}^2)$. The Weyl algebra $A_{2n}$ is the quotient of the complex polynomial algebra ${\mathbb C}[\textsc{X},\textsc{P}]$ by the commutation relations $$\textsc{P}^{\m}\textsc{X}^{\n}-\textsc{X}^{\n}\textsc{P}^{\m}=i\,\eta^{\m\n}\,. \label{Heisenberg}$$ A basis of the Lie algebra $\mathfrak{iso}(n-1,1)$ can be realized by the set $\{\textsc{P}^\m,\textsc{M}^{\n\r}\}\subset A_{2n}$. The “off-shell" Minkowski higher-spin Lie algebra ${\tt hs}_\infty(\mathfrak{iso}(n-1,1))$ is defined by endowing the realization of ${\cal U}(\mathfrak{iso}(n-1,1))$ in $A_{2n}$ with a Lie algebra structure by means of the commutator. If one considers only traceless tensors $\widehat{\z}$ in the expression (\[representatives\]) for the representatives of $\bigvee(\mathfrak{iso}(n-1,1))/R$, then, by making use of the commutation relation (\[Heisenberg\]), the corresponding representatives $S(\textsc{M},\textsc{P})$ can be rewritten as the normal-ordered polynomial $$H(\textsc{X},\textsc{P}):=\sum\limits_{s\geq 1}\sum\limits_{t=0}^{s-1} \widehat{\l}_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}\textsc{X}^{\n_1}\ldots\textsc{X}^{\n_t} \textsc{P}^{\m_1}\ldots\textsc{P}^{\m_{s-1}}\,,\label{representativesl}$$ where the constant tensors $\widehat{\l}_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ belong to the irreducible $\mathfrak{o}(n-1,1)$-module labeled by the Young diagram (85,15)(0,2) (10,7.5)(13.5,0)[1]{}(45,7)(60,7.5)[$s-1$]{} (10,0)(13.5,0)[1]{}(25,7)(41,-1)[$t$]{} . The elements (\[representativesl\]) of the Weyl algebra are of physical interest because they commute with any Lorentz scalar $\textsc{K}(\textsc{P}^2)$ built out of $\textsc{P}^2$. Given a unitary representation of the Weyl algebra $A_{2n}$ defined by (\[Heisenberg\]) such that $\textsc{X}^\m$ and $\textsc{P}^\n$ are Hermitian, the Hermitian representatives (\[representativesl\]) generate global symmetries of the quadratic action $\langle\phi|\,\textsc{K}|\phi\rangle$ with kinetic operator $\textsc{K}=\textsc{K}(\textsc{P}^2)$. For instance, the Hermitian operators $\textsc{X}^\m:=x^\m$ and $\textsc{P}^\m:=i\partial^\m$ define a unitary representation of the Weyl algebra $A_{2n}$ on the functional space of square-integrable complex bosonic scalar fields $\phi$. The realization of the enveloping algebra ${\cal U}(\mathfrak{iso}(n-1,1))$ in $A_{2n}$ quotiented by the traces admits a faithful unitary representation acting as the global infinitesimal transformations $$\d_{\widehat{\varepsilon}}\phi\,=\, i^{s-2}\,\widehat{\varepsilon}_{\n_1\ldots\n_{s-1}}(x)\,\partial^{\n_1}\ldots\partial^{\n_{s-1}} \phi\,,\label{glinftr}$$ where $\widehat{\varepsilon}(x)$ is an arbitrary traceless off-shell Killing tensor field. The higher-order global symmetries (\[glinftr\]) should naturally give rise via the Noether theorem to the list of higher-spin conserved currents explicitly constructed in [@Vasiliev:1999]. It is clear that the infinitesimal transformations (\[glinftr\]) are related to the mappings (\[isomorfism\]) between Killing tensor fields and polynomials in $\textsc{X}$ and $\textsc{P}$ in the constrained approach. As explained in Subsection \[Killingsect\], for a scalar field, the trace in the parameter correspond to global transformations leaving the scalar field invariant on-shell. Factoring out the traces of ${\tt hs}_\infty(\mathfrak{iso}(n-1,1))$ leads to the “on-shell" Minkowski higher-spin Lie algebra denoted by ${\tt hs}(\mathfrak{iso}(n-1,1))$. Theorem \[Killing\] implies that, as suggested in the introduction, the infinite tower of Killing tensors in flat space-time is related to some Minkowski higher-spin algebra: \[coro3\] The elements of the off-shell higher-spin algebra ${\tt hs}_\infty(\mathfrak{iso}(n-1,1))$ are in one-to-one correspondence with the off-shell Killing tensor fields $\varepsilon_{\n_1\ldots\n_{s-1}}$ of the unconstrained approach, while the elements of the on-shell higher-spin algebra ${\tt hs}(\mathfrak{iso}(n-1,1))$ are in one-to-one correspondence with the non-trivial on-shell Killing tensor fields $\widehat{\varepsilon}_{\n_1\ldots\n_{s-1}}$ of the constrained approach. The bijection is more manifest in the frame-like formulation. The constructions of the previous higher-spin algebras (together with their relationship with Killing tensors) are the analogue of the case of bosonic anti de Sitter / conformal higher-spin algebras [@Mik; @Eastwood; @Valgebra; @Sagnotti:2005] (see also Section 5 of [@BCIV]) except that the Lie algebra $\mathfrak{iso}(n-1,1)$ must be replaced by $\mathfrak{so}(n-1,2)$. The former Lie algebra can be obtained from the latter as an Inönü-Wigner contraction in the flat limit $\Lambda\rightarrow 0$, therefore this also holds for the corresponding higher-spin algebras. During the redaction of the present paper, it has been shown by M.A. Vasiliev [@Vunf] that the algebra ${\tt hs}_\infty(\mathfrak{iso}(n-1,1))$ corresponds to the global symmetries of Minkowski vacuum solution of the off-shell unfolded equations for the bosonic higher-spin gauge field theory [@Valgebra]. This result confirms the relevance of our analysis.\ [**Note added:**]{} After having submitted the paper to [arXiv]{}, we received the work [@Nazim] where the on-shell Killing tensors and the characteristic cohomology groups in form degrees $\leq n-2$ are computed for the constrained spin-$3$ theory; and we have been informed that slightly more general versions of the point (ii) of Theorem \[Killing\] and of Corollary \[Hdeltamod\] have been obtained independently by G. Barnich and N. Bouatta [@BaBo]. Acknowledgements {#acknowledgements .unnumbered} ================ We thank L. Gualtieri for early discussions on the spin-$3$ case. Irreducible representations {#irreps} =========================== Most of the textbook material reviewed in this section can be extracted from [@Littlewood]. In conformity with the mathematical literature, we adopt the Euclidean signature convention. Young diagrams {#Young} -------------- Partition of integers play a key role in labeling the irreducible representations (irreps) of the general linear and orthogonal groups. The partition of the positive integer $|\l|$ into $r$ integer parts $\l_1$, $\l_2$, ..., $\l_r$ with $\l_1+\l_2+\ldots+\l_r=|\l|$ and $\l_1\geq \l_2\geq \ldots\geq \l_r>0$ is denoted by $\l=(\l_1,\l_2, \ldots, \l_r)$. Each such partition $\l$ specifies a *Young diagram* consisting of $|\l|$ boxes arranged in $r$ left-adjusted rows, where the length of the $i$th row is $\l_i$ ($i=1,2,\ldots,r$). Let $Y$ be the Abelian group made of all formal finite sums of Young diagrams with integer coefficients. This group is $\mathbb N$-graded by the number $|\l|$ of boxes: $Y=\sum_{n\in\mathbb N} Y_n $. The famous “Littlewood-Richardson rule" defines a multiplication law which endows $Y$ with a structure of graded commutative ring. The product of two Young diagrams $\l$ and $\m$ is defined as the bilinear mapping to $$\l\cdot\m=\sum_\nu m_{\l\m\,|\,\n}\,\n\,,$$where the coefficients $m_{\l\m\,|\,\n}=m_{\m\l\,|\,\n}$ are the number of distinct labeling of the Young diagram $\n$ obtained from the following procedure: 1. : Label the Young diagram $\m$ by writing the letter “a" in all boxes of its first row, the letter “b" in all boxes of its second row, the letter “c" in all boxes of its third row, [*etc*]{}. 2. : Add the labeled boxes of the Young tableau $\m$ to $\l$ in Latin alphabetic order, one letter at a time and in such a way that at every stage: (i) : The resulting diagram is a Young diagram, (ii) : No two identical letters appear in the same column, (iii) : Reading from right to left across each row in turn from top to bottom (like in Arabic), the number of a’s read should always be $\geq$ the number of b’s read $\geq$ the number of c’s read, [*etc*]{}. As one can see, $|\l\cdot\m|=|\l|+|\m|$. A related operation in $Y$ is the “division" of $\n$ by $\m$ defined as the bilinear mapping to $$\n/\m=\sum_\l m_{\l\,\m\,|\,\n}\,\l\,,$$where the sum is over Young diagrams $\l$ such that the product $\l\cdot\m$ contains the term $\n$ (with coefficient $m_{\l\m\,|\,\n}$). The following obvious lemma will be used many times to simplify the computation of the cohomology of the differential $\g$ along the gauge orbits: \[twocolYd\] Let $m$ and $n$ be two strictly positive integers such that $m\geq n$. The product of two rows of respective lengths $m$ and $n$ is the sum of the two-row Young diagram obtained by putting the shortest row on the bottom of the longest and the product of two rows of respective lengths $m+1$ and $n-1$. $$\begin{picture}(130,15)(-5,2) \put(25,12){$m$}\multiframe(0,0)(10.5,0){1}(60,10){} \put(70,1){$\cdot$} \put(95,12){$n$}\multiframe(80,0)(10.5,0){1}(40,10){} \end{picture} = \begin{picture}(75,15)(-5,2) \put(25,17){$m$} \multiframe(0,5)(10.5,0){1}(60,10){} \multiframe(0,-5)(10.5,0){1}(40,10){} \put(16,-13){$n$} \end{picture} + \begin{picture}(120,15)(-5,2) \put(20,12){$m+1$}\multiframe(0,0)(10.5,0){1}(65,10){} \put(75,1){$\cdot$} \put(90,12){$n-1$}\multiframe(85,0)(10.5,0){1}(35,10){} \end{picture}$$ Kronecker products and branching rules {#Kronecker} -------------------------------------- Irreps of $\mathfrak{gl}(n)$ may be labeled by $\{\l\}$ where the partition $\l$ serves to specify the symmetry properties of the corresponding rank-$|\,\l|$ covariant (or contravariant) tensors forming the basis of this irreducible Lie algebra module. Let $Y_+$ be the Abelian monoid made of all formal finite sums of Young diagrams with non-negative integer coefficients. Finite direct sums of irreps of $\mathfrak{gl}(n)$ may therefore be labeled by $Y_+$ via the rule$$\{\,m\,\m\,+\,n\,\n\,\}\,=\,m\,\{\m\}\,\oplus\, n\,\{\n\}\,,$$ where the positive integer coefficients $m,n\in\mathbb N$ must be interpreted as the multiplicity of the corresponding representation. The evaluation of the Kronecker product of two $\mathfrak{gl}(n)$-irreps $\{\l\}$ and $\{\m\}$ can be done by means of the Littlewood-Richardson rule which gives $$\{\l\}\otimes\{\m\}=\{\l\cdot\m\}=\bigoplus_\nu\, m_{\l\m\,|\,\n}\{\n\}\,.\label{LRgl}$$ A related operation is that of contraction of one set of contravariant indices of symmetry $\m$ with a subset of a set of covariant tensor indices of symmetry $\n$ to yield a sum of covariant tensors with indices of symmetry $\l$ given by the division rule$$\{\n\}/\{\m\}=\{\n/\m\}=\bigoplus_\l \,m_{\l\m\,|\,\n}\,\{\l\}\,.$$ The irreps of $\mathfrak{gl}(n)$ may be reduced to irreps of $\mathfrak{o}(n)$ by extracting all possible trace terms formed by contraction with the metric tensor $\eta$ and its inverse. The corresponding irreps are labeled by $[\s]$. The reduction is given by the branching rule$$\mathfrak{gl}(n)\downarrow \mathfrak{o}(n)\quad:\quad \{\l\}\downarrow [\l/\Delta]\,,\label{reduction}$$ where $\Delta$ is the formal infinite sum$$\Delta=\,1\,+\begin{picture}(30,15)(-5,2) \multiframe(0,0)(10.5,0){2}(10,10){}{} \end{picture} + \begin{picture}(55,15)(-5,2) \multiframe(0,0)(10.5,0){4}(10,10){}{}{}{} \end{picture} + \begin{picture}(30,15)(-5,2) \multiframe(0,5)(10.5,0){2}(10,10){}{} \multiframe(0,-5)(10.5,0){2}(10,10){}{} \end{picture} + \begin{picture}(80,15)(-5,2) \multiframe(0,0)(10.5,0){6}(10,10){}{}{}{}{}{} \end{picture} + \begin{picture}(55,15)(-5,2) \multiframe(0,5)(10.5,0){4}(10,10){}{}{}{} \multiframe(0,-5)(10.5,0){2}(10,10){}{} \end{picture} + \begin{picture}(30,15)(-5,2) \multiframe(0,10)(10.5,0){2}(10,10){}{} \multiframe(0,0)(10.5,0){2}(10,10){}{} \multiframe(0,-10)(10.5,0){2}(10,10){}{} \end{picture} +\ldots$$ corresponding to the sum of all possible plethysms of the metric tensor. The decomposition (\[reduction\]) actually has a useful converse $$\mathfrak{o}(n)\uparrow \mathfrak{gl}(n)\quad:\quad [\l]\uparrow \{\l/\Delta^{-1}\}\,,\label{extension}$$ because the series $\Delta$ has an inverse $$\Delta^{-1}=\,1\,-\begin{picture}(30,15)(-5,2) \multiframe(0,0)(10.5,0){2}(10,10){}{} \end{picture} + \begin{picture}(43,15)(-5,2) \multiframe(0,5)(10.5,0){3}(10,10){}{}{} \multiframe(0,-5)(10.5,0){1}(10,10){} \end{picture} - \begin{picture}(55,15)(-5,2) \multiframe(0,10)(10.5,0){4}(10,10){}{}{}{} \multiframe(0,0)(10.5,0){1}(10,10){} \multiframe(0,-10)(10.5,0){1}(10,10){} \end{picture} - \begin{picture}(43,15)(-5,2) \multiframe(0,5)(10.5,0){3}(10,10){}{}{} \multiframe(0,-5)(10.5,0){3}(10,10){}{}{} \end{picture} +\ldots$$(The material reviewed in this paragraph finds its origin in the book [@King].) The operation (\[reduction\]) leads to a formal finite sum of irreps, some of which with strictly negative integer coefficients that have to be interpreted as constraints on some trace of the corresponding tensor basis. (Remark: These constraints are not preserved by the full $\mathfrak{gl}(n)$ algebra.) We introduce the notation $$\{\l-\m\}\equiv\{\l\}\ominus\{\m\}\,,$$ for later convenience. Proof of Theorem \[Hgamma\] {#proof1} =========================== Our proof makes a decisive use of the simple fact that contractible pairs drop out from the (co)homology. The plan of our proof is therefore very simple: provide a new set of generators for which the contractible pairs are manifest. This convenient set of generators is identified via the decomposition of the jet bundle in irreducible modules of either $\mathfrak{gl}(n)$ or $\mathfrak{o}(n)$ algebras. This strategy follows the lines of previous computations of $H^0(\g)$ for other gauge theories, such as completely antisymmetric [@Henneaux:1998rp], spin-two [@Boulanger:2000rq] and two-column mixed symmetry [@mixed] gauge fields. Contractible pairs ------------------ Let $\cal A$ be the supercommutative differential algebra over the field $\mathbb K$ that is (i) freely generated by the variables $x^i$, $y^a$ and $z^a$ whose respective Grassmann parities are related by $$\epsilon(z^a)=\epsilon(y^a)+1\,,$$ and (ii) endowed with the differential $\Delta$ defined via $$\Delta x^i=0\,,\quad \Delta y^a=z^a\,,\quad\Delta z^a=0\,,$$ and the Leibnitz rule. The differential superalgebra ${\cal A}$ is graded by the degree of homogeneity in the variables $y^a$ for which the differential $\Delta$ is of degree minus one. The pairs $(y^a,z^a)$ are called [*contractible pairs*]{}. This terminology follows from the well-known lemma[^9]: \[contractpair\] The differential superalgebra $({\cal A},\Delta)$ provides a homological resolution of the polynomial algebra ${\mathbb K}[x^i]$. More precisely, the homology $H(\Delta,{\cal A})$ decomposes according to the degree of homogeneity in the variables $y^a$ as follows: $$H_0(\Delta,{\cal A})= {\mathbb K}[x^i]\,,\quad H_k(\Delta,{\cal A})=0\,,\quad k\neq 0\,.$$ Computation of $H(\gamma)$ -------------------------- The lemma \[twocolYd\] combined with the rule (\[LRgl\]) leads to the following rule for the Kronecker product $$\label{Krule} \{\begin{picture}(60,15)(-5,2) \put(25,11){r} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\otimes\{\begin{picture}(50,15)(-5,2) \put(20,11){s} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\} = \{\begin{picture}(60,15)(-5,2)\put(9,17){\tiny{max(r,s)}} \multiframe(0,6.5)(10.5,0){1}(50,7){} \multiframe(0,-1)(10.5,0){1}(40,7){}\put(5,-9){\tiny{min(r,s)}} \end{picture} \}\,\oplus\,\Big(\,\{\begin{picture}(70,15)(-5,2) \put(10,12){\tiny{max(r,s)+1}} \multiframe(0,2)(10.5,0){1}(60,7){} \end{picture}\}\otimes\{\begin{picture}(40,15)(-5,2) \put(0,12){\tiny{min(r,s)-1}} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}\,\Big)$$ that will be very useful in the sequel. ### Unconstrained case The set of generators are the variables $[\Phi]$. The variables $[\varphi]$ and $[\xi]$ must be compared in the $\mathfrak{gl}(n)$-module of covariant tensors of the same rank. The covariant tensors $\partial^k\varphi$ belonging to the Kronecker product $\{\begin{picture}(60,15)(-5,2) \put(25,10){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\otimes\{\begin{picture}(50,15)(-5,2) \put(20,10){s} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\}$ must be compared with the covariant tensors $\partial^{k+1}\xi$ belonging to the $\mathfrak{gl}(n)$-module labeled by $\{\begin{picture}(70,15)(-5,2) \put(24,10){k+1} \multiframe(0,2)(10.5,0){1}(60,7){} \end{picture}\}\otimes\{\begin{picture}(40,15)(-5,2) \put(10,10){s-1} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}$. When doing this comparison, we remember that they are related via the formula (\[gammaformula\]) and look for a decomposition of the variables $[\Phi]$ into a set $\{x^i,y^a,z^a\}$ in analogy with Lemma \[contractpair\]. Two distinct cases have to be addressed: $\diamond$ : $\underline{-1\leq k\leq s-2}$: This means that $0\leq k+1\leq s-1$, therefore the covariant tensors $\partial^{k+1}\xi$ decomposes as the direct sum $$\label{k+1.s-1} \{\begin{picture}(50,15)(-5,2) \put(12,11){k+1} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\}\otimes\{\begin{picture}(60,15)(-5,2) \put(18,11){s-1} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\} = \{\begin{picture}(60,15)(-5,2)\put(15,17){s-1} \multiframe(0,6.5)(10.5,0){1}(50,7){} \multiframe(0,-1)(10.5,0){1}(40,7){}\put(10,-11){k+1} \end{picture} \}\,\oplus\,\Big(\,\{\begin{picture}(40,15)(-5,2) \put(14,12){k} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}\otimes\{\begin{picture}(70,15)(-5,2) \put(28,12){s} \multiframe(0,2)(10.5,0){1}(60,7){} \end{picture}\}\,\Big)\,,$$ where the second term on the r.h.s. is absent for $k=-1$. The first term corresponds to irreducible covariant tensors $d^{k+1}\xi$ ($k+1\leq s-1$) which are $x^i$ variables for the longitudinal differential $\g$ according to the notation of Lemma \[contractpair\]. The second term on the r.h.s. of (\[k+1.s-1\]) corresponds to covariant tensors $z^a$ in the Kronecker product $\{\begin{picture}(40,15)(-5,2) \put(14,12){k} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}\otimes\{\begin{picture}(50,15)(-5,2) \put(18,12){s} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\}$ that form contractible pairs with the variables $y^a\equiv \partial^k\varphi$ for $k\geq 0$. $\diamond$ : $\underline{k\geq s-1}$: The Kronecker product corresponding to $\partial^k\varphi$ decomposes as$$\label{k.s} \{\begin{picture}(60,15)(-5,2) \put(25,11){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\otimes\{\begin{picture}(50,15)(-5,2) \put(20,11){s} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\} = \{\begin{picture}(60,15)(-5,2)\put(18,17){k} \multiframe(0,6.5)(10.5,0){1}(50,7){} \multiframe(0,-1)(10.5,0){1}(40,7){}\put(18,-10){s} \end{picture} \}\,\oplus\,\Big(\,\{\begin{picture}(70,15)(-5,2) \put(20,12){k+1} \multiframe(0,2)(10.5,0){1}(60,7){} \end{picture}\}\otimes\{\begin{picture}(40,15)(-5,2) \put(8,12){s-1} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}\,\Big)$$ The first term on the r.h.s. corresponds to $(k-s)$th partial derivatives of the de Wit–Freedman tensor $\partial^{k-s}{\cal R}\equiv x^j$. (The Bianchi identities are responsible for the fact that the partial derivatives of the linearized curvature tensor belongs to the irrep. labeled by two-row Young diagrams.) Eventually, the second term on the r.h.s. of (\[k.s\]) correspond to variables $y^b$ forming a contractible pair together with the variables $z^b\equiv \partial^{k+1}\xi$. The application of Lemma \[contractpair\] is straightforward since the pairs have been explicitly separated. (The antifield variables $[\Phi^*]$ are inert under the action of the exterior differential.) This leads to the first part of Theorem \[Hgamma\]. ### Constrained case: change of basis The basis we start with corresponds to $[\overline{\Phi}]$. The basis elements $[\widetilde{\varphi}]$ and $[\widehat{\xi}]$ must be compared in $\mathfrak{gl}(n)$-modules of covariant tensors of the same rank. To do so, we have first to apply the “converse" branching rule (\[extension\]) since the fields $\overline{\Phi}$ satisfy some trace constraints. Firstly, $\widetilde{\varphi}$ are double-traceless covariant tensors of rank $s$, thus they span the direct sum $[\begin{picture}(60,15)(-5,2) \put(25,10){s} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}]\oplus[\begin{picture}(50,15)(-5,2) \put(15,10){s-2} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}]$ of irreducible $\mathfrak{o}(n)$-modules, which may also be labeled by $\{\begin{picture}(60,15)(-5,2) \put(25,10){s} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\ominus\{\begin{picture}(30,15)(-5,2) \put(4,10){s-4} \multiframe(0,2)(10.5,0){1}(20,7){} \end{picture}\}$. Secondly, $\widehat{\xi}$ are traceless covariant tensors of rank $s-1$, thus they span the irreducible $\mathfrak{o}(n)$-module labeled by $[\begin{picture}(50,15)(-5,2) \put(15,10){s-1} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}]\,\,\uparrow\,\{\begin{picture}(50,15)(-5,2) \put(15,10){s-1} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\}\ominus\{\begin{picture}(40,15)(-5,2) \put(10,10){s-3} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}$. Consequently, the comparison between the tensors $\partial^k\widetilde{\varphi}$ and $\partial^{k+1}\widehat{\xi}$ is more involved in the constrained approach. Actually, it is practical to consider the gauge fields as unconstrained and impose the constraints separately. More specifically, when $k\geq 2$, it turns out to be convenient (and of physical significance) to split the basis elements $\partial^k\widetilde{\varphi}$ into the tensors $\widehat{\partial^k}\varphi$ and $\partial^{k-2}{\cal F}$ subject to the constraints $\widehat{\partial^k}\varphi^{\prime\prime}=0$. Let us now explain what is meant by this splitting and how it is performed. - : To start with, we consider the product of $k$ partial derivatives $\partial_{\m_1}\ldots\partial_{\m_k}$ as covariant tensor in the irreducible $\mathfrak{gl}(n)$-module labeled by $\{\begin{picture}(60,15)(-5,2) \put(25,10){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}$ which can be reduced to the sum of its traceless part $\widehat{\partial^k}$ labeled by $\{\begin{picture}(60,15)(-5,2) \put(25,10){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\ominus\{\begin{picture}(40,15)(-5,2) \put(8,10){k-2} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}\,\downarrow\, [\begin{picture}(60,15)(-5,2) \put(25,10){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}]$ and a part proportional to its trace $\Box\partial^{k-2}$ labeled by $\{\begin{picture}(40,15)(-5,2) \put(8,10){k-2} \multiframe(0,2)(10.5,0){1}(30,7){} \end{picture}\}$. - : The space spanned by the tensors $\partial^k\widetilde{\varphi}$ is by definition exactly the same as the space spanned by the tensors $\partial^k\varphi$ quotiented by the space spanned by $\partial^k\varphi^{\prime\prime}$ that is labeled by the Kronecker product $\{\begin{picture}(60,15)(-5,2) \put(25,10){k} \multiframe(0,2)(10.5,0){1}(50,7){} \end{picture}\}\otimes\{\begin{picture}(35,15)(-5,2) \put(7,10){s-4} \multiframe(0,2)(10.5,0){1}(25,7){} \end{picture}\}$. - : Then we perform the following invertible change of basis $\{\partial^k\varphi\}\longleftrightarrow \{\widehat{\partial^k}\varphi\,,\,\partial^{k-2}\Box\varphi\}$. Moreover, it is possible to perform a triangular invertible linear change of variables $\Box\varphi\longleftrightarrow {\cal F}$ relating the Laplacian of the gauge field to the Fronsdal tensor (\[Fronsdaltensor\]). - : Putting all these remarks together, we have proved that the span of $\{\partial^k\widetilde{\varphi}\}$ is the quotient of the span $\{\widehat{\partial^k}\varphi\,,\,\partial^{k-2}{\cal F}\}$ divided by the span of $\{\partial^k\varphi^{\prime\prime}\}$. To summarize, we must compare the following four Kronecker products&:&({ (60,15)(-5,2) (25,10)[k]{} (0,2)(10.5,0)[1]{}(50,7) }{ (40,15)(-5,2) (8,10)[k-2]{} (0,2)(10.5,0)[1]{}(30,7) }){ (60,15)(-5,2) (25,10)[s]{} (0,2)(10.5,0)[1]{}(50,7) }\[k-2hatvarphi\]\ \^k\^&:&{ (60,15)(-5,2) (25,10)[k]{} (0,2)(10.5,0)[1]{}(50,7) }{ (30,15)(-5,2) (4,10)[s-4]{} (0,2)(10.5,0)[1]{}(20,7) } \[kphidoubletrace\]\ \^[k-2]{}[F]{}&:&{ (40,15)(-5,2) (8,10)[k-2]{} (0,2)(10.5,0)[1]{}(30,7) }{ (60,15)(-5,2) (25,10)[s]{} (0,2)(10.5,0)[1]{}(50,7) }\[k-2F\]\ \^[k+1]{}&:&{ (70,15)(-5,2) (25,10)[k+1]{} (0,2)(10.5,0)[1]{}(60,7) }({ (50,15)(-5,2) (15,10)[s-1]{} (0,2)(10.5,0)[1]{}(40,7) }{ (40,15)(-5,2) (10,10)[s-3]{} (0,2)(10.5,0)[1]{}(30,7) (50,5)[.]{} })\[k+1C\] ### Constrained case: decomposition of Kronecker products Three cases have to be distinguished: $\ast$ : $\underline{-1\leq k\leq s-4}$: The Kronecker product corresponding to the $(k+1)$th partial derivative of the trace constraint on the ghost $\xi$ can be reexpressed as the direct sum \[k+1.s-3\] { (40,15)(-5,2) (6,10)[k+1]{} (0,2)(10.5,0)[1]{}(30,7) } { (50,15)(-5,2) (16,10)[s-3]{} (0,2)(10.5,0)[1]{}(40,7) } &=& { (50,15)(-5,2)(12,16)[s-3]{} (0,6.5)(10.5,0)[1]{}(40,7) (0,-1)(10.5,0)[1]{}(30,7)(6,-11)[k+1]{} } { (55,15)(-5,2)(12,16)[s-2]{} (0,6.5)(10.5,0)[1]{}(45,7) (0,-1)(10.5,0)[1]{}(25,7)(12,-11)[k]{} } { (60,15)(-5,2)(15,16)[s-1]{} (0,6.5)(10.5,0)[1]{}(50,7) (0,-1)(10.5,0)[1]{}(20,7)(3,-11)[k-1]{} }\ \ &&({ (30,15)(-5,2)(3,11)[k-2]{} (0,2)(10.5,0)[1]{}(20,7) } { (70,15)(-5,2) (30,11)[s]{} (0,2)(10.5,0)[1]{}(60,7) } ), where this equation is obtained by three successive application of the rule (\[Krule\]) when $k\geq 2$. (If $k<2$ then the formula (\[k+1.s-3\]) remains valid but the terms that would not be well-defined are absent.) Combining (\[k+1.s-1\]) with (\[k+1.s-3\]) leads to the following decomposition of the Kronecker product (\[k+1C\]): && { (70,15)(-5,2)(23,16)[s-1]{} (0,6.5)(10.5,0)[1]{}(60,7) (0,-1)(10.5,0)[1]{}(30,7)(6,-11)[k+1]{} } { (50,15)(-5,2)(12,16)[s-3]{} (0,6.5)(10.5,0)[1]{}(40,7) (0,-1)(10.5,0)[1]{}(30,7)(6,-11)[k+1]{} } { (55,15)(-5,2)(12,16)[s-2]{} (0,6.5)(10.5,0)[1]{}(45,7) (0,-1)(10.5,0)[1]{}(25,7)(12,-11)[k]{} } { (60,15)(-5,2)(15,16)[s-1]{} (0,6.5)(10.5,0)[1]{}(50,7) (0,-1)(10.5,0)[1]{}(20,7)(3,-11)[k-1]{} }\ \ && (({ (35,15)(-5,2)(12,11)[k]{} (0,2)(10.5,0)[1]{}(25,7) }{ (30,15)(-5,2)(3,11)[k-2]{} (0,2)(10.5,0)[1]{}(20,7) }) { (70,15)(-5,2) (30,10)[s]{} (0,2)(10.5,0)[1]{}(60,7) }). \[complement\]Therefore the span of tensors $\partial^{k+1}\widehat{\xi}$ when $k+1$ ranges from $0$ till $s-1$ contains the irreducible $\mathfrak{o}(n)$-modules labeled by $$\{\begin{picture}(70,15)(-5,2)\put(23,16){s-1} \multiframe(0,6.5)(10.5,0){1}(60,7){} \multiframe(0,-1)(10.5,0){1}(30,7){}\put(6,-11){k+1} \end{picture}\} \ominus \{\begin{picture}(50,15)(-5,2)\put(12,16){s-3} \multiframe(0,6.5)(10.5,0){1}(40,7){} \multiframe(0,-1)(10.5,0){1}(30,7){}\put(6,-11){k+1} \end{picture}\} \ominus \{\begin{picture}(55,15)(-5,2)\put(12,16){s-2} \multiframe(0,6.5)(10.5,0){1}(45,7){} \multiframe(0,-1)(10.5,0){1}(25,7){}\put(12,-11){k} \end{picture}\} \ominus \{\begin{picture}(60,15)(-5,2)\put(15,16){s-1} \multiframe(0,6.5)(10.5,0){1}(50,7){} \multiframe(0,-1)(10.5,0){1}(20,7){}\put(3,-11){k-1} \end{picture}\}\quad\uparrow\quad[\begin{picture}(70,15)(-5,2)\put(23,16){s-1} \multiframe(0,6.5)(10.5,0){1}(60,7){} \multiframe(0,-1)(10.5,0){1}(30,7){}\put(6,-11){k+1} \end{picture}]$$ corresponding to the set of variables ${\mbox{\textlquill}}\,\widehat{\xi}\,{\mbox{\textrquill}}$ introduced in Subsection \[Th1\]. The complement of the space spanned by the variables ${\mbox{\textlquill}}\,\widehat{\xi}\,{\mbox{\textrquill}}$ is the module labeled by the diagram in the second line of (\[complement\]) the basis elements of which form contractible pairs together with the tensors $\widehat{\partial^k}\varphi$ described in (\[k-2hatvarphi\]). The basis elements of the jet space that remain are therefore in the cohomology since they cannot be associated with any derivative of ghost. They are the derivatives $\partial^{k-2}\cal F$ (\[k-2F\]) of the Fronsdal tensor subject to the constraints $\partial^k\varphi^{\prime\prime}=0$ (\[kphidoubletrace\]). $\ast$ : $\underline{s-3\leq k\leq s+1}$: This case contains five subcases each of which must be treated separately. The final result reproduces the corresponding general results of the two other cases, so we leave the explicit check as an exercise for the reader. $\ast$ : $\underline{k\geq s+2}$: This is equivalent to $k-2\geq s$ therefore we apply the rule (\[Krule\]) three times on the Kronecker product (\[k-2F\]), associated with the derivatives $\partial^{k-2}\cal F$ of the Fronsdal operator, and get as a result \[k-2.s\] { (50,15)(-5,2) (14,10)[k-2]{} (0,2)(10.5,0)[1]{}(40,7) } { (40,15)(-5,2) (14,10)[s]{} (0,2)(10.5,0)[1]{}(30,7) } &=& { (50,15)(-5,2)(12,16)[k-2]{} (0,6.5)(10.5,0)[1]{}(40,7) (0,-1)(10.5,0)[1]{}(30,7)(13,-11)[s]{} } { (55,15)(-5,2)(13,16)[k-1]{} (0,6.5)(10.5,0)[1]{}(45,7) (0,-1)(10.5,0)[1]{}(25,7)(6,-11)[s-1]{} } { (60,15)(-5,2)(18,16)[k]{} (0,6.5)(10.5,0)[1]{}(50,7) (0,-1)(10.5,0)[1]{}(20,7)(3,-11)[s-2]{} }\ \ &&( { (70,15)(-5,2) (20,11)[k+1]{} (0,2)(10.5,0)[1]{}(60,7) } { (30,15)(-5,2)(3,11)[s-3]{} (0,2)(10.5,0)[1]{}(20,7) } ). The Damour-Deser identity relates the trace of the curvature tensor to $s-2$ curls of the Fronsdal tensor ${\cal R}^\prime\propto d^{s-2}\cal F$ [@DD; @BB2]. Consequently, the right-hand-side of the first line of (\[k-2.s\]) can be expressed entirely in terms of derivatives of the trace ${\cal R}^\prime$ of the Riemann tensor. Combining (\[k-2.s\]) with (\[k.s\]) we get that the $k$th derivatives of the double-traceless gauge field of rank $s$ decompose as follows &&{ (60,15)(-5,2) (23,10)[k]{} (0,2)(10.5,0)[1]{}(50,7) } ( { (40,15)(-5,2) (14,10)[s]{} (0,2)(10.5,0)[1]{}(30,7) } { (30,15)(-5,2) (4,10)[s-4]{} (0,2)(10.5,0)[1]{}(20,7) })\ \ &&= { (70,15)(-5,2)(28,16)[k]{} (0,6.5)(10.5,0)[1]{}(60,7) (0,-1)(10.5,0)[1]{}(30,7)(13,-11)[s]{} } { (50,15)(-5,2)(12,16)[k-2]{} (0,6.5)(10.5,0)[1]{}(40,7) (0,-1)(10.5,0)[1]{}(30,7)(13,-11)[s]{} } { (55,15)(-5,2)(13,16)[k-1]{} (0,6.5)(10.5,0)[1]{}(45,7) (0,-1)(10.5,0)[1]{}(25,7)(6,-11)[s-1]{} } { (60,15)(-5,2)(18,16)[k]{} (0,6.5)(10.5,0)[1]{}(50,7) (0,-1)(10.5,0)[1]{}(20,7)(3,-11)[s-2]{} }\ \ &&({ (50,15)(-5,2) (14,10)[k-2]{} (0,2)(10.5,0)[1]{}(40,7) } { (40,15)(-5,2) (14,10)[s]{} (0,2)(10.5,0)[1]{}(30,7) }) ({ (60,15)(-5,2) (25,10)[k]{} (0,2)(10.5,0)[1]{}(50,7) }{ (30,15)(-5,2) (4,10)[s-4]{} (0,2)(10.5,0)[1]{}(20,7) })\ \ &&({ (70,15)(-5,2) (25,10)[k+1]{} (0,2)(10.5,0)[1]{}(60,7) }({ (50,15)(-5,2) (15,10)[s-1]{} (0,2)(10.5,0)[1]{}(40,7) }{ (40,15)(-5,2) (10,10)[s-3]{} (0,2)(10.5,0)[1]{}(30,7) })) \[complements\] The last line of (\[complements\]) is paired with the ghost sector (\[k+1C\]) which disappear from the cohomology. The second line of (\[complements\]) states that the traceless component $\widehat{\,[{\cal R }]\,}$ of the derivatives of the curvature tensors are in the cohomology since they span the irreducible $\mathfrak{o}(n)$-modules labeled by $[\begin{picture}(80,15)(-5,2) \multiframe(0,6.5)(10.5,0){1}(60,7){}\put(65,6.5){k} \multiframe(0,-1)(10.5,0){1}(30,7){}\put(50,-2){s} \end{picture}]$ while the third line of (\[complements\]) together with (\[kphidoubletrace\]) and (\[k-2F\]) show that the derivatives of the Fronsdal tensor depending on $\widetilde{\varphi}$ complete the generators of the cohomology of $\gamma$. Proof of Theorem \[Killing\] {#proof2} ============================ Unconstrained off-shell Killing tensor fields {#shproof} --------------------------------------------- If $\varepsilon_{\m_1\ldots\m_{s-1}}(x)$ is a formal power series in $x\,$, then, by definition, $$\label{formpow}\varepsilon_{\m_1\ldots\,\m_{s-1}}(x)=\sum\limits_{t=0}^\infty \l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}\,x^{\n_1}\ldots \,x^{\n_t}\,,$$ where the tensors $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ are all constant and each set of indices is symmetrized. If one acts with $s-1$ partial derivatives on both sides of the off-shell Killing-like equation (\[Killing\]), then the resulting equation is equivalent to$$\partial_{\n_1}\ldots\partial_{\n_s}\varepsilon_{\m_1\ldots\m_{s-1}}= 0 \label{Ks}$$ because all terms in the decomposition of the tensor product $\{\begin{picture}(60,15)(-5,2) \put(25,11){s} \multiframe(0,2)(10.5,0){1}(8.5,7){$\cdot$} \multiframe(9,2)(10.5,0){1}(30,7){$\cdots$} \multiframe(39.5,2)(10.5,0){1}(8.5,7){$\cdot$} \end{picture}\}\otimes\{\begin{picture}(50,15)(-5,2) \put(16,11){s-1} \multiframe(0,2)(10.5,0){1}(40,7){} \end{picture}\} $ (where a dot in a box stands for a partial derivative) always contain, as a subdiagram, the Young diagram $\begin{picture}(60,15)(-5,2) \put(25,11){s} \multiframe(0,2)(10.5,0){1}(40,7){} \multiframe(40.5,2)(10.5,0){1}(8.5,7){$\cdot$} \end{picture}$ corresponding to the left-hand-side of (\[Killingequ\]). In other words, the left-hand-side of (\[Ks\]) depends linearly on the left-hand-side of (\[Killingequ\]). Inserting (\[formpow\]) in (\[Ks\]) leads to the fact that the tensors $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ are zero for $t \geq s$. Then, substituting the resulting polynomial for $\varepsilon_{\m_1\ldots\m_{s-1}}$ in the original Killing-like equation (\[Killingequ\]) gives the system of equations $$\l_{(\m_1\ldots\,\m_{s-1}\,,\,\,\n_1)\n_2\ldots\,\n_t}=0\quad(t<s) \label{systequ}$$ which implies that the constant tensor $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ belongs to the $\mathfrak{gl}(n)$-module labeled by $\{\begin{picture}(90,15)(0,2) \multiframe(10,7.5)(10.5,0){1}(45,7){}\put(60,7.5){$s-1$} \multiframe(10,0)(10.5,0){1}(25,7){}\put(41,-1){$t$} \end{picture}\}$. This proves Lemma \[lem\]. Analogously, the point (i) of Theorem \[Killing\] is a straightforward consequence of the relations (\[rel\]) applied to any monomial of the schematic form $\textsc{M}^t \textsc{P}^{s-1-t}$ which leads to a result identical to the previous one in terms of irreducible $\mathfrak{gl}(n)$-modules. The main point is that the $\mathfrak{gl}(n)$-irreducibility conditions (\[rel\]) are expressed in the “antisymmetric" convention for Young diagrams. Alternatively, the isomorphism (i) follows from the fact that the algebra of Killing tensors on any constant curvature space is generated by the corresponding Killing vectors [@Killingconstcurvv] (which are of course in one-to-one correspondence with the generators of the isometry algebra). Constrained on-shell Killing tensor fields ------------------------------------------ The proof of the point (ii) of Theorem \[Killing\] requires more work and uses some general results on local Koszul-Tate cohomology groups. Let us consider the on-shell (also called “weak") Killing-like equation (\[weaK\]) for the constrained free spin-$s$ gauge field theory: $$\partial_{(\m_1}\widehat{\varepsilon}_{\m_2\ldots\m_s)}\approx 0\,.\label{weaKc}$$ As explained in the previous subsection, if one acts with $s-1$ partial derivatives on both sides of (\[weaKc\]), then the resulting equation is equivalent to[^10]$$\partial_{\n_1}\ldots\partial_{\n_s}\widehat{\varepsilon}_{\m_1\ldots\m_{s-1}}\approx 0\,,\label{weakKs}$$ because the partial derivative preserves the stationary surface since the Fronsdal equation (\[Fronsdalequ\]) does not depend explicitly on $x$. An application of the general theorems 6.2 and 6.4 of [@Barnich:2000zw] to the constrained free spin-$s$ gauge field theory of Fronsdal leads to the fact that the “characteristic cohomology group" in form degree zero is represented by the constants (see for instance [@Henneaux:1998rp; @Boulanger:2000rq; @mixed]). In other words, $$\partial_\m f(x,[\widetilde{\varphi}])\approx 0\quad\Longleftrightarrow \quad f(x,[\widetilde{\varphi}])\approx cst\,,$$ where “$cst$" stands for a constant independent of $x$ and the (partial derivatives of the) gauge field $\widetilde{\varphi}$. Therefore, (\[weakKs\]) is equivalent to $$\partial_{\n_1}\ldots\partial_{\n_{s-1}}\widehat{\varepsilon}_{\m_1\ldots\m_{s-1}}\approx \l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_{s-1}}\,,$$ where the tensor $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_{s-1}}$ is constant, each set of indices is symmetrized and the indices $\mu$ are traceless. Consequently, the tensor $$\widehat{\varepsilon}^{\,\prime}_{\m_1\ldots\m_{s-1}}:=\widehat{\varepsilon}_{\m_1\ldots\m_{s-1}}- \l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_{s-1}}\,x^{\m_1}\ldots x^{\m_{s-1}} \,,$$ satisfies $$\partial_{\n_1}\ldots\partial_{\n_{s-1}}\widehat{\varepsilon}^{\,\prime}_{\m_1\ldots\m_{s-1}}\approx 0\,.$$ By repeating the previous argument $s-1$ times, we arrive at the conclusion that $$\label{pow}\widehat{\varepsilon}_{\m_1\ldots\,\m_{s-1}}(x,[\widetilde{\varphi}])\approx \sum\limits_{t=0}^{s-1} \l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}\,x^{\n_1}\ldots \,x^{\n_t}\,,$$ where the tensors $\l_{\m_1\ldots\,\m_{s-1}\,,\,\,\n_1\ldots\,\n_t}$ are constant, symmetric in each set of indices and traceless in the indices $\mu$. Substituting (\[pow\]) in the (weak) Killing-like equation (\[weaKc\]), we obtain $$\sum\limits_{t=0}^{s-2} (t+1)\,\l_{(\m_1\ldots\,\m_{s-1}\,,\,\,\m_s)\,\n_1\ldots\,\n_t}\,x^{\n_1}\ldots \,x^{\n_t}\approx 0\,. \label{equw}$$ Taking into account the fact that the left-hand-side of the (weak) equality (\[equw\]) depends only on the space-time variable $x$ while the Fronsdal equation (\[Fronsdalequ\]) is linear in the jet space variables $[\widetilde{\varphi}]$, we arrive at the conclusion that the constant tensors $\l$ satisfy the (strong) conditions (\[systequ\]). To end up the proof of the last point of Theorem \[Killing\], we use the general property that tensors irreducible under $\mathfrak{gl}(n)$ and traceless in the indices corresponding to the largest row are traceless in all their indices (see also Lemma 4.5 of [@BGST] for another proof of the symmetry properties of traceless Killing tensors). [99]{} C. Fronsdal, Phys. Rev. D [**18**]{} (1978) 3624. X. Bekaert, S. Cnockaert, C. Iazeolla and M.A. Vasiliev, “Nonlinear Higher Spin Theories in Various Dimensions”, in the proceedings of the First Solvay Workshop on Higher-Spin Gauge Theories (Brussels, Belgium, May 2004) \[[hep-th/0503128]{}\]. B. de Wit and D. Z. Freedman, Phys. Rev. D [**21**]{} (1980) 358. M. A. Vasiliev, Sov. J. Nucl. Phys. [**32**]{} (1980) 439 \[Yad. Fiz.  [**32**]{} (1980) 855\]; Fortsch. Phys.  [**35**]{} (1987) 741;\ V. E. Lopatin and M. A. Vasiliev, Mod. Phys. Lett. A [**3**]{} (1988) 257. D. Francia and A. Sagnotti, Phys. Lett. B [**543**]{} (2002) 303 \[[hep-th/0207002]{}\]. D. Sorokin, “Introduction to the classical theory of higher spins”, in the proceedings of XIX Max Born Symposium (Wroclaw, Poland, October 2004) pp. 172-202 \[[hep-th/0405069]{}\];\ N. Bouatta, G. Compere and A. Sagnotti, “An introduction to free higher-spin fields”, in the proceedings of the First Solvay Workshop on Higher-Spin Gauge Theories (Brussels, Belgium, May 2004) \[[hep-th/0409068]{}\]. T. Damour and S. Deser, Annales Poincaré Phys. Théor.  [**47**]{} (1987) 277. P.J. Olver, “Differential hyperforms [I]{}", Univ. of Minnesota report 82-101;\ M. Dubois-Violette and M. Henneaux, Lett. Math. Phys. [**49**]{} (1999) 245 \[[math.qa/9907135]{}\]; Commun. Math. Phys.  [**226**]{} (2002) 393 \[[math.qa/0110088]{}\]. X. Bekaert and N. Boulanger, Phys. Lett. B [**561**]{} (2003) 183 \[[hep-th/0301243]{}\]; “Mixed symmetry gauge fields in a flat background”, in the proceedings of the International Seminar on Supersymmetries and Quantum Symmetries “SQS 03" (Dubna, Russia, July 2003) \[[hep-th/0310209]{}\]. M. Henneaux and C. Teitelboim, *Quantization of Gauge Systems* (Princeton University Press, 1992). X. Bekaert, N. Boulanger and S. Cnockaert, in preparation. X. Bekaert and N. Boulanger, Commun. Math. Phys.  [**245**]{} (2004) 27 \[[hep-th/0208058]{}\]; Class. Quantum Grav.  [**20**]{} (2003) S417 \[[hep-th/0212131]{}\]. O. V. Shaynkman and M. A. Vasiliev, Theor. Math. Phys.  [**123**]{} (2000) 683 \[Teor. Mat. Fiz. (2000) 323\] \[[hep-th/0003123]{}\]; Theor. Math. Phys.  [**128**]{} (2001) 1155 \[Teor. Mat. Fiz. (2001) 378\] \[[hep-th/0103208]{}\];\ O. A. Gelfond and M. A. Vasiliev, \[[hep-th/0304020]{}\]. L. P. Eisenhart, *Riemannian Geometry* (Princeton University Press, 1966) p. 128. G. Thompson, J. Math. Phys. [**27**]{} (1986) 2693. T. Wolf, Comp. Phys. Comm. [**115**]{} (1998) 316;\ R. G. McLenaghan, R. Milson and R. G. Smirnov, C. R. Acad. Sci. Paris, Ser. [**I 339**]{} (2004) 621. G. Barnich, F. Brandt and M. Henneaux, Phys. Rept.  [**338**]{} (2000) 439 \[[hep-th/0002245]{}\]. M. G. Eastwood, “Higher symmetries of the Laplacian” \[[hep-th/0206233]{}\]. M. A. Vasiliev, Phys. Lett. B [**567**]{} (2003) 139 \[[hep-th/0304049]{}\]. F. A. Berends, G. J. H. Burgers and H. van Dam, Nucl. Phys. B [**271**]{} (1986) 429. S. R. Coleman and J. Mandula, Phys. Rev.  [**159**]{} (1967) 1251;\ R. Haag, J. T. Lopuszanski and M. Sohnius, Nucl. Phys. B [**88**]{}, 257 (1975). X. Bekaert, in preparation. M. Henneaux, Nucl. Phys. Proc. Suppl. [**18A**]{} (1990) 47. G. Barnich, M. Grigoriev, A. Semikhatov and I. Tipunin, “Parent field theory and unfolding in BRST first-quantized terms” \[[hep-th/0406192]{}\]. A. Mikhailov, “Notes On Higher Spin Symmetries” \[[hep-th/0201019]{}\]. A. Sagnotti, E. Sezgin and P. Sundell, “On higher spins with a strong $Sp(2,{\mathbb R})$ condition”, in the proceedings of the First Solvay Workshop on Higher-Spin Gauge Theories (Brussels, Belgium, May 2004) \[[hep-th/0501156]{}\]. D. Anselmi, Nucl. Phys. B [**541**]{} (1999) 323 \[[hep-th/9808004]{}\]; Class. Quantum Grav.  [**17**]{} (2000) 1383 \[[hep-th/9906167]{}\];\ M. A. Vasiliev, “Higher spin gauge theories: Star-product and AdS space”, \[[hep-th/9910096]{}\] in M. Shifman ed., *The many faces of the superworld* (World Scientific, 2000) Section 2. M. A. Vasiliev, “Actions, charges and off-shell fields in the unfolded dynamics approach” \[[hep-th/0504090]{}\]. N. Bouatta, “Approche BRST des systèmes invariants de jauge à spins élevés”, Master Degree Thesis defended at the Free University of Brussels (Academic year 2003-2004). G. Barnich, private communication. M. Henneaux and B. Knaepen, Nucl. Phys. B [**548**]{} (1999) 491 \[[hep-th/9812140]{}\]. N. Boulanger, T. Damour, L. Gualtieri and M. Henneaux, Nucl. Phys. B [**597**]{} (2001) 127 \[[hep-th/0007220]{}\]. J. A. Garcia and B. Knaepen, Phys. Lett. B [**441**]{} (1998) 198 \[[hep-th/9807016]{}\];\ X. Bekaert, N. Boulanger and M. Henneaux, Phys. Rev. D [**67**]{} (2003) 044010 \[[hep-th/0210278]{}\];\ N. Boulanger and S. Cnockaert, JHEP [**0403**]{} (2004) 031 \[[hep-th/0402180]{}\];\ X. Bekaert, N. Boulanger and S. Cnockaert, J. Math. Phys. [**46**]{} (2005) 012303 \[[hep-th/0407102]{}\]. D.E. Littlewood, [*The theory of group characters*]{} (Clarendon, 1940); G. R. E. Black, R. C. King and B. G. Wybourne, J. Phys. A: Math. Gen. [**16**]{} (1983) 1555; S. A. Fulling, R. C. King, B. G. Wybourne and C. J. Cummins, Class. Quantum Grav.  [**9**]{} (1992) 1151. R.C. King, J. Phys. A: Math. Gen. [**8**]{} (1975) 429. [^1]: E-mail address: bekaert@ihes.fr [^2]: Chargé de Recherches FNRS (Belgium), E-mail address: nicolas.boulanger@umh.ac.be [^3]: Throughout this text, the spin $s$ is taken to be a strictly positive integer. Note that the half-integer spin case is also covered here since all the results apply to fermions by simply replacing the spin $s$ by its integer part $[s]$. [^4]: For example, $\varphi_{(\m_1\ldots\,\,\m_s)}\equiv \varphi_{\m_1\ldots\,\,\m_s}$. [^5]: Our results are independent of the choice of signature because they rely on algebraic considerations only. For physical reasons, we choose the Lorentzian signature. [^6]: $\tt hs$ stands at the same time for “higher-spin" and “higher-symmetry". [^7]: For reviews on the Lagrangian ([*i.e.*]{} antifield) BRST method, we refer the reader to the book [@Henneauxbook] and the lectures [@Henneaux:1989]. [^8]: In the notation $\Upsilon_0({\cal R},{\cal F},\overline{\Phi}^*)$, there is some redundacy in the sense that the variables $[{\cal R}]$ and $[{\cal F}]$ are not entirely independent. [^9]: For a proof, see [*e.g.*]{} Section 8.3.2 of [@Henneauxbook]. [^10]: The equation (\[weakKs\]) was conjectured for arbitrary spin $s$ in [@Nazim].

--- author: - 'M. Servillat' - 'A. Dieball' - 'N. A. Webb' - 'C. Knigge' - 'R. Cornelisse' - | \ D. Barret - 'K. S. Long' - 'M. M. Shara' - 'D. R. Zurek' date: 'Received -; accepted -' title: 'Combined Chandra, XMM-Newton and Hubble Space Telescope observations of the Galactic globular cluster ' --- [Using new Chandra X-ray observations and existing XMM-Newton X-ray and Hubble far ultraviolet observations, we aim to detect and identify the faint X-ray sources belonging to the Galactic globular cluster in order to understand their role in the evolution of globular clusters.]{} [We present a Chandra X-ray observation of the Galactic globular cluster . We classify the X-ray sources associated with the cluster by analysing their colours and variability. Previous observations with XMM-Newton and far ultraviolet observations with the Hubble Space Telescope are re-investigated to help identify the Chandra sources associated with the cluster. We compare our results to population synthesis models and observations of other Galactic globular clusters.]{} [We detect 113 sources, of which 16 fall inside the half-mass radius of and are concentrated towards the cluster core. From statistical analysis, these 16 sources are very likely to be linked to the cluster. We detect short-term (1 day) variability in X-rays for 7 sources, of which 2 fall inside the half-mass radius, and long-term (28 months) variability for 10 further sources, of which 2 fall inside the half-mass radius. Ultraviolet counterparts are found for 8 Chandra sources in the core, of which 2 have good matching probabilities and have ultraviolet properties expected for cataclysmic variables. We find one likely neutron star-quiescent low-mass X-ray binary and 7 cataclysmic variable candidates in the core of . The other 8 sources are cataclysmic variable candidates, but some could possibly be active binaries or millisecond pulsars. We find a possible deficit of X-ray sources compared to which could be related to the metallicity content and the complexity of the evolution of .]{} Introduction ============ Globular clusters (GCs) are old, gravitationally bound stellar systems which can have extremely high stellar densities, especially in their core regions. In such an environment, dynamical interactions between the cluster members are inevitable, leading to a variety of close binary (CB) systems and other exotic stellar objects. The observed overabundance of neutron star (NS) low-mass X-ray binaries (LMXBs) in GCs relative to the Galactic field was explained by the dynamical processes occurring in the dense cores of GCs [@fabian]. In contrast, evolution of a primordial binary into an LMXB in a GC is considered to be much less likely [@VH87]. Observations also support the fact that quiescent LMXBs (qLMXBs) in GCs scale with the cluster encounter rate [@GBW03; @Pooley+03], implying that qLMXBs are formed through dynamical processes in the dense cores. As white dwarfs (WDs) are far more common than NSs, we would then also expect many more close binaries containing an accreting WD primary, i.e. cataclysmic variables (CVs). The dynamically-formed CBs are expected to be found in the cores of GCs, where the stellar densities are at a maximum. The less dense regions outside the cores might be populated by CBs that evolved from primordial binaries [e.g. @Davies97] which are unlikely to survive in the dense core region. @HAS07 found that the combined effects of new binary creation and mass segregation exceed the destruction of primordial binaries in the central region of GCs, leading to a marked increase of the binary fraction in the central regions. Thus, we expect the majority of CBs, which are more massive than the mean stellar mass, to be located inside the half-mass radius. Outside the half-mass radius, the primordial binary fraction is well preserved [@HAS07]. CBs are important for our understanding of GC evolution, since the binding energy of a few, very close binaries can rival that of a modest-sized GC [e.g. @EHI87; @Hut+92; @Hut+03 and references therein]. In the core, binaries are subject to encounters and hard binaries become harder while transferring their energy to passing stars. Thus, CBs can significantly affect the dynamical evolution of the cluster. If there are only a few CBs, thermal processes dominate the cluster evolution, leading to core collapse followed by GC disruption on a timescale shorter than the mean age of GCs, estimated to be $12.9\pm2.9$ Gyr [@CGCFP00]. In contrast, the presence of many CBs leads to violent interactions, which heat the cluster, delay the core collapse, and promote its expansion. This depends critically on the number of CBs, which is still poorly known. Finding and studying these systems has proven to be extremely difficult, since the spatial resolution and detection limits of most available telescopes are insufficient for their detection. Only with the improved sensitivity and imaging quality of XMM-Newton and Chandra in the X-ray [e.g. @Webb+04; @WWB06; @Heinke+03; @Heinke+06] and HST in the ultraviolet (UV) to infrared (IR) wavebands [@GHEM01; @Albrow+01; @EGHG03; @EGHG03b; @Knigge+02; @Knigge+03 and references therein] has it become possible to finally detect significant numbers of CB systems in GCs. The 13 bright X-ray sources found in the $\sim$150 known Galactic GCs are LMXBs showing type I X-ray bursts , whereas the faint sources belonging to the clusters are qLMXBs, CVs, active binaries (ABs, generally RS CVn systems), or millisecond pulsars (MSPs). Multiwavelength studies can be used to identify the faint X-ray sources. For example, qLMXBs are usually identified by their soft blackbody-like or hydrogen atmosphere X-ray spectra , CVs can be confirmed by their blue, variable optical counterpart with hydrogen emission lines in their spectra , ABs by their main-sequence, variable optical counterparts , and MSPs by their radio counterpart . Here, we present an X-ray study of the massive () GC (). This intermediate metallicity GC [, @Walker99] lies at a distance of 9.6 kpc and is reddened by $E_{B-V}=0.22\pm0.01$ [@Harris96]. An absorption column of N${_H = 1.2\times10^{21}\mathrm{~cm^{-2}}}$ is derived from the reddening with the relation computed by @PS95. The cluster has a very dense and compact core ($0.26\arcmin$), a half-mass radius of $0.76\arcmin$, a tidal radius of $15.55\arcmin$, and a half-mass relaxation time of $1.35\times10^{9}$ yrs [@Harris96]. has received considerable attention in the literature and has been observed in the optical in detail as this GC is one of the most extreme examples with an unusual horizontal branch (HB) morphology, as first noted by @Harris74. It shows a bimodal HB and one of the longest blue HB tails, the so-called extreme HB (EHB), with prominent gaps between the red HB (RHB), blue HB (BHB) and EHB [see also @Bedin+00; @Carretta+06]. Recently, @Piotto+07 found that ’s main sequence (MS) is separated into three branches, which might be associated with the complex HB morphology and abundance distribution, and might be due to successive rounds of star formation with different helium abundances. is proposed as a good candidate to harbour an intermediate mass black hole (IMBH) in its core, due to its optical luminosity profile and EHB morphology [@Miocchi07]. The core of has been imaged with the Space Telescope Imaging Spectrograph (STIS) on board the Hubble Space Telescope (HST) in the far-UV (FUV) and the near-UV (NUV). @Dieball+05 used the data set with an emphasis on the dynamically-formed stellar populations like CVs and blue stragglers (BSs) and young WDs. They found $\sim$40 WD, $\sim$60 BS and $\sim$60 CV candidates in the field of view that covers the core of the cluster. Two of the CV candidates are variable (FUV sources 222 and 397). has also been observed with XMM-Newton in Feb. 2005. @Servillat+08-a found 96 sources in the field of view (equivalent to the tidal radius), of which five fall inside the half-mass radius and are likely to be linked to the cluster. One qLMXB candidate and four CV candidates were discovered in the core of . However, several sources remained unresolved. In Sect. \[obs\], we present the new Chandra X-ray data, and then compare them to the XMM-Newton observations (Sect. \[xmm\]). We present HST and XMM-Newton Optical Monitor UV counterparts in Sect. \[uv\]. We finally discuss our results in Sect. \[disc\]. X-ray observations {#obs} ================== was observed with the Chandra X-ray Observatory and the Chandra Advanced CCD Imaging Spectrometer-Imager () at its focus on 2007 June 19–21 (28 months after the XMM-Newton observation) for two distinct exposures of 46 and 11 kilo seconds (ks). The four front-illuminated chips were used as well as the front-illuminated chip and the back-illuminated chip at the edge of the field of view. The data were taken in faint, timed-exposure mode. The core of falls inside the chip, where the highest resolution is achieved. Data reduction -------------- Data reduction was performed using the CIAO v3.4 software[^1] [@Fruscione+06] and the CALDB v3.4.0 set of calibration files (gain maps, quantum efficiency, quantum efficiency uniformity, effective area). We reprocessed the level 1 event files of both observations without including the pixel randomization that is added during standard processing. This method slightly improves the point-spread function (PSF). We removed cosmic-ray events which could be detected as spurious faint sources using the tool *acis\_detect\_afterglow*, and identified bad pixels with the tool *acis\_run\_hotpix*. The event lists were then filtered for grades, status, and good time intervals (GTIs), as it is done in the standard processing. We selected events within the energy range , where good sensitivity is achieved. The two observations are successive, the fields of view and the roll angles are similar, and the PSFs from one observation to another are also similar. Therefore the two epochs can be processed as a single observation. We thus reprojected the event list of the second observation to match the first observation, and merged the two event lists using the thread *merge\_all*. We generated an exposure map of the field of view using the same thread. This led to 56.9 ks of clean observation (sum of GTIs). Source detection {#sec:sourcedetect} ---------------- {width="\textwidth"} In order to obtain a list of source candidates, we employed the CIAO wavelet-based *wavdetect* tool for source detection in the field of view covered by the four chips. Two energy bands were used, the 0.3–10.0 keV band with all events which allows the detection of the faintest sources, and 0.5–6.0 keV with a higher signal to noise ratio which gives secure detections. We selected scales of 1.0, 1.4, 2.0, 2.8, 4.0, and 5.6 pixels. The scales were chosen to look for narrow PSF sources on-axis and ensure optimal separation, and larger PSF sources at the edge of the detectors, where the PSF is degraded. We selected a threshold probability of $10^{-6}$, designed to give one false source per $10^{6}$ pixels. This led to the detection of 88 source candidates of which $\sim6$ may be false. Of these sources, eleven fall inside the half-mass radius of . To this list, we added other source candidates in order to test their significance in further processing. We added 14 X-ray sources previously detected with XMM-Newton [@Servillat+08-a]. Seven of them fall on the two chips where we did not use *wavdetect* as the PSF is too large, and seven others were not detected by *wavdetect* on the chips and have possibly varied in flux. Several other faint sources can be picked out by eye, but are not identified by *wavdetect*, possibly because of crowding on-axis and lower signal to noise at the edge of the detector due to the vignetting. We thus added 22 source candidates, of which 10 are inside the half-mass radius of . In total we listed 124 source candidates. We then used ACIS Extract[^2] [@acisextract] to refine the positions of the sources and to estimate the significance of each candiate source. We excluded sources with a probability of being a source lower than 99.99%. This probability is estimated from the number of source candidate counts compared to the surrounding background counts, taking into account Poisson statistics [@acisextract §5.9.3]. Our final list has 113 sources, of which 16 are located inside the half-mass radius of . All of these sources have more than four counts. They are numbered according to their offset from the center of NGC 2808. Their properties are listed in Table \[tab:chandra\_sources\], and a source map is presented in Fig. \[fig:chandra\_map\]. The core image is compared to the XMM-Newton image of @Servillat+08-a in Sect. \[sec:Chandra\_vs\_XMM\]. From the faintest sources detected, the limiting flux of the observation is ${L_{[0.5-8keV]}\sim0.9\times10^{-15}\mathrm{~erg~cm^{-2}~s^{-1}}}$, corresponding to a limiting luminosity of ${L_{[0.5-8keV]}\sim1.0\times10^{31}\mathrm{~erg~s^{-1}}}$ (at the distance of the cluster). The $1\sigma$ position error reported in Table \[tab:chandra\_sources\] is used in this paper to look for matching XMM-Newton and HST FUV sources. One may estimate the absolute position error by adding in quadrature the pointing accuracy of Chandra ($0.4\arcsec$, $1\sigma$ error) and the spatial distortion error over the detectors ($0.1\arcsec$, $1\sigma$ error). Members of ----------- ----- ---------- -------------- ---------- -------------- ---- Expected Detected Expected Detected 0 1.8 $ 4.0\pm0.8$ 20 $ 3.4\pm0.8$ 12 1.8 3.8 $12.7\pm2.0$ 20 $ 9.8\pm2.0$ 11 3.8 7.0 $36.2\pm6.0$ 43 $27.2\pm6.0$ 30 0 0.76 $ 0.8\pm0.8$ 16 $ 0.6\pm0.8$ 11 ----- ---------- -------------- ---------- -------------- ---- : Expected background sources and detected X-ray sources in field of view. Results are presented assuming a detection limit and a completeness limit (see text).[]{data-label="tab:lnls"} We estimated the number of background X-ray sources we expect to detect in our observation in order to deduce the number of sources likely to be linked to the cluster. We divided the field of view into three annuli to account for vignetting, and to include in each region at least 20 detected sources. We used the $log(N)$–$log(S)$ relation calculated by @HMS05 [see also @Giacconi+01; @Hasinger+01] to convert in each annulus our minimum detectable fluxes ($S$) into the number of background sources expected ($N$). This relation was derived from a survey of soft X-ray active galactic nuclei (AGN) in the energy range 0.5–2.0 keV performed with both XMM-Newton and Chandra. We took into account two errors on the value read from the $log(N)$–$log(S)$ diagram which were added in quadrature: the error on the Chandra flux estimate (see Table \[tab:chandra\_sources\]) converted to an error in $N$ while reading the $log(N)$–$log(S)$ diagram, and the precision of the relation which includes Poisson noise [see error bars in @Giacconi+01 their Fig. 3]. For each annulus, we estimated the minimum detectable unabsorbed flux of a point source in the energy range 0.5–2.0 keV using WebPIMMS[^3] v3.9b [@Mukai93]. We assumed for the source a power law model of photon index 1.5 (mean of the detected sources) and the absorption of the cluster. We assumed a minimum detectable count rate corresponding to two cases: the detection limit (the count rate of the faintest source in each annulus) and the completeness limit (twice the count rate of the faintest source). These values were corrected for the vignetting in each annulus using the exposure map. The estimates might be slightly overestimated for the detection limit as we have an incomplete sample of sources. The results are shown in Table \[tab:lnls\]. ![Colour-flux diagram of Chandra sources in the direction of . For clarity, we labelled only the core sources (filled circles), the brightest sources, and variable sources (see Fig. \[fig:chandra\_lcvar\]). The values for all sources can be found in Table \[tab:chandra\_sources\]. Some error bars are shown which are representative of the error bars at the same flux. Black lines with diamonds show the colours (with an arbitrary flux) of different models with an absorption of $1.2\times10^{21}\mathrm{~cm^{-2}}$: **PO**: power law with photon indices 3, 2.5, 2, 1.5, 1, 0.5, 0. **BR**: thermal bremsstrahlung with temperatures 1, 5, 10, 15, 20, 50 keV. **BB**: blackbody spectrum with temperatures 0.1, 0.5, 1, 1.5, 2 keV. []{data-label="fig:chandra_cmd"}](10188f2.ps){width="\columnwidth"} ![Lightcurves of variable Chandra sources. The data is shown with filled circles and error bars, and the corresponding background extracted in an annulus around the source is shown with diamonds. The first (left) and second (right) observations are separated by 86 ks. The zero time is 2007 June 19 at 15h55min53.5s. []{data-label="fig:chandra_lcvar"}](10188f3.eps){width="\columnwidth"} An excess is clearly seen in the center. In the annulus 1.8–3.8, an excess is seen if we assume the detection limit (Table \[tab:lnls\]), however, with the completeness limit this excess is not confirmed. In the last annulus, we detect the expected number of background sources within the errors. We performed the same estimation inside the half-mass radius of and $0.8\pm0.8$ sources are expected. Therefore, the 16 sources inside the half-mass radius are very likely to be cluster sources, with perhaps one background source aligned fortuitously. As has a low Galactic latitude [$b=-11.3\degr$, @Harris96], we may also expect foreground sources such as active stars or field CVs. Spectral and variability analysis --------------------------------- We used the ACIS Extract procedure *ae\_standard\_extraction* to extract spectra and lightcurves for each source. We estimated for each source an optimal extraction region to enclose 90% of the PSF, and estimated the background by selecting 100 surrounding counts outside source regions. None of the extraction regions overlap. The fluxes are then estimated from the count rates in several bands, and the flux is derived. Hardness ratios (colours) are estimated from two energy bands, 0.5–2 and 2–8 keV. The response files of the detector were computed with ACIS Extract for each source using the CIAO tasks *mkarf* and *mkacisrmf* with the associated gain files. We performed a generic spectral fitting of the sources with a power law model and the absorption of the cluster using Xspec v12.2 [@Arnaud96] through the procedure *acis\_extract* (FIT\_SPECTRA stage). The unabsorbed flux is estimated from the best fit of this model. All these results are reported in Table \[tab:chandra\_sources\], and we show a colour-flux diagram of all the sources in Fig. \[fig:chandra\_cmd\]. {width="\textwidth"} A Kolmogorov-Smirnov (KS) test was performed on the extracted and unbinned lightcurve in order to detect variable sources. Seven sources are found to have a KS probability of constancy lower than $10^{-2}$, of which two are located in the core of (sources 3 and 16), and we confirmed the variability of these sources using Poisson statistics. We extracted their lightcurves and the background lightcurves in an annulus around the source with CIAO *dmextract* task. The binned lightcurves are shown in Fig. \[fig:chandra\_lcvar\]. Comparison with XMM-Newton observations {#xmm} ======================================= The XMM-Newton X-ray observation of , performed on 2005 February $1^{st}$ [@Servillat+08-a], was reprocessed in order to match our Chandra energy bands (0.5–2 and  keV). We used the XMM-Newton Science Analysis System[^4] (SAS v7.1) and the most recent calibration data files. The data reduction is detailed in @Servillat+08-a. The 96 detected sources in the XMM-Newton field of view were reprocessed with *emldetect*, without refining the position, in order to extract fluxes and hardness ratios. These values are reported in Table \[tab:chandra\_sources\] for the sources that are inside the Chandra field of view. We took into account the different sensitivities of the instruments in the energy bands used. Using a power law model of photon indices 0, 1, 2 or 3 and the absorption of the cluster, we converted with WebPIMMS a given flux into Chandra ACIS-I and XMM-Newton pn count rates, and compared the hardness ratios. The following conversion factors from XMM-Newton pn to Chandra ACIS-I count rates were deduced: C$_{0.5-2}=0.75$ and C$_{2-8}=1.00$. The corrected hardness ratios for the different models used are found to match with a maximum error of $0.07$, comparable to the $1\sigma$ errors on hardness ratios (see Table \[tab:chandra\_sources\]). Hereafter we refer to the corrected hardness ratios for XMM-Newton sources. Inside the half-mass radius {#sec:Chandra_vs_XMM} --------------------------- ![Colour-luminosity diagram of Chandra and XMM-Newton core sources in . Filled red circles represent Chandra sources, and open blue triangles XMM-Newton sources. XMM-Newton sources are resolved into several Chandra sources which are linked with a line. For clarity, only a few error bars are shown. Black lines correspond to the models described in Fig. \[fig:chandra\_cmd\]. []{data-label="fig:chandra_xmm_cmd_core"}](10188f5.ps){width="\columnwidth"} The better resolution of Chandra allowed us to resolve the core sources previously detected with XMM-Newton. At first, in order to compare these different observations, we degraded the Chandra image with a Gaussian filter to enlarge the PSF to the XMM-Newton PSF size. The result is presented in Fig. \[fig:n2808\_core\]. We note that the images are in general very similar. XMM-Newton source C5 is clearly missing in the Chandra observation, and Chandra source 16 was not detected with XMM-Newton. We extracted a spectrum and generated response files at the position of Chandra source 16 in the XMM-Newton observation to estimate a flux detection limit for this source. We found that Chandra source 16 has varied by a factor of at least $\sim5$. XMM-Newton source C5 has also varied by a factor of at least $\sim5$. We show a colour-luminosity diagram for Chandra core sources with corresponding XMM-Newton core sources in Fig. \[fig:chandra\_xmm\_cmd\_core\]. We note that the observations are well correlated, taking into account the different resolutions. XMM-Newton source C1 is resolved into four Chandra sources (5, 7, 10 and 11) with consistent colours. XMM-Newton source C2 matches with Chandra source 1. This source appeared slightly harder than the Chandra source, probably because it is overlapped by XMM-Newton source C1 which is hard, as shown in @Servillat+08-a. XMM-Newton source C3 is matching with Chandra source 12, and XMM-Newton source C4 has consistent parameters with Chandra source 13. We also note that Chandra sources 2 and 15 are related to $2\sigma$ detections in the XMM-Newton observation, and their fluxes are consistent with the detection limit of the latter observation [@Servillat+08-a]. We extracted spectra and response files for XMM-Newton core sources C1, C3 and C4, and fitted with Xspec a power law with the absorption of the cluster, and a small contribution from close sources as the PSFs overlap. We also extracted a combined spectrum for Chandra sources 5, 7, 10 and 11, and spectra for Chandra sources 12 and 13, and performed a similar fitting. The results are presented in Table \[tab:xmmch\_core\]. XMM-Newton source C3 is more luminous than its matching Chandra source (12, $2.3\sigma$ variation). XMM ID $\Gamma$ Flux -------------- --------------- ------------- C1 $1.41\pm0.12$ $7.3\pm0.7$ C3 $1.36\pm0.20$ $2.1\pm0.5$ C4 $1.73\pm0.30$ $1.1\pm0.5$ Chandra ID $\Gamma$ Flux 5, 7, 10, 11 $1.39\pm0.15$ $5.7\pm1.2$ 12 $1.33\pm0.85$ $0.4\pm0.2$ 13 $1.48\pm0.50$ $0.8\pm0.3$ : Fit results for XMM-Newton core sources C1, C3 and C4, and corresponding Chandra sources or association of sources. The fit was performed with a power law model with the absorption in the direction of the cluster. The photon index ($\Gamma$) and the unabsorbed flux in the 0.5–8 keV energy band in \[${\times10^{-14}\mathrm{erg~cm^{-2}~s^{-1}}}$\] are given.[]{data-label="tab:xmmch_core"} Other sources ------------- Overall, the sources are found to have consistent fluxes and colours in both observations (see Table \[tab:chandra\_sources\]). Chandra source 99 has faded by a factor 5 and has become harder. Chandra sources 31, 73 and 86 appear softer, and 109 and 111 harder than in the XMM-Newton observation ($3\sigma$ variations). \[sec:xmmundet\] For the seven XMM-Newton sources outside the half-mass radius and XMM-Newton source C5, that were not detected with Chandra, we estimated the count rate expected with Chandra. Then, we converted these count rates into counts according to the exposure map of our Chandra observation, i.e. correcting for the vignetting. We estimated a detection threshold by looking at the number of counts of the faintest source detected with the same vignetting. The results are given in Table \[tab:xmmonly\]. We conclude that XMM-Newton sources 34, 66 and C5 should have been detected, unless they have varied between XMM-Newton and Chandra observations. In particular, C5 must have varied by a factor of at least 5 in flux. XMM ID XMM flux XMM HR Chip Exp. Thresh. -------- ----------------- -------------------- -------- ---------------- -------------- 34 1.31 $\pm$ 0.44 $-$0.39 $\pm$ 0.13 ACIS-I **70 $\pm$ 8** 20 $\pm$ 4 48 2.01 $\pm$ 0.69 $-$0.06 $\pm$ 0.15 ACIS-S 115 $\pm$ 10 150 $\pm$ 12 65 1.04 $\pm$ 0.47 $-$0.93 $\pm$ 0.14 ACIS-S 51 $\pm$ 7 150 $\pm$ 12 66 0.36 $\pm$ 0.22 $-$0.67 $\pm$ 0.20 ACIS-I **21 $\pm$ 4** 6 $\pm$ 2 69 0.66 $\pm$ 0.35 $-$0.30 $\pm$ 0.25 ACIS-I 29 $\pm$ 5 25 $\pm$ 5 72 0.78 $\pm$ 0.46 $-$0.79 $\pm$ 0.21 ACIS-S 74 $\pm$ 8 150 $\pm$ 12 84 0.27 $\pm$ 0.28 $-$0.38 $\pm$ 1.45 ACIS-I 10 $\pm$ 3 30 $\pm$ 5 C5 1.07 $\pm$ 0.38 0.57 $\pm$ 0.26 ACIS-I **28 $\pm$ 5** 4 $\pm$ 2 : XMM-Newton sources undetected with Chandra. The XMM-Newton ID, the flux in ${\times10^{-14}\mathrm{erg~cm^{-2}~s^{-1}}}$ and the hardness ratio are shown for each source. We give the Chandra chip type where the source was expected to be detected, the expected number of counts, and the detection threshold in counts. The expected counts that are significantly higher than the threshold are in bold face.[]{data-label="tab:xmmonly"} Counterparts to the X-ray sources {#uv} ================================= Ultraviolet counterparts in the core {#sec:uv_core} ------------------------------------ The core of has been observed with the Space Telescope Imaging Spectrograph (STIS) on board the [*HST*]{} in January/February 2000 using the F25QTZ filter, centered at 159 nm in the FUV band, and the F25CN270 filter, centered at 270 nm in the NUV band. The mosaic of FUV images has a radius of $\sim50\arcsec$ and only covers the core region of . We searched for FUV counterparts to our Chandra sources, using the FUV catalogue provided by @Dieball+05. As a first step, we simply overplotted the Chandra positions on the FUV mosaic. Chandra source 7 and FUV source 222 have a close positional match with a distance of $0.61\arcsec$. This offset agrees with the $0.4\arcsec$ absolute pointing error of Chandra ($1\sigma$ error) and the additional HST absolute pointing accuracy of $0.1-2\arcsec$. FUV source 222 is the best CV candidate in the FUV catalogue, as it is a variable source located between the MS and WD cooling sequence in the FUV–NUV color-magnitude diagram (CMD), which is the expected location for CVs [@Dieball+05]. This source was expected to show X-ray emission, and we are confident that Chandra source 7 and FUV source 222 are the same object. The emission of Chandra source 7 is also consistent with CV emission (see §\[disc\]). We therefore shifted the Chandra positions so that Chandra source 7 and FUV source 222 match exactly. No significant rotation of the field of view is expected [see for instance @EGHG03]. This appears to be the most likely shift in order to align the FUV and X-ray images. We searched for FUV counterparts within a maximum tolerance radius of $3\sigma$ of the corresponding Chandra source, assuming only detection errors. Note that the full width half maximum (FWHM) of the FUV PSF is $< 0.074\arcsec$, i.e. much smaller than the $3\sigma$ error circles for all Chandra sources (and in most cases smaller than the $1 \sigma$ error circle, see Table \[tab:chandra\_sources\]). Therefore we adopted the larger Chandra error circles as the maximum matching radius. {width="10cm"} ----- ----- ------------------ ---------------- ------------ ------------------ ----------------- ------------ ----------- ----------------- ------ ------- ------- Ch. FUV $F_X$ $F_X$ $\sigma$ Prob. Prob. $3\sigma$ ID ID \[mag\] \[flux\] /$F_{NUV}$ \[mag\] \[flux\] /$F_{FUV}$ \[pixel\] \[$\arcsec$\] $\%$ $\%$ 4 400 19.738$\pm$0.024 3.64$\pm$0.08 38 22.042$\pm$0.118 0.55$\pm$0.06 253 20.43 0.505 1.74 32.65 60.53 4 392 – – – 21.743$\pm$0.098 0.73$\pm$0.06 192 25.94 0.641 2.21 47.96 7 222 20.824$\pm$0.030 1.34$\pm$0.04 2168 20.906$\pm$0.048 1.58$\pm$0.07 1839 0.00 0.000 0.00 – 5.26 8 111 17.428$\pm$0.012 3.05$\pm$0.34 5 18.374$\pm$0.015 16.23$\pm$0.22 8 31.25 0.772 2.42 40.82 50.00 9 476 19.583$\pm$0.034 4.20$\pm$0.13 29 21.881$\pm$0.161 0.64$\pm$0.10 187 31.94 0.789 2.47 53.06 73.68 9 457 – – – 21.646$\pm$0.125 0.80$\pm$0.09 151 37.33 0.922 2.89 62.25 10 182 18.353$\pm$0.009 13.03$\pm$0.11 200 17.039$\pm$0.008 55.51$\pm$0.41 47 6.07 0.150 2.17 5.10 2.63 11 252 17.009$\pm$0.004 44.92$\pm$0.17 13 16.400$\pm$0.006 100.00$\pm$0.55 6 16.14 0.399 2.57 6.12 18.42 12 492 18.217$\pm$0.016 14.76$\pm$0.22 29 16.783$\pm$0.019 70.27$\pm$1.22 6 17.91 0.443 2.47 20.41 34.21 14 446 24.139$\pm$0.668 0.06$\pm$0.04 1425 22.966$\pm$0.226 0.24$\pm$0.05 381 7.08 0.175 0.48 4.08 78.95 ----- ----- ------------------ ---------------- ------------ ------------------ ----------------- ------------ ----------- ----------------- ------ ------- ------- ------------ ------------ -------- --------- ------------- --------- ------------- ----------- --------------- ------ ----- Chandra ID Optical ID FUV ID V $F_X$/$F_V$ B $F_X$/$F_B$ $\sigma$ CMD \[mag\] \[mag\] \[pixel\] \[$\arcsec$\] 3 6881 – 18.445 224 18.600 56 2.521 0.116 0.74 BS 4 5315 400 17.947 22 18.122 6 10.949 0.504 1.72 BS 8 6805 – 15.481 2 15.491 1 16.890 0.777 2.41 BHB 8 13062 – 20.648 265 20.928 74 14.739 0.678 2.10 MS 8 7222 – 20.444 220 20.686 59 19.410 0.893 2.77 MS 9 3064 476 17.729 15 17.997 4 17.251 0.794 2.46 BS 11 2486 252 16.156 18 16.006 3 8.763 0.403 2.57 BHB 11 10872 – 21.260 1996 21.578 579 2.789 0.128 0.82 MS 11 2529 – 20.998 1568 21.308 452 3.593 0.165 1.06 MS 12 3165 – 18.543 117 18.563 25 9.697 0.446 2.47 BS ------------ ------------ -------- --------- ------------- --------- ------------- ----------- --------------- ------ ----- Fig. \[fig:fuvima\] shows close-ups on the FUV mosaic. In total, we found 10 possible FUV counterparts to 8 X-ray sources, which are listed in Table \[tab:uvcounterparts\]. Their location in the FUV–NUV CMD is indicated in Fig. \[fig:fuvnuvcmd\]. As can be seen from Fig. \[fig:fuvima\], the $3\sigma$ errors can be quite large for some Chandra sources. The probability of a false match between a Chandra and a FUV source correlates with the size of the Chandra error circle. In order to estimate the probability of such coincidental matches, we repeatedly shifted the set of Chandra sources by 90 STIS pixels or $2.2\arcsec$ to the top and left and right. We only allowed for shifts that kept all Chandra sources within the FUV field of view, so that the statistics for all Chandra sources are the same. After each shift, we again checked for FUV counterparts to each Chandra source. The last two columns in Table \[tab:uvcounterparts\] gives the probability that, just by chance, at least one FUV source can be found within the same distance as the FUV source that was found without a random shift, or within the three sigma error circle of that Chandra source. We used a Monte-Carlo simulation to produce 1000 fields with the detected number of UV and X-ray sources, and we deduced the number of matches expected by chance. Using $1\sigma$, $2\sigma$ and $3\sigma$ error circles for the 16 Chandra sources, we found that $0.6\pm0.7$, $2.2\pm1.4$ and $5.2\pm2.4$ matches may be spurious respectively. We thus expect $\sim2-3$ matches to be real among the 10 UV counterparts. In addition to source 7, two other sources (10 and 14) have probabilities that imply that they may be associated with a FUV source (Table \[tab:uvcounterparts\]). ![FUV–NUV CMD, taken from @Dieball+05. The variable FUV sources are marked with diamonds and their corresponding FUV ID. Additionally, we marked the FUV sources that constitute possible Chandra counterparts with crosses and numbers. For orientation purposes, we include a theoretical WD cooling sequence, a zero-age main sequence (ZAMS), and a zero-age HB track (ZAHB). Note that the sources located along the ZAMS are BS stars, which are located above the main sequence turn-off and slightly to the red. EHB stars are clustered at the bright end of the ZAHB, and the clump of bright sources slightly to the blue and somewhat fainter are the blue hook stars [@Brown+01]. See also [@Dieball+05] for an explanation of the theoretical sequences and a discussion of the CMD.\[fig:fuvnuvcmd\]](10188f7.ps){width="\columnwidth"} ![Optical CMD. Blue crosses (’x’) denote FUV sources seen in optical, green ’+’ denote optical counterparts to X-ray sources, red crosses (’x’) denote the selected optical counterparts listed in Table \[tab:opticalcounterparts\], and the red circled crosses denote the optical counterparts to both FUV and X-ray sources. \[fig:optcmd\]](10188f8.eps){width="\columnwidth"} Optical counterparts -------------------- was observed with the HST Wide Field Planetary Camera 2 (WFPC2) in the F439W and F555W bands, with the Planetary Camera (PC) centered on the cluster’s center. The field of view of the PC is $36\arcsec\times36\arcsec$. In order to check for possible optical counterparts to the 16 core X-ray sources, we used the optical catalogue presented by @Piotto+02. We searched for optical counterparts by transforming all the FUV coordinates, including the 16 shifted Chandra positions, to the PC cartesian system, using the *geomap* and *geoxytran* tasks running under IRAF[^5] and the FUV–optical matches in the @Dieball+05 catalogue as a database. We then looked for optical counterparts within a radius $\le 3\sigma$ of the transformed Chandra errors. However, nearly all of these are either located on the MS, red giant branch or red HB. Two of the optical sources are on the BHB and four sources are in the optical BS region above the MS turn-off. There are four sources that are faint and blue with $B-V < 0.4$ and $V > 20.4$ mag. An optical counterpart to a CV is expected to be very blue. All the above mentioned optical sources are listed in Table \[tab:opticalcounterparts\] and a colour-magnitude diagram is presented in Fig. \[fig:optcmd\]. Of these, four are optical counterparts to the FUV sources that might constitute possible matches to the Chandra sources, see Table \[tab:uvcounterparts\]. Ultraviolet counterparts outside the core ----------------------------------------- was observed with the XMM-Newton Optical Monitor (OM) using the UVM2 filter, centered at 231 nm in the UV band [@Servillat+08-a]. We reprocessed the data in the same way as @Servillat+08-a, in order to use the latest version of the SAS (v7.1), and looked for UV counterparts to the Chandra sources. Because of the crowding of sources, we could not use this data inside the half-mass radius. A bright source is seen in the field of view outside the crowded area of the core of , which is consistent with the star at RA$_{2000}$ $9^h11^m33.293^s$, Dec$_{2000}$ $-$645103.28. This star is an A0V star with magnitudes ${B=10.42}$ and ${V=10.15}$, compatible with the UVM2 magnitude we found. We thus shifted the coordinates of the OM observation to match this source. After the shift, the number of possible matches significantly increased from 2 to 16. We used a Monte-Carlo simulation to deduce the number of matches expected by chance, and estimated that $1.6\pm1.2$ matches may be spurious (using $3\sigma$ error circles). This indicates that the nearly all of the matches are real. Moreover, a rotation of the field of view does not seem necessary. We took into account the $1\sigma$ positional error of the Chandra sources, and the $1\sigma$ positional error of the UV sources. The 16 possible counterparts found outside the half-mass radius are listed in Table \[tab:uvom\], nine of which are within the $1\sigma$ error circle. From the nine UV counterparts found for XMM-Newton sources [@Servillat+08-a], eight are consistent with the Chandra sources. The XMM-Newton source 29 counterpart is ruled out as the corresponding Chandra source, 33, has a smaller $3\sigma$ error circle which does not match the proposed UV counterpart. Chandra source 36 has a UV counterpart and a matching XMM-Newton source, but the match between UV and XMM-Newton sources was not found [@Servillat+08-a] because this UV source was not detected with the older version of the SAS. As they are found outside the half-mass radius of , these sources are likely to be background or foreground sources (see Sect. \[sec:fore\] and \[sec:back\]). ID Distance $\sigma$ Offset UVM2 -------- ---------- ---------- -------- ------------------ 17 1.77 2.17 0.89 17.00 $\pm$ 0.06 19 2.15 2.64 1.30 15.18 $\pm$ 0.02 21 1.36 1.71 1.95 17.25 $\pm$ 0.05 31$^*$ 0.43 0.54 3.19 10.95 $\pm$ 0.01 35 0.81 0.94 3.36 19.52 $\pm$ 0.41 36$^*$ 1.34 1.52 3.53 19.55 $\pm$ 0.40 39$^*$ 0.53 0.66 3.77 17.83 $\pm$ 0.07 41$^*$ 0.52 0.65 3.86 18.29 $\pm$ 0.10 45 2.44 2.83 4.02 17.28 $\pm$ 0.04 48$^*$ 2.39 2.85 4.10 17.23 $\pm$ 0.04 49 1.26 1.23 4.14 19.50 $\pm$ 0.43 51$^*$ 0.83 0.96 4.29 18.99 $\pm$ 0.17 53$^*$ 0.09 0.12 4.42 14.70 $\pm$ 0.01 63 0.29 0.33 5.15 18.86 $\pm$ 0.15 74$^*$ 0.11 0.13 5.82 17.99 $\pm$ 0.07 99$^*$ 0.79 0.96 8.53 18.29 $\pm$ 0.09 : XMM-Newton OM UV counterparts to Chandra X-ray sources. The first column corresponds to the Chandra ID and a star indicates that the source was detected with XMM-Newton. In column 2 we give the distance between the X-ray source and the possible UV counterpart and in column 3 this distance is divided by the $1\sigma$ error. The offset from the center of is given in column 4 and in the last column, the UVM2 Vega magnitude.[]{data-label="tab:uvom"} Discussion {#disc} ========== In order to identify the X-ray sources linked to , we first discuss the properties of X-ray sources and their possible counterparts. The brightest sources with ${L_{[0.5-8keV]}\gtrsim10^{32}\mathrm{~erg~s^{-1}}}$ are likely to be qLMXBs if they are very soft and well fitted by hydrogen atmosphere model, with masses and radii consistent with a neutron star [@HGLE03; @GBW03b; @GBW03]. They are likely to be CVs if they are harder, with $kT>3$ keV [@BWO05], or in this work a photon index less than 2. X-ray sources with lower luminosities are most likely either CVs, ABs or MSPs, although we note that there are also qLMXBs which have luminosities around ${L_{[0.5-8keV]}\sim5\times10^{31}\mathrm{~erg~s^{-1}}}$ [e.g. @HGE05]. ABs and MSPs are in general expected to be soft [@DLSF93; @Bogdanov+06]. The brightest AB observed in a GC reached ${L_{[0.5-8keV]}\sim10^{32}\mathrm{~erg~s^{-1}}}$ [@Heinke+05]. Their spectrum is generally well fitted by a two temperature model with mean $T_{high}\sim10^7$ K and $T_{low}\sim10^6$ K, and they have luminosities of ${L_{X}\sim10^{29}-10^{31}\mathrm{~erg~s^{-1}}}$ [@DLSF93]. MSPs in GCs are well described by a thermal (blackbody or neutron star hydrogen atmosphere) spectrum with a temperature $T_{eff}\sim(1-3)\times10^6$ K, emission radius $R_{eff}\sim0.1-3$ km, and luminosity ${L_{X}\sim10^{30}-10^{31}\mathrm{~erg~s^{-1}}}$ [@Bogdanov+06]. Some of them show a significant or dominant non-thermal (power law) component, with spectral photon index $\Gamma\sim1-1.5$, and are more luminous [@Bogdanov+06]. The brightest MSP observed in a GC reached ${L_{[0.5-8keV]}\sim1.3\times10^{33}\mathrm{~erg~s^{-1}}}$ [in M28, @Becker+03]. In the UV, CVs are expected to be brighter than qLMXBs, ABs and MSPs, and are located between the MS and the WD cooling sequence in the CMD [e.g. @Dieball+05]. ABs, on the other hand, are expected to be redder than MS stars or on the MS [e.g. @EGHG03]. CVs can show high variability in X-rays on different timescales [e.g. @1995cvs..book.....W]. Large variations of a factor of 5–10, and timescales of hours to days are most commonly associated with dwarf nova outbursts [@BWO05; @WMM03]. However, such variation can also appear during the quiescent state of some CVs [@BWO05]. ABs show flaring activity. The decay time of a compact flare is $\sim1$ ks, and for two-ribbon flares it reaches 5–20 ks [@PPK88]. The rise time is generally shorter than 2 ks [@PPK88]. Longer flares have been observed, with a noticeable decay [110 ks, @FPT01], and very exceptional flares can last several days [e.g. @KS96]. X-ray sources with a possible UV counterpart in the core -------------------------------------------------------- ### Chandra source 7: a CV Chandra source 7 has X-ray colour and flux compatible with CV emission (see Table \[tab:chandra\_sources\]). Furthermore, its UV counterpart is located in the gap region between the MS and the WD cooling sequence in the FUV–NUV CMD (see Fig. \[fig:fuvnuvcmd\]). The X-ray/UV ratios are high for this source (see Table \[tab:uvcounterparts\]), similar to those observed for intermediate polars (see Sect. \[sec:XUV\_CVs\]). ### Chandra source 14: a possible CV Chandra source 14 is fainter, with X-ray colour and flux compatible with CV, AB or MSP emission. Its possible UV counterpart is located in the CV region and the X-ray/UV ratios are compatible with the emission of CVs (see Sect. \[sec:XUV\_CVs\]). The probability of finding by chance a FUV source within a $3\sigma$ error circle is high ($\sim80$%). However, the probability of finding a FUV source just by chance within a $0.5\sigma$ error circle of Chandra source 14 is 4% (see Table \[tab:uvcounterparts\]), therefore it seems likely that FUV source 446 is the true counterpart, making Chandra source 14 a good CV candidate. From simulations, the probability of having a FUV source which is also one of the $\sim60$ CV candidates is 0.5% . ### Chandra source 10: another CV? Chandra source 10 has very similar X-ray parameters to Chandra source 7, a priori suggesting that this source also constitutes a CV (Fig. \[fig:chandra\_cmd\]). Its possible UV counterpart is bright in FUV and NUV, and it is located in the faint part of the EHB region in the FUV–NUV CMD. Although field CVs in outburst can be as bright as this source in FUV [see Fig. 3 in @Dieball+05], in this case the source would have to be in outburst for $\sim30$ days, or during each FUV observation [2000 January 18 and 19, and February 16 and 20, @Dieball+05], which is unlikely as CV outbursts are expected to be rare in GCs [@Shara+96]. It is also possible that its true UV counterpart is masked by the bright, close UV source. ### Chandra source 11 The flux and colour of Chandra source 11 are consistent with a CV or an AB candidate. It possibly matches with a bright UV source, located in the BHB clump in the FUV–NUV CMD, and an optical source which also suggests that this source is a BHB star (Table \[tab:uvcounterparts\]). X-ray emission is not expected from such HB stars, and this source could constitute another kind of exotic binary. The matching distance between the X-ray and FUV coordinates is $2.6\sigma$. It is possible that the UV source is simply a mismatch. One of the optical counterparts (\#10872, Table \[tab:opticalcounterparts\]) is blue, as expected for a CV. ### Chandra source 12 Chandra source 12 has a similar flux and colour to Chandra source 11, consistent with the CV or AB hypothesis. Its luminosity also possibly varied between XMM-Newton and Chandra observations ($2\sigma$). The possible UV counterpart is located in the EHB clump in the FUV–NUV CMD. However, the match to this FUV source is within $2.5\sigma$, with a probability of $\sim18\%$ that this match is a chance superposition. This might well be a mismatch and the true FUV counterpart was not detected in our FUV images. One of the possible optical counterparts is in the BS region (see Table \[tab:opticalcounterparts\] and Fig. \[fig:optcmd\]). ### Chandra source 8 Chandra source 8 can be either a CV, an AB, or a MSP candidates. It possibly matches with a bright FUV source, located on the HB sequence in the FUV–NUV CMD. This source is thus not expected to reach such a luminosity in X-rays and it could be a mismatch. We found two possible optical counterparts compatible with the expected emission of a CV (see Table \[tab:opticalcounterparts\] and Fig. \[fig:optcmd\]). ### Chandra sources 4 and 9 Chandra sources 4 and 9 are fainter X-ray sources, giving weak constraints about their nature. They are either CV, AB, or MSP candidates. Both Chandra sources 4 and 9 have two possible FUV counterparts. Of these, FUV sources 400 and 476 seem to be the likely counterparts to Chandra source 4 and 9, respectively, as they are closer (Table \[tab:uvcounterparts\]). These FUV sources are located at the faint end of the BS sequence. The discrimination between BS and CV candidates is difficult, and both populations might well overlap in the FUV–NUV CMD. Thus, it is possible that X-ray sources 4 and 9 are CVs, or they might be X-ray emitting BSs. Such an object was found in 47 Tuc by @Knigge+06. They suggested that this exotic object is a detached binary consisting of a BS primary with an X-ray-active MS companion whose formation would involve at least three stars. @EGHG03 and @Heinke+05 also identified such objects in 47 Tuc. The coincidence probability of these sources is high (Table \[tab:uvcounterparts\]) and they could be mismatches. X-ray sources without a UV counterpart in the core -------------------------------------------------- ### Chandra source 1: a qLMXB candidate Chandra source 1 corresponds to XMM-Newton source C2, which was presented as a qLMXB candidate by @Servillat+08-a. The Chandra flux and colour of this source also support the idea that it is a qLMXB. Moreover, the spectrum of Chandra source 1 is well fitted with a hydrogen atmosphere model with parameters consistent with a neutron star. No hard tail is detected in the spectrum, as in general in GC qLMXBs [@Heinke+03a]. The upper limit is ${F_{[0.5-8keV]}\sim5\times10^{-16}\mathrm{~erg~cm^{-2}~s^{-1}}}$ ($<2.5$% of the total flux), assuming a power law model with photon index 1–1.5 [@Campana+98], and 1 photon detected in the band 2–8 keV. @Servillat+08-a estimated that $3\pm1$ qLMXBs are expected in the core of , using the empirical correlation between the number of qLMXBs and the star encounter rate. We found no evidence in our Chandra observation for other qLMXB candidates. ### Chandra source 16: a variable CV Chandra source 16 is a variable X-ray source. This variability is observed on different time scales: during the first Chandra observation (few hours), between the two Chandra observations (1 day), and between XMM-Newton and Chandra observations (28 months). Its flux, colour and variability in X-rays suggest it could be a CV, possibly showing the signature of a CV outburst in X-rays [e.g. @BWO05]. An AB interpretation is less probable as even in long flares, the decay of the flare would be detectable over the first 45 ks of Chandra observation [as in @FPT01 for a 110 ks flare]. ### Chandra source 3: possible CV Chandra source 3 is very hard and shows short-term variability. The lightcurve indicates a count increase by a factor $\sim5$ between the beginning of the first Chandra observation and the second Chandra observation (Fig. \[fig:chandra\_lcvar\]). The rise time is longer than 10 ks (end of first Chandra observation, Fig. \[fig:chandra\_lcvar\]), exceptional for an AB flare [e.g. @PPK88]. This kind of fluctuation would be more typical of a CV, and it is possibly the signature of a CV outburst in X-rays [e.g. @BWO05]. We note that a possible optical counterpart falls in the BS region in Fig. \[fig:optcmd\], however it could be a mismatch. ### XMM-Newton source C5 XMM-Newton source C5, which was not detected during our Chandra observation, was hard, and is variable over 28 months. The flux varied by a factor of at least 5 over this period. From its flux and colour, it could be a variable CV or an AB which showed a flare. No variability was detected from this source during the XMM-Newton observation of $\sim30$ ks with the *pn* detector [@Servillat+08-a]. Such a long flare would be exceptional for an AB, although it cannot be excluded [e.g. @KS96]. CVs routinely show fluctuations in X-rays, some of them being associated with CV outbursts [e.g. @BWO05]. ### Chandra source 17 We note that Chandra source 17, which is close to the half-mass radius of , has colours and fluxes consistent with those of a CV, an AB or a MSP. It shows variability that is possibly consistent with a CV fluctuation or an AB flare, as it lasted $\sim10$ ks (Fig. \[fig:chandra\_lcvar\]). This source could therefore be linked to the cluster, and could be a CV or an AB. If it is a CV, its offset from the core would make it a primordial CV candidate [@HAS07]. ### Other Chandra core sources From their luminosities and colours, Chandra sources 2, 5, 13, 15 are consistent with the CV or AB hypothesis. Similar X-ray sources are identified as CVs or ABs by @Heinke+05 in . However, we did not find FUV counterparts, nor detected variability, and we cannot confirm their nature. Chandra source 6 is fainter, and is consistent with being a CV, an AB, or even a bright MSP candidate. Variable sources in the core ---------------------------- Several authors have claimed that there is a lack of CV outbursts in GCs [e.g. @Shara+96; @EGHG03b; @DLM06]. To explain this deficit, it was proposed that most GC CVs are moderately magnetic with low accretion rates, systems for which very few outbursts are expected [@Grindlay99; @Ivanova+06; @DLM06]. @Shara+96 looked for dwarf nova outbursts in 47 Tuc using HST observations for a cumulative time of $\sim130$ ks. They found only one outburst and showed that they should have detected at least 12 from simulations, implying a lack of outbursts from GC CVs. Outbursts of GC CVs have also been observed in UV for two magnetic CVs in [@Shara+05]. The cumulative time of their observation was $\sim66$ ks. Our X-ray observations lasted $\sim90~ks$ (XMM-Newton and Chandra), and could therefore lead to the detection of dwarf nova outburst signatures. An optical outburst of a dwarf nova can be linked to an increase of X-ray emission (factor $\sim5$), followed by a drop (factor $\sim3$) due to the variation of thickness of the boundary layer at different accretion rates [@WMM03; @1995cvs..book.....W]. However, X-ray fluctuations of CVs are also detected during the quiescent state [@BWO05]. Our observation indicates that some sources in the core show such fluctuations in X-rays. Chandra sources 3, 16 and XMM-Newton source C5 are likely to be CVs that possibly showed signatures of outbursts in X-rays. Simultaneous X-ray and UV (or optical) observations should be performed for several GCs to confirm these possible dwarf nova outbursts and assess the rate of outbursts in GCs. Expected X-ray sources linked to {#sec:expected} --------------------------------- Assuming a completeness in the detection of sources at a luminosity of ${L_{[0.5-8keV]}\sim2\times10^{31}\mathrm{~erg~s^{-1}}}$, we can estimate the number of expected X-ray sources in . To estimate the number of expected X-ray sources in , we use a general approach based on the correlations given by @PH06, where they used the results of observations of 22 GCs with Chandra. Taking into account the specific encounter frequency of , as defined in @PH06, and our completeness limit, we expect $30\pm6$ X-ray sources in . The error is extrapolated from the $1\sigma$ error in @PH06. Of these sources, $3\pm1$ qLMXBs are expected, and $17\pm3$ CVs with a luminosity greater than ${L_{[0.5-8keV]}\sim4.25\times10^{31}\mathrm{~erg~s^{-1}}}$ [corresponding to populations I and II respectively, as defined in @PH06]. However, we detected only 11 sources in the half-mass radius of and above the completeness limit. Looking only at CVs, we can compare our results to the population synthesis simulation performed by @Ivanova+06. They estimated the number of CVs and detectable CVs expected in a GC similar to (their standard model): same core density, relaxation time and metallicity, but velocity dispersion and escape velocity somewhat lower in their model. Their simulation lead to 209 CVs formed after 10 Gyr, of which 47 have a luminosity higher than ${L_{[0.5-8keV]}\sim3.6\times10^{30}\mathrm{~erg~s^{-1}}}$. The number of detectable CVs seems stable over an interval of ages of 7–14 Gyr with a standard deviation of $4$. We can adapt this result to our completeness limit by applying a factor $\sim0.36$ [obtained from the number of sources detected in with the corresponding limiting luminosities in @Heinke+05]. This leads to $17\pm4$ detectable CVs predicted for our observation. We detected 8 (and possibly up to 15) CV candidates, which seems less than the predicted number of CVs from simulations. In the same way, @Knigge+08 found that there are only a few CVs among the bright FUV sources located between the MS and WD cooling sequence in , indicating a possible deficit of CVs compared to what is expected. If this is true, our understanding of CV formation and evolution in GCs might have to be revised. We compared our results to the observation of [@Heinke+05]. is similar in mass, density and concentration to [@Harris96]. The specific encounter frequency of is comparable to that of . has an intermediate metallicity [ corresponding to ${Z\sim0.001}$, @Yi+01 Table 2], while has a higher metallicity ($\rm{[Fe/H]} = -0.76$, or ${Z\sim0.0035}$). About $31\pm3$ X-ray sources above our completeness limit [the error corresponds to the $1\sigma$ error on the luminosity in @Heinke+05] were detected in the half-mass radius of , of which: 2–5 qLMXBs, 16–19 CVs, 4–5 ABs, and 1 MSP. , has been observed with Chandra [@Heinke+03]. Its metallicity is lower ($\rm{[Fe/H]} = -1.75$, or ${Z\sim0.0004}$), and its mass, density and concentration are close to values [@Harris96]. A total of $17\pm2$ sources were detected above a similar luminosity threshold to our observation, of which 2 are qLMXB candidates, and 5 possible CVs. This is similar to our number of detections in . The possible deficit of X-ray sources could therefore be linked to specific parameters of . Metallicity seems to be a key parameter that could highly affect the number of X-ray sources in GCs at a given age. Due to the lower opacity, metal poor stars are generally hotter and more compact. This could lead to the formation of WDs with different properties, and will also determine if, when and how mass transfer occurs in binaries [@demink+07]. , and have very different metallicities, which could explain the differences in the number of detectable X-ray sources. In the same way, @KMZ07 found that metal-rich extragalactic GCs host three times as many LMXBs than metal-poor ones. presents unusual features in the optical: an extended BHB with clumps [@Bedin+00], several MS corresponding to different star populations [@Piotto+07], and abundances anomalies in HB stars [@Pace+06]. All these particular features are linked in some way to the metallicity of the stars. @Piotto+07 proposed that several populations of stars were formed successively, increasing the helium and metal content of the cluster material at each round. The possible deficit of X-ray sources we observed is possibly linked to this specific evolution of which modified its metallicity content. X-ray and UV emission from CVs {#sec:XUV_CVs} ------------------------------ ![Cumulative radial profiles of different populations in . The populations of WDs and CVs correspond to FUV selected sources [@Dieball+05]. \[fig:radial\_profiles\]](10188f9.eps){width="\columnwidth"} @Knigge+08 showed that no more than half of the objects lying in the CV region in the FUV-NUV diagram of are confirmed as CVs. If the types of object in the gap in have similar proportions as in , then we expect that $\sim30$ of the $\sim60$ CV candidates detected [@Dieball+05] will ultimately turn out to be CVs. With Chandra, we obtained at most 15 CV candidates. If we take into account only significant matches, we found two UV counterparts that have UV properties clearly compatible with the CV hypothesis (Chandra sources 7 and 14). Even if we take into account the incompleteness of our observations, it seems that X-ray and UV emission from CVs are decorrelated, as the brightest X-ray sources in are generally not the brightest FUV sources. The four CVs confirmed in FUV by @Knigge+02 in also have very different X-ray to FUV ratios, strengthening this idea. For instance, AKO 9 is brighter than V1 in FUV [magnitudes $\sim16$ and $\sim18$ respectively, @Knigge+02], but fainter in X-rays [${L_{\mathrm{0.5-6~keV}}\sim5\times10^{31}}$ and ${6.8\times10^{32}\mathrm{~erg~s^{-1}}}$ respectively, @Heinke+05]. The cumulative radial profiles of the X-ray population and the FUV CV candidates are shown in Fig. \[fig:radial\_profiles\]. The spatial distributions do not appear to be significantly different with a KS test. As there is a mass segregation effect in the cluster, this would indicate that the mean mass of the systems is not a dominant parameter to explain the different properties of these populations. UV and X-ray photons come from different processes. The UV emission is mainly due to the accretion disk [non-magnetic CVs, @Godon+08], accretion curtains (intermediate polars) or accretion streams (polars), and possibly the hot spot. In systems with low accretion rates or in magnetic CVs, the WD also contributes to the UV emission [@MS84]. On the other hand, the X-ray emission arises from close to the surface of the WD [@WW03], produced by the boundary layer [@Pringle77; @PR85] or the shock above the WD magnetic pole. Therefore long period non-magnetic nova-like CVs (large accretion disk with high accretion rate) tend to be relatively X-ray faint and UV bright. Non-magnetic systems should generally be bright in UV due to their dominant accretion disc. ROSAT observations of field CVs in X-ray and optical strengthen this idea [@Verbunt+97]. Contrary to this, magnetic CVs seem to be more luminous in X-rays than in UV [e.g. @Verbunt+97; @EGHG03b]. Part of the UV emission is expected to be suppressed in these systems due to the truncation of the inner portion of the accretion disc of intermediate polars [@Grindlay99]. This is also observed for polars which have no accretion disks [e.g. @1995cvs..book.....W]. FUV source 397 is variable in FUV, as is FUV source 222, and is a CV candidate. However, no counterpart is detected in X-ray, contrary to FUV source 222 which matches with Chandra source 7. These CV candidates could therefore be in separate CV classes. FUV source 397 is likely to possess a UV-bright accretion disk as expected for non-magnetic systems. Chandra source 7 is brighter in X-ray as expected for magnetic systems, and variable and bright in UV due to the probable presence of an accretion disk. It is therefore likely to be an intermediate polar. We estimated a $F_X$/$F_{NUV}$ ratio for several CVs belonging to different classes, as estimated in Table \[tab:uvcounterparts\], where $F_{NUV}$ is the flux density between 2500–3000 Å, and $F_X$ the flux in the band 0.5–8 keV. Polars have ratios greater than 5000 [ratios extrapolated from @RC03; @Ramsay+04; @Vogel+08]. Intermediate polars appear to have a ratio greater than 2000 [@HMZ02; @deMartino+05; @deMartino+06]. The detection limit of the NUV observation is ${6\times10^{-19}\mathrm{~erg~cm^{-2}~s^{-1}~\AA^{-1}}}$, and the limit in X-rays is ${9\times10^{-16}\mathrm{~erg~cm^{-2}~s^{-1}}}$. Therefore, the X-ray/NUV ratio for the CV candidates detected in UV is lower than $\sim1500$. The $\sim30$ CV candidates detected in UV and not in X-rays are thus likely to be mostly non-magnetic systems [such as the dwarf nova YZ Cnc with a ratio of $\sim500$, @Hakala+04]. Most intermediate polars in the field are more luminous than ${10^{31}\mathrm{~erg~cm^{-2}~s^{-1}}}$ in X-rays [@Verbunt+97 see also the Intermediate Polar Home Page[^6] maintained by K. Mukai, where 12 out of 14 have luminosities above this limit]. We should have detected most of these in our Chandra observation if their emission is similar to intermediate polars in the field. This would lead to a maximum of $\sim14$ intermediate polars (we exclude Chandra source 14 whose X-ray/NUV ratio is lower than 2000, and Chandra source 1 which is a qLMXB candidate). The proportion derived is $\sim30\%$ of the detected CV candidates, and $\sim7\%$ of the expected GC CV population, estimated to be $\sim200$ CVs [@Ivanova+06 see also §\[sec:expected\]]. This is coherent with the proportion of intermediate polar candidates in the field, which can be estimated to $\sim5\%$ from the catalogue of @RK03 [updated Feb. 2008]. Due to the incompleteness of our observations, this result does not allow us to confirm or rule out a possible excess of intermediate polars in . However, with a deeper sample, this method could allow us to better quantify the proportion of intermediate polars. Constraint on the mass of an intermediate mass black-hole in ------------------------------------------------------------- Following @Servillat+08-a, we can constrain the possible mass of an intermediate mass black-hole (IMBH) in . No sources are detected at the center of mass of , where such a massive object is expected to be found. We assumed that the BH is fed by intracluster gas with a density of $\sim0.5\mathrm{~cm^{-3}}$ derived from @PR01 with parameters. We considered that the IMBH is radiatively inefficient as for an optically thin advection-dominated accretion flow [ADAF, see e.g. @NMQ98]. Following @GHEM01, our limiting luminosity implies an upper limit of $\sim140\mathrm{~M_{\sun}}$ for a central IMBH in . @MS08 used ATCA radio observations to put a mass limit on an IMBH in (370 M$_{\sun}$). They cast doubt on suggestions that globular clusters may follow the same $M_{BH}-\sigma$ relation as galaxies. Foreground sources {#sec:fore} ------------------ ### Chandra source 31 Chandra source 31 has a bright UV and optical counterpart compatible with the A0V star . The distance estimate of to this source [@Servillat+08-a] leads to an X-ray luminosity of ${L_{\mathrm{0.5-8~keV}}=7.3\pm1.9\times10^{29}\mathrm{~erg~s^{-1}}}$. The X-ray emission is soft and showed some variation since the XMM-Newton observation.The luminosity is in agreement with the observed luminosity of X-ray bright A-type stars [@SS07]. However, the source of the emission may come from a late-type active star companion, as the majority of A-type stars are expected to be particularly X-ray dark [@SS07]. Using VLT/FORS1 spectra (see Fig. \[fig:hd79548\]), we note that a contribution from a late-type star is possible, and difficult to detect as its flux could be at least 100 times lower than the A-star. The 12 radial velocity measurements of the H$\beta$ line do not show evidence for a period, so we cannot confirm the presence of a companion. ### Chandra source 113 Chandra source 113 showed a flare in its X-ray lightcurve (Fig. \[fig:chandra\_lcvar\]), which reached at least three times the mean flux of the source. The decay timescale was $\sim5$ ks. The photon index of the spectrum indicates a soft source (2.29, Table \[tab:chandra\_sources\]). These characteristics are in agreement with those of a chromospherically active star [@BP03]. Background sources {#sec:back} ------------------ ### Sources with radio counterparts was observed with the Australia Telescope Compact Array (ATCA) on 24 January 1992. The data processing is described in @MS08. Chandra source 73 has a radio counterpart. It is the most luminous X-ray source of the field, its X-ray emission is hard and absorbed, and it showed a colour variation between XMM-Newton and Chandra observations. Chandra sources 50 and 89 also have a radio counterpart, and possibly Chandra source 107 and XMM-Newton source 8 (outside the field of view of Chandra). All these sources are hard and absorbed in X-ray (Table \[tab:chandra\_sources\]), as expected for AGN [@Mainieri+07]. ### Chandra source 99 This source varied in luminosity between the XMM-Newton and Chandra observations. It has been proposed to be a background AGN based on its hard, absorbed X-ray spectrum [@Servillat+08-a]. The Chandra observation is consistent with this hypothesis, as is the UV emission detected with the XMM-Newton OM. We detect an infrared counterpart in the Spitzer data found in the archives[^7] (AOR 11586048), which is not compatible with the black body emission of a star. These properties are consistent with the emission of a galaxy, possibly an ultraluminous infrared galaxy [@LFS06; @Braito+04]. Conclusions =========== We presented Chandra observations of coupled with previous XMM-Newton observations [@Servillat+08-a], HST FUV observations [@Dieball+05], VLT and ATCA observations. We have shown that 16 Chandra sources are likely to be linked to the cluster, with possibly a 17th close to the half mass radius. One of these is consistent with the X-ray emission of a qLMXB, confirming the previous detection with XMM-Newton. Two Chandra sources (7 and 14) have FUV counterparts that show emission compatible with a CV. Chandra source 10 is likely to be a CV from its X-ray emission, but no UV counterparts was found to confirm its nature. Another highly variable source (16) in the core is likely to be a CV, as well as two other variable sources (Chandra 3 and XMM-Newton C5). Two other Chandra sources (8 and 11) have optical counterparts compatible with the expected emission of CVs. We have thus identified 7 CV candidates (plus XMM-Newton source C5) and the observations indicate that there may be as many as 15 in the Chandra observations (although some of the faintest may be ABs or MSPs), along with $\sim30$ CV candidates in the HST UV observations. This significant population of close binaries is likely to play an important role in slowing down the core collapse of this cluster. Compared to the number of X-ray sources detected in and expected from dynamical formation, we found a possible deficit of X-ray sources in . This might indicate a true deficit of CVs, which is possibly linked to the metallicity content and the complexity of the evolution of . This paper has strongly benefitted from the careful review by the referee, C. Heinke, for which we are very grateful. MS is grateful to the University of Southampton and the Astronomy group for hosting him for two months, where part of this work was done. MS thanks T. J. Maccarone for interesting discussions and for the processing of ATCA data which improved the content of this article. This research has made use of data obtained from the Chandra Data Archive and software provided by the Chandra X-ray Center. This work is also based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA. We thank the CNES for support of the operational phase of this mission. Part of this work is based on: FORS1 observations collected with the Very Large Telescope at the European Southern Observatory, Cerro Paranal, Chile; observations with the Australia Telescope Compact Array, founded by the Commonwealth of Australia for operation as a National Facility managed by the CSIRO; and observations made with the Spitzer Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. [90]{} natexlab\#1[\#1]{} , M. D., [Gilliland]{}, R. L., [Brown]{}, T. M., [et al.]{} 2001, , 559, 1060 , K. A. 1996, in ASP Conf. Ser. 101: Astronomical Data Analysis Software and Systems V, ed. G. H. [Jacoby]{} & J. [Barnes]{}, 17 , D. S., [Wheatley]{}, P. J., & [Osborne]{}, J. P. 2005, , 357, 626 , W., [Swartz]{}, D. A., [Pavlov]{}, G. G., [et al.]{} 2003, , 594, 798 , L. R., [Piotto]{}, G., [Zoccali]{}, M., [et al.]{} 2000, , 363, 159 , S., [Grindlay]{}, J. E., [Heinke]{}, C. O., [et al.]{} 2006, , 646, 1104 , V., [Della Ceca]{}, R., [Piconcelli]{}, E., [et al.]{} 2004, , 420, 79 , K. R. & [Pye]{}, J. P. 2003, , 345, 714 Broos, P., Townsley, L., Getman, K., & Bauer, F. 2002, ACIS Extract, An ACIS Point Source Extraction Package, Pennsylvania State University, http://www.astro.psu.edu/xray/docs/TARA/ae\_users\_guide.html , T. M., [Sweigart]{}, A. V., [Lanz]{}, T., [Landsman]{}, W. B., & [Hubeny]{}, I. 2001, , 562, 368 , S., [Colpi]{}, M., [Mereghetti]{}, S., [Stella]{}, L., & [Tavani]{}, M. 1998, , 8, 279 , E., [Bragaglia]{}, A., [Gratton]{}, R. G., [et al.]{} 2006, , 450, 523 , E., [Gratton]{}, R. G., [Clementini]{}, G., & [Fusi Pecci]{}, F. 2000, , 533, 215 , M. B. 1997, , 288, 117 , D., [Matt]{}, G., [Mukai]{}, K., [et al.]{} 2006, , 454, 287 , D., [Matt]{}, G., [Mukai]{}, K., [et al.]{} 2005, , 437, 935 , S. E., [Pols]{}, O. R., & [Yoon]{}, S. . 2007, ArXiv e-prints, 710 , R. C., [Linsky]{}, J. L., [Schmitt]{}, J. H. M. M., & [Fleming]{}, T. A. 1993, , 413, 333 , A., [Knigge]{}, C., [Zurek]{}, D. R., [Shara]{}, M. M., & [Long]{}, K. S. 2005, , 625, 156 , A., [Lasota]{}, J.-P., & [Menou]{}, K. 2006, , 640, 288 , P. D., [Gilliland]{}, R. L., [Heinke]{}, C. O., & [Grindlay]{}, J. E. 2003, , 596, 1177 , P. D., [Gilliland]{}, R. L., [Heinke]{}, C. O., & [Grindlay]{}, J. E. 2003, , 596, 1197 , R., [Hut]{}, P., & [Inagaki]{}, S. 1987, , 25, 565 , A. C., [Pringle]{}, J. E., & [Rees]{}, M. J. 1975, , 172, 15P , E., [Pallavicini]{}, R., & [Tagliaferri]{}, G. 2001, , 375, 196 , A., [McDowell]{}, J. C., [Allen]{}, G. E., [et al.]{} 2006, in SPIE Conference, Vol. 6270, Observatory Operations: Strategies, Processes, and Systems. Edited by Silva, David R.; Doxsey, Rodger E. , B., [Barret]{}, D., & [Webb]{}, N. A. 2003, , 403, L11 , B., [Barret]{}, D., & [Webb]{}, N. A. 2003, , 400, 521 , R., [Rosati]{}, P., [Tozzi]{}, P., [et al.]{} 2001, , 551, 624 , P., [Sion]{}, E. M., [Barrett]{}, P. E., [et al.]{} 2008, , 679, 1447 , J. E. 1999, in Astronomical Society of the Pacific Conference Series, Vol. 157, Annapolis Workshop on Magnetic Cataclysmic Variables, ed. C. [Hellier]{} & K. [Mukai]{}, 377 , J. E., [Heinke]{}, C., [Edmonds]{}, P. D., & [Murray]{}, S. S. 2001, Science, 292, 2290 , F., [Motch]{}, C., & [Zickgraf]{}, F.-J. 2002, , 387, 201 , P., [Ramsay]{}, G., [Wheatley]{}, P., [Harlaftis]{}, E. T., & [Papadimitriou]{}, C. 2004, , 420, 273 , W. E. 1974, , 192, L161 , W. E. 1996, , 112, 1487 , G., [Altieri]{}, B., [Arnaud]{}, M., [et al.]{} 2001, , 365, L45 , G., [Miyaji]{}, T., & [Schmidt]{}, M. 2005, , 441, 417 , C. O., [Grindlay]{}, J. E., & [Edmonds]{}, P. D. 2005, , 622, 556 , C. O., [Grindlay]{}, J. E., [Edmonds]{}, P. D., [et al.]{} 2005, , 625, 796 , C. O., [Grindlay]{}, J. E., [Edmonds]{}, P. D., [et al.]{} 2003, , 598, 516 , C. O., [Grindlay]{}, J. E., [Lloyd]{}, D. A., & [Edmonds]{}, P. D. 2003, , 588, 452 , C. O., [Grindlay]{}, J. E., [Lugger]{}, P. M., [et al.]{} 2003, , 598, 501 , C. O., [Wijnands]{}, R., [Cohn]{}, H. N., [et al.]{} 2006, , 651, 1098 , J. R., [Aarseth]{}, S. J., & [Shara]{}, M. M. 2007, , 665, 707 , P., [McMillan]{}, S., [Goodman]{}, J., [et al.]{} 1992, , 104, 981 , P., [Shara]{}, M. M., [Aarseth]{}, S. J., [et al.]{} 2003, New Astronomy, 8, 337 , N., [Heinke]{}, C. O., [Rasio]{}, F. A., [et al.]{} 2006, , 372, 1043 , C., [Dieball]{}, A., [Maiz Apellaniz]{}, J., [et al.]{} 2008, , in press/ArXiv:0805.0140 , C., [Gilliland]{}, R. L., [Dieball]{}, A., [et al.]{} 2006, , 641, 281 Knigge, C., Zurek, D. R., Shara, M. M., & Long, K. S. 2002, apj, 579, 752 , C., [Zurek]{}, D. R., [Shara]{}, M. M., [Long]{}, K. S., & [Gilliland]{}, R. L. 2003, , 599, 1320 , M. & [Schmitt]{}, J. H. M. M. 1996, , 311, 211 , A., [Maccarone]{}, T. J., & [Zepf]{}, S. E. 2007, , 662, 525 , R. L. 1993, VizieR Online Data Catalog, 6039, 0 , W. H. G. & [Joss]{}, P. C. 1983, in Accretion-Driven Stellar X-ray Sources, ed. W. H. G. [Lewin]{} & E. P. J. [van den Heuvel]{}, 41 , C. J., [Farrah]{}, D., & [Smith]{}, H. E. 2006, Ultraluminous Infrared Galaxies (Astrophysics Update 2), 285 Maccarone, T. J. & Servillat, M. 2008, mnras, 843 , V., [Hasinger]{}, G., [Cappelluti]{}, N., [et al.]{} 2007, , 172, 368 , M. & [Szkody]{}, P. 1984, , 89, 863 Miocchi, P. 2007, mnras, 381, 103 , K. 1993, Legacy, vol. 3, p.21-31, 3, 21 , R., [Mahadevan]{}, R., & [Quataert]{}, E. 1998, in Theory of Black Hole Accretion Disks, ed. M. A. [Abramowicz]{}, G. [Bjornsson]{}, & J. E. [Pringle]{}, 148 , G., [Recio-Blanco]{}, A., [Piotto]{}, G., & [Momany]{}, Y. 2006, , 452, 493 Patterson, J. & Raymond, J. C. 1985, apj, 292, 535 , E. & [Rappaport]{}, S. 2001, , 550, 172 , G., [Bedin]{}, L. R., [Anderson]{}, J., [et al.]{} 2007, , 661, L53 , G., [King]{}, I. R., [Djorgovski]{}, S. G., [et al.]{} 2002, , 391, 945 , G., [Pallavicini]{}, R., & [Kopp]{}, R. A. 1988, , 201, 93 , D. & [Hut]{}, P. 2006, , 646, L143 , D., [Lewin]{}, W. H. G., [Anderson]{}, S. F., [et al.]{} 2003, , 591, L131 , P. & [Schmitt]{}, J. H. M. M. 1995, , 293, 889 , J. E. 1977, , 178, 195 , G., [Bridge]{}, C. M., [Cropper]{}, M., [et al.]{} 2004, , 354, 773 , G. & [Cropper]{}, M. 2003, , 338, 219 , H. & [Kolb]{}, U. 2003, , 404, 301 , C. & [Schmitt]{}, J. H. M. M. 2007, , 475, 677 , M., [Webb]{}, N. A., & [Barret]{}, D. 2008, , 480, 397 , M. M., [Bergeron]{}, L. E., [Gilliland]{}, R. L., [Saha]{}, A., & [Petro]{}, L. 1996, , 471, 804 , M. M., [Hinkley]{}, S., [Zurek]{}, D. R., [Knigge]{}, C., & [Dieball]{}, A. 2005, , 130, 1829 , F., [Bunk]{}, W. H., [Ritter]{}, H., & [Pfeffermann]{}, E. 1997, , 327, 602 , F. & [Hut]{}, P. 1987, in IAU Symp. 125: The Origin and Evolution of Neutron Stars, ed. D. J. [Helfand]{} & J.-H. [Huang]{}, 187 , J., [Byckling]{}, K., [Schwope]{}, A., [et al.]{} 2008, , 485, 787 , A. R. 1999, , 118, 432 Warner, B. 1995, Cataclysmic variable stars (Cambridge Astrophysics Series, Cambridge, New York: Cambridge University Press, |c1995) , N. A., [Serre]{}, D., [Gendre]{}, B., [et al.]{} 2004, , 424, 133 , N. A., [Wheatley]{}, P. J., & [Barret]{}, D. 2006, , 445, 155 , P. J., [Mauche]{}, C. W., & [Mattei]{}, J. A. 2003, , 345, 49 , P. J. & [West]{}, R. G. 2003, , 345, 1009 , S., [Demarque]{}, P., [Kim]{}, Y.-C., [et al.]{} 2001, , 136, 417 [^1]: http://cxc.harvard.edu/ciao [^2]: http://www.astro.psu.edu/xray/docs/TARA/ae\_users\_guide.html [^3]: http://heasarc.gsfc.nasa.gov/Tools/w3pimms.html [^4]: http://xmm.vilspa.esa.es/sas [^5]: IRAF (Image Reduction and Analysis Facility) is distributed by the National Astronomy and Optical Observatory, which is operated by AURA, Inc., under cooperative agreement with the National Science Foundation. [^6]: http://asd.gsfc.nasa.gov/Koji.Mukai/iphome/iphome.html [^7]: http://ssc.spitzer.caltech.edu/archanaly/archive.html

--- author: - 'A. Fernández-Martín [^1] , E. Pérez-Montero , J.M. Vílchez ,' - 'A. Mampaso .' title: Chemical distribution of HII regions towards the Galactic anticentre --- [The study of the radial variations of metallicity across the Galactic disc is a powerful method for understanding the history of star formation and chemical evolution of the Milky Way. Although several studies about gradients have been performed so far, the knowledge of the Galactic antincentre is still poor.]{} [This work aims to determine accurately the physical and chemical properties of a sample of [H<span style="font-variant:small-caps;">ii</span>]{} regions located at R$_G>$11 kpc and to study the radial distribution of abundances in the outermost part of the Galaxy disc.]{} [We carried out new optical spectroscopic observations of nine [H<span style="font-variant:small-caps;">ii</span>]{} regions with the William Herschel Telescope covering the spectral range from 3500 [Å]{} to 10100[Å]{}. In addition, we increased the sample by searching the literature for optical observations of regions towards the Galactic anticentre, re-analysing them to obtain a single sample of 23 objects to be processed in a homogeneous and consistent manner. The total sample distribution covers the Galactocentric radius from 11 kpc to 18 kpc.]{} Emission line ratios were used to determine accurate electron densities and temperatures of several ionic species in 13 [H<span style="font-variant:small-caps;">ii</span>]{} regions. These physical parameters were applied to the spectra to determine direct total chemical abundances. For those regions without direct estimations of temperature, chemical abundances were derived by performing tailor-made photoionisation models and/or by using an empirical relation obtained from radio recombination and optical temperatures. We performed weighted least-squares fits to the distribution of the derived abundances along the Galactocentric distances to study the radial gradients of metallicity across the outermost part of the MW. The distributions O/H, N/H, S/H, and Ar/H towards the anticentre can be represented by decreasing linear radial gradients, while in the case of N/O abundances the radial distribution is better fitted with a two-zone model. The He/H radial gradient is presented here for the first time; we find a slope that is not significantly different from zero. The derived gradient for oxygen shows a clear decrease with distance with a slope of -0.053$\pm$0.009 dex kpc$^{-1}$. Although a shallower slope at large Galactocentric distances is suggested by our data, the flattening of the distribution cannot be confirmed and more objects towards the anticentre need to be studied in order to establish the true form of the metallicity gradient. Introduction {#sect:intro} ============ The chemical evolution of the interstellar medium (ISM) varies between galaxies and is both position and time dependent within a galaxy. Since the nucleosynthesis of different elements occurs in stars of different masses, the study of chemical abundances in the discs of spiral galaxies is a powerful method for understanding the history of star formation and evolution of galaxies. A knowledge of the radial variations of metallicity across the galactic disc (i.e. abundance gradients) is central to our understanding of a wide variety of observed phenomena, including the physics of star formation rates [@Phillipps1991], initial mass function [@Guesten1982], the radial inflows/outflows of gas [@Mayor1981], and the stellar evolution and the process of nucleosythesis itself. Galactic abundance gradients in the ISM were first described by @Searle1971 in a survey of [H<span style="font-variant:small-caps;">ii</span>]{} regions in six late-type spiral galaxies. Since then, considerable effort has been made to establish the chemical distribution in the Milky Way (MW) by studying many sources such as supernova remnants (SNR) [@Binette1982], molecular clouds [@Rudolph1996], open clusters [@Twarog1997; @Andreuzzi2011], cepheids [@Luck2003; @Korotin2014], OB stars [@Rolleston2000], and planetary nebulae (PNe) [@Maciel1994; @Henry2010]. However, when deriving the abundances representative of the current ISM values, [H<span style="font-variant:small-caps;">ii</span>]{} regions provide the most accessible probe of abundances gradients. Since they are bright and hot they emit strongly in many lines observable over much of the MW. Unlike stars, [H<span style="font-variant:small-caps;">ii</span>]{} regions probe the current state of abundances, and unlike PNe and SNR, do not contaminate the surrounding ISM.\ The existence of a large-scale gradient in the MW with [H<span style="font-variant:small-caps;">ii</span>]{} regions was established by the pioneering work of @Shaver1983, who found a decrease of metallicity with Galactic distances. Subsequently, several studies with [H<span style="font-variant:small-caps;">ii</span>]{} regions [@Hawley1978; @Talent1979; @Fich1991; @Hunter1992; @Afflerbach1997; @Vilchez1996; @Deharveng2000; @Rudolph2006] have been carried out, firmly establishing the existence of a negative gradient of abundance of the elements heavier than helium in the disc of the MW. [l c c c c c ]{}\ [H<span style="font-variant:small-caps;">ii</span>]{} region$^{(a)}$ & $\alpha_{J2000}$ & $\delta_{J2000}$ & Airmass$^{(b)}$ & Exp. Time $^{(c)}$ & Date\ & (h m s) & ($^{\rm o~'~''}$) & & (s) &\ \ S83 & 19:24:30.77 & +20:47:45.92 & 1.18 & 3$\times$1800 & July, 13, 2010\ S132 & 22:19:08.26 & +56:05:12.98 & 1.37 & 3$\times$1800 & July, 13, 2010\ S156 & 23:05:08.33 & +60:14:46.29 & 1.23 & 3$\times$700 / 3$\times$200 $^{(d)}$ & July, 13, 2010\ S162 & 23:20:43.94 & +61:12:27.06 & 1.20 & 500+900+1000 / 18$\times$100 $^{(d,e)}$ & July, 13, 2010\ S207 & 4:19:49.49 & +53:09:34.91 & 1.27 & 3$\times$1200 & Dec., 19, 2009\ S208 & 4:19:32.39 & +52:58:38.98 & 1.21 & 3$\times$1500 & Dec., 19, 2009\ S212 & 4:40:36.50 & +50:27:44.31 & 1.14 & 3$\times$1500 & Dec., 19, 2009\ S228 & 5:13:23.37 & +37:27:19.63 & 1.02 & 3$\times$1500 & Dec., 19, 2009\ S270 & 6:10:12.96 & +12:48:37.38 & 1.24 & 1200+1300+1500 $^{(e)}$ & Dec., 19, 2009\ Nonetheless, the sampling of the whole Galactic disc is still poor. In particular, relatively few [H<span style="font-variant:small-caps;">ii</span>]{} regions have been studied towards the Galactic anticentre region. Optically, there are only ten Galactic [H<span style="font-variant:small-caps;">ii</span>]{} regions which have been observed at R$_G>$11 kpc and for which direct measurements of the electron temperature are available. This is also a handicap for the MW studies, compared with those for nearby external galaxies, and limits the application of models of galactic evolution. In particular, it limits those aspect of the models that describe the evolution of the outermost parts of the disc, which might be considered – at least from the point of view of the chemical evolution – to be much closer to the pre-galactic/early conditions in the MW. Moreover, the possibility of variations in the slope of the gradients in the outer disc has been a subject of debate in the past years. Some authors have claimed that radial abundance gradients of some elements may flatten out at the outer parts of the Galactic disc [@Fich1991; @Vilchez1996], while other authors do not support such flattening [@Deharveng2000; @Rudolph2006]. Therefore, extending the measurement of abundances to large Galactocentric distances in the MW is essential to our understanding of the metallicity gradient and the chemical evolution of our Galaxy.\ This paper has two main purposes. The first is to present new long-slit observations in a wide spectral range of [H<span style="font-variant:small-caps;">ii</span>]{} regions located towards the Galactic anticentre with Galactocentric distances from 11 kpc to 17 kpc. The second is to present a re-determination of chemical abundances, from direct electron temperatures and from physical modelling, of all the optical outermost [H<span style="font-variant:small-caps;">ii</span>]{} regions observations in the literature in a homogeneous and consistent manner to perform a self-consistent study of the chemical gradients towards the outermost disc of the MW. The wide optical spectral range covered, together with the consistent chemical analysis of the sample, allowed us to give an accurate description of the radial chemical distribution towards the Galactic anticentre up to a distance of 18 kpc. In the following section we present the data sample including our own observations and other data from the literature. Section \[sect:results\] describes the analysis and the results of the study, while Sect. \[sect:discussion\] is devoted to discussing the resulting abundance gradients and their implications in the context of Galactic evolution. Finally, we describe the main conclusions of this work in Sect. \[sect:conclusions\]. {width="\textwidth"} Sample of [H<span style="font-variant:small-caps;">ii</span>]{} regions {#sect:sample} ======================================================================= This work aims to study the chemical abundances of [H<span style="font-variant:small-caps;">ii</span>]{} regions located towards the Galactic anticentre. Therefore we selected Galactic [H<span style="font-variant:small-caps;">ii</span>]{} regions in the direction of the anticentre with R$_G>$ 11 kpc from catalogues of @Fich1984, @Deharveng2000, and @Quireza2006. The mother sample includes 89 anticentre [H<span style="font-variant:small-caps;">ii</span>]{} regions observable from the northern hemisphere. Selection and observations of WHT data {#sect:owndata} -------------------------------------- We selected for our observations those [H<span style="font-variant:small-caps;">ii</span>]{} regions from the mother sample located at the furthest distance from the Galactic centre and, if possible, with available information on their thermal structure (e.g. a radio continuum temperature). Nine [H<span style="font-variant:small-caps;">ii</span>]{} regions located at 11 kpc $ < $ R$_G< $ 17 kpc were observed; their names and positions are listed in Table \[table:logISIS\]. The observations were carried out in December 2009 and July 2010 using the ISIS double-armed spectrograph at the 4.2m William Herschel Telescope (WHT) at the Observatorio del Roque de los Muchachos (La Palma, Spain). A dichroic was set at $\sim$5336 [Å]{} permitting simultaneous observation in the blue and red arms, which are optimised for their respective wavelength ranges. The blue arm detector, EEV12, is an array of 4096$\times$2048 (13.5 $\mu$m) pixels giving a spatial resolution of 0.2 arcsec/pix, while the CCD for the ISIS red arm, RED+, is a red-sensitive array of 4096$\times$2048 (15.0 $\mu$m) with a spatial resolution of 0.22 arcsec/pix. The gratings were chosen in order to obtain as much information as possible on the most important emission lines of the optical range. In the blue arm, the R300B grating was centred at 4400 [Å]{} covering the effective spectral range from 3200 [Å]{} to 5800 [Å]{} with a dispersion of 0.86 [Å]{}/pix, giving a spectral resolution of $\mathrm{R=\lambda/\delta\lambda\sim}$1070 at 4400 [Å]{}. In the red arm, the R158R was centred at 7800 [Å]{} covering the effective spectral range from 5300 [Å]{} to 10000 [Å]{} with a dispersion of 1.81 [Å]{}/pix and a spectral resolution of R$\sim$1010 at 7800 [Å]{}. Table \[table:logISIS\] also shows the observational log for the [H<span style="font-variant:small-caps;">ii</span>]{} regions of the observed sample. At least three exposures were taken in each region with different exposition times depending on the object. The slit width was set at 1 arcsec throughout the observation and oriented in parallactic angle to avoid differential atmospheric refraction problems. The necessary bias frames, continuum, arcs, and spectrophotometric standard stars were also acquired.\ The spectra were reduced using [IRAF]{}[^2] by following the standard procedure for 2D spectroscopic observations (bias subtraction, flat-fielding, and cosmic ray rejection). The wavelength calibration was made using CuNe+CuAr arcs taken shortly after the object spectrum was taken. Two spectrophotometric standard stars were observed each night (G191-B2B and HR345 in December 2009, and BD+262603 and BD+174708 in July 2010) so that the spectra could be flux calibrated. For each exposure, a background region was selected in order to avoid nebular emission or stars. Pixels of these regions were combined with a median in a single spectrum and a 2D background spectrum was created with the value of the combined spectrum in each column. This sky background was subsequently subtracted from every science object. This method was used in all the [H<span style="font-variant:small-caps;">ii</span>]{} regions except for S162 where the S156-sky was used for the sky subtraction since S162 was too extended and no area without nebular emission was found. The sky subtraction worked well overall; only the strongest sky-lines in the far red spectra ($\lambda >$9000 [Å]{}) left some residuals. In order to select the zones to extract the 1D spectra we examined the spatial profiles of the most important emission lines, whenever possible selecting areas with emission of the auroral lines necessary to estimate the electron temperature. The spectra were traced and extracted using the [APALL]{} task in [IRAF]{} with an aperture optimised for a good S/N which was able to minimise the contamination of adjacent pixels; in this process we took into account the different spatial scales between blue and red arms. Eighteen spectra were finally extracted over the sample. In Fig. \[fig:spectra\] we present four representative examples showing the auroral lines necessary to estimate electron temperatures. Additional data from the literature {#sect:extended} ----------------------------------- To realise a more extensive study of the Galactic anticentre, the sample was increased with [H<span style="font-variant:small-caps;">ii</span>]{} regions from previous works. With this aim, we carried out an exhaustive bibliographical review, selecting those [H<span style="font-variant:small-caps;">ii</span>]{} regions located at R$_G>$11 kpc observed in the optical range and with measurements of the electron temperature sensitive lines or, at least, with information about the [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$9068 line to obtain chemical abundances by means of individual photoionozation models. As a result of this search we found 14 [H<span style="font-variant:small-caps;">ii</span>]{} regions that satisfy the stipulated requirements: S98, S127, S128, S209, S219, S266 (B), S283, and S203 (BFS31) from @Vilchez1996; S255 (c), S301 (RCW6), and S311 (RCW16a) from @Shaver1983; S158 (P1) from @Talent1979; S206 from @Caplan2000; and S298 (B2) from @Esteban1990. Some authors provide information about regions in several pointings; in these cases we chose those with more emission lines measured (in parentheses we indicate the pointing selected for our study and the identification given by each author). [l c c c c c c c c c]{}\ Region & l & b$^{[1]}$ & R$_G$ & [d$_{\odot}$]{}& z$^{[7]}$ & $\Theta^{[1]}$ & Spectral type of dominant & t$_e^*$ & Other\ & ($^o$) & ($^o$) & (kpc) & (kpc) & (pc) & (arcsec) & exciting star & ($\times$10$^4$K) & names\ \ S83 & 55.12 & 2.42 & 15.2 $^{[2]}$ & 18.7 $^{[2]}$ & - & 2 & - & - & -\ S98 & 68.15 & 1.02 & 12.8 $^{[3]}$ & 13.3 $^{[3]}$ & - & 15 & O4V$^{[8]}$ & 1.08$^{[3]}$ & -\ S127 & 96.29 & 2.60 & 13.9 $^{[2]}$ & 12.2 $^{[2]}$ & 698 & 2 & O8V$^{[9]}$ & 1.14$^{[3]}$ & -\ S128 & 97.50 & 3.16 & 12.7 $^{[2]}$ & 9.0 $^{[2]}$ & 484 & 1 & O7V$^{[9]}$ & 1.04$^{[3]}$ & -\ S132 & 102.79 & -0.65 & 11.3 $^{[4]}$ & 5.8 $^{[4]}$ & -59 & 90 & O8.5V$^{[13]}$/B0III$^{[14]}$/WN6$^{[15]}$ & - & -\ S156 & 110.11 & 0.06 & 11.5 $^{[4]}$ & 5.3 $^{[4]}$ & 6 & 2 & O8V$^{[10]}$/O6.5V:$^{[15]}$/O7V$^{[16]}$ & 0.92$^{[3]}$/0.91$^{[4]}$ & -\ S158 & 111.55 & 0.82 & 12.4 $^{[4]}$ & 6.4 $^{[4]}$ & 38 & 10 & O9V$^{[10]}$ & 0.85$^{[3]}$/0.82$^{[4]}$ & NGC7538\ S162 & 112.22 & 0.23 & 11.1 $^{[4]}$ & 4.7 $^{[4]}$ & 13 & 40 & O7I$^{[9]}$/O6.5IIIf$^{[11,16]}$ & 0.86$^{[3]}$/0.81$^{[4]}$ & NGC7635\ S203 & 143.81 & -1.57 & 11.3 $^{[5]}$ & 3.3 $^{[5]}$ & - & 2 & B2V$^{[10]}$ & - & BFS31\ S206 & 150.58 & -0.94 & 11.1 $^{[2]}$ & 2.8 $^{[2]}$ & -44 & 50 & O4$^{[15]}$/O5V$^{[16]}$/O5neb$^{[17]}$ & 1.00$^{[3]}$/0.97$^{[4]}$ & -\ S207 & 151.19 & 2.13 & 16.8 $^{[2]}$ & 4.3 $^{[2]}$ & - & 4 & O9V$^{[12,13]}$/O9.5IV$^{[9,17]}$ & - & -\ S208 & 151.31 & 1.99 & 16.8 $^{[2]}$ & 4.1 $^{[2]}$ & 261 & 1 & O9.5V$^{[9]}$/B0V$^{[12,17]}$ & - & -\ S209 & 151.60 & -0.25 & 16.9 $^{[4]}$ & 8.9 $^{[4]}$ & -51 & 14 & O9III$^{[9]}$ & 1.08$^{[3]}$/1.05$^{[4]}$ & -\ S212 & 155.35 & 2.60 & 16.7 $^{[4]}$ & 8.6 $^{[4]}$ & 278 & 5 & O6I$^{[9]}$/O7f$^{[15]}$/O5.5neb$^{[17]}$ & 1.03$^{[3]}$/1.05$^{[4]}$ & -\ S219 & 159.36 & 2.60 & 13.3 $^{[2]}$ & 4.5 $^{[2]}$ & 189 & 3 & B2.5V$^{[10]}$/B0V$^{[16,17]}$ & - & -\ S228 & 169.21 & -0.90 & 13.8 $^{[4]}$ & 5.3 $^{[4]}$ & - & 8 & O8Ve$^{[9,13]}$/B0V$^{[15]}$ & 0.94$^{[3]}$/0.97$^{[4]}$ & -\ S255 & 192.64 & 0.00 & 11.0 $^{[5]}$ & 2.5 $^{[5]}$ & -2 & 3 & B0V$^{[15]}$/B0IIIneb$^{[17]}$ & - & -\ S266 & 195.66 & -0.09 & 17.9 $^{[6]}$ & 9.6 $^{[6]}$ & -15 & 1 & BeI? $^{[18]}$ & - & -\ S270 & 196.83 & -3.10 & 14.2 $^{[6]}$ & 5.9 $^{[6]}$ & -478 & 1 & - & - & -\ S283 & 210.81 & -2.56 & 17.0 $^{[6]}$ & 9.1 $^{[6]}$ & -407 & 3 & B3V$^{[10]}$/O7V$^{[10]}$/B1V$^{[10]}$/B0:V:$^{[17]}$ & - & -\ S298 & 227.75 & -0.15 & 12.4 $^{[4]}$ & 5.0 $^{[4]}$ & -7 & 22 & WN5$^{[14,17]}$ & 1.25$^{[3]}$/1.36$^{[4]}$ & NGC2359/RCW5\ S301 & 231.45 & -4.41 & 12.9 $^{[4]}$ & 5.8 $^{[4]}$ & -439 & 9 & O6V$^{[16]}$/O7$^{[17]}$/B1III$^{[17]}$ & 0.91$^{[3]}$/0.97$^{[4]}$ & RCW6\ S311 & 243.16 & 0.36 & 12.0 $^{[4]}$ & 5.4 $^{[4]}$ & 32 & 45 & O5V$^{[13]}$/O6.5V$^{[13]}$/O6:$^{[14]}$ & 1.05$^{[3]}$/1.02$^{[4]}$ & RCW16\ Final sample {#sect:totalsample} ------------ The final sample to be analysed in this work includes both samples described above (Sects. \[sect:owndata\] and \[sect:extended\]) and consists of 23 [H<span style="font-variant:small-caps;">ii</span>]{} regions located in a range of Galactoncentric distances from 11 kpc to 18 kpc. In Table \[table:infoRHII\] we present a summary of the fundamental physical parameters of the [H<span style="font-variant:small-caps;">ii</span>]{} regions analysed in this study. Each [H<span style="font-variant:small-caps;">ii</span>]{} region is identified in Col. 1 by its Sharpless number [@Sharpless1959]. Columns 2 and 3 give the Galactic coordinates (l,b). Columns 4 and 5 give the Galactocentric and heliocentric distances (R$_G$ and [d$_{\odot}$]{}), both of which are analysed and justified in Sect. \[sect:distances\]. Column 6 shows the distance above the Galactic plane (z), and Col. 7 gives the angular size of the nebula ($\Theta$). In Col. 9 we show the spectral type of the dominant exciting star. Column 10 gives the electron temperature derived from radio observations (t$_e^*$) and, finally, Col. 11 indicates other identification names. The corresponding references are given in the footnotes.\ Analysis and results {#sect:results} ==================== [l c c c c c c]{}\ [H<span style="font-variant:small-caps;">ii</span>]{} region & n$_e$ & t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}) & t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) & t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) & t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}) & t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{})\ & (cm$^{-3}$) & (10$^4$ K) & (10$^4$ K) & (10$^4$ K) & (10$^4$ K) & (10$^4$ K)\ \ S83 & 256 $\pm$ 62 & 1.16 $\pm$ 0.03 $_E$ & 1.14 $\pm$ 0.05 & 1.13 $\pm$ 0.02 & - & -\ S127 & 325 $\pm$ 75 & - & 1.01 $\pm$ 0.13 $_E$ & 0.88 $\pm$ 0.15 $_E$ & 0.91 $\pm$ 0.05 & -\ S128 & $<$100 & - & 0.92 $\pm$ 0.15 $_E$ & 0.77 $\pm$ 0.18 $_E$ & 0.98 $\pm$ 0.12 & -\ S132 & 315 $\pm$ 15 & 0.96 $\pm$ 0.05 & 0.82 $\pm$ 0.07 $_E$ & 0.81 $\pm$ 0.02 & - & 0.84 $\pm$ 0.09\ S156 & 1133 $\pm$ 29 & 0.96 $\pm$ 0.04 & 0.92 $\pm$ 0.03 & 0.80 $\pm$ 0.01 & - & 0.87 $\pm$ 0.18\ S158 & 1323 $\pm$ 265 & 0.99 $\pm$ 0.15 & 0.87 $\pm$ 0.21 $_E$ & 0.72 $\pm$ 0.25 $_E\dagger $ & - & -\ S162 & 1440 $\pm$ 23 & 0.93 $\pm$ 0.03 & 0.94 $\pm$ 0.02 & 0.80 $\pm$ 0.03 & - & 0.81 $\pm$ 0.03\ S206 & 357 $\pm$ 161 & - & 0.95 $\pm$ 0.02 & - & 0.87 $\pm$ 0.05 $_E$ & -\ S212 & 211 $\pm$ 29 & 1.11 $\pm$ 0.06 & 1.21 $\pm$ 0.04 & 1.18 $\pm$ 0.03 & - & -\ S255 & 321 & 0.82 & 0.65 $_E$ & - & - & -\ S298 & $<$100 & - & 1.20 $\pm$ 0.03 & 1.11 $\pm$ 0.04 $_E$ & 1.17 $\pm$ 0.08 $_E$ & -\ S301 & $<$100 & - & $<$0.91 & $<$0.77 $_E$ & $<$0.96 $_E$ & -\ S311 & 130 & 0.95 & 0.87 & 0.72 $_E$ & - & 0.88\ \ Line intensities and reddening correction {#sect:intensities} ----------------------------------------- We now report on the method used to obtain the line intensities of the regions observed with ISIS (our own sample). Line fluxes were measured by integrating all the flux in the line between two given limits and over a fitted local continuum. Partially blended lines, such [H$\alpha$]{} and [\[N<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda\lambda$6548,6584, were deblended with two or three Gaussian profiles to measure the individual line fluxes. All the measurement were made with the [SPLOT]{} task of [IRAF]{}. The statistical errors associated with the observed fluxes were calculated using the expression $$\sigma_{\mathrm{1}}=\sigma_{\mathrm{c}} N^{1/2} [1+EW/(N\Delta)]^{1/2},$$ where $\sigma_{\mathrm{1}}$ represents the error in the observed line flux, N is the number of pixels used to measure the line, EW the line equivalent width, $\sigma_{\mathrm{c}} $ the standard deviation of the continuum in a box near the line, and $\Delta$ represents the dispersion in [Å]{}/pix [@PerezMontero2003]. The reddening coefficient c([H$\beta$]{}) was derived from the observed flux ratios of the brightest Balmer lines [H$\alpha$]{}/[H$\beta$]{}, [H$\gamma$]{}/[H$\beta$]{}, and [H$\delta$]{}/[H$\beta$]{} as compared with the theoretical values obtained from the public software of @Storey1995 assuming Case B recombination and following an iterative process with the n$_e$ and t$_e$ estimations (Sect. \[sect:parameters\]). To minimise the reddening effect and to avoid error propagation due to differences between red and blue grating calibrations, the measured lines were reddening corrected relative to [H$\beta$]{}, [H$\alpha$]{}, or P10 (if S/N$>$5) depending on the spectral range and later normalised to [H$\beta$]{} flux using the corresponding theoretical ratios. Table \[table:sample\_intensities\] lists the reddening-corrected intensities of the emission lines measured for every observed [H<span style="font-variant:small-caps;">ii</span>]{} region labelled with their standard identification. The third column gives the adopted reddening function, f($\lambda$), using the extinction law by @Cardelli1989 with $R_{\mathrm{V}}=3.1$. Errors in the emission line intensities were derived by propagating the observational errors in the fluxes and the reddening constant uncertainties. The estimated fluxes and errors were normalised to F([H$\beta$]{})=1000. The values obtained for c([H$\beta$]{}) are also presented in the last row of Table \[table:sample\_intensities\]. At zero redshift the [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$9530[Å]{} line fluxes are strongly affected by a sky absorption band. Therefore, this line was not used for the derivation of the [\[S<span style="font-variant:small-caps;">iii</span>\]]{} electron temperature neither for the S$^{++}$/H$^+$ ionic abundance. For these quantities a theoretical ratio [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$9530/[\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$9069=2.44 was assumed.\ The methodology used to obtain the intensities of the [H<span style="font-variant:small-caps;">ii</span>]{} regions selected from the literature was different since in these regions we did not measure the fluxes in spectra. In this case, we calculated the original reddened fluxes considering the extinction law and c([H$\beta$]{}) used by each author and later we re-derived the intensities following the same self-consistent method as in our regions: theoretical Balmer lines with Case B recombination, iterative process with n$_e$ and t$_e$, and reddening function using the extinction law by @Cardelli1989 with $R_{\mathrm{V}}=3.1$. Finally, we obtained consistent intensities derived exactly under the same conditions as in our sample. Data from these 14 [H<span style="font-variant:small-caps;">ii</span>]{} regions were added to our new observations reported here to create a data set of 23 [H<span style="font-variant:small-caps;">ii</span>]{} regions with 11 kpc $< R_G<$ 18 kpc. Hereafter we work with all of these objects as a single sample following the same methodology for both samples. Physical parameters {#sect:parameters} ------------------- To obtain the physical conditions of the gas, we performed an iterative process for each region until an agreement was achieved between electron density (n$_e$) and electron temperature (t$_e$). The values for t$_e$ and n$_e$ derived for each [H<span style="font-variant:small-caps;">ii</span>]{} region are shown in Table \[table:physical\_parameters\]. Those regions without measurements of auroral lines necessary to estimate at least one electron temperature were discarded temporarily. Physical parameters and chemical abundances were not derived with this direct method for any of them and, therefore, they are not included in Table \[table:physical\_parameters\] (but see Sects. \[sect:models\] and \[sect:tradio\]).\ Electron density, n$_e$, was calculated for all the observed regions from the [\[S<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda\lambda$6717,6731 line ratio using the [IRAF]{} package [TEMDEN]{} based on a five-level statistical equilibrium model [@deRobertis1987; @Shaw1995]. Electron temperatures, t$_e$, were derived using the appropriate line ratios $\mathrm{R_{N2}, R_{O3}, R_{S3}, R_{O2}}$, and $\mathrm{R_{S2}' }$ following the equations below for each ion: $$\label{eq:Te_conNIIbis} t_e({[N\,\textsc{ii}]})\,=\,0.537\,+\,0.000253~R_{N2}\,+\,42.126/R_{N2};$$ $$\label{eq:Te_conOIIIbis} t_e({[O\,\textsc{iii}]})\,=\,0.8254\,-\,0.0002415~R_{O3}\,+\,47.77/R_{O3};$$ $$\label{eq:Te_conSIII} t_e({[S\,\textsc{iii}]})\,=\left[R_{S3}\,+\,36.4\right]/\left[1.8~R_{S3}\,-\,3.01\right];$$ $$\begin{aligned} \label{eq:Te_conOII} t_e({[O\,\textsc{ii}]})\,=&\,a_0(n_e)\,+&\,a_1(n_e)~R_{O2}\,+\,a_2(n_e)/R_{O2},\nonumber\\ \rm{where} & a_0(n_e) =& 0.23-0.005~n_e-0.17/n_e \nonumber\\ & a_1(n_e)=& 0.007+0.000009~n_e+0.0064/n_e \nonumber\\ & a_2(n_e)=& 38.3-0.021~n_e-1.64/n_e; \end{aligned}$$ $$\begin{aligned} \label{eq:Te_conSII} t_e({[S\,\textsc{ii}]})\,=&a_0(n_e)\,+&\,a_1(n_e)~R_{S2}'\,+\,a_2(n_e)/R_{S2}'\,+\,a_3(n_e)/R_{S2}'^2, \nonumber\\ \rm{where} & a_0(n_e)=&1.92-0.0017~n_e+0.848/n_e \nonumber\\ & a_1(n_e)=&-0.0375+4.038~10^{-5}~n_e-0.0185/n_e \nonumber\\ & a_2(n_e)=&-14.15+0.019~n_e-10.4/n_e \nonumber\\ & a_3(n_e)=&105.64+0.019~n_e+58.52/n_e. \end{aligned}$$ In all the equations, electron density is in units of cm$^{-3}$ and electron temperatures in units of 10${^4}$ K.\ {width="6cm"} {width="6cm"} {width="6cm"} {width="6cm"} {width="6cm"} Our criterion in the determination of the electron temperature was to derive at least one temperature characteristic for each ionisation degree: t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) as the high-ionisation zone, t$_e($[\[S<span style="font-variant:small-caps;">iii</span>\]]{}) as the medium-ionisation zone, and t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}), t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{}), or t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}) as the low-ionisation zone. However, not all the necessary auroral lines were detected in all the regions (we did not estimate t$_e$ for all the ions). For S206, S298, and S301 no auroral line associated with a low-excitation ion was observed and we resorted to the expression $$t_e({[O\,\textsc{ii}]})=\frac{1.2~+~0.002~n_e~+~4.2/n_e}{t_e({[O\,\textsc{iii}]})^{-1}~+~0.08~+~0.003~n_e~+~2.5/n_e}, \label{eq:teOIIconOIII}$$ which relates t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}) with t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) [@PerezMontero2009]. Equation \[eq:teOIIconOIII\] was also used for S127 and S128, in which we inferred t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) from t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}). In the case of S132, S158, and S255 we estimated t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) by using the relation $$t_e({[O\,\textsc{iii}]})=\frac{t_e({[N\,\textsc{ii}]})}{1.85~-~0.72~t_e({[N\,\textsc{ii}]})} \label{eq:teOIIIconNII}$$ obtained from @PerezMontero2009. A similar scenario occurs when deriving t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) in regions in which [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$6312 and [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda\lambda$9068,9530 were not measured simultaneously (S127, S128, S158, S298, S301, and S311). For theses objects we resorted to the expression proposed by @PerezMontero2003 that relates t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) with t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}): $$t_e({[S\,\textsc{iii}]})=1.19~t_e({[O\,\textsc{iii}]})~-~0.32 . \label{eq:teSIIIconOIII}$$ Equations \[eq:teOIIconOIII\] and \[eq:teOIIIconNII\] are based on photoionisation models, while equation \[eq:teSIIIconOIII\] was obtained with observational data. Electron temperatures estimated with these methods are quoted in Table \[table:physical\_parameters\] labelled with the subscript E.\ Although the intensities of the necessary auroral lines are relatively reliable, the estimation of the electron temperature is not exempt from problems. There are several sources of uncertainty that need to be mentioned because the variations in the temperature would involve offsets in the chemical abundances. @Kennicutt2003 compare the consistency of electron temperatures measured from different ions based on observation of [H<span style="font-variant:small-caps;">ii</span>]{} regions in M101. They find that the temperatures derived from the [\[O<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$7325 line show a large scatter and are nearly uncorrelated with temperatures derived from other ions. The source of this disagreement remains unresolved, but they speculate on possible sources such as dielectronic recombination, collisional de-excitation, radiative transfer effects, and observational uncertainties. On the other hand, @Binette2012 analyse pairs of [\[O<span style="font-variant:small-caps;">ii</span>\]]{} and [\[S<span style="font-variant:small-caps;">ii</span>\]]{} temperatures in [H<span style="font-variant:small-caps;">ii</span>]{} regions finding that T$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) appears to be higher than T$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) in objects with T$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) lower than 14000 K. They perform photoionisation models to look for an explanation for this trend and conclude that models with metallicity inhomogeneities or with shock waves that propagate in the photoionised gas successfully reproduce the observed excess in T$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) temperatures.\ In order to determine the uncertainties in physical parameters and to analyse possible discrepancies between electron temperatures associated with different ions, we performed diagnostic diagrams that show t$_e$ dependences with n$_e$ and vice versa. We represented extreme values of t$_e$ and n$_e$ characterised by the flux ratios and their observational errors obtaining diagnostic bands that allow us to visualise the electron temperature structure in the ionised gas. Figure \[fig:diagdiag\] shows the diagnostic diagrams generated for the WHT regions with direct estimation of t$_e$: S83, S132, S156, S162, and S212. We show the electron density and its dependence with t$_e$ in blue. Yellow, red, and green bands represent electron temperatures independent of density, respectively t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}), t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), and t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}). Finally, we represent the single electron temperature sensitive to density, t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{}), in purple. The criterion used to paint the parameters is arbitrary and the parameters are represented giving preference to some of them following this order: t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) - t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) - t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}) - t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{}) - n$_e$. For this reason, in some regions certain temperatures are not clearly observed because they are hidden behind others. Analysing the density variations, we found that S83, S132, and S212 always show medium densities with values between 150 cm$^{-3}$ and 400 cm$^{-3}$, while S156 and S162 present higher densities with a wider range of values, from 800 cm$^{-3}$ to 2000 cm$^{-3}$, for a given temperature. In general, the agreement between electron temperatures is very good, and it is possible to appreciate the consistency both for t$_e$ ranges of different ions at estimated n$_e$ and for values presented in Table \[table:physical\_parameters\]. The electron temperature associated with [\[S<span style="font-variant:small-caps;">iii</span>\]]{} is the only one that shows lower values than the others in three regions for a fixed density; this has been taken into account for the derivation of ionic abundances, as we show in the next section. Direct chemical abundances {#sect:abundances} -------------------------- The ionic chemical abundances of the different species were estimated from the forbidden-to-hydrogen emission line ratios of the strongest available emission lines detected in the analysed spectra. We resorted to the functional forms given by @Hagele2008 which are based on the package [IONIC]{} of [IRAF]{}. To determine the single ionised helium abundance we used the equations proposed by @Olive2004. We checked that results for He$^+$/H$^+$ with this method do not differ more than $\sim$3 per cent from values estimated with recent emissivities proposed by @Porter2013. We also calculated the mean weighted by the errors of the five observed [He<span style="font-variant:small-caps;">i</span>]{} lines ([He<span style="font-variant:small-caps;">i</span>]{}$\lambda$4026, [He<span style="font-variant:small-caps;">i</span>]{}$\lambda$4471, [He<span style="font-variant:small-caps;">i</span>]{}$\lambda$5875, [He<span style="font-variant:small-caps;">i</span>]{}$\lambda$6678, and [He<span style="font-variant:small-caps;">i</span>]{}$\lambda$7065). The ionic abundances of the different elements with respect to ionised hydrogen along with their corresponding errors are given in Table \[table:abundances\]. To estimate the chemical abundances, electron density and temperature were required. We derived each ionic abundance with its corresponding electron temperature, if it was possible. In this way we estimated O$^{+}$ with t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}) , O$^{2+}$ with t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), S$^{+}$ with t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{}), S$^{2+}$ with t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}), and N$^{+}$ with t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}). In the case of argon we assumed t$_e$([\[Ar<span style="font-variant:small-caps;">iii</span>\]]{})=t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) [@Garnett1992], while for iron, helium, and neon the ionic abundances were derived under the assumption that t$_e$([\[Fe<span style="font-variant:small-caps;">iii</span>\]]{})=t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), t$_e$([He<span style="font-variant:small-caps;">i</span>]{})=t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), and t$_e$([\[Ne<span style="font-variant:small-caps;">iii</span>\]]{})=t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) [@Peimbert1969]. For those regions where not all the electron temperatures were derived we assumed an ionisation structure: the adopted criterion was to derive abundances with electron temperatures related with ions with the same excitation degree and similar ionisation potential. Nonetheless, we must bear in mind that assuming that temperatures are uniform in a given ionisation zone, our results can be subject to slight changes if we consider the presence of temperature inhomogeneities [@Esteban2009; @MesaDelgado2010; @Stasinska2013]. The temperature fluctuation paradigm was proposed by @Peimbert1967 and it is characterised by the mean square of the spatial distribution of temperature (the so-called temperature fluctuation parameter t$^2$). Our direct chemical abundances are derived using the standard method based on collisional excited lines (CELs), which have a strong dependence on T$_e$; therefore, temperature fluctuations can result in an underestimation of the absolute abundances of the ionised gas. This will be considered in the analysis of the chemical gradients.\ The total abundances were derived by taking into account, when required, the unseen ionisation stages of each element, using the appropriate ionisation correction factor (ICF) for each species. The O/H was obtained by directly adding the two ionic abundances (O/H$\sim$O$^{+}$/H$^{+}$+O$^{2+}$/H$^{+}$). The N/O can be approximated to N$^{+}$/O$^{+}$ and the N/H ratio as N/H$\sim$(N/O)$\times$(O/H). Due to the spectral range of our observations, no [\[S<span style="font-variant:small-caps;">iv</span>\]]{} line was observed, but – by taking the excitation condition into account – a relatively important contribution from S$^{3+}$ may be expected in these [H<span style="font-variant:small-caps;">ii</span>]{} regions. Therefore, the total sulfur abundance was inferred using the ICF(S$^+$+S$^{2+}$) proposed by @Barker1980. The total abundances of neon and argon were calculated using the ionisation correction factors ICF(Ne$^{2+}$) and ICF(Ar$^{2+}$), respectively, given by @PerezMontero2007. The total abundance of iron was inferred using the ICF(Fe$^{2+}$) proposed by @Rodriguez2004. Finally, in the case of helium we do not detect the [He<span style="font-variant:small-caps;">ii</span>]{}$\lambda$4686 line (belonging to the observed spectral range) in any region and we can assume that He$^{2+}$=0. On the other hand, it is expected that the contribution of the unobservable neutral helium was important. In this work, we adopted the ICF(He$^{+}$+He$^{2+}$) proposed by @Peimbert1992 to consider the He$^{0}$ contribution. In Table \[table:abundances\] we present the total direct abundances and the ICFs derived along with their corresponding errors.\ Tailor-made photoionisation models {#sect:models} ---------------------------------- For those regions without a detection of any auroral temperature-sensitive emission line (e.g. [\[N<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$5755, [\[O<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$4363) we were not able to derive chemical abundances following the direct method. Instead, it is possible to apply methods based on strong emission line ratios, such as R$_{23}$, O$_3$N$_2$, or S$_{23}$. However, most of these methods have been calibrated for massive star-forming complexes whose scales and properties do not match those in our sample of nebulae ionised by single stars in the disc of the Galaxy. As an alternative, we can resort to tailor-made photoionisation models covering the observed properties of these objects. In Table \[table:infoRHII\] we showed the spectral type of the dominant exciting stars of the sample. These spectral types were derived by different authors using different methods, and great discrepancies among authors in spectral types for the same star are found. For this reason, and in order to work with a self-consistent sample, we decided to estimate representative effective temperatures resorting to the $\eta'$ parameter instead of that derived from the spectral type in Table \[table:infoRHII\]. The spectral energy distributions (SEDs) of the ionising stars can be obtained from the observed emission line ratios using appropriate diagnostic diagrams that yield a stellar effective temperature. This is the case of the $\eta$ parameter [@Vilchez1988], defined as $$\mathrm{ \eta=\frac {O^{+}/O^{2+}} {S^{+}/S^{2+}} }, \label{eq:eta}$$ which is sensitive to the hardness of the ionising spectral energy distribution. When the ionic abundances cannot be derived, it is possible to use the $\eta'$ parameter based on emission lines: $$\mathrm{ \eta'=\frac {{[O\,\textsc{ii}]}3727/ {[O\,\textsc{iii}]}4959,5007} {{[S\,\textsc{ii}]}6717,6731/{[S\,\textsc{iii}]}9069,9532} }. \label{eq:etap}$$ Although this parameter mainly correlates with the stellar effective temperature, it also has an additional dependence on electron temperature and, hence, on chemical abundance. We then used the calibration of the $\eta'$ parameter presented by @PerezMontero2014 based on fittings to models with WM-Basic [@Pauldrach2001] single stars at a metallicity Z=0.2 Z$_\odot$ close to the expected values for this subsample. The $\eta'$ criterion limits the use of models when deriving chemical abundances to those [H<span style="font-variant:small-caps;">ii</span>]{} regions with the four involved emission lines. All the [H<span style="font-variant:small-caps;">ii</span>]{} regions with measures of [\[O<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$3727, [\[O<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$5007, [\[S<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda\lambda$6717,6731, and [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$9069 are shown along with the models used to derive the effective temperature in Fig. \[fig:teff\].\ ![[\[S<span style="font-variant:small-caps;">ii</span>\]]{}/[\[S<span style="font-variant:small-caps;">iii</span>\]]{} versus [\[O<span style="font-variant:small-caps;">ii</span>\]]{}/[\[O<span style="font-variant:small-caps;">iii</span>\]]{} in logarithm units compared with a grid of WM-Basic models at Z=0.2 Z$_\odot$. The solid lines represent models with the same effective temperature as indicated. Circles show [H<span style="font-variant:small-caps;">ii</span>]{} regions with reliable line emission measurements and triangles regions with upper limit intensities. Filled symbols represent data from our own sample (observed regions) while open symbols represent regions taken from the literature.[]{data-label="fig:teff"}](figs/Teff_etap.pdf){width="7cm"} ---------- -------------------- ------------------------ -------- ------------- ---------- [T$_{\rm eff}$ ]{} $\log$ L([H$\beta$]{}) Radius 12+log(O/H) log(N/O) (kK) (erg/s) (pc) S83$^*$ 40 34.15 0.1 8.16 -0.95 S98 41 35.02 4.4 8.16 -1.06 S127$^*$ 37 34.33 0.3 8.29 -1.34 S128$^*$ 41 34.57 0.7 8.21 -1.00 S132$^*$ 38 33.02 0.1 8.32 -0.98 S156$^*$ 36 34.11 0.1 8.26 -0.95 S162$^*$ 37 33.57 3.0 8.30 -0.86 S207 40 34.23 2.0 8.13 -1.06 S208 39 33.90 1.7 $>$8.31 $<$-0.96 S212$^*$ 39 34.76 4.1 8.15 -1.06 S219 42 33.35 3.7 $>$7.92 $<$-0.86 S228 40 34.33 1.4 8.23 -0.92 S266 42 33.21 0.6 $>$8.11 $<$-0.78 S270 42 34.53 1.1 $>$7.72 $<$-0.66 S283 42 33.17 1.0 8.20 -1.19 ---------- -------------------- ------------------------ -------- ------------- ---------- : Obtained properties from the resulting tailor-made photoionisation models.[]{data-label="table:models"} ![Upper and middle rows show a comparison between observed quantities and the found solutions of the tailor-made models, while in the last row we compare the chemical abundances obtained from models with those derived with the direct method. Red squares represent [H<span style="font-variant:small-caps;">ii</span>]{} regions with direct estimations of T$_e$ and black circles represent regions without a detection of any auroral temperature-sensitive emission line.[]{data-label="fig:models"}](figs/comp_obs_mod.pdf){width="\columnwidth"} Once the ionising SEDs were established, we used the photoionisation code Cloudy v13.03 [@Ferland2013] using as input conditions a constant density in the gas according to the value derived from the ratio of [\[S<span style="font-variant:small-caps;">ii</span>\]]{} emission lines and a dust-to-gas ratio with a standard value of 2$\times$10$^{-3}$. All models have a radiation-bounded geometry and the stopping criterion is that 99% of the H atoms are ionised. The models for each region are repeated in an iterative process varying Q(H), the inner radius, the number of hydrogen ionising photons, the abundance of all elements scaled to oxygen, and the nitrogen-to-oxygen ratio that most closely reproduces the observed [\[O<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$3727/[H$\beta$]{}, [\[O<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$5007/[H$\beta$]{}, [\[N<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$6584/[\[O<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda$3727, and the H$\beta$ luminosity. In the S208, S219, S266, and S270 regions, the [\[O<span style="font-variant:small-caps;">iii</span>\]]{}$\lambda$5007 line is not detected or it has a very weak flux, so we considered an upper limit of this line two times the uncertainty associated with the continuum at the same wavelength. The typical number of iterations before a solution is found is around 50, with the exception of S203 and S209 where no reliable solutions were found. In the former region, the relatively very high [\[O<span style="font-variant:small-caps;">iii</span>\]]{}/[H$\beta$]{} ratios is possibly due to a density-bounded geometry and in the latter to possible aperture effects. The obtained properties from the resulting models, including L([H$\beta$]{}), external radius, oxygen abundance, and N/O ratios are summarised in Table \[table:models\]. Figure \[fig:models\] shows the agreement between the four observed quantities and the found solutions. With the exception of [\[O<span style="font-variant:small-caps;">ii</span>\]]{}/[H$\beta$]{} in S98 and S283, errors in the accuracy of the adjustments are less than 1 per cent for H$\beta$ luminosity, 5 per cent for [\[O<span style="font-variant:small-caps;">iii</span>\]]{}/[H$\beta$]{} and [\[N<span style="font-variant:small-caps;">ii</span>\]]{}/[\[O<span style="font-variant:small-caps;">ii</span>\]]{}, and 10 per cent for [\[O<span style="font-variant:small-caps;">ii</span>\]]{}/[H$\beta$]{}. In S98 and S283 the observed very high value of [\[O<span style="font-variant:small-caps;">ii</span>\]]{}/[H$\beta$]{} is possibly due to some contamination from the diffuse gas. In the last row of Fig. \[fig:models\] we also show the comparison between chemical abundances obtained from models (Table \[table:models\]) with those derived with the direct method (Table \[table:abundances\]). As can be seen, the results from the two methods are in good agreement showing that the two treatments are consistent. Relation between radio and optical electron temperatures {#sect:tradio} -------------------------------------------------------- In this subsection we perform a comparative analysis of the optical and radio electron temperatures to derive a relation between them. Since there is a large sample of nebulae widely distributed across the Galactic disc with estimations of electron temperatures based on radio recombination lines (RRL) and continuum data, this relation could be used to obtain the optical temperature and to calculate chemical abundances in those cases without detectable optical auroral lines. Most of the observed emission lines in ionised nebulae are collisionally excited and their intensities depend exponentially on temperature. This temperature can be determined from appropriate line ratios. However, these ratios involve the detection and measurement of auroral lines which are intrinsically weak and remain undetected in many objects. On the other hand, most of the radio emission observed from [H<span style="font-variant:small-caps;">ii</span>]{} regions is continuum radiation produced by free-free thermal bremsstrahlung in the plasma. At high frequencies the nebular gas is optically thin, and the ratio between the brightness temperature of a RRL and that of a free-free emission continuum depends on the radio frequency and the gas temperature. Thus, the observed RRL-to-continuum ratio can be used to estimate the electron temperature of the [H<span style="font-variant:small-caps;">ii</span>]{} regions [@Rohlfs2000]. Because the RRLs are not obscured by interstellar dust, relatively faint [H<span style="font-variant:small-caps;">ii</span>]{} regions at large distances from the Sun can be detected, allowing radio electron temperatures t${_e^*}$ to be estimated more easily than with optical auroral lines.\ ![Comparison of optical electron temperatures derived from the [\[O<span style="font-variant:small-caps;">iii</span>\]]{} lines, t${_e}$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), with those derived from the radio recombination lines, t${_e^*}$. The red solid line represents the fit performed weighted by errors.[]{data-label="fig:TeTr"}](figs/Te_Tr.pdf){width="\columnwidth"} With the intention of establishing a relation to convert t${_e^*}$ into optical electron temperatures t${_e}$ which are relevant in abundance determinations, we first checked the literature for studies with reliable estimations of radio temperatures (see Col. 9 of Table \[table:infoRHII\]). Later, we compared these values with our optical electron temperatures derived from the [\[O<span style="font-variant:small-caps;">iii</span>\]]{} lines (see Col. 4 of Table \[table:physical\_parameters\]) for [H<span style="font-variant:small-caps;">ii</span>]{} regions in common. Figure \[fig:TeTr\] shows optical electron temperatures, t${_e}$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), derived in this work as a function of the radio temperatures, t${_e^*}$, adopted from the literature. Each point represents an [H<span style="font-variant:small-caps;">ii</span>]{} region with radio temperatures being a weighted mean of determinations from @Quireza2006 and @Balser2011. Despite the reduced number of regions, the obtained relation is described by the weighted fit $$\mathrm{t{_e}}(\mathrm{{[O\,\textsc{iii}]}}) =(0.175\pm 0.080)+(0.792\pm0.081)\times \mathrm{t_e^*} \label{eq:TrTe}$$ [l c c c ]{}\ & S98 & S209 & S228\ t$_e^*$ (10$^4$ K) & 1.08 $\pm$ 0.02 & 1.06 $\pm$ 0.01 & 0.94 $\pm$ 0.01\ t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) (10$^4$ K) & 1.03 $\pm$ 0.12 & 1.02 $\pm$ 0.12 & 0.92 $\pm$ 0.11\ t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}) (10$^4$ K) & 1.10 $\pm$ 0.07 & 1.09 $\pm$ 0.07 & 1.02 $\pm$ 0.07\ t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) (10$^4$ K) & 0.91 $\pm$ 0.14 & 0.89 $\pm$ 0.14 & 0.78 $\pm$ 0.13\ n$_e$ (cm$^{-3}$) & $<$100 & 423 $\pm$ 49 & 233 $\pm$ 12\ \ 12+log(O$^+$/H$^+$) & 8.23 $\pm$ 0.12 & 7.84 $\pm$ 0.12 & 8.13 $\pm$ 0.13\ 12+log(O$^{2+}$/H$^+$) & 7.06 $\pm$ 0.21 & 7.94 $\pm$ 0.17 & 7.27 $\pm$ 0.19\ 12+log(S$^+$/H$^+$) & 5.94 $\pm$ 0.08 & 5.47 $\pm$ 0.08 & 6.07 $\pm$ 0.07\ 12+log(S$^{2+}$/H$^+$) & 6.14 $\pm$ 0.18 & 6.43 $\pm$ 0.17 & 6.51 $\pm$ 0.18\ 12+log(N$^+$/H$^+$) & 7.21 $\pm$ 0.08 & 6.70 $\pm$ 0.08 & 7.26 $\pm$ 0.08\ log(N$^+$/O$^+$) & -1.01 $\pm$ 0.07 & -1.10 $\pm$ 0.08 & -0.85 $\pm$ 0.06\ (He$^+$/H$^+$) 5875 & - & 0.08 $\pm$ 0.01 & -\ (He$^+$/H$^+$) 6678 & - & 0.08 $\pm$ 0.03 & 0.09 $\pm$ 0.01\ (He$^+$/H$^+$) 7065 & - & 0.07 $\pm$ 0.02 & 0.07 $\pm$ 0.01\ (He$^+$/H$^+$) & - & 0.08 $\pm$ 0.01 & 0.08 $\pm$ 0.01\ \ ICF(S$^+$,S$^{2+}$) & 1.00 $\pm$ 0.01 & 1.11 $\pm$ 0.07 & 1.00 $\pm$ 0.01\ ICF(He$^+$) & - & 1.10 $\pm$ 0.04 & 1.36 $\pm$ 0.16\ \ 12+log(O/H) & 8.26 $\pm$ 0.12 & 8.19 $\pm$ 0.11 & 8.18 $\pm$ 0.12\ 12+log(S/H) & 6.36 $\pm$ 0.12 & 6.52 $\pm$ 0.16 & 6.65 $\pm$ 0.13\ 12+log(N/H) & 7.24 $\pm$ 0.19 & 7.05 $\pm$ 0.19 & 7.29 $\pm$ 0.19\ log(N/O) & -1.01 $\pm$ 0.07 & -1.10 $\pm$ 0.08 & -0.85 $\pm$ 0.06\ He/H & - & 0.09 $\pm$ 0.01 & 0.11 $\pm$ 0.02\ This result differs from that established by @Shaver1983, who found that the weighted average of the ratio of these temperatures is t${_e}$([\[O<span style="font-variant:small-caps;">iii</span>\]]{})/t${_e^*}$=1.06$\pm$ 0.02, with the result that the slope derived here is smaller than their value. Their relation was obtained from only six [H<span style="font-variant:small-caps;">ii</span>]{} regions located close to the Sun, but we have not found other studies in the literature that can be compared with our results.\ We applied the relation obtained to those regions of the sample without direct estimations of optical temperatures, but with information of radio temperature from the literature (S98, S209, and S228). Once t${_e}$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}) was estimated for each region using Eq. \[eq:TrTe\], we calculated temperatures associated with other ions with different ionisation degrees (t${_e}$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}) with Eq. \[eq:teOIIIconNII\] and t${_e}$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}) with Eq. \[eq:teSIIIconOIII\]). Finally, we inferred chemical abundances following the same procedure as for regions with observable auroral lines (Sect. \[sect:abundances\]). Electron temperatures, ionic abundances, and chemical abundances derived with this method are shown in Table \[table:Tr\_abundances\]. For two regions, S98 and S288, chemical abundances were also estimated resorting to the photoionisation models described in the previous section; as can be seen (Tables \[table:models\] and \[table:Tr\_abundances\]), both derivations are consistent within errors. Discussion {#sect:discussion} ========== Galactocentric distances of the [H<span style="font-variant:small-caps;">ii</span>]{} regions {#sect:distances} ---------------------------------------------------------------------------------------------- ![Distribution of the [H<span style="font-variant:small-caps;">ii</span>]{} regions analysed in this study projected onto the Galactic plane. The numbers close to the circles show distances to the Galactic centre in kpc. Red circles denote regions with abundances estimated with direct methods, black squares represent results from tailor-made models, and blue squares plot the abundances obtained from the t$_e$- t$_e^*$ relation. Regions with chemical abundances estimated with two methods are represented twice. The Sun and M42 are also represented at R$_G$=8.5 kpc (black solar symbol) and R$_G$=8.95 kpc (black cross), respectively.[]{data-label="fig:positions"}](figs/gal_plane_new.pdf){width="6cm"} The study of abundance gradients in galaxies is strongly affected by the uncertainty in galactocentric distance: it can affect the slope and, therefore, the interpretation of the results. The main difficulty in establishing the parameters shown in Cols. 4 and 5 of Table \[table:infoRHII\] was to decide what Galactocentric radius of the [H<span style="font-variant:small-caps;">ii</span>]{} regions to adopt in this work. We searched in the literature for the most accurate distances for our sample, and found several previous works which used different methods: stellar photometry, kinematical, or mixed techniques. In order to generate a self-consistent sample and to minimise the effects of inaccuracies in distances, we selected Galactic and heliocentric distances from the same catalogue for all the regions whenever possible. Taking these considerations into account, in a first attempt, we adopted the Galactocentric distances from @Caplan2000 and @Quireza2006. From these works we obtained information about R$_G$ and [d$_{\odot}$]{} for 17 [H<span style="font-variant:small-caps;">ii</span>]{} regions (77$\%$ of the sample). Both groups of authors use the same method, deriving distances kinematically from the observed nebular recombination line local standard of rest (LSR) velocity, assuming the @Brand1993 Galactic rotation curve and placing the Sun at a Galactocentric distance of [R$_{\odot}$]{}=8.5 kpc orbiting the Galactic centre at a LSR circular velocity of $\theta$=220 km s$^{-1}$. The two works complement each other well since they give distances for different regions or similar estimations for those regions listed in both works. Only in the case of S212 do @Caplan2000 and @Quireza2006 differ in the derived distances, estimating R$_G$=14.8 kpc and 16.7 kpc, respectively. We adopted the latter because it agrees with the value proposed by @Balser2011, whose estimation will be used for other regions (see below). Five regions of our sample do not have distances estimated by @Caplan2000 or @Quireza2006. For these regions, we decided to use works that derive distances with the same methodology. To this end, we adopt distances derived by @Fich1991 for S98, S266, S270, and S283 and by @Balser2011 for S98. These five regions are not individually catalogued in any other study of kinematical distances. Both @Fich1991 and @Balser2011 derive kinematic distances with [R$_{\odot}$]{}=8.5 kpc and $\theta$=220 km s$^{-1}$, although @Fich1991 use a flat rotation curve instead a curve from @Brand1993. Here we have thus chosen to use only these kinematic determinations of the R$_G$ of our sources, obtaining a sample with consistent Galactocentric radius. Figure \[fig:positions\] shows the distribution of the 22 [H<span style="font-variant:small-caps;">ii</span>]{} regions with derived chemical abundances projected onto the Galactic plane and plotted as function of the Galactic longitude and Galactic radius. They cover the anticentre range of distances from $\sim$11 kpc to $\sim$18 kpc. Abundance gradients of O/H, N/H, N/O, S/H, Ar/H, and He/H towards the Galactic anticentre {#sect:gradients} ----------------------------------------------------------------------------------------- [l c ]{}\ Data used & Final abundance gradients\ \ All data & (9.006 $\pm$ 0.112) - (0.053 $\pm$ 0.009) R$_G$\ Direct & (9.113 $\pm$ 0.139) - (0.061$\pm$ 0.011 ) R$_G$\ \ All data &(8.260 $\pm$ 0.258) - (0.080 $\pm$ 0.019) R$_G$\ Direct & (8.331 $\pm$ 0.274 ) - (0.085 $\pm$ 0.021) R$_G$\ \ All data & - (0.478 $\pm$ 0.086) - (0.041 $\pm$0.006) R$_G$\ Direct & - (0.399 $\pm$ 0.095) - (0.047 $\pm$ 0.007) R$_G$\ \ All data &(8.162 $\pm$ 0.088) - (0.106 $\pm$ 0.006) R$_G$\ Direct &(8.194 $\pm$ 0.089) - (0.108 $\pm$ 0.006) R$_G$\ \ All data &(7.178 $\pm$ 0.073) - (0.074 $\pm$ 0.006) R$_G$\ \ All data &(0.0968 $\pm$ 0.0259) - (0.0005 $\pm$ 0.0019) R$_G$\ Direct &(0.0982 $\pm$ 0.0299) - (0.0007 $\pm$ 0.0022) R$_G$\ From the derived abundances shown in Tables \[table:abundances\], \[table:models\], and \[table:Tr\_abundances\], we can study the radial distribution of chemical abundances along the Galactic anticentre. Figures \[fig:OHgrad\] through \[fig:HeHgrad\] show the total chemical abundances of O/H, N/H, N/O, S/H, Ar/H, and He/H, respectively, plotted against Galactocentric distances for the [H<span style="font-variant:small-caps;">ii</span>]{} regions with chemical abundances inferred in this work. To make it easier to distinguish between the three methods used to determine abundances, the regions with direct T$_e$ are plotted as red circles, results from tailor-made models are represented with black stars, and abundances obtained from the t$_e$-t$_e^*$ relation are shown as blue squares. Regions with abundances that are upper or lower limits are represented by arrows in the figures, but they have not been taken into account in any fitting. To determine linear gradients we performed a least-squares fitting weighted by abundance uncertainties. Data from the three methods (except limit abundances) were combined into a single fit and plotted as solid lines in the figures. We also computed gradients using only abundances derived for the [H<span style="font-variant:small-caps;">ii</span>]{} regions with direct determinations of T$_e$; however, for simplicity, they are not represented in the figures. The final abundance gradients are presented in Table \[table:gradientes\]. Regions with chemical abundances derived with two methods (S98 and S228) are represented twice in the figures, but are only consider once in the performed fits; as the representative value we took the weighted mean by errors of the two methods. In addition, it is interesting to note that those abundances estimated using the electron temperature derived from the relation between optical and radio temperatures (blue squares) follow the same tendency as abundances derived with the direct method and, in general, they agree with the fits obtained. To compare our results with other works, in Table \[table:grad\_literature\] we present a summary of all the gradients obtained by different authors from [H<span style="font-variant:small-caps;">ii</span>]{} regions in the inner and outer Galaxy in several spectral ranges. The extrapolated radial gradients obtained by @Rudolph2006 with optical data are also represented in graphics as dashed lines. These come from a selection of previous works for a large range of Galactocentric distances shown here for the sake of comparison with a suggestion of flattening in the outer parts of the disc.\ ![Radial distribution of 12+log(O/H) (in dex) plotted versus Galactocentric radius (R$_G$ in kpc). Red circles denote regions with abundances estimated with direct methods, black squares represent results from tailor-made models, and blue squares plot the abundances obtained from t$_e$- t$_e^*$ relation. Regions with abundances that are upper or lower limits are represented by arrows. The solid line represents the least-squares weighted fit performed to all the data (except limits abundances), while the extrapolated gradient from @Rudolph2006 is represented by a dashed line.[]{data-label="fig:OHgrad"}](figs/OHgrad.pdf){width="\columnwidth"} ![Radial distribution of 12+log(N/H) plotted versus Galactocentric radius. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:NHgrad"}](figs/NHgrad.pdf){width="\columnwidth"} The radial distribution of the oxygen abundance is presented in Fig. \[fig:OHgrad\] and it clearly shows a decrease with Galactocentric distance. Within the range of distances between 11 and 18 kpc, the maximum variation goes from 12+log(O/H)=8.44 to 12+log(O/H)=8.05. The results obtained by computing the fit with all the regions and by using oly the abundances derived with the direct method (Table \[table:abundances\]) are consistent, taking the associated uncertainties into account; the slope of the latter is slightly steeper. When comparing our results with those for the inner parts of the Galaxy, we find that our slope is clearly shallower than the one obtained by @Shaver1983, but in general their fits have steeper slopes than many others in the literature. In addition, our result is in very good agreement with the gradient obtained by @Afflerbach1997 with infrared data. The slope obtained by @Esteban2005 is smoother than ours, even though they performed the study of gradients based on recombination lines instead of collisional lines. On the other hand, our gradient is in good agreement with those performed in a wider range of distances (0 to18 kpc) and, as can be seen in Fig. \[fig:OHgrad\], our slope agrees with @Rudolph2006. Gradients obtained by @Deharveng2000, @Quireza2006, and @Balser2011 are slightly shallower than ours, but are consistent within errors. @Vilchez1996 are the only authors who have focused their study on the outer part of the Galaxy. They provide three different oxygen gradients, and our slope agrees with the one obtained from all of their data (third option in Table \[table:grad\_literature\]), but not with the other fits.\ Figure \[fig:NHgrad\] shows the radial distribution of the final N/H abundances. The derived nitrogen abundances present values between 12+log(N/H)=7.44 and 6.88 and their distribution shows a clear decrease with distance. Observed values follow closely the performed fit, although those derived from radio temperatures (blue squares) are slightly higher, but still in agreement within measured errors. The resulting fits, for all the regions and for direct abundances alone, are in agreement, taking the errors into account. Within the uncertainties this result for the nitrogen gradient is not different from those found by @Shaver1983, @Simpson1995, @Rudolph1997, @Afflerbach1997, and @Rudolph2006. However, our slope differs from that obtained by @Vilchez1996 for a similar range of R$_G$: we do not see any clear steepening of nitrogen in the outer part of the Galaxy disc.\ ![Radial distribution of log(N/O) plotted versus Galactocentric radius. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:NOgrad"}](figs/NOgrad.pdf){width="\columnwidth"} ![Radial distribution of 12+log(S/H) plotted versus Galactocentric radius. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:SHgrad"}](figs/SHgrad.pdf){width="\columnwidth"} Figure \[fig:NOgrad\] shows the behaviour of N/O versus Galactocentric distance for the studied regions. The derived values of N/O range from log(N/O)=-0.87 to -1.19. In this case almost no difference exists between the slope derived from directly derived abundances and the value considering also abundances derived using model fits, since the latter have much higher associated errors and they thus have a lower weight in the fit. Our slope is consistent with the value obtained by @Simpson1995 from infrared observation in the inner parts of the Galaxy, but differs from @Rudolph2006 who found that N/O is nearly constant at all distances. The fit of the data shown in Fig. \[fig:NOgrad\] presents a decreasing slope in N/O with Galactocentric distance, although a step function is also plausible (see next section).\ The sulphur radial gradient is plotted in Fig. \[fig:SHgrad\]. The derived sulfur abundances present values between 12+log(S/H)=7.01 and 6.37 along the range of Galactocentric distances, with the exception of the high S/H value of S311 (which has the lowest value of t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{})) and the low estimation of S/H in S98. The distribution shows a negative radial gradient with a slope that is much steeper than those derived by other authors. This may be a consequence of the large weight of regions located at large Galactic radius whose abundances have very low errors.\ ![Radial distribution of 12+log(Ar/H) plotted versus Galactocentric radius. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:ArHgrad"}](figs/ArHgrad.pdf){width="\columnwidth"} In Fig. \[fig:ArHgrad\], we present the argon gradient along the Galactocentric anticentre. The distribution ranges from 12+log(Ar/H)=6.4 to 5.9, decreasing clearly with distance. In this case all the abundances represented were derived using the direct method (direct measures of electron temperature for all the regions) and we only show one fit for the gradient in Table \[table:gradientes\]. This is the first time that the distribution of argon towards the Galactic anticentre is shown. @Shaver1983 derived a gradient of d(log(Ar/H))/dR$_G$=-0.06$\pm$0.015 in the inner part of the Galaxy, which is in good agreement with our slope, taking the fit uncertainties into account.\ ![Radial distribution of He/H plotted versus Galactocentric radius. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:HeHgrad"}](figs/HeHgrad.pdf){width="\columnwidth"} Finally, Fig. \[fig:HeHgrad\] shows the radial distribution of helium abundances. The derived values of He/H cluster around He/H$\sim$0.09 no matter their Galactocentric distances. There is no obvious gradient and the least-squares fit to the data gives a slope that is not significantly different from zero. Trying to compare our results with those of other authors, we find that no attempts have been made previously to estimate the total helium radial abundance gradient of the Galaxy, only the distribution of He$^+$/ H$^+$ has been studied before with slopes also close to zero [@Peimbert1978; @Talent1979; @Shaver1983; @Deharveng2000]. Chemical composition of the outer Galaxy {#sect:distribution} ---------------------------------------- The knowledge of the chemical distribution across the Galactic disc is a benchmark for models of chemical evolution. For this reason, considerable effort has been made to establish the abundance gradients in the MW by studying the distribution of many elements and many sources which all probe different epochs in the MW chemical evolution history. Figure \[fig:gradOHgrande\] provides comparisons of our O/H abundance gradient derived from the total [H<span style="font-variant:small-caps;">ii</span>]{} region sample towards the anticentre with many other gradients derived from studies with PNe, cepheids, B-type stars, and [H<span style="font-variant:small-caps;">ii</span>]{} regions in a wide range of Galactic distances (0-20 kpc). The gradients adopted for the different objects are shown in Table \[table:grad\_sources\] and are those provided directly by the authors. Moreover, oxygen abundances derived in the Large and Small Magellanic Clouds [@Russell1990], Orion nebula [@Tsamis2011], and the Sun [@Asplund2009] are also represented in Fig. \[fig:gradOHgrande\]. B-type stars are massive young stars ($<$10 Myr) and their abundances trace the metallicity near their current location. @Rolleston2000 studied a sample of about 80 early B-type main sequence stars located from 6 to 18 kpc. They conclude that the distribution of oxygen can be represented by a linear gradient with no evidence of a better fitting with a two-zone model. As we can see in Fig. \[fig:gradOHgrande\] and Table \[table:grad\_sources\], the results from @Rolleston2000 are the most different to ours, with a much steeper slope and the largest ordinate at origin. Planetary nebulae trace metallicity in the Galactic disc, bulge, and halo. Their ages span 1-8 Gyr and thus PNe can in principle probe metallicity with time. We compare our results with those obtained from @Henry2010 who determined chemical abundances of 124 PNe extending from 0.9 to 21 kpc in Galactocentric distances. We find an agreement in slope with their gradient as we can see in Fig. \[fig:gradOHgrande\] and Table \[table:grad\_sources\]. They found some evidence that the gradient beyond 10 kpc is steeper than it is inside this distance. Cepheids are evolved stars that can be detected at large distances and are relatively young ($<$200Myrs), hence they are primary tracers of metals in the Galactic disc. We compare our gradient with that obtained by the recent work of @Korotin2014 and find a good agreement between the two slopes. They also found some evidence that the distribution might become flatter in the outer parts of the disc. Galactic [H<span style="font-variant:small-caps;">ii</span>]{} regions are the formation sites of massive OB stars and they reveal the location of current Galactic star formation. In Fig. \[fig:gradOHgrande\] we also represent the gradient obtained by @Rudolph2006, considering it as a collection of previous studies with [H<span style="font-variant:small-caps;">ii</span>]{} regions. Their results are consistent with ours, taking the uncertainties into account.\ ![Extrapolated gradients of 12+log(O/H) along the Galactic disc for different sources. Black circles represent all the [H<span style="font-variant:small-caps;">ii</span>]{} regions studied in this work and the solid black line indicates the gradient obtained with them. Gradients derived from PNs (long dashed red line), [H<span style="font-variant:small-caps;">ii</span>]{} regions (dash-dotted blue line), B-type stars (dash-dot-dotted pink line) and Cepheids (short-dashed green line) are also represented. The Sun and the Orion nebula (M42) are shown at their respective O/H abundances located at R$_G$ =8.5 kpc and R$_G$=8.95 kpc, respectively. Horizontal arrows show chemical abundances of the Small and Large Magellanic Clouds.[]{data-label="fig:gradOHgrande"}](figs/OHcomp.pdf){width="\columnwidth"} [l c c c ]{}\ Author(s) & Object & d(log(O/H))/dR$_G$ & $\Delta$R$_G$\ & type & (dex/kpc$^{-1}$) & (kpc)\ This work & [H<span style="font-variant:small-caps;">ii</span>]{} regions & -0.053 & 11-18\ @Rolleston2000 & B stars & -0.067 & 6-18\ @Rudolph2006 & [H<span style="font-variant:small-caps;">ii</span>]{} regions & -0.060 & 5-18\ @Henry2010 & PNe & -0.058 & 0.9-21\ @Korotin2014 & Cepheids & -0.058 & 5-18\ What do we conclude from these studies? The existence of those gradients offers the opportunity to compare with models of disc evolution to understand the formation and chemical evolution of our Galaxy. All the results obtained from different sources present a negative radial gradient (although with significant differences in their slopes), thus metal abundances in the inner disc are higher than those in the outer disc. This result supports the Inside-Out formation scheme of the disc [@Hou2000; @Alibes2001; @Chiappini2003]. The inner part of the disc always forms first, and the outer part forms progressively later as gas with higher angular momentum settles into the equatorial plane at larger radii, causing the disc to grow outward with time. The metal abundance of the innermost part of the disc is higher because of the prior enrichment of the infalling gas before it reaches the equatorial plane, and the metal abundances of the outermost part of the disc is smaller because of the relatively long time scale for star formation and metal enrichment. Therefore, the negative gradients obtained in this work and results from other sources are consistent with Inside-Out models; the differences in slopes can be explained by the amount of material, yields, and star formation rates adopted by the model (e.g. @Molla2015).\ As far as the shape of the gradient is concerned, the possibility of variations in the slope towards the outer disc is still widely debated. Several works (not represented in Fig. \[fig:gradOHgrande\] for clarity) based on different tracers (cepheids, open clusters, PNe, and [H<span style="font-variant:small-caps;">ii</span>]{} regions) have found evidence that the radial abundance gradient of some elements may flatten out at the outer part of the Galaxy disc [@Fich1991; @Vilchez1996; @Maciel1999; @Andrievsky2004; @Costa2004; @Luck2006; @Carraro2007; @Pedicelli2009; @Henry2010; @Andreuzzi2011; @Korotin2014]. In addition, strong evidence for flattening at large distances of discs of other spiral galaxies has been presented in several papers [@vanZee1998; @Goddard2011; @Werk2011; @Bressolin2012]. Our sample is located towards the anticentre; in order to analyse the radial behaviour of the abundance gradient in the outermost disc, we need to make a comparison with literature gradients performed in wider radial ranges including the inner parts of the Galaxy. When doing so, our data suggest a shallower slope. Although this result does not necessarily imply a flattening of the gradient in the outer disc, what is clear is that the simple extrapolation of the inner slope towards the anticentre is not consistent with the observations. ![Radial distribution of N/O plotted versus Galactocentric radius. Black points represent N/O abundances obtained with the direct method. The two lines represent the fit to data in two distances ranges: within 12 kpc from the Galactic centre (blue dot-dashed line) and farther than this limit (red dashed line), see text for details. The Orion nebula, the Sun and Magellanic Clouds are represented as in Fig. \[fig:gradOHgrande\].[]{data-label="fig:gradNOgrande"}](figs/NOdoble.pdf){width="\columnwidth"} From the theoretical point of view, several chemical evolution models have been computed to reproduce the flattening observed. However, authors suggest different physical mechanisms to explain this behaviour: lower effective yields in the outer disc than in the inner disc [@Twarog1997], effective suppression of mixing processes near the corotation circle [@Andrievsky2002; @Luck2003; @Andrievsky2004], constant density distribution of the halo stellar component in the inner 20 kpc [@Cescutti2007], levelling out of the star formation efficiency about and beyond the isophotal radius [@Esteban2013], etc. More information about this topic can be found in @Bressolin2012, who discuss several mechanisms that can be responsible for the flattening in the outermost disc of spiral galaxies.\ The N/O abundance ratio is a useful indicator of the chemical age of the galaxy and how evolved the disc is. An interesting feature suggested by Fig. \[fig:NOgrad\] is that our N/O gradient seems to be better represented by a step function instead of a linear fit. We compute a new double fit using only those N/O abundances derived from the direct method (Table \[table:abundances\]) and differentiating those regions placed within 12 kpc from the centre and farther than this limit. The result for R$_G<$12 kpc is log(N/O) = - (0.7264 $\pm$ 1.3963) + (0.0152 $\pm$ 0.1240) R$_G$, while for R$_G>$12kpc we obtain log(N/O) = - (0.8993 $\pm$ 0.1689) + (0.0150 $\pm$ 0.0116) R$_G$. Figure \[fig:gradNOgrande\] shows the radial distribution of direct N/O abundances within the two fits performed with data at R$_G<$12 kpc (blue dot-dashed line) and R$_G>$12 kpc (red dashed line). As can be seen, both data sets are represented by lines with similar slopes (nearly constant) but different origins as the N/O ratio is uniform in two distance ranges. We conclude that the N/O distribution with Galactic distance for the disc can be better described by a step function rather than a linear gradient. Furthermore, with this double fit the statistical dispersion is obviously smaller than with the single fit. ![Top panel: Relation between nitrogen and oxygen abundances. Bottom panel: Relative abundance, log(N/O), versus the oxygen abundance. Symbols and colours are as in Fig. \[fig:OHgrad\].[]{data-label="fig:compNO"}](figs/Ndisc.pdf){width="\columnwidth"} From the point of view of chemical evolution, the origin of nitrogen is still a matter of important debate because it can have a primary or secondary origin. The overall picture of the chemical composition of galaxies seems to indicate that there are two main behaviours in the N/O versus O/H relationship depending on metallicity: a flat line with log(N/O)=-1.46 [@Garnett1990] for the data of low-mass and irregular galaxies with 12+log(O/H)$<$8.3, and a strong slope for the N/O data against oxygen abundance for metal-rich regions. The first case reflects the proportion of the true primary production of nitrogen (independent of oxygen abundance), which is believed to be synthesised in massive stars. The second situation explains the existence of secondary nitrogen (proportional to oxygen abundance) processed in the CNO burning during the stellar evolution and returned into the ISM via PNe or stellar winds of massive stars (e.g. @Edmunds1990, @Molla2006). In Fig. \[fig:compNO\] we analyse the behaviour of nitrogen abundances (top panel) and N/O ratio (bottom panel) versus oxygen abundances for the regions located towards the anticentre studied in this work. As can be seen, nitrogen follows oxygen and the N/O ratio is constant with O/H, suggesting that nitrogen has a primary origin. However, our data show that in this zone log(N/O)$>$-1.46, which is indicative of a previous enrichment by evolved intermediate mass stars (the secondary nitrogen component). In this case, the fact that N/O is constant informs us that all regions have a similar chemical age(understanding age in the sense of the time since the bulk of its star formation occurred). The observed N/O values and radial behaviour in the outer parts of the disc compared to the solar neighbourhood can be understood if we are witnessing a special time in the chemical evolution of the outer MW; in the @Edmunds1978 chemical age scenario this can be a result of the increase in nitrogen abundance from intermediate-mass stars. When comparing our nitrogen distribution with the abundances in the Magellanic Clouds, it can be seen that our N/O ratio appears larger than the value for the SMC and LMC, log(N/O)= -1.58 $\pm$ 0.16 and log(N/O)= -1.30 $\pm$ 0.30, respectively [@Russell1990], for a range in log(O/H) comparable to the abundances measured in the outer disc; this is in line with the chemical ageing scenario mentioned before for the outer disc of the MW. Summary and conclusions {#sect:conclusions} ======================= The abundances of heavy elements in the ISM provide a fossil record of the enrichment which has taken place due to the nucleosynthesis in successive generation stars. The study of the radial variations of metallicity across the Galactic disc is a powerful method for understanding the history of star formation and chemical evolution of the MW. The primary focus of this paper was to give an accurate description of the radial chemical distribution towards the Galactic anticentre up to a distance of 18 kpc. We have carried out new optical spectroscopic observations of nine [H<span style="font-variant:small-caps;">ii</span>]{} regions located towards the Galactic anticentre using the ISIS spectrograph at the WHT. To realise a more extensive study, the sample was increased by searching for optical [H<span style="font-variant:small-caps;">ii</span>]{} regions at R$_G>$11 kpc in the literature. The complete sample comprised 23 [H<span style="font-variant:small-caps;">ii</span>]{} regions extending in Galactocentric radius from 11 kpc to 18 kpc. For 13 regions of the sample we have derived accurate electron densities from the [\[S<span style="font-variant:small-caps;">ii</span>\]]{}$\lambda\lambda$6717,6731 line ratio and electron temperatures t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{}), t$_e$([\[S<span style="font-variant:small-caps;">iii</span>\]]{}), t$_e$([\[N<span style="font-variant:small-caps;">ii</span>\]]{}), t$_e$([\[O<span style="font-variant:small-caps;">ii</span>\]]{}), and t$_e$([\[S<span style="font-variant:small-caps;">ii</span>\]]{}). These physical parameters have been used to determine direct total chemical abundances. We have also performed tailor-made photoionisation models to derive chemical abundances for those regions without direct estimations of temperature. In addition, by comparing radio recombination temperatures from the literature and optical temperatures from this work, we have obtained the empirical relation t$_e$([\[O<span style="font-variant:small-caps;">iii</span>\]]{})=(0.175$\pm$0.080)+(0.792$\pm$0.081)$\times$t$_e^*$. This relation has been applied to three regions without direct temperature to obtain their physical parameters and abundances.\ To study the radial gradients of metallicity across the outermost part of the MW, we have performed weighted least-squares fits to the distribution of the O/H, N/H, N/O, S/H, Ar/H, and He/H abundances along the Galactocentric distances. We have concluded the following: 1. The radial distribution of oxygen abundance is fit well by a regression whose form is 12+log(O/H)=(9.006$\pm$0.112)-(0.053$\pm$0.009)R$_G$, where R$_G$ is the Galactocentric radius in kpc. When comparing our result with gradients performed with several objects in wider radial ranges, including the inner parts of the Galaxy, we find a suggestion for a shallower slope. Our data cannot be used to confirm the flattening of the abundance gradient towards the outer disc suggested previously, but clearly show that a linear extrapolation of the inner slope is inconsistent with the observations 2. The derived gradient for N/H abundances shows a clear decrease with distance with a slope of -0.080$\pm$0.019 dex kpc$^{-1}$. This result is in good agreement with previous works. We do not see any steepening of nitrogen in the outer part of Galaxy disc. 3. In the case of the N/O ratio, the radial distribution is better described by a step function rather than a linear gradient. The double-fit performed shows different behaviours for regions located beyond 12 kpc than for those with R$_G<$12 kpc, but both with constant slopes. The radial behaviour in the outer parts of the disc compared to the solar neighbourhood can be understood if we are witnessing a special time in the chemical evolution of the outer MW, which can be a result of the increase in nitrogen abundance from intermediate-mass stars. 4. The fit obtained for the distribution of sulphur abundance is 12+log(S/H)=(8.162$\pm$0.088)-(0.106$\pm$0.006)R$_G$. This gradient shows a slope that is much steeper than those derived by other authors, probably reflecting the more accurate values from our data at large Galactic distances. 5. The argon gradient has been derived for the first time in regions towards the Galactic anticentre with the functional form 12+log(Ar/H)=(7.178$\pm$0.073)-(0.074$\pm$0.006)R$_G$. This result is in good agreement with the value obtained in the inner part of the disc. 6. The total helium radial abundance gradient has been also presented here for the first time. The least-squares fit to the data gives a slope that is not significantly different from zero. This work has been partially funded by the project AYA2013-47742-C4-1. A.M. acknowledges funding from the Spanish AYA2007-66804 and AYA2012-35330 grants. We thank D. Díaz-Fraile, A. Sota, J. Blasco, and the ESTALLIDOS collaboration for their useful comments and scientific support. Afflerbach, A., Churchwell, E., & Werner, M. W. 1997, , 478, 190 Alib[é]{}s, A., Labay, J., & Canal, R. 2001, , 370, 1103 Andreuzzi, G., Bragaglia, A., Tosi, M., & Marconi, G. 2011, , 412, 1265 Andrievsky, S. M., Kovtyukh, V. V., Luck, R. E., et al. 2002, , 392, 491 Andrievsky, S. M., Luck, R. E., Martin, P., & L[é]{}pine, J. R. D. 2004, , 413, 159 Asplund, M., Grevesse, N., Sauval, A. J., & Scott, P. 2009, , 47, 481 Balser, D. S., Rood, R. T., Bania, T. M., & Anderson, L. D. 2011, , 738, 27 Barker, T. 1980, , 240, 99 Binette, L., Dopita, M. A., Dodorico, S., & Benvenuti, P. 1982, , 115, 315 Binette, L., Matadamas, R., H[ä]{}gele, G. F., et al. 2012, , 547, A29 Blitz, L., Fich, M., & Stark, A. A. 1982, , 49, 183 Brand, J., & Blitz, L. 1993, , 275, 67 Bresolin, F., Kennicutt, R. C., & Ryan-Weber, E. 2012, , 750, 122 Caplan, J., Deharveng, L., Pe[ñ]{}a, M., Costero, R., & Blondel, C. 2000, , 311, 317 Carraro, G., Geisler, D., Villanova, S., Frinchaboy, P. M., & Majewski, S. R. 2007, , 476, 217 Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245 Cescutti, G., Matteucci, F., Fran[ç]{}ois, P., & Chiappini, C. 2007, , 462, 943 Chiappini, C., Romano, D., & Matteucci, F. 2003, , 339, 63 Chini, R., & Wink, J. E. 1984, , 139, L5 Conti, P. S., & Leep, E. M. 1974, , 193, 113 Costa, R. D. D., Uchida, M. M. M., & Maciel, W. J. 2004, , 423, 199 Crampton, D., & Fisher, W. A. 1974, Publications of the Dominion Astrophysical Observatory Victoria, 14, 283 Crampton, D., Georgelin, Y. M., & Georgelin, Y. P. 1978, , 66, 1 De Robertis, M. M., Dufour, R. J., & Hunt, R. W. 1987, , 81, 195 Deharveng, L., Pe[ñ]{}a, M., Caplan, J., & Costero, R. 2000, , 311, 329 Edmunds, M. G. 1990, , 246, 678 Edmunds, M. G., & Pagel, B. E. J. 1978, , 185, 77P Esteban, C., Vilchez, J. M., Manchado, A., & Edmunds, M. G. 1990, , 227, 515 Esteban, C., Garc[í]{}a-Rojas, J., Peimbert, M., et al. 2005, , 618, L95 Esteban, C., Bresolin, F., Peimbert, M., et al. 2009, , 700, 654 Esteban, C., Carigi, L., Copetti, M. V. F., et al. 2013, , 433, 382 Ferland, G. J., Porter, R. L., van Hoof, P. A. M., et al. 2013, , 49, 137 Fich, M., & Blitz, L. 1984, , 279, 125 Fich, M., & Silkey, M. 1991, , 366, 107 Garnett, D. R. 1990, , 363, 142 Garnett, D. R. 1992, , 103, 1330 Georgelin, Y. P., & Georgelin, Y. M. 1970, , 6, 349 Georgelin, Y. M., Lortet-Zuckermann, M. C., & Monnet, G. 1975, , 42, 273 Goddard, Q. E., Bresolin, F., Kennicutt, R. C., Ryan-Weber, E. V., & Rosales-Ortega, F. F. 2011, , 412, 1246 Guesten, R., & Mezger, P. G. 1982, Vistas in Astronomy, 26, 159 H[ä]{}gele, G. F., D[í]{}az, [Á]{}. I., Terlevich, E., Terlevich, R., P[é]{}rez-Montero, E., & Cardaci, M. V. 2008, , 383, 209 Hawley, S. A. 1978, , 224, 417 Henry, R. B. C., Kwitter, K. B., Jaskot, A. E., et al. 2010, , 724, 748 Hou, J. L., Prantzos, N., & Boissier, S. 2000, , 362, 921 Hunter, D. A. 1992, , 79, 469 Hunter, D. A., & Massey, P. 1990, , 99, 846 Kennicutt, R. C., Jr., Bresolin, F., & Garnett, D. R. 2003, , 591, 801 Korotin, S. A., Andrievsky, S. M., Luck, R. E., et al. 2014, , 444, 3301 Luck, R. E., Gieren, W. P., Andrievsky, S. M., et al. 2003, , 401, 939 Luck, R. E., Kovtyukh, V. V., & Andrievsky, S. M. 2006, , 132, 902 Maciel, W. J., & Koppen, J. 1994, , 282, 436 Maciel, W. J., & Quireza, C. 1999, , 345, 629 Mampaso, A. 1991,  PhD Thesis,  Universidad de La Laguna Mayor, M., & Vigroux, L. 1981, , 98, 1 Mesa-Delgado, A., & Esteban, C. 2010, , 405, 2651 Moffat, A. F. J., Jackson, P. D., & Fitzgerald, M. P. 1979, , 38, 197 Moll[á]{}, M., V[í]{}lchez, J. M., Gavil[á]{}n, M., & D[í]{}az, A. I. 2006, , 372, 1069 Moll[á]{}, M., Cavichia, O., Gavil[á]{}n, M., & Gibson, B. K. 2015, , 451, 3693 Olive, K. A., & Skillman, E. D. 2004, , 617, 29 Pauldrach, A. W. A., Hoffmann, T. L., & Lennon, M. 2001, , 375, 161 Pedicelli, S., Bono, G., Lemasle, B., et al. 2009, , 504, 81 Peimbert, M. 1967, , 150, 825 Peimbert, M., & Costero, R. 1969, Boletin de los Observatorios Tonantzintla y Tacubaya, 5, 3 Peimbert, M., Rayo, J. F., & Torres-Peimbert, S. 1978, , 220, 516 Peimbert, M., Torres-Peimbert, S., & Ruiz, M. T. 1992, , 24, 155 P[é]{}rez-Montero, E., & Contini, T. 2009, , 398, 949 P[é]{}rez-Montero, E., & D[í]{}az, A. I. 2003, , 346, 105 P[é]{}rez-Montero, E., H[ä]{}gele, G. F., Contini, T., & D[í]{}az, [Á]{}. I. 2007, , 381, 125 P[é]{}rez-Montero, E., Monreal-Ibero, A., Rela[ñ]{}o, M., et al. 2014, , 566, AA12 Phillipps, S., & Edmunds, M. G. 1991, , 251, 84 Porter, R. L., Ferland, G. J., Storey, P. J., & Detisch, M. J. 2013, , L91 Quireza, C., Rood, R. T., Bania, T. M., Balser, D. S., & Maciel, W. J. 2006, , 653, 1226 Rodr[í]{}guez, M., & Rubin, R. H. 2004, Recycling Intergalactic and Interstellar Matter, 217, 188 Rohlfs, K., & Wilson, T. L. 2000, Tools of radio astronomy / K. Rohlfs, T.L. Wilson. New York : Springer, 2000. (Astronomy and astrophysics library,ISSN0941-7834) Rolleston, W. R. J., Smartt, S. J., Dufton, P. L., & Ryans, R. S. I. 2000, , 363, 537 Rudolph, A. L., Brand,J., de Geus, E. J., & Wouterloot, J. G. A. 1996, , 458, 653 Rudolph, A. L., Simpson, J. P., Haas, M. R., Erickson, E. F., & Fich, M. 1997, , 489, 94 Rudolph, A. L., Fich, M., Bell, G. R., et al. 2006, , 162, 346 Russeil, D., Adami, C., & Georgelin, Y. M. 2007, , 470, 161 Russell, S. C., & Dopita, M. A. 1990, , 74, 93 Searle, L. 1971, , 168, 327 Sharpless, S. 1959, , 4, 257 Shaver, P. A., McGee, R. X., Newton, L. M., Danks, A. C., & Pottasch, S. R. 1983, , 204, 53 Shaw, R. A., & Dufour, R. J. 1995, , 107, 896 Simpson, J. P., Colgan, S. W. J., Rubin, R. H., Erickson, E. F., & Haas, M. R. 1995, , 444, 721 Stasi[ń]{}ska, G., Morisset, C., Sim[ó]{}n-D[í]{}az, S., et al. 2013, , 551, A82 Storey, P. J., & Hummer, D. G. 1995, , 272, 41 Talent, D. L., & Dufour, R. J. 1979, , 233, 888 Tsamis, Y. G., Walsh, J. R., V[í]{}lchez, J. M., & P[é]{}quignot, D. 2011, , 412, 1367 Twarog, B. A., Ashman, K. M., & Anthony-Twarog, B. J. 1997, , 114, 2556 van Zee, L., Salzer, J. J., Haynes, M. P., O’Donoghue, A. A., & Balonek, T. J. 1998, , 116, 2805 V[í]{}lchez, J. M., & Esteban, C. 1996, , 280, 720 Vilchez, J. M., & Pagel, B. E. J. 1988, , 231, 257 Werk, J. K., Putman, M. E., Meurer, G. R., & Santiago-Figueroa, N. 2011, , 735, 71 [l c c c c c c c c c c c]{} \[table:sample\_intensities\] &&& Line & $\lambda$ ([Å]{})& f($\lambda$)& S83& S132& S156& S162& S207& S208& S212& S228& S270\ \ \ &&&\ Line & $\lambda$ ([Å]{})& f($\lambda$)& S83& S132& S156& S162& S207& S208& S212& S228& S270\ \ \ Continues.\ [\[O<span style="font-variant:small-caps;">ii</span>\]]{}& 3727.43 & 0.322 & 908.8 $\pm$ 87.0 & 4119.4 $\pm$ 393.2 & 3015.6 $\pm$ 244.0 & 2572.5 $\pm$ 85.3 & 4497.0 $\pm$ 331.0 & 3375.0 $\pm$ 270.4 & 2959.8 $\pm$ 204.3 & 3673.2 $\pm$ 311.8 & 2328.3 $\pm$ 114.1 $\dagger$\ H13 & 3734.37 & 0.320 & - & - & - & 11.1 $\pm$ 2.0 & - & - & - & - & -\ H12 & 3750.15 & 0.317 & - & 42.1 $\pm$ 8.6 & 37.5 $\pm$ 3.2 & 32.6 $\pm$ 1.8 & - & - & 26.8 $\pm$ 2.8 $\dagger$ & - & -\ H11 & 3770.63 & 0.313 & - & 51.7 $\pm$ 8.1 & 42.5 $\pm$ 3.8 & 36.6 $\pm$ 1.3 & - & - & 37.4 $\pm$ 4.7 $\dagger$ & - & -\ H10 & 3797.90 & 0.307 & - & 63.3 $\pm$ 7.3 & 58.3 $\pm$ 4.8 & 51.7 $\pm$ 1.7 & - & - & 39.4 $\pm$ 4.0 $\dagger$ & - & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 3819.61 & 0.302 & - & - & 7.0 $\pm$ 0.9 & 11.9 $\pm$ 0.8 & - & - & - & - & -\ H9+HeI & 3834.48 & 0.299 & - & 86.6 $\pm$ 9.8 & 82.2 $\pm$ 6.4 & 82.8 $\pm$ 2.3 & - & - & 94.8 $\pm$ 9.3 & - & -\ [\[Ne<span style="font-variant:small-caps;">iii</span>\]]{}& 3869.35 & 0.291 & 560.4 $\pm$ 64.2 & - & 13.9 $\pm$ 1.5 & 26.3 $\pm$ 1.2 & - & - & - & - & -\ H8+[He<span style="font-variant:small-caps;">i</span>]{}& 3889.05 & 0.286 & 284.9 $\pm$ 26.5 & 269.9 $\pm$ 28.0 & 202.6 $\pm$ 14.6 & 191.2 $\pm$ 5.2 & 248.4 $\pm$ 31.8 $\dagger$ & - & 274.3 $\pm$ 23.3 & 256.3 $\pm$ 30.5 $\dagger$ & -\ H7+[\[Ne<span style="font-variant:small-caps;">iii</span>\]]{}& 3968.77 & 0.266 & 227.7 $\pm$ 43.2 & 238.9 $\pm$ 25.5 & 205.5 $\pm$ 13.8 & 192.6 $\pm$ 4.9 & 206.5 $\pm$ 27.5 $\dagger$ & - & 234.4 $\pm$ 16.4 & 157.5 $\pm$ 20.9 $\dagger$ & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 4026.21 & 0.251 & - & - & 18.7 $\pm$ 1.5 & 20.1 $\pm$ 0.8 & - & - & - & - & -\ [\[S<span style="font-variant:small-caps;">ii</span>\]]{}& 4068.60 & 0.239 & - & 41.4 $\pm$ 4.6 & 23.6 $\pm$ 1.6 & 24.7 $\pm$ 0.9 & - & - & 11.5 $\pm$ 1.1 $\dagger$ & - & -\ [\[S<span style="font-variant:small-caps;">ii</span>\]]{}& 4076.35 & 0.236 & - & 23.0 $\pm$ 5.6 $\dagger$ & 8.6 $\pm$ 0.9 & 8.6 $\pm$ 0.6 & - & - & - & - & -\ [H$\delta$]{}& 4101.74 & 0.229 & 310.1 $\pm$ 36.1 & 322.2 $\pm$ 22.9 & 313.1 $\pm$ 18.1 & 276.9 $\pm$ 5.9 & 352.3 $\pm$ 32.7 $\dagger$ & 414.3 $\pm$ 50.7 $\dagger$ & 305.7 $\pm$ 15.4 & 312.0 $\pm$ 28.7 & -\ [H$\gamma$]{}& 4340.47 & 0.156 & 535.4 $\pm$ 28.6 & 523.7 $\pm$ 24.6 & 511.8 $\pm$ 20.1 & 478.4 $\pm$ 6.9 & 515.8 $\pm$ 37.9 & 641.1 $\pm$ 29.3 $\dagger$ & 513.7 $\pm$ 18.1 & 523.6 $\pm$ 35.5 & 495.4 $\pm$ 39.9\ [\[O<span style="font-variant:small-caps;">iii</span>\]]{}& 4363.21 & 0.149 & 54.6 $\pm$ 6.9 & - & 4.3 $\pm$ 0.6 & 7.0 $\pm$ 0.5 & - & - & 22.8 $\pm$ 2.0 & - & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 4387.93 & 0.141 & - & - & 4.6 $\pm$ 0.5 $\dagger$ & 5.0 $\pm$ 0.2 & - & - & - & - & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 4471.47 & 0.115 & - & 40.3 $\pm$ 3.4 & 38.6 $\pm$ 1.5 & 36.9 $\pm$ 0.7 & - & - & 45.1 $\pm$ 3.0 & - & -\ [\[Fe<span style="font-variant:small-caps;">iii</span>\]]{}& 4658.10 & 0.058 & - & - & 5.9 $\pm$ 0.6 & 2.1 $\pm$ 0.3 & - & - & - & - & -\ [\[Ar<span style="font-variant:small-caps;">iv</span>\]]{}+[He<span style="font-variant:small-caps;">i</span>]{}& 4712.25 & 0.042 & - & - & 4.8 $\pm$ 0.3 & 4.4 $\pm$ 0.3 & - & - & - & - & -\ [H$\beta$]{}& 4861.33 & 0.000 & 1000.0 $\pm$ 10.5 & 1000.0 $\pm$ 5.6 & 1000.0 $\pm$ 0.9 & 1000.0 $\pm$ 5.4 & 1000.0 $\pm$ 10.8 & 1000.0 $\pm$ 24.7 & 1000.0 $\pm$ 5.2 & 1000.0 $\pm$ 11.2 & 1000.0 $\pm$ 20.9\ [He<span style="font-variant:small-caps;">i</span>]{}& 4921.93 & -0.016 & - & 10.8 $\pm$ 1.1 & 10.8 $\pm$ 0.7 & 11.7 $\pm$ 0.6 & - & - & 13.7 $\pm$ 2.2 $\dagger$ & - & -\ [\[O<span style="font-variant:small-caps;">iii</span>\]]{}& 4958.91 & -0.026 & 1829.7 $\pm$ 30.9 & 73.4 $\pm$ 2.5 & 308.2 $\pm$ 2.7 & 453.9 $\pm$ 1.7 & 101.8 $\pm$ 5.7 & - & 661.5 $\pm$ 3.8 & 143.1 $\pm$ 5.3 & -\ [\[O<span style="font-variant:small-caps;">iii</span>\]]{}& 5006.84 & -0.038 & 5804.3 $\pm$ 55.0 & 235.3 $\pm$ 4.2 & 937.2 $\pm$ 9.0 & 1384.4 $\pm$ 4.9 & 307.4 $\pm$ 6.2 & - & 1997.0 $\pm$ 16.7 & 416.2 $\pm$ 8.0 & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 5015.68 & -0.040 & - & 16.2 $\pm$ 2.3 & 24.5 $\pm$ 1.0 & 27.0 $\pm$ 0.3 & - & - & 35.5 $\pm$ 4.0 & - & -\ [\[Cl<span style="font-variant:small-caps;">iii</span>\]]{}& 5517.71 & -0.145 & - & - & 4.3 $\pm$ 0.3 & 5.8 $\pm$ 0.2 & - & - & - & - & -\ [\[Cl<span style="font-variant:small-caps;">iii</span>\]]{}& 5537.88 & -0.149 & - & - & 4.6 $\pm$ 0.3 & 4.7 $\pm$ 0.3 & - & - & - & - & -\ [\[N<span style="font-variant:small-caps;">ii</span>\]]{}& 5754.64 & -0.185 & - & 14.2 $\pm$ 1.9 & 9.7 $\pm$ 1.1 & 8.4 $\pm$ 0.7 & - & - & 5.9 $\pm$ 0.6 & - & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 5875.64 & -0.203 & 141.0 $\pm$ 5.4 & 131.8 $\pm$ 4.9 $\dagger$ & 123.6 $\pm$ 3.1 & 116.9 $\pm$ 1.2 & - & - & 141.6 $\pm$ 7.0 $\dagger$ & - & -\ [\[O<span style="font-variant:small-caps;">i</span>\]]{}& 6300.30 & -0.263 & - & - & 4.0 $\pm$ 0.4 & 14.4 $\pm$ 0.2 & - & - & - & - & -\ [\[S<span style="font-variant:small-caps;">iii</span>\]]{}& 6312.10 & -0.264 & 15.7 $\pm$ 0.9 & 7.5 $\pm$ 1.0 & 14.4 $\pm$ 0.5 & 16.2 $\pm$ 0.5 & - & - & 16.1 $\pm$ 1.4 & - & -\ [\[Si<span style="font-variant:small-caps;">ii</span>\]]{}& 6347.11 & -0.269 & - & - & 1.8 $\pm$ 0.2 & 2.1 $\pm$ 0.1 & - & - & - & - & -\ [\[O<span style="font-variant:small-caps;">i</span>\]]{}& 6363.78 & -0.271 & - & 10.6 $\pm$ 0.5 $\dagger$ & - & 5.1 $\pm$ 0.1 & - & - & - & - & -\ [\[Si<span style="font-variant:small-caps;">ii</span>\]]{}& 6371.36 & -0.272 & - & - & - & 1.2 $\pm$ 0.2 & - & - & - & - & -\ [\[N<span style="font-variant:small-caps;">ii</span>\]]{}& 6548.03 & -0.296 & 37.7 $\pm$ 2.5 & 383.8 $\pm$ 2.6 & 250.6 $\pm$ 0.2 & 230.7 $\pm$ 0.2 & 255.2 $\pm$ 4.1 & 290.6 $\pm$ 9.9 & 110.8 $\pm$ 5.3 & 323.7 $\pm$ 6.3 & 268.2 $\pm$ 5.1\ [H$\alpha$]{}& 6562.82 & -0.298 & 2820.0 $\pm$ 2.2 & 2860.0 $\pm$ 2.8 & 2860.0 $\pm$ 4.1 & 2860.0 $\pm$ 2.6 & 2860.0 $\pm$ 5.3 & 2860.0 $\pm$ 11.4 & 2820.0 $\pm$ 2.5 & 2860.0 $\pm$ 5.2 & 2860.0 $\pm$ 7.2\ [\[N<span style="font-variant:small-caps;">ii</span>\]]{}& 6583.41 & -0.300 & 111.0 $\pm$ 1.6 & 1157.5 $\pm$ 2.3 & 771.4 $\pm$ 0.9 & 712.3 $\pm$ 0.7 & 745.1 $\pm$ 7.2 & 919.5 $\pm$ 12.1 & 335.0 $\pm$ 3.9 & 966.7 $\pm$ 5.7 & 813.8 $\pm$ 8.4\ [He<span style="font-variant:small-caps;">i</span>]{}& 6678.15 & -0.313 & 38.1 $\pm$ 0.8 & 28.0 $\pm$ 0.5 & 30.4 $\pm$ 0.4 & 29.4 $\pm$ 0.2 & - & - & 33.0 $\pm$ 0.5 & 34.4 $\pm$ 2.3 & -\ [\[S<span style="font-variant:small-caps;">ii</span>\]]{}& 6716.47 & -0.318 & 36.8 $\pm$ 0.8 & 394.1 $\pm$ 2.5 & 97.7 $\pm$ 0.6 & 125.2 $\pm$ 0.5 & 213.4 $\pm$ 6.1 & 502.1 $\pm$ 4.6 & 104.4 $\pm$ 1.6 & 288.9 $\pm$ 2.9 & 357.4 $\pm$ 7.8\ [\[S<span style="font-variant:small-caps;">ii</span>\]]{}& 6730.85 & -0.320 & 31.0 $\pm$ 0.9 & 344.2 $\pm$ 2.4 & 115.9 $\pm$ 0.7 & 159.5 $\pm$ 0.6 & 164.0 $\pm$ 7.2 & 382.3 $\pm$ 5.0 & 85.1 $\pm$ 1.2 & 240.4 $\pm$ 3.3 & 350.7 $\pm$ 7.8\ [He<span style="font-variant:small-caps;">i</span>]{}& 7065.28 & -0.365 & 45.7 $\pm$ 0.7 & 18.9 $\pm$ 0.7 & 37.4 $\pm$ 0.7 & 24.3 $\pm$ 0.2 & 23.6 $\pm$ 1.1 $\dagger$ & - & 19.4 $\pm$ 0.5 & 21.0 $\pm$ 1.3 & -\ [\[Ar<span style="font-variant:small-caps;">iii</span>\]]{}& 7135.78 & -0.374 & 116.7 $\pm$ 2.8 & 79.2 $\pm$ 1.9 & 119.4 $\pm$ 2.3 & 102.7 $\pm$ 0.7 & 72.4 $\pm$ 2.0 & - & 118.6 $\pm$ 2.0 & 107.1 $\pm$ 2.6 & -\ [C<span style="font-variant:small-caps;">ii</span>]{}& 7231.34 & -0.387 & - & - & 3.1 $\pm$ 0.4 $\dagger$ & - & - & - & - & - & -\ [He<span style="font-variant:small-caps;">i</span>]{}& 7281.35 & -0.393 & - & - & 6.2 $\pm$ 0.4 & - & - & - & - & - & -\ [\[O<span style="font-variant:small-caps;">ii</span>\]]{}& 7319.46 & -0.399 & 26.1 $\pm$ 1.4 $\dagger$ & 38.3 $\pm$ 3.3 $\dagger$ & 45.6 $\pm$ 1.2 $\dagger$ & 42.5 $\pm$ 0.5 & 66.3 $\pm$ 5.4 $\dagger$ & 124.9 $\pm$ 7.1 $\dagger$ & 27.6 $\pm$ 1.1 $\dagger$ & - & -\ [\[O<span style="font-variant:small-caps;">ii</span>\]]{}& 7330.21 & -0.400 & 14.5 $\pm$ 1.5 $\dagger$ & 31.5 $\pm$ 3.6 $\dagger$ & 37.4 $\pm$ 1.0 $\dagger$ & 35.3 $\pm$ 0.7 & 50.4 $\pm$ 4.3 $\dagger$ & 79.8 $\pm$ 6.6 $\dagger$ & 20.0 $\pm$ 1.0 $\dagger$ & - & -\ [\[Ar<span style="font-variant:small-caps;">iii</span>\]]{}& 7751.10 & -0.455 & 33.5 $\pm$ 1.5 & 29.8 $\pm$ 2.3 & 25.3 $\pm$ 1.3 & 27.6 $\pm$ 0.6 & - & - & 32.3 $\pm$ 1.9 & - & -\ P25 & 8323.42 & -0.525 & - & - & 1.0 $\pm$ 0.1 & - & - & - & - & - & -\ P24 & 8333.78 & -0.526 & - & - & 1.5 $\pm$ 0.1 & - & - & - & - & - & -\ P21 & 8374.48 & -0.531 & - & - & 2.2 $\pm$ 0.1 & - & - & - & - & - & -\ P20 & 8392.40 & -0.533 & - & - & 2.4 $\pm$ 0.2 & - & - & - & - & - & -\ P19 & 8413.20 & -0.535 & - & - & 3.2 $\pm$ 0.1 & - & - & - & - & - & -\ P18 & 8437.95 & -0.538 & - & - & 3.8 $\pm$ 0.1 & - & - & - & - & - & -\ [\[O<span style="font-variant:small-caps;">i</span>\]]{}& 8446.00 & -0.538 & - & - & 4.1 $\pm$ 0.1 & - & - & - & - & - & -\ P17 & 8467.26 & -0.541 & - & - & 4.8 $\pm$ 0.1 & - & - & - & - & - & -\ P16 & 8502.48 & -0.544 & - & - & 5.3 $\pm$ 0.1 & - & - & - & - & - & -\ P15 & 8545.38 & -0.549 & - & - & 6.1 $\pm$ 0.1 & 7.3 $\pm$ 0.2 & - & - & - & - & -\ P14 & 8598.39 & -0.554 & 7.8 $\pm$ 0.7 & - & 7.4 $\pm$ 0.1 & 7.7 $\pm$ 0.2 & - & - & 6.7 $\pm$ 0.5 & - & -\ P12 & 8750.47 & -0.568 & - & - & 11.3 $\pm$ 0.5 & - & - & - & - & - & -\ P11 & 8862.79 & -0.578 & - & - & 15.8 $\pm$ 0.5 $\dagger$ & 16.6 $\pm$ 0.7 & - & - & - & - & -\ P10 & 9015.30 & -0.590 & 18.0 $\pm$ 2.2 & 18.4 $\pm$ 2.7 & 18.4 $\pm$ 0.4 & 18.4 $\pm$ 0.6 & - & - & 17.8 $\pm$ 2.0 $\dagger$ & 18.4 $\pm$ 1.2 & -\ [\[S<span style="font-variant:small-caps;">iii</span>\]]{}& 9068.60 & -0.594 & 212.2 $\pm$ 5.3 & 198.0 $\pm$ 4.0 & 370.8 $\pm$ 1.1 & 420.8 $\pm$ 0.8 & 211.2 $\pm$ 14.5 & 105.8 $\pm$ 7.1 & 201.4 $\pm$ 12.8 & 145.9 $\pm$ 1.2 & 73.5 $\pm$ 6.1\ P9 & 9229.70 & -0.604 & 32.8 $\pm$ 1.1 & 32.1 $\pm$ 0.7 $\dagger$ & - & - & 25.8 $\pm$ 2.8 & - & 26.8 $\pm$ 2.0 & 27.9 $\pm$ 0.8 & -\ [\[S<span style="font-variant:small-caps;">iii</span>\]]{}$^{a}$ & 9530.60 & -0.618 & 422.8 $\pm$ 4.8 & 273.7 $\pm$ 16.3 & 557.7 $\pm$ 3.9 & 661.6 $\pm$ 2.0 & 491.8 $\pm$ 40.2 & 195.5 $\pm$ 18.8 & 468.3 $\pm$ 32.1 & 460.8 $\pm$ 5.2 $\dagger$ & 192.1 $\pm$ 12.1\ P8 & 9546.2 & -0.618 & 44.4 $\pm$ 3.3 $\dagger$ & - & 35.0 $\pm$ 0.6 & - & - & - & 29.7 $\pm$ 4.2 & - & -\ P7 & 10049 & -0.624 & - & - & 52.9 $\pm$ 2.4 & 47.8 $\pm$ 1.5 & - & - & 41.9 $\pm$ 4.0 $\dagger$ & - & -\ \ \ F([H$\beta$]{})$^{b}$ & & & 5.41 $\pm$ 1.30 & 2.63 $\pm$ 0.78 & 39.49 $\pm$ 9.93 & 17.55 $\pm$ 1.61 & 23.61 $\pm$ 5.25 & 7.73 $\pm$ 1.12 & 68.26 $\pm$ 14.59 & 61.83 $\pm$ 15.67 & 27.98 $\pm$ 3.74\ c([H$\beta$]{}) & & & 2.66 $\pm$ 0.10 & 1.36 $\pm$ 0.13 & 1.57 $\pm$ 0.11 & 0.98 $\pm$ 0.04 & 1.50 $\pm$ 0.10 & 1.41 $\pm$ 0.06 & 1.42 $\pm$ 0.09 & 1.54 $\pm$ 0.11 & 1.91 $\pm$ 0.06\ [l c c c c c c c c c c c c c]{}\ &S83 & S127 & S128 & S132 & S156 & S158 & S162 & S206 & S212 & S255 & S298 & S301 & S311\ \ 12+log(O$^+$/H$^+$) & 7.30 $\pm$ 0.06 & 8.36 $\pm$ 0.12 & 8.06 $\pm$ 0.24 & 8.30 $\pm$ 0.11 & 8.23 $\pm$ 0.09 & 8.10 $\pm$ 0.30 & 8.25 $\pm$ 0.07 & 8.17 $\pm$ 0.12 & 7.88 $\pm$ 0.10 & 8.87 & 7.82 $\pm$ 0.11 & 8.08 & 8.07\ 12+log(O$^{2+}$/H$^+$) & 8.09 $\pm$ 0.06 & 7.59 $\pm$ 0.19 & 8.10 $\pm$ 0.26 & 7.21 $\pm$ 0.15 & 7.62 $\pm$ 0.05 & 8.14 $\pm$ 0.40 & 7.75 $\pm$ 0.03 & 8.11 $\pm$ 0.03 & 7.56 $\pm$ 0.04 & 7.47 & 8.22 $\pm$ 0.03 & 7.80 & 8.19\ 12+log(S$^+$/H$^+$) & 5.05 $\pm$ 0.03 & 5.77 $\pm$ 0.07 & 5.68 $\pm$ 0.14 & 6.43 $\pm$ 0.13 & 5.88 $\pm$ 0.25 & 5.41 $\pm$ 0.16 & 6.11 $\pm$ 0.05 & - & 5.53 $\pm$ 0.05 & 6.10 & 5.93 $\pm$ 0.06 & 5.83 & 5.93\ 12+log(S$^{2+}$/H$^+$) & 6.33 $\pm$ 0.02 & 6.69 $\pm$ 0.18 & 6.89 $\pm$ 0.26 & 6.60 $\pm$ 0.03 & 6.89 $\pm$ 0.01 & 7.39 $\pm$ 0.88 $\dagger$ & 6.94 $\pm$ 0.04 & - & 6.29 $\pm$ 0.03 & - & 6.63 $\pm$ 0.10 & - & 7.32\ 12+log(N$^+$/H$^+$) & 6.18 $\pm$ 0.03 & 7.11 $\pm$ 0.08 & 6.89 $\pm$ 0.15 & 7.40 $\pm$ 0.06 & 7.22 $\pm$ 0.05 & 6.94 $\pm$ 0.17 & 7.23 $\pm$ 0.04 & - & 6.70 $\pm$ 0.06 & 7.45 & 6.78 $\pm$ 0.07 & 7.03 & 6.94\ log(N$^+$/O$^+$) & -1.09 $\pm$ 0.04 & -1.21 $\pm$ 0.06 & -1.16 $\pm$ 0.11 & -0.87 $\pm$ 0.05 & -0.91 $\pm$ 0.04 & -1.04 $\pm$ 0.14 & -0.90 $\pm$ 0.03 & - & -1.16 $\pm$ 0.04 & -1.39 & -1.03 $\pm$ 0.05 & -1.04 & -1.11\ 12+log(Ne$^{2+}$/H$^+$) & 7.57 $\pm$ 0.09 & - & - & - & 6.33 $\pm$ 0.08 & - & 6.57 $\pm$ 0.04 & - & - & - & 7.82 $\pm$ 0.04 & 7.04 & 7.43\ 12+log(Ar$^{2+}$/H$^+$) & 5.91 $\pm$ 0.02 & 6.05 $\pm$ 0.21 & 6.32 $\pm$ 0.31 & 6.10 $\pm$ 0.03 & 6.29 $\pm$ 0.02 & 6.28 $\pm$ 0.47 $\dagger$ & 6.23 $\pm$ 0.05 & - & 5.88 $\pm$ 0.02 & - & 6.09 $\pm$ 0.04 & 6.08 & 6.39\ 12+log(Fe$^{2+}$/H$^+$) & - & - & - & - & 5.70 $\pm$ 0.07 & - & 5.21 $\pm$ 0.07 & - & - & - & - & - &\ (He$^+$/H$^+$) 4026 & - & - & - & - & 0.08 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & - & 0.05 $\pm$ 0.01 & - & - & - & 0.09\ (He$^+$/H$^+$) 4471 & - & - & - & 0.08 $\pm$ 0.01 & 0.08 $\pm$ 0.01 & - & 0.08 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & - & - & 0.09 & 0.11\ (He$^+$/H$^+$) 5875 & 0.10 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.06 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & 0.10 $\pm$ 0.02 & 0.09 $\pm$ 0.01 & 0.09 $\pm$ 0.01 & - & - & 0.09 $\pm$ 0.01 & 0.09 & -\ (He$^+$/H$^+$) 6678 & 0.10 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.10 $\pm$ 0.02 & 0.07 $\pm$ 0.01 & 0.08 $\pm$ 0.01 & - & 0.08 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & - & 0.10 $\pm$ 0.02 & 0.08 & 0.10\ (He$^+$/H$^+$) 7065 & 0.11 $\pm$ 0.01 & 0.05 $\pm$ 0.01 & 0.08 $\pm$ 0.02 & 0.07 $\pm$ 0.01 & 0.09 $\pm$ 0.01 & - & 0.06 $\pm$ 0.01 & - & 0.05 $\pm$ 0.01 & - & - & 0.08 & 0.08\ (He$^+$/H$^+$) & 0.10 $\pm$ 0.01 & 0.06 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.08 $\pm$ 0.01 & 0.10 $\pm$ 0.02 & 0.08 $\pm$ 0.01 & 0.09 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & 0.09 & 0.09\ \ ICF(S$^+$,S$^{2+}$) & 1.60 $\pm$ 0.09 & 1.00 $\pm$ 0.01 & 1.09 $\pm$ 0.10 & 1.00 $\pm$ 0.01 & 1.01 $\pm$ 0.01 & 1.09 $\pm$ 0.15 & 1.01 $\pm$ 0.01 & - & 1.02 $\pm$ 0.01 & - & 1.26 $\pm$ 0.07 & - & 1.12\ ICF(Ne$^{2+}$) & 1.07 $\pm$ 0.01 & - & 1.15 $\pm$ 0.09 & - & 1.65 $\pm$ 0.16 & - & 1.49 $\pm$ 0.09 & - & - & - & 1.09 $\pm$ 0.01 & 1.30 & 1.13\ ICF(Ar$^{2+}$) & 1.28 $\pm$ 0.07 & 1.26 $\pm$ 0.03 & 1.12 $\pm$ 0.04 & 1.29 $\pm$ 0.01 & 1.24 $\pm$ 0.02 & 1.13 $\pm$ 0.06 & 1.22 $\pm$ 0.01 & - & 1.19 $\pm$ 0.02 & - & 1.12 $\pm$ 0.02 & 1.18 & 1.12\ ICF(Fe$^{2+}$) & - & - & - & - & 1.41 $\pm$ 0.04 & - & 1.46 $\pm$ 0.04 & - & - & - & - & - &\ ICF(He$^+$) & 1.03 $\pm$ 0.01 & 1.12 $\pm$ 0.05 & 1.06 $\pm$ 0.04 & 1.66 $\pm$ 0.36 & 1.10 $\pm$ 0.06 & - & 1.14 $\pm$ 0.02 & - & 1.17 $\pm$ 0.03 & - & 1.15 $\pm$ 0.04 & - & 1.04\ \ 12+log(O/H) & 8.16 $\pm$ 0.05 & 8.42 $\pm$ 0.11 & 8.38 $\pm$ 0.18 & 8.33 $\pm$ 0.10 & 8.33 $\pm$ 0.07 & 8.42 $\pm$ 0.25 & 8.37 $\pm$ 0.05 & 8.44 $\pm$ 0.06 & 8.05 $\pm$ 0.07 & $<$8.89 & 8.37 $\pm$ 0.04 & $>$8.26 & 8.44 $\pm$ 0.04\ 12+log(S/H) & 6.56 $\pm$ 0.03 & 6.74 $\pm$ 0.16 & 6.96 $\pm$ 0.24 & 6.83 $\pm$ 0.06 & 6.93 $\pm$ 0.03 & 7.43 $\pm$ 0.88 $\dagger$ & 7.01 $\pm$ 0.04 & - & 6.37 $\pm$ 0.03 & - & 6.81 $\pm$ 0.09 & - & 7.39 $\pm$ 0.10\ 12+log(N/H) & 7.04 $\pm$ 0.08 & 7.18 $\pm$ 0.18 & 7.21 $\pm$ 0.33 & 7.44 $\pm$ 0.16 & 7.32 $\pm$ 0.12 & 7.26 $\pm$ 0.43 & 7.35 $\pm$ 0.09 & - & 6.87 $\pm$ 0.13 & $>$7.46 & 7.33 $\pm$ 0.14 & $>$7.22 & 7.31 $\pm$ 0.14\ log(N/O) & -1.09 $\pm$ 0.04 & -1.21 $\pm$ 0.06 & -1.16 $\pm$ 0.11 & -0.87 $\pm$ 0.05 & -0.91 $\pm$ 0.04 & -1.04 $\pm$ 0.14 & -0.90 $\pm$ 0.03 & - & -1.16 $\pm$ 0.04 & $>$-1.39 & -1.03 $\pm$ 0.05 & $<$-1.04 & -1.11 $\pm$ 0.06\ 12+log(Ne/H) & 7.60 $\pm$ 0.09 & - & - & - & 6.55 $\pm$ 0.09 & - & 6.74 $\pm$ 0.05 & - & - & - & 7.86 $\pm$ 0.04 & $>$7.15 & 7.49 $\pm$ 0.04\ 12+log(Ar/H) & 6.02 $\pm$ 0.03 & 6.15 $\pm$ 0.21 & 6.37 $\pm$ 0.31 & 6.21 $\pm$ 0.03 & 6.39 $\pm$ 0.02 & 6.33 $\pm$ 0.47 $\dagger$ & 6.31 $\pm$ 0.05 & - & 5.96 $\pm$ 0.03 & - & 6.14 $\pm$ 0.04 & $>$6.15 & 6.43 $\pm$ 0.04\ 12+log(Fe/H) & - & - & - & - & 5.85 $\pm$ 0.07 & - & 5.38 $\pm$ 0.07 & - & - & - & - & - & -\ He/H & 0.10 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.07 $\pm$ 0.01 & 0.12 $\pm$ 0.03 & 0.09 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01 & - & 0.08 $\pm$ 0.01 & - & 0.10 $\pm$ 0.01 & - & 0.09 $\pm$ 0.01\ \ [^1]: e-mail: alba@iaa.es [^2]: The Image Reduction and Analysis Facility IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.

--- abstract: 'In this paper we investigate both theoretically and numerically the forward in time (FIT) and backward in time (BIT) dispersion of fluid and inertial particle pairs in isotropic turbulence. Fluid particles are known to separate faster BIT than FIT in three-dimensional turbulence, and we find that inertial particles do the same. However, we find that the irreversibility in the inertial particle dispersion is in general much stronger than that for fluid particles. For example, the ratio of the BIT to FIT mean-square separation can be up to an order of magnitude larger for inertial particles than for the fluid particles. We also find that for both the inertial and fluid particles the irreversibility becomes stronger as the scale of their separation decreases. Regarding the physical mechanism for the irreversibility, we argue that whereas the irreversibility of fluid particle-pair dispersion can be understood in terms of a directional bias arising from the energy transfer process in turbulence, inertial particles experience an additional source of irreversibility arising from the non-local contribution to their velocity dynamics, a contribution which vanishes in the limit ${St\to0}$, where $St$ is the particle Stokes number. For each given initial (final, in the backward in time case) separation $\bm{r}^0$ there is an optimum value of $St$ for which the dispersion irreversibility is strongest, as such particles are optimally affected by both sources of irreversibility. We derive analytical expressions for the BIT, mean-square separation of inertial particles and compare the predictions with numerical data obtained from a $Re_{\lambda}\approx 580$ DNS of particle-laden isotropic turbulent flow. The small-time theory, which in the dissipation range is valid for times $\leq\max[St\tau_\eta,\tau_\eta]$ (where $\tau_\eta$ is the Kolmogorov timescale), we find excellent agreement between the theoretical predictions and the DNS. The theory for long-times is in good agreement with the DNS provided that $St$ is small enough so that the inertial particle motion at long-times may be considered as a perturbation about the fluid particle motion, a condition which would in fact be satisfied for arbitrary $St$ at sufficiently long-times in the limit ${Re_{\lambda}\to\infty}$.' author: - 'Andrew D. Bragg' - 'Peter J. Ireland' - 'Lance R. Collins' bibliography: - 'refs\_co12.bib' title: Forward and backward in time dispersion of fluid and inertial particles in isotropic turbulence --- Introduction ============ The relative dispersion of fluid particles in turbulent flows has been a subject of intense investigation since the pioneering studies of Taylor [@taylor22] and Richardson [@richardson26]. The subject has attracted great interest both because of the theoretical challenges it poses and also because of its importance in environmental problems such as the way pollutants in the atmosphere and in oceans spread out [@csanady; @berloff02] (see [@bourgoin06] for further examples). The traditional scenario involves forward in time (FIT) dispersion, that is, the variation in time of pairs of fluid particles which have a given initial separation. Much of the work has focused on the mean-square separation $\langle\vert\bm{r}^{f}(t)\vert^{2}\rangle_{\bm{r}^\prime}$, where $\bm{r}^{f}(t)$ is the fluid particle relative separation vector and $\langle\cdot\rangle_{\bm{r}^\prime}$ denotes an ensemble average conditioned on $\bm{r}^{f}(t^\prime)=\bm{r}^\prime$ with $t^\prime\leq t$. Several theoretical predictions for $\langle\vert\bm{r}^{f}(t)\vert^{2}\rangle_{\bm{r}^\prime}$ for varying $\bm{r}^\prime$ and $t$ have been developed, which we shall discuss in §\[FPDT\]. For extensive reviews of this topic see [@sawford01; @salazar09]. In [@sawford05] the backward in time (BIT) dispersion of fluid particles was investigated and compared with the FIT dispersion. BIT dispersion concerns the behavior of particle pairs which *arrive* at a given location at a given time and which were dispersed at times in the past (i.e. a given end condition, in contrast to FIT dispersion where it is a given initial condition). The BIT mean-square separation may be denoted as $\langle\vert\bm{r}^{f}(t^\prime)\vert^{2}\rangle_{\bm{r}}$, where $\bm{r}^{f}(t^\prime)$ is the fluid particle relative separation vector and $\langle\cdot\rangle_{\bm{r}}$ denotes an ensemble average conditioned on $\bm{r}^{f}(t)=\bm{r}$ with $t^\prime\leq t$. The simulations in [@sawford05] of fluid particle relative dispersion in 3D Navier-Stokes turbulence show that BIT is faster than FIT dispersion, which has also been found in experiments [@berg06a]. A point worth emphasizing to avoid confusion is that in BIT dispersion, the underlying dynamical system is not actually evolving backward in time (a scenario which would be physically uninteresting since time runs forward in physical systems). The dispersion is BIT only in the sense that one is considering the positions of particles at earlier times $t^\prime$, given their position at a later time $t$, but the dynamical evolution according to which the particle state evolved from time $t^\prime$ to time $t$ is the standard forward in time evolution. Sawford *et al.* [@sawford05] note that it is BIT dispersion, not FIT dispersion that is connected to turbulent mixing processes, which serves to emphasize the physical relevance of studying BIT dispersion. Compared to the relative dispersion of fluid particles, that of inertial particles has only recently begun to be investigated. The most comprehensive study to date is that of Bec *et al.* [@bec10b] where they used DNS data to investigate the FIT dispersion of inertial particles and also developed mean-field theoretical descriptions of the dispersion process. They found that for small-times the inertial particles undergo a ballistic separation, driven by their initial velocities. In the long-time limit, they found that the dispersion tends to the fluid particle Richardson $t^3$ law, but with an inertial correction that decays like $t^{-1}$. An experimental study [@gibert10] also observed a ballistic separation for the inertial particles at small-times, but they were unable to measure the dispersion at long-times. Theoretical work on the FIT dispersion of inertial particles has also been done for limiting cases such as $St\gg1$ [@fouxon08] and time-uncorrelated flows [@bec08]. The present study is inspired by [@bec10b] and takes the study to a next step by considering the BIT dispersion of inertial particles, comparing this to the FIT dispersion and seeking to provide both theoretical predictions and physical explanations for irreversability of the inertial particle-pair dispersion. As already noted, in [@sawford05] it was emphasized that turbulent mixing problems are physically related to BIT, not FIT dispersion. A motivation for the present work is therefore that it will lead to advances in our understanding concerning the way particle inertia affects mixing processes in turbulent velocity fields. Another motivation for this work is that it provides insight into how the inertial particle relative velocity theory in [@pan10] might be improved. Since inertial particles posses a memory timescale, their relative velocities are influenced by the fluid velocity field that they have encountered along their path-history, and this depends upon the location of the particle pairs at times in the past, i.e. their BIT dispersion. However, the authors in [@pan10] note that an investigation into the BIT dispersion of inertial particle-pairs has not yet been undertaken, and therefore in their theory they approximate the BIT dispersion by the known FIT dispersion which was examined in [@bec10b]. Under the assumption that FIT and BIT dispersion are not equivalent for inertial particles, in [@bragg14c] it was suggested that the approximation of their equivalence in [@pan10] could be a source of error in the relative velocity predictions from the theory. This highlights the need to understand and predict the BIT dispersion of inertial particles in turbulence. The outline of the paper is as follows. In §\[GEGS\] we construct the general expressions which describe the mean-square separation of inertial particles in a turbulent flow field that will be used in subsequent sections to derive closed, analytic predictions for the particle dispersion. In §\[FPD\] we consider the relative dispersion of fluid particles both theoretically and using DNS simulations. In §\[IPD\] we present the additional irreversibility mechanism that inertia introduces to the dispersion, derive theoretical predictions for the mean-square dispersion at small and long-times, and then consider DNS data for these quantities, against which we test the theoretical predictions. Governing equations and general solutions {#GEGS} ========================================= In this section we construct the exact, but unclosed, expressions for the FIT and BIT mean square separation of the inertial particles that will be used in subsequent sections as the basis from which closed, analytical expressions for these quantities are derived. We consider the relative dispersion of monodisperse inertial particles that are small, $d/\eta\ll1$ (where $d$ is the particle diameter, and $\eta$ is the Kolmogorov lengthscale), dense $\rho_p/\rho_f\gg1$ (where $\rho_p$ and $\rho_f$ are the particle and fluid material densities, respectively) and subject to linear drag forces only. The latter approximation is generally considered to be appropriate for describing the dynamics of water droplets in air (e.g. [@shaw03]). It is possible that for some of the larger particles considered in this present study, the linear drag approximation may not be valid. However, the linear drag approximation will serve as a first approximation for understanding and predicting the BIT dispersion of inertial particles, and we hope that in future work more realistic equations of motion could be considered. Furthermore, excluding the range of very large particle Reynolds numbers, we expect that non-linear drag effects will only change the dispersion quantitatively and that the essential physical aspects of BIT dispersion will not be qualitatively affected by the liner drag approximation. The equation governing the relative motion of particles satisfying the aforementioned requirements is then obtained from the simplified form of the Maxey-Riley equation [@maxey83] $$\begin{aligned} \ddot{\bm{r}}^{p}(t)=\dot{\bm{w}}^{p}(t)=\frac{1}{\tau_{p}}\Big(\Delta\bm{u}^{p}(t)-\bm{w}^{p}(t)\Big), \label{reom}\end{aligned}$$ where $\bm{r}^{p}(t),\bm{w}^{p}(t)$ are the inertial particle-pair relative position and relative velocity vectors, $\tau_{p}$ is the momentum response time of the particles (we will also use the Stokes number later in the paper $St\equiv\tau_p/\tau_\eta$, where $\tau_\eta$ is the fluid Kolmogorov timescale), $\Delta\bm{u}^{p}(t)\equiv\Delta\bm{u}(\bm{x}^{p}(t),\bm{r}^{p}(t),t)$ is the difference between the fluid velocity field evaluated at the positions of the two particles and $\bm{x}^{p}(t)$ is the position of the reference particle. The formal solution to (\[reom\]) may be written as (for isotropic $\Delta\bm{u}$) $$\begin{aligned} \bm{r}^{p}(t)&=\bm{r}^{p}(t^\prime)+G(t-t^\prime)\bm{w}^{p}(t^\prime)+\tau_{p}^{-1}\int\limits_{t^\prime}^{t}G(t-s)\Delta\bm{u}^{p}(s)ds,\quad t^\prime\leq t,\label{rpsol}\\ \bm{w}^{p}(t)&=\dot{G}(t-t^\prime)\bm{w}^{p}(t^\prime)+\tau_{p}^{-1}\int\limits_{t^\prime}^{t}\dot{G}(t-s)\Delta\bm{u}^{p}(s)ds,\quad t^\prime\leq t, \label{wpsol}\end{aligned}$$ where $G$ is the Green function for the equation of motion for $\bm{r}^{p}(t)$ $$\begin{aligned} G(t-t^\prime)&\equiv\tau_{p}\Big(1-\exp[-\tau_{p}^{-1}(t-t^\prime)]\Big), \label{G}\\ \dot{G}(t-t^\prime)&\equiv\exp[-\tau_{p}^{-1}(t-t^\prime)]. \label{Gdot}\end{aligned}$$ The FIT dispersion PDF is defined as $$\begin{aligned} \varrho^{F}(\bm{r},t\vert\bm{r}^\prime,t^\prime)\equiv\Big\langle\delta(\bm{r}^{p}(t)-\bm{r})\Big\rangle_{\bm{r}^\prime}, \label{fitPDF}\end{aligned}$$ and for BIT dispersion $$\begin{aligned} \varrho^{B}(\bm{r}^\prime,t^\prime\vert\bm{r},t)\equiv\Big\langle\delta(\bm{r}^{p}(t^\prime)-\bm{r}^\prime)\Big\rangle_{\bm{r}}, \label{bitPDF}\end{aligned}$$ and in each case $t^\prime\leq t$. The notation $\langle{\cdot}\rangle_{\bm{r}^\prime}$ and $\langle{\cdot}\rangle_{\bm{r}}$ in (\[fitPDF\]) and (\[bitPDF\]) denote conditional ensemble averaging; conditioned on $\bm{r}^{p}(t^\prime)=\bm{r}^\prime$ in the FIT case (‘initial-time conditioning’) and $\bm{r}^{p}(t)=\bm{r}$ in the BIT case (‘end-time conditioning’). In this paper we are interested in the mean-square separation behavior rather than the full dispersion PDF. From (\[fitPDF\]) and (\[bitPDF\]) we may define the FIT and BIT mean-square separation $$\begin{aligned} \Big\langle\vert\bm{r}^{p}(t)\vert^{2}\Big\rangle_{\bm{r}^\prime}\equiv\int\limits_{\bm{r}}\bm{r}\bm{\cdot}\bm{r}\varrho^{F}(\bm{r},t\vert\bm{r}^\prime,t^\prime)d\bm{r}, \label{MSfit}\end{aligned}$$ $$\begin{aligned} \Big\langle\vert\bm{r}^{p}(t^\prime)\vert^{2}\Big\rangle_{\bm{r}}\equiv\int\limits_{\bm{r}^\prime}\bm{r}^{\prime}\bm{\cdot}\bm{r}^{\prime}\varrho^{B}(\bm{r}^\prime,t^\prime\vert\bm{r},t)d\bm{r}^{\prime}. \label{MSbit}\end{aligned}$$ We may construct an exact expression for (\[MSfit\]) using (\[rpsol\]) $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(t)\vert^{2}\Big\rangle_{\bm{r}^\prime}=\,&\bm{r}^\prime\bm{\cdot}\bm{r}^\prime+2G(t-t^\prime)\bm{r}^\prime\bm{\cdot}\Big\langle \bm{w}^{p}(t^\prime)\Big\rangle_{\bm{r}^\prime}+2\tau_{p}^{-1}\bm{r}^\prime\bm{\cdot}\int\limits_{t^\prime}^{t}G(t-s)\Big\langle\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}^\prime}ds\\ &+G^{2}(t-t^\prime)\Big\langle\bm{w}^{p}(t^\prime)\bm{\cdot}\bm{w}^{p}(t^\prime)\Big\rangle_{\bm{r}^\prime}+2\tau_{p}^{-1}G(t-t^\prime)\int\limits_{t^\prime}^{t}G(t-s)\Big\langle\bm{w}^{p}(t^\prime)\bm{\cdot}\Delta \bm{u}^{p}(s)\Big\rangle_{\bm{r}^\prime}ds\\ &+\tau_{p}^{-2}\int\limits_{t^\prime}^{t}\int\limits_{t^\prime}^{t}G(t-s)G(t-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^\prime}ds^\prime\,ds, \label{FITmsr} \end{split}\end{aligned}$$ and rearranging (\[rpsol\]) for $\bm{r}^{p}(t^\prime)$ we may construct the expression for (\[MSbit\]) $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(t^\prime)\vert^{2}\Big\rangle_{\bm{r}}=\,&\bm{r}\bm{\cdot}\bm{r}-2G(t-t^\prime)\bm{r}\bm{\cdot}\Big\langle \bm{w}^{p}(t^\prime)\Big\rangle_{\bm{r}}-2\tau_{p}^{-1}\bm{r}\bm{\cdot}\int\limits_{t^\prime}^{t}G(t-s)\Big\langle\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}}ds\\ &+G^{2}(t-t^\prime)\Big\langle\bm{w}^{p}(t^\prime)\bm{\cdot}\bm{w}^{p}(t^\prime)\Big\rangle_{\bm{r}}+2\tau_{p}^{-1}G(t-t^\prime)\int\limits_{t^\prime}^{t}G(t-s)\Big\langle \bm{w}^{p}(t^\prime)\bm{\cdot}\Delta \bm{u}^{p}(s)\Big\rangle_{\bm{r}}ds\\ &+\tau_{p}^{-2}\int\limits_{t^\prime}^{t}\int\limits_{t^\prime}^{t}G(t-s)G(t-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}}ds^\prime\,ds. \label{BITmsr} \end{split}\end{aligned}$$ Since we are interested in statistically stationary systems, we may set the ‘conditioning time’ to zero and consider the dispersion behavior as a function of time separation. In the FIT case this amounts to setting $t^\prime=0$ and in the BIT case setting $t=0$. Further, since $t^\prime\leq t$ we may re-write (\[FITmsr\]) and (\[BITmsr\]) as $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}=\,&\bm{r}^{0}\bm{\cdot}\bm{r}^{0}+2G(\mathcal{T})\bm{r}^{0}\bm{\cdot}\Big\langle \bm{w}^{p}(0)\Big\rangle_{\bm{r}^0}+2\tau_{p}^{-1}\bm{r}^{0}\bm{\cdot}\int\limits_{0}^{\mathcal{T}}G(\mathcal{T}-s)\Big\langle\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &+G^{2}(\mathcal{T})\Big\langle\bm{w}^{p}(0)\bm{\cdot}\bm{w}^{p}(0)\Big\rangle_{\bm{r}^0}+2\tau_{p}^{-1}G(\mathcal{T})\int\limits_{0}^{\mathcal{T}}G(\mathcal{T}-s)\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &+\tau_{p}^{-2}\int\limits_{0}^{\mathcal{T}}\int\limits_{0}^{\mathcal{T}}G(\mathcal{T}-s)G(\mathcal{T}-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds, \label{FITmsr2} \end{split}\end{aligned}$$ where $\mathcal{T}=t-t^\prime$ with $t^\prime=0$, $t^\prime\leq t$ so that $\mathcal{T}\geq0$ and $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}=\,&\bm{r}^{0}\bm{\cdot}\bm{r}^{0}-2G(\mathcal{T})\bm{r}^{0}\bm{\cdot}\Big\langle \bm{w}^{p}(-\mathcal{T})\Big\rangle_{\bm{r}^0}-2\tau_{p}^{-1}\bm{r}^{0}\bm{\cdot}\int\limits^{0}_{-\mathcal{T}}G(-s)\Big\langle\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &+G^{2}(\mathcal{T})\Big\langle\bm{w}^{p}(-\mathcal{T})\bm{\cdot}\bm{w}^{p}(-\mathcal{T})\Big\rangle_{\bm{r}^0}+2\tau_{p}^{-1}G(\mathcal{T})\int\limits^{0}_{-\mathcal{T}}G(-s)\Big\langle\bm{w}^{p}(-\mathcal{T})\bm{\cdot}\Delta \bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &+\tau_{p}^{-2}\int\limits^{0}_{-\mathcal{T}}\int\limits^{0}_{-\mathcal{T}}G(-s)G(-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds, \label{BITmsr2} \end{split}\end{aligned}$$ where again $\mathcal{T}=t-t^\prime$, but now with $t=0$, $t^\prime\leq t$ so that $\mathcal{T}\geq0$ and $t^\prime=-\mathcal{T}$, and $\langle{\cdot}\rangle_{\bm{r}^0}$ denotes an ensemble average conditioned on $\bm{r}^{p}(\mathcal{T}=0)=\bm{r}^0$. In going from (\[BITmsr\]) to (\[BITmsr2\]), we note that$$G(t-t^\prime)=\tau_{p}\Big(1-\exp[-\tau_{p}^{-1}(t-t^\prime)]\Big)=\tau_{p}\Big(1-\exp[-\tau_{p}^{-1}\mathcal{T}]\Big)=G(\mathcal{T}),$$ and$$G(t-s)=\tau_{p}\Big(1-\exp[-\tau_{p}^{-1}(t-s)]\Big)=\tau_{p}\Big(1-\exp[\tau_{p}^{-1}s]\Big)=G(-s).$$We may develop theoretical descriptions for the FIT and BIT dispersion of inertial particles by applying closure approximations to (\[FITmsr2\]) and (\[BITmsr2\]) to construct closed analytical solutions. Note that we have not constructed (\[BITmsr2\]) using the solutions to the time-reversed form of the equations of motion, which is commonly done when analyzing BIT problems. We have chosen not to construct the BIT results via this method because we believe it hinders the physical understanding of the problem since, as explained earlier, in BIT dispersion the dynamical system is not actually evolving backward in time. Rather, we have therefore constructed (\[BITmsr2\]) in a manner consistent with how the same statistics would be obtained in an experiment or DNS where the BIT statistics are constructed by recording the trajectories of the particles (which are being evolved using the standard forward-in-time equations of motion) and then subsequently evaluate the BIT statistics based on the particle trajectory histories. In addition we will also consider the dispersion of fluid particles whose equation of relative motion is simply $$\begin{aligned} \dot{\bm{r}}^f(t)=\Delta\bm{u}^f(t), \end{aligned}$$ with solution $$\begin{aligned} \bm{r}^f(t)=\bm{r}^f(t')+\int\limits_{t'}^t\Delta\bm{u}^f(s)\,ds,\label{rfsol} \end{aligned}$$ where $\bm{r}^f(t)$ is the relative separation between two fluid particles and $\Delta\bm{u}^{f}(t)\equiv\Delta\bm{u}(\bm{x}^{f}(t),\bm{r}^{f}(t),t)$ is the difference in the fluid velocity evaluated at the positions of the two particles, $\bm{x}^f(t)$ being the position of the reference fluid particle. Here and throughout the superscript ‘$p$’ denotes that the variable is defined along inertial particle trajectories, and superscript‘$f$’ is used to denote that the variable is defined along fluid particle trajectories. Following the same steps as was used to derive (\[FITmsr2\]) and (\[BITmsr2\]) we obtain for the fluid particles $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}=&\bm{r}^{0}\bm{\cdot}\bm{r}^{0}+2\bm{r}^{0}\bm{\cdot}\int\limits_{0}^{\mathcal{T}}\Big\langle\Delta\bm{u}^{f}(s)\Big\rangle_{\bm{r}^0}ds+\int\limits_{0}^{\mathcal{T}}\int\limits_{0}^{\mathcal{T}}\Big\langle\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{f}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds, \label{FITmsr2FP} \end{split}\end{aligned}$$ $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}=&\bm{r}^{0}\bm{\cdot}\bm{r}^{0}-2\bm{r}^{0}\bm{\cdot}\int\limits^{0}_{-\mathcal{T}}\Big\langle\Delta\bm{u}^{f}(s)\Big\rangle_{\bm{r}^0}ds+\int\limits^{0}_{-\mathcal{T}}\int\limits^{0}_{-\mathcal{T}}\Big\langle\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{f}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds. \label{BITmsr2FP} \end{split}\end{aligned}$$ Fluid particle dispersion {#FPD} ========================= We will first consider the case of fluid particle dispersion before considering inertial particle dispersion, which is the main contribution of this paper. Our purpose in this section is not to derive new results but to consider various results and explanations that have been previously proposed. This will be especially helpful when we consider in §\[IPD\] inertial particle dispersion, which introduces additional complexities compared to the fluid particle dispersion. Irreversibility mechanisms {#FPIM} -------------------------- The FIT and BIT dispersion of fluid particles in turbulence has been considered in several studies, using theoretical, computational and experimental techniques (e.g. [@sawford05; @berg06a; @jucha14]). These studies have revealed that in 3D turbulence, BIT dispersion is faster than FIT dispersion. Different explanations have been given for this observed irreversibility. In [@sawford05] the behavior of the odd-moments of the fluid velocity increments $\Delta\bm{u}$ in turbulence under time-reversal was used to provide an explanation for the difference between FIT and BIT dispersion. In [@berg06a] the authors appealed to the behavior of the eigenvalues of the strain-rate tensor $\bm{\mathcal{S}}(\bm{x},t)\equiv (1/2)[\bm{\nabla_x}\bm{u}+(\bm{\nabla_x}\bm{u})^\top]$ under time-reversal to explain the difference. The largest eigenvalue of $\bm{\mathcal{S}}(\bm{x},-t)$ is greater than the largest eigenvalue of $\bm{\mathcal{S}}(\bm{x},t)$, thus explaining, not only why FIT and BIT dispersion are different, but also why BIT dispersion is faster than FIT dispersion. They also argued that since the course-grained version of $\bm{\mathcal{S}}$ exhibits similar dynamics, the same argument also applies for dispersion in the inertial range of the turbulence. We may further clarify the irreversibility mechanism by considering the equation governing the relative separation of the fluid particle-pair which is simply $\dot{\bm{r}}^f(t)=\Delta\bm{u}(\bm{r}^f(t),t)$. FIT dispersion corresponds to particles separating (on average) as time increases (i.e. $\dot{\bm{r}}^f(t)>\bm{0}$), whereas BIT can be thought of as particles approaching each other as time increases (i.e. $\dot{\bm{r}}^f(t)<\bm{0}$). Since the PDF of $\Delta\bm{u}$ is negatively skewed in 3D turbulence because of the energy transfer to the small scales, the particle-pairs move together faster than they move apart, and hence BIT is faster than FIT dispersion. This type of energy flux argument can also be quantified in the inertial range by considering a small-time expansion of the dispersion process, and such an analysis indeed predicts that BIT dispersion should be faster than FIT dispersion in 3D turbulence [@jucha14]. Consistent with these arguments, fluid particle dispersion in kinematically simulated flow fields, where $\Delta\bm{u}$ has a Gaussian distribution, exhibits FIT and BIT symmetry [@flohr05]. These arguments for FIT and BIT asymmetry based on the energy flux and associated asymmetry in the PDF for $\Delta\bm{u}$ also suggest that in 2D turbulence where there is a flux of energy towards the large scales (yielding a positively skewed PDF for $\Delta\bm{u}$), FIT dispersion should be faster than BIT dispersion, something that has been shown numerically in [@faber09]. What each of these explanations share in common is that the time-irreversibility of fluid particle dispersion in turbulence arises, fundamentally, because of the intrinsic time directionality in turbulence dynamics, a consequence of its dissipative nature. However, although this is the physical origin of the irreversibility in Navier-Stokes turbulence, any model flow field which generates asymmetric probability density functions for $\Delta\bm{u}(\bm{r},t)$ would give rise to irreversible fluid particle-pair dispersion. Theoretical results {#FPDT} ------------------- Having considered how FIT and BIT differ in turbulence, we now turn to consider various theoretical predictions that have been made to describe the fluid particle-pair dispersion. Note that throughout this paper we are considering the statistically stationary state of incompressible flow where the fluid particles are assumed to be fully mixed in the system. The FIT and BIT mean square dispersion of fluid particles is given by (\[FITmsr2FP\]) and (\[BITmsr2FP\]). We now introduce the turnover timescale of $\Delta\bm{u}$ at separation $r^0\equiv\vert\bm{r}^0\vert$, known in the context of dispersion studies as the Batchelor timescale $\tau_{r^{0}}$. For initial separations in the dissipation regime, we take $\tau_{r^{0}}=\tau_{\eta}$, and in the inertial range $\tau_{r^{0}}=(\vert\bm{r}^{0}\vert^{2}/\langle\epsilon\rangle)^{1/3}$, where $\langle\epsilon\rangle$ is the turbulent kinetic energy dissipation rate. In the regime ${\mathcal{T}\ll\tau_{r^0}}$ we may make the approximation $$\begin{aligned} \Delta\bm{u}^{f}(\mathcal{T})\approx\Delta\bm{u}^{f}(0)+\mathcal{O}(\mathcal{T}/\tau_{r^0}),\label{uTexp}\end{aligned}$$ (and similarly for the terms in (\[FITmsr2\]) and (\[BITmsr2\]) involving $s$ and $s^\prime$ in the time arguments) and introducing this approximation into (\[FITmsr2FP\]) and (\[BITmsr2FP\]), we obtain $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\Big\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}+\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\label{Bal}. \end{aligned}$$ Notice that in the ballistic regime the fluid-particle dispersion is time-reversible: this is a consequence of (\[uTexp\]), which ignores the dynamical evolution of $\Delta\bm{u}$ along the particle trajectories, and it is the nature of the dynamical evolution of the turbulence that gives rise to irreversibility in the dispersion process for fluid particles. In [@ouellette06c] the authors consider the importance of correctly including the terms involving $\bm{r}^{0}$ in the description of the mean square separation (this is also discussed in [@salazar09]). Specifically they consider the difference between (\[Bal\]) and $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^{0}\vert^{2}\Big\rangle_{\bm{r}^0}\approx\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\label{Bal2}. \end{aligned}$$ We may write $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^{0}\vert^{2}\Big\rangle_{\bm{r}^0}=\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-2\bm{r}^{0}\bm{\cdot}\Big\langle\bm{r}^{f}(\mathcal{T})\Big\rangle_{\bm{r}^0}+\vert\bm{r}^0\vert^{2},\end{aligned}$$ showing that a $\mathcal{T}$ dependent difference between (\[Bal\]) and (\[Bal2\]) arises when $\langle\bm{r}^{f}(\mathcal{T})\rangle_{\bm{r}^0}\neq\bm{r}^{0}$. From (\[rfsol\]) we may derive the result for ${\mathcal{T}\ll\tau_{r^0}}$ $$\begin{aligned} \Big\langle\bm{r}^{f}(\mathcal{T})\Big\rangle_{\bm{r}^0}=\bm{r}^{0}+\mathcal{T}\Big\langle\Delta\bm{u}^{f}(0)\Big\rangle_{\bm{r}^0}+\mathcal{O}(\mathcal{T}/\tau_{r^0}),\end{aligned}$$ giving $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^{0}\vert^{2}\Big\rangle_{\bm{r}^0}=\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^{0}\vert^{2}-2\mathcal{T}\bm{r}^{0}\bm{\cdot}\Big\langle\Delta\bm{u}^{f}(0)\Big\rangle_{\bm{r}^0}+\mathcal{O}(\mathcal{T}/\tau_{r^0}).\label{I1}\end{aligned}$$ In [@ouellette06c] the authors find that if they plot their experimental data for$$\Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^{0}\vert^{2}\Big\rangle_{\bm{r}^0}\Big/\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0},$$they find a good collapse for ${\mathcal{T}<\tau_{r^0}}$, verifying the validity of the ballistic prediction. However when they plot $$\Big(\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^{0}\vert^{2}\Big)\Big/\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0},$$they do not find a good collapse of the data indicating that the term of difference between the two expressions is important. From (\[I1\]) we see that the difference between these two expressions depends upon $\langle\Delta\bm{u}^{f}(0)\rangle_{\bm{r}^0}$ and for fully mixed fluid particles in incompressible, isotropic turbulence $$\begin{aligned} \Big\langle\Delta\bm{u}^{f}(0)\Big\rangle_{\bm{r}^0}\equiv\frac{1}{\varphi(\bm{r}^0,t)}\Big\langle\Delta\bm{u}^{f}(0)\delta\Big(\bm{r}^f(0)-\bm{r}^0\Big)\Big\rangle=\Big\langle\Delta\bm{u}(\bm{r}^0,0)\Big\rangle=\bm{0},\end{aligned}$$ where $\varphi(\bm{r}^0,t)\equiv\langle\delta(\bm{r}^f(0)-\bm{r}^0)\rangle$. Consequently, under the ballistic approximation $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^{0}\vert^{2}\Big\rangle_{\bm{r}^0}=\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^{0}\vert^{2}.\end{aligned}$$ The results in [@ouellette06c] for the ballistic case are therefore surprising, since in this regime there should be no difference between (\[Bal\]) and (\[Bal2\]). However, the data presented in Fig. 6 of [@ouellette06c] implies $$\Big(\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert \bm{r}^{0}\vert^{2}\Big)\Big/\tau_{\eta}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert\Big\rangle_{\bm{r}^{0}}\to\text{finite value as $\mathcal{T}\to0$},$$ which cannot be correct (by definition $\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}-\vert \bm{r}^{0}\vert^{2}\to0$ as $\mathcal{T}\to0$). One explanation for this may be errors introduced by the relatively large size of the bins used to construct the statistics from their experimental data, having widths $\approx 43\eta$. In the case where ${\mathcal{T}\geq\mathcal{O}(\tau_{r^0})}$, $\langle\bm{r}^{f}(\mathcal{T})\rangle_{\bm{r}^0}$ depends upon $\int^\mathcal{T}_0 \langle\Delta\bm{u}^{f}(s)\rangle_{\bm{r}^0}\,ds$, and $\langle\Delta\bm{u}^{f}(s)\rangle_{\bm{r}^0}\neq\bm{0}$ even for isotropic turbulence. The reason is the conditional nature of the average; particle-pairs which were at $\bm{r}^{0}$ at $\mathcal{T}=0$ and are on average separating will be experiencing positive velocity differences on average, i.e. $\langle\Delta\bm{u}^{f}(s)\rangle_{\bm{r}^0}\geq\bm{0}$. Nevertheless, we expect that the effect of this on the prediction of $\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ will be small compared to the higher-order moment terms in its evolution equation. Similar arguments also describe the BIT case, only in that case, since fluid particles are on average approaching each other towards $\bm{r}^{0}$, then $\langle\Delta\bm{u}^{f}(s)\rangle_{\bm{r}^0}\leq\bm{0}$. Having considered the small-time behavior that gives rise to the $\mathcal{T}^2$ ballistic relationship, we now consider the finite $\mathcal{T}$ behavior, at which point the irreversibility of the disperson process becomes manifest. The simplest way to proceed is to consider the contribution from higher order terms in the $\mathcal{T}/\tau_{r^{0}}$ expansion in (\[uTexp\]). Accounting for the second term in the expansion ($\propto\mathcal{T}$) gives rise to the first term describing the break in time-reversibility of the dispersion, and involves the correlation between the fluid relative velocity and relative acceleration measured at ${\mathcal{T}=0}$ [@jucha14]. However, since this expansion becomes formally divergent for ${\mathcal{T}\geq\mathcal{O}(\tau_{r^0})}$, we seek alternative approximations to describe the finite $\mathcal{T}/\tau_{r^{0}}$ behavior of the dispersion. We begin by considering the case for $\bm{r}^0$ in the dissipation range, and then for $\bm{r}^0$ in the inertial range. In the dissipation regime $\Delta\bm{u}^f(\mathcal{T})\approx\bm{\Gamma}^f(\mathcal{T})\bm{\cdot}\bm{r}^f(\mathcal{T})$, where $\bm{\Gamma}^f(\mathcal{T})\equiv\bm{\nabla_x}\bm{u}(\bm{x}^f(\mathcal{T}),\mathcal{T})$. In this case we have $\dot{\bm{r}}^f(\mathcal{T})=\bm{\Gamma}^f(\mathcal{T})\bm{\cdot}\bm{r}^f(\mathcal{T})$ whose solution may be expressed using the time-ordered exponential $\exp_{\mathrm{T}}[\cdot]$ as [@falkovich01] $$\begin{aligned} \bm{r}^f(\mathcal{T})=\bm{r}^f(0)\exp_{\mathrm{T}}\Bigg(\int_{0}^{\mathcal{T}}\bm{\Gamma}^f(s)\,ds\Bigg).\end{aligned}$$ Based upon this observation that the pair separation grows exponentially in time in the dissipation range, Batchelor [@batchelor52a] gave an order-of-magnitude estimate for $\vert\bm{r}^{f}(\mathcal{T})\vert$ for large $\mathcal{T}/\tau_\eta$ which gives rise to the prediction $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}=\vert\bm{r}^{0}\vert^{2}\exp[2B\tau_{\eta}^{-1}\mathcal{T}], \label{BatchelorSmallr} \end{split}\end{aligned}$$ where various values for $B$ have been given (see [@salazar12a]). It is important to emphasize that (\[BatchelorSmallr\]) is only supposed to be valid for large $\mathcal{T}/\tau_\eta$; expanding the exponential gives $$\begin{aligned} \begin{split} \exp[2B\tau_{\eta}^{-1}\mathcal{T}]=1+2B\tau_{\eta}^{-1}\mathcal{T}+2B^2\tau_{\eta}^{-2}\mathcal{T}^2+\mathcal{O}(\mathcal{T}^3), \label{BatchExpan} \end{split}\end{aligned}$$ which shows that (\[BatchelorSmallr\]) is not consistent with the ballistic behavior that should be obtained in the regime $\mathcal{T}/\tau_\eta\ll 1$. In [@ni13] experimental evidence for (\[BatchelorSmallr\]) is given, and it is claimed that the result is validated for $\mathcal{T}<\tau_\eta$ and $\vert\bm{r}^0\vert$ in in the dissipation range. However, their results appear to be problematic for several reasons. The main issue is that the experimental results in Fig 5(b) of [@ni13] show that the fluid particles undergo an exponential type growth (described by (\[BatchelorSmallr\])) for $\mathcal{T}<\tau_\eta$ and $\vert\bm{r}^0\vert$ in the dissipation range, and then *subsequently* undergo a ballistic type growth for $\mathcal{T}>\tau_\eta$. This behavior cannot be correct since the ballistic law is exact in the limit ${\mathcal{T}/\tau_{r^0}\to 0}$, provided that the fluid particles are fully mixed (so that $\langle\Delta\bm{u}^{f}(0)\rangle_{\bm{r}^0}=\bm{0}$). The exponential growth can only occur subsequent to this in the dissipation range when finite $\mathcal{T}$ contributions to $\Delta\bm{u}^f(\mathcal{T})=\bm{\Gamma}^f(\mathcal{T})\bm{\cdot}\bm{r}^f(\mathcal{T})$ become significant. We will consider these issues further in the next section when we consider DNS data for the fluid particle dispersion. We now consider the case where $\bm{r}^{0}$ lies in the inertial range of scales, i.e. $\eta\ll\vert\bm{r}^{0}\vert\ll L$, where $L$ is the integral lengthscale. In this case the Lagrangian behavior of $\Delta\bm{u}$ is more complex and is no longer linearly proportional to $\bm{r}$. The standard approach for describing dispersion in the inertial range is to use Kolmogorov’s K41 theory to describe the growth of the fluid velocity differences in (\[FITmsr2FP\]) and (\[BITmsr2FP\]) and obtain the result for $\mathcal{T}\gg\tau_{r^0}$ in the inertial range $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}&=\mathfrak{g}^{F}\langle\epsilon\rangle\mathcal{T}^{3},\label{ROf}\\ \Big\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}&=\mathfrak{g}^{B}\langle\epsilon\rangle\mathcal{T}^{3},\label{ROb}\end{aligned}$$ where $\mathfrak{g}^{F}$ and $\mathfrak{g}^{B}$ are the FIT and BIT Richardson’s constants, estimated from experimental data to be $\mathfrak{g}^{F}\approx 0.55$ and $\mathfrak{g}^{B}\approx 1.15$ [@berg06a]. The result in (\[ROf\]) is from Batchelor’s work [@batchelor50]. The result in (\[ROb\]), the BIT equivalent of (\[ROf\]), was proposed by Sawford *et al.* [@sawford05]. It is also conventional to refer to the $\mathcal{T}^{3}$ scaling law as the Richardson-Obukhov (RO) law. Note that the use of K41 does not lead to a prediction of the relative rate of FIT and BIT dispersion; that $\mathfrak{g}^{B}>\mathfrak{g}^{F}$ in 3D turbulence is simply an empirical finding. However, based upon the arguments given in §\[FPIM\], we would in fact expect that $\mathfrak{g}^{B}>\mathfrak{g}^{F}$ in 3D turbulence. If $\vert\bm{r}^{0}\vert>L$ then the dispersion (following the initial ballistic separation) is diffusive and time-reversible. At these separations the two particles experience no correlation between their motion, and so the dispersion becomes directly related to the one-particle dispersion which in stationary, homogeneous turbulence is time-reversible. This is simply a result of the system symmetries; if $\bm{x}^f$ is the single particle position then $\langle(\bm{x}^f(t)-\bm{x}^f(0))^2\rangle$ becomes under time reversal ${\langle(\bm{x}^f(-t)-\bm{x}^f(0))^2\rangle}$. However, because of homogeneity and stationarity, applying a time shift gives $\langle(\bm{x}^f(-t)-\bm{x}^f(0))^2\rangle=\langle(\bm{x}^f(0)-\bm{x}^f(t))^2\rangle=\langle(\bm{x}^f(t)-\bm{x}^f(0))^2\rangle$. DNS results {#FPS} ----------- We now consider results from a DNS of statistically stationary, homogeneous, isotropic turbulence against which we will test the theoretical results discussed in §\[FPDT\]. We use a pseudospectral method to solve the incompressible Navier-Stokes equations for statistically stationary isotropic turbulence in a three-dimensional periodic cube of length $2 \pi$, $$\partial_t\bm{u} +\bm{\omega}\times\bm{u} +\bm{\nabla_x}\left(\frac{p}{\rho} + \frac{\bm{u}\cdot\bm{u}}{2}\right) = \nu\bm{\nabla_x}^2\bm{u} + \bm{f}, \label{eq:N-S}$$ where $\bm{u}(\bm{x},t)$ is the fluid velocity, $\bm{\omega}(\bm{x},t)$ is the vorticity, $p(\bm{x},t)$ is the pressure, $\rho$ is the fluid density, $\nu$ is the kinematic viscosity, and $\bm{f}(\bm{x},t)$ is a large-scale forcing function that is added to achieve stationary turbulence. For this simulation, forcing was added to the first two wavenumbers in Fourier space. Time integration is performed through a second-order, explicit Runge-Kutta scheme with aliasing errors removed by means of a combination of spherical truncation and phase-shifting. The time step was chosen to achieve a CFL number of about $0.5$. The fluid field was solved on a grid with $2048^3$ grid points on $16,384$ processors on the Yellowstone cluster at the U.S. National Center for Atmospheric Research [@yellowstone]. The three-dimensional fast Fourier transforms required for the pseudospectral solution of (\[eq:N-S\]) are performed in parallel with MPI using the P3DFFT library [@p3dfft]. The Taylor microscale Reynolds number $R_\lambda$ for our flow is about $580$ and the ratio $L/\eta \approx 800$. The viscosity was chosen to achieve a small-resolution $k_\mathrm{max} \eta \approx 1.7$ (where $k_\mathrm{max} = 2048 \sqrt{2}/3$ is the maximum resolved wavenumber magnitude). This initial flow field was then evolved for about $5$ large eddy turnover times until the flow statistics were statistically stationary. For comparison with the theory to be presented in §\[IPD\], the inertial particle equation of motion is the simplified form of the Maxey-Riley equation [@maxey83] $$\begin{aligned} \ddot{\bm{x}}^{p}=\dot{\bm{v}}^{p}= \frac{1}{\tau_{p}}\Big(\bm{u}(\bm{x}^{p}(t),t)-\bm{v}^{p}(t)\Big), \label{eq:velocity}\end{aligned}$$ where $\bm{x}^{p}(t)$ and $\bm{v}^{p}(t)$ are the particle position and velocity vectors and $\bm{u}(\bm{x}^{p}(t),t)$ is the fluid velocity at the particle position which is calculated using an eight-point B-spline interpolation [@vanhinsberg12]. Fluid particles were also tracked by solving $\dot{\bm{x}}^{f}(t)=\bm{u}(\bm{x}^{f}(t),t)$. A total of $18$ different particle classes were simulated, with Stokes numbers ranging from $0$ to $30$. About $17$ million particles were tracked for each value of $St$, for a total of $300$ million particles. At the initial time, particles were injected in the flow with a uniform distribution. The particles were allowed to equilibrate with the statistically stationary flow field for about $5$ large-eddy turnover times before we began gathering statistics. Measurement of the particle radial distributions and velocities confirmed that the particle field had reached a statistically stationary state after this development time. The mean-square separation calculations are carried out over a total time of $100 \tau_\eta$, or about $1.6$ large eddy turnover times, and particle positions and velocities were stored approximately every $0.1 \tau_\eta$. Since we are considering an isotropic system, the statistics depend only upon on the separation magnitude $r^0\equiv\vert\bm{r}^0\vert$. The results in Fig. \[FP\_ratio\] show the ratio of the BIT to FIT mean square separation, and the results clearly show that the fluid particle dispersion is irreversible, with BIT faster than FIT dispersion. The results also show that the peak in the ratio increases with decreasing $r^0$, indicating that the irreversibility in the dispersion becomes stronger as one goes to smaller scales. This is consistent with the explanation given for the irreversibility in §\[FPIM\] since, as shown in Figure \[Skew\_plot\], the skewness of the field $\Delta\bm{u}(\bm{r},t)$ becomes stronger with decreasing $\bm{r}$. It is worth pointing out that in our DNS, the dissipation spectra peaks at wavenumber $k\approx 0.12/\eta$, which roughly translates to $r\approx 37\eta$, yet the results in figure \[FP\_ratio\] show that the strength of the irreversibility of the dispersion continues to increase as $r^0$ is decreased below $\eta$. This serves to emphasize that it is not the amount of local dissipation, *per-se*, that controls the irreversibility of the dispersion, but rather the local asymmetry of the distribution of $\Delta\bm{u}$. The results in Fig. \[FP\_ratio\] also show that the time it takes for the ratio to begin to increase from unity increases with increasing $r^0$. This is because the initial ballistic motion, in which the separation is time reversible, persists for longer times as $r^0$ is increased, because $\tau_{r^0}$ increases with increasing $r^0$. For $r^0>L$, where $L$ is the integral length scale of the flow ($L/\eta\approx 800$ for this DNS), the ratio would be unity as explained earlier. In Fig. \[Bal\_scaled\] we plot the DNS data for $(\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}-\vert \bm{r}^{0}\vert^{2})/\tau_{\eta}^{2}\langle\vert\Delta\bm{u}^{f}(0)\vert\rangle_{\bm{r}^{0}}$ and $(\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}-\vert \bm{r}^{0}\vert^{2})/\tau_{\eta}^{2}\langle\vert\Delta\bm{u}^{f}(0)\vert\rangle_{\bm{r}^{0}}$ (here and throughout, we use the DNS data for $\langle\vert\Delta\bm{u}^{f}(0)\vert\rangle_{\bm{r}^{0}}$). The data shows a good collapse for small $\mathcal{T}$ for both the FIT and BIT cases, demonstrating the accuracy of (\[Bal\]) for small $\mathcal{T}$, and also that the duration of the ballistic regime increases with increasing $r^{0}$. Note, however, that in agreement with Fig. 4 in [@ouellette06c] we find in Fig. \[Bal\_scaled\] (a) that for the larger initial separations, the data shows that the growth slows down for a time after an initially ballistic separation. In Fig. \[Bal\_scaled\_tB\] we plot the fluid particle mean-square separation (with the initial separation subtracted) scaled by the ballistic prediction and plotted against $\mathcal{T}/\tau_{r^{0}}$, i.e. with time scaled by the Batchelor timescale. In agreement with the experimental data in [@ouellette06c], we find in Fig. \[Bal\_scaled\_tB\] (a) that the FIT dispersion of the particles is sub-ballistic for some time after $\mathcal{T}=\mathcal{O}(\tau_{r^{0}}/10)$. The authors in [@ouellette06c] argue that this slowing down is not due to higher-order correction terms in the small-time series expansion, but argue that it is more likely explained in terms of the effect of the large scales on the separation. In our case the influence of the large scales cannot be the explanation: At these times the particle separations are much smaller than the integral length scale and our data shows a temporary slowing down but then a speeding up towards a faster than ballistic separation (clearly observable in Fig. \[Bal\_scaled\_tB\] (a) for $r^{0}\in[8\eta,10\eta]$ and $r^{0}\in[20\eta,25\eta]$) that cannot be caused by the large scales. A possible explanation for the temporary slowing down of the separation is the effect of the decorrelation of $\Delta\bm{u}^f$ along the pair trajectory (which the ballistic regime neglects through its use of ${\Delta\bm{u}^{f}(\mathcal{T})\approx\Delta\bm{u}^{f}(0)}$), which is subsequently overcome by the growth of the autocovariances of $\Delta\bm{u}^{f}$ in the inertial regime. The BIT results in Fig. \[Bal\_scaled\_tB\] (b) also show the slowing down but to a lesser degree. The BIT data for $r^{0}\in[40\eta,50\eta]$ and $r^{0}\in[80\eta,100\eta]$ shows that after the initial ballistic separation their separation slows down, speeds up again to a faster than ballistic separation behavior and then finally slows down again. This final stage of separation is likely due to the large scales since at these times the particle separations exceed the integral length scale, at which point the growth follows a diffusive law (corresponding to a line $\propto\mathcal{T}^{-1}$ in Fig. \[Bal\_scaled\_tB\]). It is however also possible that at the largest values of $\mathcal{T}$, the results are affected by the finite box size and the periodic boundary conditions used in the DNS. We now consider in more depth the case where $\bm{r}^{f}(\mathcal{T})$ lies in the dissipation range. In §\[FPD\] we argued that for $\bm{r}^{f}(\mathcal{T})$ in the dissipation range, the pair separation begins with a ballistic separation growth and then may transition to some form of exponential growth. This is in contrast to the experimental results in [@ni13] where they argue based on their data that the growth is first exponential and then ballistic. The results in figure \[EXP\_test\] show that for $\mathcal{T}\leq\mathcal{O}(\tau_\eta)$ the dispersion in the dissipation range is described well by the ballistic law. There are slight departures for very small $\mathcal{T}/\tau_\eta$, however, these could be caused by noise in the data which is amplified by the fact that the denominator in the plotted expression tends to zero in the limit $\mathcal{T}\to0$. These results should be contrasted with those of Fig. 5(b) of [@ni13] where the same quantity is plotted: they observe that for $r=\mathcal{O}(\eta)$ and $\mathcal{T}=0.1\tau_\eta$, the quantity is $\mathcal{O}(10)$, whereas in our data it is $\mathcal{O}(1)$. In figure \[EXP\_test2\] we plot$$\dfrac{d}{d\mathcal{T}}\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\Big/\Big\langle\vert\Delta\bm{u}^{f}(0)\vert\Big\rangle_{\bm{r}^{0}},$$(and also the BIT equivalent) as a further test to examine departures from the ballistic law at small-times. Recall that the Batchelor-type exponential growth grows with $\mathcal{T}$ in the small-$\mathcal{T}$ regime, and so the plotted quantity would be a constant at small-times if the Batchelor exponential growth were correct. The data for this quantity shows a strong $\propto\mathcal{T}$ scaling at small-times, confirming the ballistic law and ruling out an exponential growth of the Batchelor kind at small-times, consistent with theoretical expectations. These results then call into question the findings in [@ni13]. While we are uncertain as to the full explanation for the discrepancies, we note that the discrepancy cannot arise simply as a consequence of the difference between the types of turbulent flow field that we are considering (their experimental flow is a turbulent thermal convective flow), since the ballistic law follows in the dissipation range from the small-time approximation $\Delta\bm{u}^f(\mathcal{T})\approx\Delta\bm{u}^f(0)+\mathcal{O}(\mathcal{T}/\tau_\eta)$, for arbitrary statistical properties of the field $\Delta\bm{u}(\bm{r},t)$. One possible explanation for the $\propto\mathcal{T}$ growth observed in [@ni13] concerns whether or not the tracer particles in their experiment were fully-mixed, that is, whether ${\bm{\nabla}_{\bm{r}^0}}\langle\delta(\bm{r}^f(0)-\bm{r}^0)\rangle=\bm{0}$. In the general case where ${\bm{\nabla}_{\bm{r}^0}}\langle\delta(\bm{r}^f(0)-\bm{r}^0)\rangle\neq\bm{0}$, the FIT small-time fluid particle mean-square dispersion is (under the approximation $\Delta\bm{u}^f(\mathcal{T})\approx\Delta\bm{u}^f(0)+\mathcal{O}(\mathcal{T}/\tau_\eta)$) $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}+2\mathcal{T}\bm{r}^0\bm{\cdot}\Big\langle\Delta\bm{u}^{f}(0)\Big\rangle_{\bm{r}^0}+\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\label{BalNFM}. \end{aligned}$$ For fully-mixed fluid particles, $\langle\Delta\bm{u}^{f}(0)\rangle_{\bm{r}^0}=\bm{0}$, and (\[BalNFM\]) reduces to (\[Bal\]). For non-fully-mixed fluid particles, $\langle\Delta\bm{u}^{f}(0)\rangle_{\bm{r}^0}\neq\bm{0}$ which gives$$\lim_{\mathcal{T}/\tau_\eta \to0}\Bigg[\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^0\vert^{2}\Bigg]\propto\mathcal{T},$$consistent with the observation in [@ni13]. Furthermore, in [@ni13] they also consider the quantity $\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\rangle_{\bm{r}^0}$. Applying the approximation $\Delta\bm{u}^f(\mathcal{T})\approx\Delta\bm{u}^f(0)+\mathcal{O}(\mathcal{T}/\tau_\eta)$ to the evolution equation governing $\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\rangle_{\bm{r}^0}$ we obtain $$\begin{aligned} \Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\Big\rangle_{\bm{r}^0}\approx\mathcal{T}^{2}\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0},\label{BalNFM2} \end{aligned}$$ which applies *irrespective* of whether the fluid particles are fully-mixed. The quantity $\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$ is in principle different for fully-mixed and non-fully-mixed fluid particles, however this does not change the $\mathcal{T}$ dependence of $\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\rangle_{\bm{r}^0}$. The important point then is that whereas in the fully-mixed case (see §\[FPDT\])$$\lim_{\mathcal{T}/\tau_\eta \to0}\Bigg[\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^0\vert^{2}\Bigg]\propto\mathcal{T}^2,\quad \lim_{\mathcal{T}/\tau_\eta \to0}\Bigg[\Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\Big\rangle_{\bm{r}^0}\Bigg]\propto\mathcal{T}^2$$ in the non-fully mixed case$$\lim_{\mathcal{T}/\tau_\eta \to0}\Bigg[\Big\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-\vert\bm{r}^0\vert^{2}\Bigg]\propto\mathcal{T},\quad \lim_{\mathcal{T}/\tau_\eta \to0}\Bigg[\Big\langle\vert\bm{r}^{f}(\mathcal{T})-\bm{r}^0\vert^{2}\Big\rangle_{\bm{r}^0}\Bigg]\propto\mathcal{T}^2.$$These non-fully-mixed predictions seem very close to the scalings observed for the two quantities in fig.5a and fig.5b of [@ni13] for $r^0$ in the dissipation range. This suggests the possibility that the scaling they observed is not evidence of exponential growth in the dissipation range at small-times, but is rather a scaling arising from the dispersion of fluid particles that are not fully-mixed. In the experiments of [@ni13], after the particles were introduced into the system, a time period of approximately $100$ large scale eddy turnover times was allowed to elapse before the dispersion statistics were recorded, which would be expected to provide sufficient time for the tracer particles to fully mix throughout the flow [@ni15]. However, in their experiments, thermal plumes can rise from the bottom of the system and may bring with them an increased concentration of particles into the observation volume, thus destroying the well-mixedness of the system [@ni15]. Given that the results in Fig.5 of [@ni13] which appear to be affected by non-well-mixedness are for $r^0=\mathcal{O}(\eta)$, and that the thermal plumes in the experiment have cross-sectional sizes $\mathcal{O}(\eta)$, this may well provide a plausible explanation for the aforementioned discrepancies. Future experimental efforts are required to test whether or not this is in fact the case. In order to consider whether the fluid pairs undergo a Batchelor-type exponential growth subsequent to the initial ballistic growth, in figure \[EXP\_test3\] we plot $\mathcal{T}^{-1}\ln(\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\vert\bm{r}^{0}\vert^{2})$ (and the equivalent BIT version), which would be constant in a Batchelor type exponential growth regime. The results do not reveal any evidence of an exponential growth, for either the FIT and BIT case. It is, however, possible that our initial separations are simply not small enough in order for the pairs to remain in the dissipation range at $\mathcal{T}/\tau_\eta\geq\mathcal{O}(1)$. We now consider the case where at sufficiently large $\mathcal{T}$ the particle separation $\bm{r}^f(\mathcal{T})$ lies in the inertial range, in which case the mean-square separation is predicted to follow the RO $\mathcal{T}^3$ law for arbitrary $r^0$ when $Re_\lambda\to\infty$. In Fig. \[RO\_scaling\] we plot $\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\epsilon\rangle\mathcal{T}^{3}$ and $\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\epsilon\rangle\mathcal{T}^{3}$ for various $r^{0}$ in order to see whether the data shows an approach to RO scaling. For $r^{0}\in[3\eta,4\eta]$ the data shows a clear convergence to RO scaling in both the FIT and BIT cases, yielding values of $\mathfrak{g}^{F}$ and $\mathfrak{g}^{B}$ in excellent agreement with experimental data [@berg06a]. For $r^{0}\leq\eta$ the data indicates that for $\mathcal{T}\gtrsim5\tau_{\eta}$ the particles separate faster than RO scaling (indicated by the positive slope for $\mathcal{T}\gtrsim5\tau_{\eta}$). We expect this is due to the influence of their separation in the dissipation range because of the finite temporal correlation radius of the field $\Delta\bm{u}$. For separations larger than $r^{0}\in[3\eta,4\eta]$ the fluid mean square separation is slower than RO scaling (indicated by the negative slope) throughout the range of $\mathcal{T}$ for which we have data. The curves do however seem to be tending to RO scaling at the largest values of $\mathcal{T}$. For $r^{0}$ in the inertial regime we would expect an initial ballistic separation followed by RO with a transition region in between. For separations larger than $r^{0}\in[3\eta,4\eta]$ it is likely that our time span of $0\leq\mathcal{T}\leq100\tau_{\eta}$ only extends to the transition region and hence we do not observe RO scaling. To observe RO scaling over a larger range of $r^{0}$ we would need a DNS with significantly larger $Re_{\lambda}$. The results in Fig. \[RO\_scaling\] (a) are very similar to those in Fig. 4 (a) of [@sawford08]. Note also that our data agrees with the findings in [@bitane12] that $r^{0}\approx 4\eta$ could be an “optimal choice” for the initial separation to observe RO scaling. Inertial particle dispersion {#IPD} ============================ Having considered the dispersion of fluid particles, we now consider the more complex scenario of the dispersion of inertial particles. Irreversibility mechanisms {#IPIM} -------------------------- Before proceeding to derive predictions for the BIT inertial particle dispersion, we first consider how particle inertia gives rise to an additional source of irreversibility in the dispersion process. The equations governing the time evolution of the fluid and inertial particle-pair separations may be written as (ignoring initial conditions) $$\begin{aligned} \dot{\bm{r}}^f(t)&=\Delta\bm{u}(\bm{r}^f(t),t),\label{rdotF}\\ \dot{\bm{r}}^p(t)&=(St\tau_\eta)^{-1}\int\limits_0^t\dot{G}(t-s)\Delta\bm{u}(\bm{r}^p(s),s)\,ds.\label{rdotP}\end{aligned}$$ Recall that in FIT dispersion the pairs are (on average) going to greater separations as time increases (i.e. $\dot{\bm{r}}^f(t)>\bm{0}$), whereas in BIT dispersion the pairs are going to smaller separations as time increases (i.e. $\dot{\bm{r}}^f(t)<\bm{0}$). Just as for the fluid particles, inertial particles experience a local source of irreversibility related to the asymmetry in the probability distribution of $\Delta\bm{u}$ (see §\[FPIM\]). However, the inertial particle separation described by (\[rdotP\]) contains an additional effect. Whereas $\dot{\bm{r}}^f(t)$ is entirely governed by the local turbulent field $\Delta\bm{u}$, (\[rdotP\]) shows that $\dot{\bm{r}}^p(t)$ is influenced by the behavior of $\Delta\bm{u}$ along the path-history of the particle pair for times up to $t-s=\mathcal{O}(\tau_p)$ in the past. This non-local dependence of $\dot{\bm{r}}^p(t)$ on $\Delta\bm{u}$ gives rise to an additional source of irreversibility: In FIT dispersion (statistically speaking) $\dot{\bm{r}}^f(t)>\bm{0}$ and ${\bm{r}}^f(t)>{\bm{r}}^f(s)$, while in BIT dispersion $\dot{\bm{r}}^f(t)<\bm{0}$ and ${\bm{r}}^f(t)<{\bm{r}}^f(s)$. Since $\langle\vert\Delta\bm{u}(\bm{r},t)\vert^2\rangle$ increases with $\bm{r}$, (\[rdotP\]) implies that FIT-separating pairs are influenced by their memory of smaller $\Delta\bm{u}$ in their path-history, whereas BIT-separating pairs are influenced by their memory of larger $\Delta\bm{u}$ in their path-history. This enhances the discrepancy between FIT and BIT dispersion. Note that this irreversibility mechanism is intimately connected to the non-local clustering mechanism that dominates the clustering of inertial particles for ${St\gtrsim\mathcal{O}(1)}$ [@bragg14b; @bragg14d]. In general the inertial particles are affected by both sources of irreversibility, and we expect that there will be a value of the particle inertia for which their dispersion is optimally affected by both sources of irreversibility, likely $\tau_p/\tau_{r^0}=\mathcal{O}(1)$. We now make several remarks and observations about this non-local irreversibility mechanism. First, it vanishes in the limit $\tau_p\to0$ where the particles separation is entirely governed by the local turbulence. This is represented in (\[rdotP\]) through the memory kernel $\dot{G}(t-s)$, which vanishes for $t\neq s$ in the limit $\tau_p\to0$. Second, this mechanism operates and generates irreversibility in inertial particle dispersion even in fluid velocity fields which are reversible (i.e. symmetric PDF for $\Delta\bm{u}$). The operation of the mechanism only depends upon finite $\tau_p$ and $\bm{\nabla_r}\langle\vert\Delta\bm{u}(\bm{r},t)\vert^2\rangle\neq\bm{0}$, the latter being true for any spatially correlated fluid velocity field at $\vert\bm{r}\vert<L$. Third, for $\vert\bm{r}\vert>L$ the motion of the two particles becomes uncorrelated and the pair dispersion becomes proportional to the one-particle dispersion, and this dispersion is time-reversible in a stationary, homogeneous velocity field. The non-local irreversibility mechanism is consistent with the reversibility of the one-particle dispersion, since at $\vert\bm{r}\vert>L$, $\Delta\bm{u}(\bm{r},t)$ is statistically independent of $\bm{r}$, and therefore the path-history interaction of separating and approaching pairs with $\Delta\bm{u}$ become statistically equivalent, restoring FIT/BIT symmetry in this regime. This also serves to emphasize that the statistical time-irreversibility of the particle pair dispersion does not arise simply as a consequence of the dissipative dynamics of the particles, since this would suggest that one-particle dispersion should be irreversible in stationary, homogeneous turbulence, which it is not. Fourth, since $\bm{\nabla_r}\langle\vert\Delta\bm{u}(\bm{r},t)\vert^2\rangle>\bm{0}$ for $\vert\bm{r}\vert<L$ is simply a consequence of spatial decorrelation in the fluid velocity field, the non-local irreversibility mechanism always generates faster BIT than FIT dispersion, whether the turbulence be 2D or 3D. However, the irreversibility mechanism associated with the turbulence dynamics depends upon the direction of the energy transfer in the velocity field, as manifested in the nature of the asymmetry of the PDF of $\Delta\bm{u}$. As discussed earlier, since $\Delta\bm{u}$ is positively skewed in 2D turbulence because of its inverse energy transfer process, then the dispersion of fluid particles is faster for FIT than for BIT, as confirmed in the numerical simulations in [@faber09]. This then leads to an interesting prediction for relative dispersion in 2D turbulence, namely, that below some critical $St$, FIT is faster than BIT dispersion, but then beyond this critical $St$ value, BIT is faster than FIT dispersion. This critical $St$ marks the point at which the non-local irreversibility mechanism begins to dominate over the irreversibilty mechanism associated with the turbulence dynamics. We are currently testing this prediction. Theoretical results {#theoretical-results} ------------------- Having considered the additional irreversibility mechanism that arises because of the particle inertia, we now construct theoretical predictions to describe the inertial particle dispersion. In deriving the results we will need to know something about the particle relative velocity statistics. In the following, we derive results for the dispersion given only the particle velocity statistics at $\mathcal{T}=0$. The results then describe, given the statistical state of the particles at $\mathcal{T}=0$, how the pairs disperse as a function of $\mathcal{T}$. Since the FIT behavior has already been analyzed in [@bec10b], we shall focus on developing a theory for the BIT mean-square dispersion. We begin by considering the regime $\mathcal{T}/\tau_{r^0}\ll1$, for which we may invoke the approximation $$\begin{aligned} \Delta\bm{u}^{p}(\mathcal{T})\approx\Delta\bm{u}^{p}(0)+\mathcal{O}(\mathcal{T}/\tau_{r^0}),\label{uTexp2}\end{aligned}$$ (and similarly for the terms in (\[FITmsr2\]) and (\[BITmsr2\]) involving $s$ and $s^\prime$ in the time arguments), which was also used to derive the fluid particle small-time ballistic separation prediction (see §\[FPDT\]). Introducing this approximation into (\[FITmsr2\]) and (\[BITmsr2\]) and solving the integrals, we obtain $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}&+G^{2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}+2G(\mathcal{T})\Big[\mathcal{T}-G(\mathcal{T})\Big]\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}\\ &+\Big[\mathcal{T}^{2}-2\mathcal{T}G(\mathcal{T})+G^{2}(\mathcal{T})\Big]\Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}, \label{DispFIT_r0} \end{split}\end{aligned}$$ and $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}&+G^{2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}+2G(\mathcal{T})\Big[\mathcal{T}-G(\mathcal{T})\Big]\Big\langle\bm{w}^{p}(-\mathcal{T})\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}\\ &+\Big[\mathcal{T}^{2}-2\mathcal{T}G(\mathcal{T})+G^{2}(\mathcal{T})\Big]\Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}. \label{DispBIT_r0} \end{split}\end{aligned}$$ Like the fluid particle ballistic result in (\[Bal\]), (\[DispFIT\_r0\]) and (\[DispBIT\_r0\]) are valid for any initial separation $\bm{r}^0$. We will return shortly to consider the range of $\mathcal{T}$ for which these results should be valid. In [@bec10b] the FIT result $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}+G^{2}(\mathcal{T})\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$ was derived. Our result in (\[DispFIT\_r0\]) contains this contribution, but is more general, capturing the influence of the local fluid velocity field on the dispersion, which is important for $St\lesssim\mathcal{O}(1)$. We must now verify that the results in (\[DispFIT\_r0\]) and (\[DispBIT\_r0\]) obey the necessary limiting cases. First, it is simple to confirm that (\[DispFIT\_r0\]) and (\[DispBIT\_r0\]) reduce to (\[Bal\]) in the case when $\tau_{p}=0$ (for which $G(\mathcal{T})=0$ and $\Delta\bm{u}^{p}=\Delta\bm{u}^{f}$). Second, in the limit $Re_{\lambda}\to\infty$ and sufficiently large $\bm{r}^{0}$, the inertial particle behavior should tend to that for fluid particles, since for a given $\tau_{p}$, $\tau_{p}/\tau_{r^{0}}\to0$ as $\vert\bm{r}^{0}\vert\to\infty$. In this limit, we would have $\bm{w}^{p}\to\Delta\bm{u}^{f}$ and $\Delta\bm{u}^{p}\to\Delta\bm{u}^{f}$; if we introduce these into (\[DispFIT\_r0\]) and (\[DispBIT\_r0\]), we once again find that the results reduce to (\[Bal\]). The result in (\[DispBIT\_r0\]) contains $\bm{w}^{p}(-\mathcal{T})$; we wish to derive theoretical descriptions that require only knowledge of the particle velocity statistics at $\mathcal{T}=0$. The solution for $\bm{w}^{p}(-\mathcal{T})$ is $$\begin{aligned} \bm{w}^{p}(-\mathcal{T})=\dot{G}^{-1}(\mathcal{T})\bm{w}^{p}(0)-\tau_{p}^{-1}\dot{G}^{-1}(\mathcal{T})\int\limits^{0}_{-\mathcal{T}}\dot{G}(-s)\Delta\bm{u}^{p}(s)ds, \label{wsol}\end{aligned}$$ where $\dot{G}(-s)=\exp[\tau_{p}^{-1}s]$. From (\[wsol\]) we obtain the following $$\begin{aligned} \begin{split} \Big\langle\vert\bm{w}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}&=\dot{G}^{-2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}-2\tau_{p}^{-1}\dot{G}^{-2}(\mathcal{T})\int\limits^{0}_{-\mathcal{T}}\dot{G}(-s)\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &\quad+\tau_{p}^{-2}\dot{G}^{-2}(\mathcal{T})\int\limits^{0}_{-\mathcal{T}}\int\limits^{0}_{-\mathcal{T}}\dot{G}(-s)\dot{G}(-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds. \label{w2sol} \end{split}\end{aligned}$$ Introducing into this (\[uTexp2\]) and evaluating the integrals, we obtain $$\begin{aligned} \begin{split} \Big\langle\vert\bm{w}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}&\approx\dot{G}^{-2}(\mathcal{T})\Big\langle\bm{w}^{p}(0)\bm{\cdot}\bm{w}^{p}(0)\Big\rangle_{\bm{r}^0}-2\tau_{p}^{-1}\dot{G}^{-2}(\mathcal{T})G(\mathcal{T})\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}\\ &\quad+\tau_{p}^{-2}\dot{G}^{-2}(\mathcal{T})G^{2}(\mathcal{T})\Big\langle\Delta\bm{u}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}. \label{w2sol2} \end{split}\end{aligned}$$ In a similar manner, we also obtain $$\begin{aligned} \begin{split} \Big\langle\bm{w}^{p}(-\mathcal{T})\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}&\approx\dot{G}^{-1}(\mathcal{T})\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}-\tau_{p}^{-1}\dot{G}^{-1}(\mathcal{T})G(\mathcal{T})\Big\langle\Delta\bm{u}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}. \label{wusol} \end{split}\end{aligned}$$ Using these results in (\[DispBIT\_r0\]), we obtain $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}&+G^{2}(-\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}-2G(-\mathcal{T})\Big(G(-\mathcal{T})+\mathcal{T}\Big)\Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}\\ &+\Big[G^{2}(-\mathcal{T})+2\mathcal{T}G(-\mathcal{T})+\mathcal{T}^{2}\Big]\Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}. \label{DispBIT_r02a} \end{split}\end{aligned}$$ In addition, since we do not in general know the statistics of $\Delta\bm{u}^p$, we make the approximation $$\begin{aligned} \Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}&\approx\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0},\label{upufA}\end{aligned}$$ and also $$\begin{aligned} \Big\langle\bm{w}^{p}(0)\bm{\cdot}\Delta\bm{u}^{p}(0)\Big\rangle_{\bm{r}^0}&\approx\sqrt{\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\sqrt{\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}},\end{aligned}$$ which finally gives us $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\vert\bm{r}^0\vert^{2}&+G^{2}(-\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}+\Big[G^{2}(-\mathcal{T})+2\mathcal{T}G(-\mathcal{T})+\mathcal{T}^{2}\Big]\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\\ &-2G(-\mathcal{T})\Big[G(-\mathcal{T})+\mathcal{T}\Big]\sqrt{\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\sqrt{\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}. \label{STT} \end{split}\end{aligned}$$ Note that in the expression $\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$, the conditionality is $\bm{r}^f(0)=\bm{r}^{0}$ (not $\bm{r}^p(0)=\bm{r}^{0}$), such that ${\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}=\langle\vert\Delta\bm{u}(\bm{r}^0,0)\vert^{2}\rangle}$ since the fluid particles are fully mixed. It is possible that (\[STT\]) may not be accurate for small $St$, since we have partially removed the effect of the preferential sampling of $\Delta\bm{u}$ by the inertial particles by invoking approximation (\[upufA\]) (some of the effect is captured within $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$). We consider the effects of this on (\[STT\]) for low $St$ particles in §\[IPS\]. The result in (\[STT\]) implies that the time dependance of the mean-square dispersion BIT depends not only upon $St$ but also $\bm{r}^0$, which is distinct from the ${St=0}$ case where the mean-square dispersion in the small-time regime grows as $\mathcal{T}^2$ for any $\bm{r}^0$. In particular, in the dissipation regime where $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}\propto\vert\bm{r}^0\vert^{\xi}$ with $\xi(St)\leq 2$ [@gustavsson11; @salazar12a], then for a given finite $St$, $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ may grow like $\mathcal{T}^2$ at larger separations (where $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}=\mathcal{O}(1)$) but grow like $G^2(-\mathcal{T})$ in the limit $\vert\bm{r}^0\vert/\eta\to0$. Formally, the range of $\mathcal{T}$ over which (\[STT\]) should remain valid is determined by the approximation in (\[uTexp2\]). Since this approximation leads to a good description of the fluid particle dispersion up to ${\mathcal{T}=\mathcal{O}(\tau_{r^0})}$, then we may expect that (\[STT\]) should remain valid for ${\mathcal{T}\leq\mathcal{O}(\tau_{r^0})}$. There is, however, an exception to this: In regions where $\bm{w}^p\gg\Delta\bm{u}^p$, which we refer to as ‘caustic regions’ [@wilkinson05; @bec10a; @salazar12a], $\Delta\bm{u}^p$ is irrelevant to the particle dispersion process, and so the range of the validity of (\[uTexp2\]) does not control the range of the validity of (\[STT\]). In this case, the range of the validity of (\[STT\]) is controlled by how long $\bm{w}^p\gg\Delta\bm{u}^p$ along the pair trajectory. We expect that this time should be $\mathcal{O}(\tau_p)$, reflecting the time it takes for the particles to dissipate their excess kinetic energy relative to that of the local fluid velocity difference field. Therefore, (\[STT\]) should remain valid for ${\mathcal{T}\leq\widehat{\mathcal{T}}}$ where $$\begin{aligned} \widehat{\mathcal{T}}= \left\{ \begin{array}{ll} \tau_p & \mbox{if } \langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}\gg\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}\\ \tau_{r^0} & \mbox{otherwise.} \end{array} \right.\end{aligned}$$ We now consider the regime ${\mathcal{T}>\widehat{\mathcal{T}}}$. If we ignore the terms involving the mean of $\bm{w}^p$ and $\Delta\bm{u}^p$, under the assumption that their contribution is small relative to the other terms, we may re-write (\[BITmsr2\]) for $\mathcal{T}>\widehat{\mathcal{T}}$ as $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx\, &\Big\langle\vert\bm{r}^{p}(-\widehat{\mathcal{T}})\vert^{2}\Big\rangle_{\bm{r}^0}+G^{2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}-G^{2}(\widehat{\mathcal{T}})\Big\langle\vert\bm{w}^{p}(-\widehat{\mathcal{T}})\vert^{2}\Big\rangle_{\bm{r}^0}\\ &+2\tau_{p}^{-1}G(\mathcal{T})\int\limits^{-\widehat{\mathcal{T}}}_{-\mathcal{T}}G(-s)\Big\langle\bm{w}^{p}(-\mathcal{T})\bm{\cdot}\Delta \bm{u}^{p}(s)\Big\rangle_{\bm{r}^0}ds\\ &+\tau_{p}^{-2}\int\limits^{-\widehat{\mathcal{T}}}_{-\mathcal{T}}\int\limits^{-\widehat{\mathcal{T}}}_{-\mathcal{T}}G(-s)G(-s^\prime)\Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}ds^\prime\,ds. \label{BITmsr2LargeT} \end{split}\end{aligned}$$ Determining the appropriate closure approximations to apply to (\[BITmsr2LargeT\]) depends upon both $St$ and $r^0$. For example, for $r^0\ll\eta$, the particles may be either still in the dissipation regime or in the inertial regime at time $\widehat{\mathcal{T}}$, depending upon the value of $St$. We will assume that at time $\widehat{\mathcal{T}}$ the pairs are in the inertial regime and leave the other case for future work (especially since, as shown in §\[FPS\], even the fluid particle separation in the dissipation range at finite $\mathcal{T}$ is not fully understood). Let us define a time-dependent Stokes number in the inertial range as $St_r(t)\equiv \tau_p/\tau_{r}$ where $\tau_{r}\equiv (\vert\bm{r}^p(t)\vert^2/\langle\epsilon\rangle)^{1/3}$, which satisfies ${St_r(t)\to0}$ for ${\vert\bm{r}^p(t)\vert/\eta\to\infty}$ in the limit ${Re_\lambda\to\infty}$. This implies that the effect of the particles’ inertia becomes perturbative as their separation growth increases such that ${\bm{w}^{p}(t)\approx\Delta\bm{u}^{p}(t)+\mathcal{O}(St_r(t))}$. We will make the approximation that $\bm{w}^{p}\approx\Delta\bm{u}^{p}$ for $\mathcal{T}>\widehat{\mathcal{T}}$. We then need to describe $\Delta\bm{u}^{p}$ in the regime $St_r(t)\ll1$, and to do this we introduce the expansion $$\begin{aligned} \Delta\bm{u}^p(t)=\Delta\bm{u}^{[0]}(t)+St_r(t)\Delta\bm{u}^{[1]}(t)+\mathcal{O}([St_r(t)]^2),\label{DupPer}\end{aligned}$$ where the superscripts $[0],[1]$ denote the order of the perturbation term and $\Delta\bm{u}^{[0]}(t)\equiv\Delta\bm{u}^{f}(t)$. Under the approximation $\bm{w}^{p}\approx\Delta\bm{u}^{p}$, the unknown terms in (\[BITmsr2LargeT\]) all involve autocovariances of $\Delta\bm{u}^{p}$, and using (\[DupPer\]), these are expressed as $$\begin{aligned} \begin{split} \Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}=&\quad\Big\langle\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{f}(s^\prime)\Big\rangle_{\bm{r}^0}+\Big\langle[\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{[1]}(s^\prime)]St_r(s^\prime)\Big\rangle_{\bm{r}^0}\\ &+\Big\langle[\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{[1]}(s^\prime)]St_r(s)\Big\rangle_{\bm{r}^0}+\mathcal{O}\Big(St_r(s)St_r(s^\prime)\Big).\label{DupPerCov} \end{split} \end{aligned}$$ The terms involving $\mathcal{T}$ instead of $s$ in the time arguments are treated similarly. Making the crude approximation that ${\vert\bm{r}^p(s)\vert^2\approx-\mathfrak{g}^{B}\langle\epsilon\rangle s^3}$ (for ${s<0}$), we have ${St_r(s)\approx -St\tau_\eta[\mathfrak{g}^{B}]^{-1/3}s^{-1}}$. Using K41 arguments for the Lagrangian behavior of $\Delta\bm{u}$ in the inertial range, we have $$\begin{aligned} \Big\langle\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{f}(s^\prime)\Big\rangle_{\bm{r}^0}\approx&-(1/2)\mathcal{A}^{[0]}\langle\epsilon\rangle(s+s^\prime),\quad\text{for}\quad s<0,\, s^\prime<0,\label{LK41a}\\ \Big\langle\Delta\bm{u}^{f}(s)\bm{\cdot}\Delta\bm{u}^{[1]}(s^\prime)\Big\rangle_{\bm{r}^0}\approx&-(1/2)\mathcal{A}^{[1]}\langle\epsilon\rangle(s+s^\prime),\quad\text{for}\quad s<0,\, s^\prime<0,\label{LK41b}\end{aligned}$$ and substituting these into (\[DupPerCov\]), we have $$\begin{aligned} \begin{split} \Big\langle\Delta\bm{u}^{p}(s)\bm{\cdot}\Delta\bm{u}^{p}(s^\prime)\Big\rangle_{\bm{r}^0}\approx-(1/2)\mathcal{A}^{[0]}\langle\epsilon\rangle(s+s^\prime)+(1/2)St\tau_\eta\langle\epsilon\rangle\mathcal{A}^{[1]}[\mathfrak{g}^{B}]^{-1/3}\Bigg(\frac{s+s^\prime}{s}+\frac{s+s^\prime}{s^\prime}\Bigg),\label{DupPerCov2} \end{split} \end{aligned}$$ where $\mathcal{A}^{[0]}$ and $\mathcal{A}^{[1]}$ are positive (so that (\[LK41a\]) and (\[LK41b\]) increase backward-in-time in the inertial range, as we would expect), dimensionless constants to be determined later. If we now substitute (\[DupPerCov2\]) into (\[BITmsr2LargeT\]), invoking $\bm{w}^{p}\approx\Delta\bm{u}^{p}$ and $G(\mathcal{T}\geq\widehat{\mathcal{T}})\approx\tau_{p}$, and solve the integrals, we obtain $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^{0}}&\approx\Big\langle\vert\bm{r}^{p}(-\widehat{\mathcal{T}})\vert^{2}\Big\rangle_{\bm{r}^{0}} +\tau_{p}^{2}\mathcal{A}^{[0]}\langle\epsilon\rangle\Big(\mathcal{T}-\widehat{\mathcal{T}}\Big)+\tau_{p}(1/2)\mathcal{A}^{[0]}\langle\epsilon\rangle\Big(3\mathcal{T}^{2}-\widehat{\mathcal{T}}^{2}-2\widehat{\mathcal{T}}\mathcal{T}\Big)\\ &\quad+(St\tau_\eta)^2\langle\epsilon\rangle\mathcal{A}^{[1]}[\mathfrak{g}^B]^{-1/3}\Big(2(\mathcal{T}-\widehat{\mathcal{T}})+(1/2\mathcal{T})(\mathcal{T}^2-\widehat{\mathcal{T}}^2)+\mathcal{T}\ln[\mathcal{T}/\widehat{\mathcal{T}}]\Big)\\ &\quad+(1/2)St\tau_\eta\langle\epsilon\rangle\mathcal{A}^{[1]}[\mathfrak{g}^B]^{-1/3}\Big(2(\mathcal{T}-\widehat{\mathcal{T}})^2+(\mathcal{T}^2-\widehat{\mathcal{T}}^2)\ln[\mathcal{T}/\widehat{\mathcal{T}}]\Big)\\ &\quad+(1/2)\mathcal{A}^{[0]}\langle\epsilon\rangle\Big(\mathcal{T}^{3}+\widehat{\mathcal{T}}^{3}-\widehat{\mathcal{T}}^{2}\mathcal{T}-\widehat{\mathcal{T}}\mathcal{T}^{2}\Big),\quad\text{for $\mathcal{T}>\widehat{\mathcal{T}}$}. \end{split} \label{LTT}\end{aligned}$$ Taking the limit $\tau_{p}\to0$ and considering the regime $\mathcal{T}\gg\widehat{\mathcal{T}}$, we may identify $(1/2)\mathcal{A}^{[0]}$ in (\[LTT\]) as the backward-in-time Richardson constant $\mathfrak{g}^{B}$. The constant $\mathcal{A}^{[1]}$ is independent of $St$ to leading order and we will later estimate its value from DNS. Note that by its essentially perturbative construction, (\[LTT\]) is, like the fluid particle RO law, free from intermittency corrections, since it depends linearly upon the kinetic energy dissipation rate of the fluid [@novikov90; @boffetta99; @schmitt05]. In the regime $\mathcal{T}\gg\widehat{\mathcal{T}}$, the result in (\[LTT\]) simplifies to $$\begin{aligned} \Big\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\Big\rangle_{\bm{r}^0}\approx \mathfrak{g}^B\langle\epsilon\rangle\mathcal{T}^3\Big[1+St\mathcal{B}\mathcal{T}^{-1}\ln[\mathcal{T}/\widehat{\mathcal{T}}]\Big],\quad\text{for $\mathcal{T}\gg\widehat{\mathcal{T}}$},\label{LTT2}\end{aligned}$$ where $\mathcal{B}\equiv(1/2)\tau_\eta\mathcal{A}^{[1]}[\mathfrak{g}^{B}]^{-4/3}\geq0$. The result in (\[LTT2\]) is quite similar to the equivalent FIT result derived in [@bec10b], but differs in one important respect. Whereas the FIT result derived in [@bec10b] predicts that in the limit $\mathcal{T}/\widehat{\mathcal{T}}\to\infty$, $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ approaches unity from below, (\[LTT2\]) suggests that in the BIT case $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ approaches unity from above. In other words, in the BIT case the perturbative effect of the particle inertia is to increase the separation rate relative to that of fluid particles, whereas in the FIT case it decreases the separation rate. DNS results {#IPS} ----------- Before considering the results for the particle-pair mean-square separation, we first present the DNS data for $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$. This statistic is useful to consider both because it will aid in the understanding the dispersion results, and because it features in the small-time theory (\[STT\]). Comparing the $St>0$ results in Figure \[w2\_DNS\] with the $St=0$ results (i.e. fluid particles, for which $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}=\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$), we see that at larger separations the particle inertia gives rise to $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}<\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$, whereas at the smaller separations, $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}>\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$. The separation at which the transition in behavior occurs is a strong function of $St$. For a detailed explanation of the role of inertia on these statistics see [@bragg14c; @salazar12a]; here, we summarize. The predominant effects of inertia at these Stokes numbers are the filtering and non-local effects (preferential sampling of $\Delta\bm{u}$ has a role mainly for particles with $0<St\lesssim0.4$ and separations in the dissipation range [@salazar12a]). The inertia of the particles causes them to filter out the high frequency fluctuations of $\Delta\bm{u}$, and since the inertia gives the particles a memory, their velocity dynamics at a given separation are strongly influenced by their path-history interactions with the turbulent velocity field. For a given $\tau_{p}$, the non-local contribution becomes less important as one goes to larger and larger separations causing the filtering effect to dominate; this gives rise to $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}<\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$. At the smaller scales, the non-local effect of inertia dominates and gives rise to $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}>\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$. \ Fig. \[FIT\_BIT\_DNS\] shows results for $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ and $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ from DNS for ${r^{0}\in[0.25\eta,0.5\eta]}$, ${r^{0}\in[3\eta,4\eta]}$ and various $St$. The results confirm the expected behavior based on the irreversibility mechanism described in §\[IPIM\], namely that BIT dispersion is faster than FIT dispersion for the inertial particles. The results also show that in addition to quantitative differences, significant qualitative differences exist between the FIT and BIT dispersion, and these may also be understood in terms of the mechanism described in §\[IPIM\], as we now explain. The FIT dispersion is explained as follows: For small $r^{0}$ and sufficiently large $St$, the inertial particles initially separate much faster than the fluid particles owing to the presence of caustics in their relative velocities at small $r^{0}$, giving rise to $\bm{w}^{p}\gg\Delta\bm{u}^{p}$ (see Fig. \[w2\_DNS\]). This continues up to $\mathcal{T}\sim\tau_{p}$, after which the fluid velocity field begins to significantly affect their motion. After a few multiples of $\tau_{p}$, the fluid particles overtake the inertial particles because they are now separating faster than the inertial particles. The inertial particle separation begins to lag behind that of the fluid particles for two reasons: First, inertia filters out high-frequency modes of $\Delta\bm{u}^{p}$ [@salazar12a]. Second, because the particles are on average separating, they carry a memory of smaller fluid velocity differences in their path history. Consequently, if we consider a set of inertial particles with a range of $\tau_{p}$, then for $\mathcal{T}\gg\max[\tau_{p}]$, $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ decreases with increasing $\tau_{p}$. This behavior is observable in Fig. \[FIT\_BIT\_DNS\] (c). The BIT behavior is explained as follows: At large separations, the particle relative velocities decrease relative to the fluid with increasing $St$, as shown in Fig. \[w2\_DNS\]. Consequently, for very large $\mathcal{T}$ (corresponding to pairs at large-scale separations), the rate at which particle pairs are approaching each other decreases as $St$ increases. Note, however, that the range of $\mathcal{T}$ in our DNS data is too limited to observe this decrease. However, as the particle pairs begin to enter the inertial range, the fluid velocity differences they experience begin to decrease (statistically), and the rate of approach for fluid particle-pairs decreases. Since inertial particles have some memory, however, when they are in the inertial range, they will still carry a memory of their interaction with larger-scale turbulent motions in their path history. Therefore, at some separation and some time $\mathcal{T}$, they begin to decouple from the fluid turbulence and approach each other with relative velocities greater than the fluid particle pair at the same separation. Because BIT separating pairs retain, on average, a memory of larger-scale fluid velocity differences in their path-history, they approach each other more quickly than the corresponding fluid particle pair at both small and intermediate times, in contrast to the FIT case as discussed earlier. Thus the particles’ memory of their path-history interactions with the turbulence not only gives rise to faster BIT than FIT separation, but also gives rise to qualitative differences in their separation behavior relative to that of the corresponding fluid particle pair. In order to show the difference between FIT and BIT dispersion more clearly, in Figure \[BIT\_FIT\_ratio\] (a)-(c) we plot the ratio of the BIT to FIT mean-square separation for various $r^0$ and $St$. The results clearly show that $\max[\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}]$ is a strong function of $St$ and $r^0$, and that the irreversibility in particle dispersion is much stronger for inertial particles than for fluid particles. \ Refer to §\[IPIM\] for the physical explanation. At each $r^0$, there is an optimum value of $St$ for which the irreversibility is strongest, which corresponds to the value for which the particles experience optimally both the irreversibility associated with the turbulence dynamics and also that arising because of the history effect of their particle inertia. As $r^0$ is increased, this optimum value of $St$ increases because the effects of inertia (for a given $\tau_p$) decrease with increasing separation. As explained in §\[IPIM\], for $r^0>L$ FIT and BIT dispersion become equivalent, and become related to the one-particle dispersion problem which is reversible in stationary, homogeneous turbulence. The results in Figure \[BIT\_FIT\_ratio\] (d) for the single particle dispersion in the DNS confirm this expectation (where $\bm{x}^p(t)$ is the position of a single particle at time $t$). Having illustrated the effect of particle inertia on the dispersion irreversibility, we plot the ratio of the inertial particle to fluid particle mean-square separation in Figure \[Particle\_Fluid\_ratio\], both FIT and BIT, in order to highlight the differences between the dispersion rates of inertial and fluid particles. The results demonstrate the dramatic effect of inertia on the dispersion, with the inertial particle dispersion often being orders of magnitude greater than the fluid particle dispersion. Notice also, especially in comparing Fig. \[BIT\_FIT\_ratio\] (a) with Fig. \[Particle\_Fluid\_ratio\] (a), that whereas the time of the peak value of $\max[\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}]$ varies significantly with $St$, the time of the peak value of $\max[\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}]$ is approximately independent of $St$. In the latter case, this peak time roughly corresponds to $\tau_{r^0}$. At $\mathcal{T}<\tau_{r^0}$ the fluid particle dispersion is, relatively speaking, minimal. The results in Fig. \[Particle\_Fluid\_ratio\] also confirm the prediction of (\[LTT2\]) that in the large $\mathcal{T}$-regime, the inertial particles separate faster than the fluid particles BIT. This is in contrast to the FIT case where the inertial particles lag behind the fluid particles in the large $\mathcal{T}$-regime. We now compare the DNS results with the theoretical prediction of (\[STT\]) in figure \[DNS\_Theory\]. The DNS data for $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$ and $\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\rangle_{\bm{r}^0}$ was used in (\[STT\]). The results in Fig. \[DNS\_Theory\] show that (\[STT\]) describes the DNS very accurately for $\mathcal{T}\leq\widehat{\mathcal{T}}$, both qualitatively and quantitatively, capturing the effects of $St$ and $r^0$ on the separation behavior. It is interesting to note that both the DNS results and the predictions of (\[STT\]) show that the value of $\mathcal{T}$ for which $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ becomes approximately independent of $r^0$ decreases as $St$ increases. This might seem counter-intuitive, since one might expect that the smaller $St$ is, the faster the particles forget their initial conditions. The reason for the observed behavior, however, is due to the fact that as $St$ increases, $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$ becomes approximately independent of $\bm{r}^0$ in the dissipation range (see fig. \[w2\_DNS\]). For smaller $St$, $\langle\vert\bm{w}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$ remains dependent upon $\bm{r}^0$, and so particles with smaller $\bm{r}^0$ begin with both smaller separations *and* smaller relative velocities, and hence the $\bm{r}^0$-dependence of $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ is apparent for longer times. \ The results in Fig. \[DNS\_Theory\] are for $St\geq\mathcal{O}(1)$ particles; in the regime ${St\ll1}$, where the particle motion tends to be dominated by their local interaction with $\Delta\bm{u}$, one might expect that the preferential sampling of the turbulence by the particles could influence their separation behavior. In particular, since particles tend to avoid regions of the turbulence with high vorticity, and accumulate in the straining regions of the flow, one might expect that inertial particles in the regime ${St\ll1}$ might separate slower than fluid particle since the preferential sampling leads to $\langle\vert\Delta\bm{u}^p(0)\vert^2\rangle_{\bm{r}^0}<\langle\vert\Delta\bm{u}^f(0)\vert^2\rangle_{\bm{r}^0}$. The result in (\[STT\]) does not fully account for this effect since it uses the approximation $\Delta\bm{u}^p\approx\Delta\bm{u}^f$. We therefore now modify this relation to account for preferential sampling. In the regime ${St\ll1}$, we have $$\begin{aligned} \Big\langle\vert\bm{w}^p(0)\vert^2\Big\rangle_{\bm{r}^0}=\Big\langle\vert\Delta\bm{u}^p(0)\vert^2\Big\rangle_{\bm{r}^0}-2St\Big\langle\Delta\bm{u}^p(0)\bm{\cdot}\Delta\bm{a}^p(0)\Big\rangle_{\bm{r}^0}+\mathcal{O}(St^2),\label{wpertsol}\end{aligned}$$ where $\Delta\bm{a}^p$ is the fluid relative acceleration vector evaluated at the particle-pair separation. Making the replacement ${\Delta\bm{u}^f\to\Delta\bm{u}^p}$ in (\[STT\]), and introducing into the resulting expression (\[wpertsol\]), we obtain the result that accounts for the effects of preferential sampling on $\langle\vert\bm{r}^p(-\mathcal{T})\vert^2\rangle_{\bm{r}^0}$ in the regime ${St\ll1}$ $$\begin{aligned} \begin{split} \Big\langle\vert\bm{r}^p(-\mathcal{T})\vert^2\Big\rangle_{\bm{r}^0}&=\vert\bm{r}^0\vert^2+\mathcal{C}\mathcal{T}^2+2\mathcal{T}G(-\mathcal{T})\Big(\mathcal{C}-\sqrt{\mathcal{C}\mathcal{D}}\Big)+G^2(-\mathcal{T})\Big(\mathcal{C}+\mathcal{D}-2\sqrt{\mathcal{C}\mathcal{D}}\Big)+\mathcal{O}(St^2), \label{BIT_pert} \end{split}\end{aligned}$$ where $$\begin{aligned} \mathcal{C}&\equiv\Big\langle\vert\Delta\bm{u}^p(0)\vert^2\Big\rangle_{\bm{r}^0},\\ \mathcal{D}&\equiv\mathcal{C}-2 St\Big\langle\Delta\bm{u}^p(0)\bm{\cdot}\Delta\bm{a}^p(0)\Big\rangle_{\bm{r}^0}.\end{aligned}$$ Assuming that ${\langle\Delta\bm{u}^p(0)\bm{\cdot}\Delta\bm{a}^p(0)\rangle_{\bm{r}^0}<0}$ in the dissipation range, as it is in the limit ${St\to0}$ and in the inertial range [@pumir01], then $\mathcal{D}\geq\mathcal{C}$ and each term in (\[BIT\_pert\]) is $\geq0$. In view of the fluid ballistic behavior, it can be seen that whether or not $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ is less than $\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ will depend upon the competition between the various terms in (\[BIT\_pert\]), and this competition depends upon both $St$ and $\mathcal{T}$. If $St\ll1$ and $\mathcal{T}\ll\tau_p$, then the leading contribution to (\[BIT\_pert\]) is ${\langle\vert\bm{r}^p(-\mathcal{T})\vert^2\rangle_{\bm{r}^0}=\vert\bm{r}^0\vert^2+\mathcal{C}\mathcal{T}^2}$. In this case, since preferential sampling gives rise to $\langle\vert\Delta\bm{u}^p(0)\vert^2\rangle_{\bm{r}^0}<\langle\vert\Delta\bm{u}^f(0)\vert^2\rangle_{\bm{r}^0}$ then $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}<\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$. However, it is likely that $\vert\bm{r}^0\vert^2\gg\mathcal{C}\mathcal{T}^2$ in this regime, and consequently differences between $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ and $\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ may be negligible. If ${St\ll1}$ and ${\mathcal{T}=\mathcal{O}(\tau_p)}$, then the third and fourth terms will dominate over the second term since $\vert G(-\mathcal{T})\vert$ grows exponentially fast in this regime. \ We then have $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}>\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$. Consequently, the preferential sampling effect may be irrelevant for the BIT dispersion, since the explosive growth BIT associated with the time-irreversibility effect of the particle inertia may overwhelm the local preferential sampling effect. However, the FIT dispersion for $St\ll1$ contains $G(\mathcal{T})$ which grows at a rate $\leq \mathcal{T}$, i.e. not explosive like $G(-\mathcal{T})$. In this case the preferential sampling effect may be sufficient to cause $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ to be less than $\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$. In Figure \[Ratio\_Small\_St\] we show the DNS results for $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ and $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$. When viewed at this scale, we can see that indeed the preferential sampling effect does cause ${St\ll1}$ particles to separate slower than fluid particles FIT. Conversely, in the BIT case, the inertial particles always separate faster than the fluid particles, for the reasons just explained. In order to check that the FIT result is in fact the result of preferential sampling, in figure \[Theory\_Ratio\_Small\_St\](a) we plot the small-time FIT theoretical prediction for $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$, $$\begin{aligned} \begin{split} \frac{\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}}{\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}}\approx\Bigg(\vert\bm{r}^0\vert^{2}&+G^{2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}+2G(\mathcal{T})\Big[\mathcal{T}-G(\mathcal{T})\Big]\sqrt{\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\sqrt{\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\\ &+\Big[\mathcal{T}^{2}-2\mathcal{T}G(\mathcal{T})+G^{2}(\mathcal{T})\Big]\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\Bigg)\Bigg/\Bigg(\vert\bm{r}^0\vert^{2}+\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\mathcal{T}^2\Bigg), \label{DispFIT_r0_2} \end{split}\end{aligned}$$ and in figure \[Theory\_Ratio\_Small\_St\](b) we plot $$\begin{aligned} \begin{split} \frac{\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}}{\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}}\approx\Bigg(\vert\bm{r}^0\vert^{2}&+G^{2}(\mathcal{T})\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}+2G(\mathcal{T})\Big[\mathcal{T}-G(\mathcal{T})\Big]\sqrt{\Big\langle\vert\bm{w}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\sqrt{\Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}}\\ &+\Big[\mathcal{T}^{2}-2\mathcal{T}G(\mathcal{T})+G^{2}(\mathcal{T})\Big]\Big\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\Bigg)\Bigg/\Bigg(\vert\bm{r}^0\vert^{2}+\Big\langle\vert\Delta\bm{u}^{f}(0)\vert^{2}\Big\rangle_{\bm{r}^0}\mathcal{T}^2\Bigg). \label{DispFIT_r0_3} \end{split}\end{aligned}$$ The only difference between (\[DispFIT\_r0\_2\]) and (\[DispFIT\_r0\_3\]) is that (\[DispFIT\_r0\_3\]), unlike (\[DispFIT\_r0\_2\]), accounts for $\Delta\bm{u}^p\neq\Delta\bm{u}^f$, i.e. it accounts for preferential sampling effects (and we use DNS data to specify $\langle\vert\Delta\bm{u}^{p}(0)\vert^{2}\rangle_{\bm{r}^0}$). Comparing figure \[Theory\_Ratio\_Small\_St\] with figure \[Ratio\_Small\_St\] (a) reveals that $\langle\vert\bm{r}^{p}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}<1$ in the small-time regime is in fact caused by preferential sampling effects. In an experimental study of the FIT dispersion of inertial particles [@gibert10], particles with $St\leq0.5$ were considered. However contrary to our results, they did not observe a reduction in the inertial particle dispersion due to preferential sampling, but instead observed that the inertial particles with ${0 <St \leq 0.5}$ separated faster than the fluid particles at small-times. There are several possible explanations for this apparent disagreement. First, their results are for $r^0$ in the inertial range, where the preferential sampling effect and clustering is weaker than in the dissipation range [@bragg14e], and hence the effect of preferential sampling on the dispersion may be harder to observe than in the dissipation range. Second, gravitational settling in their experiments is known to reduce the preferential sampling effect by reducing the interaction time between the particles and the fluid velocity field. Third, their system consisted of heavy particles in water, with a particle-to-fluid density ratio $\leq\mathcal{O}(10)$, as compared to the $\mathcal{O}(1000)$ density ratio we consider. At this much lower density ratio, the particle dynamics in the experiment are no longer governed only by drag forces, but involve additional forces, such as added mass and the Basset history force, which were neglected in the DNS. The additional forces acting on the particles in the experiment may counteract the preferential sampling effect, possibly explaining why they did not observe the inertial particle dispersion for small $St$ to be slower than that of the fluid particles. We now turn to consider the long-time behavior of the particle separation. In figure \[Particle\_RO\_scaling\] we test the scaling prediction of (\[LTT2\]), which predicts that the inertial particle separation approaches that of fluid particles at a rate ${\propto\mathcal{T}^{-1}\ln[\mathcal{T}/\widehat{\mathcal{T}}]}$. Note that because of the limited $Re_\lambda$ of the DNS, our data for $\mathcal{Q}(\mathcal{T})\equiv\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ does not reach unity in the long-time limit and we therefore subtract $\mathcal{Q}(\mathcal{T}_{max})$, rather than $1$, from $\mathcal{Q}(\mathcal{T})$ when testing the prediction in Figure \[Particle\_RO\_scaling\] (something similar was also done in Fig.9 of [@bec10b]). The results confirm the scaling prediction in (\[LTT2\]) quite well for $St\leq\mathcal{O}(1)$ over a range of $\mathcal{T}$. The significant deviations from the predicted scaling at $\mathcal{T}=\mathcal{O}(100\tau_\eta)$ are the result of the influence of the large scales where the separation becomes diffusive in time, and the deviations for ${St=\mathcal{O}(10)}$ are because at this $Re_\lambda$, the inertial range is not large enough for $St=\mathcal{O}(10)$ to reach the regime where the particles’ inertia is perturbative (i.e. where $St_r\ll1$). From the data for $St=0.05$ in Fig. \[Particle\_RO\_scaling\] we also obtain an estimate ${\mathcal{A}^{[1]}\approx 39.13}$. Using this estimated value for ${\mathcal{A}^{[1]}}$, we may then compare the quantitative prediction of (\[LTT2\]) with the DNS data. \ The results in figure \[DNS\_Theory2\] show that the long-time theory (LTT) predictions from (\[LTT2\]) are in good agreement with the DNS data, with the LTT accurately capturing the perturbing effect of the particle inertia BIT in the regime $St_r\ll1$, which gives rise to $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}/\langle\vert\bm{r}^{f}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}\geq 1$. For $St>0.3$ the effect of $St$ on $\mathcal{A}^{[1]}$ becomes apparent and this affects the predictions of (\[LTT2\]). Figure \[DNS\_Theory3\] shows the results for $St=0.5$ and $St=1$ using the DNS data for $\mathcal{A}^{[1]}$ for these $St$ numbers. The results show that when the effect of $St$ on $\mathcal{A}^{[1]}$ is accounted for, the LTT describes $\langle\vert\bm{r}^{p}(-\mathcal{T})\vert^{2}\rangle_{\bm{r}^0}$ quite well. Conclusion {#conc} ========== In this paper we have considered the FIT and BIT dispersion of fluid and inertial particles. The FIT and BIT dispersion of inertial particles are qualitatively and quantitatively different with BIT dispersion occurring at a much greater rate. In general the irreversibility of inertial particle relative dispersion is much greater than that for fluid particles. This is because inertial particle pair relative dispersion is subject to both the irreversibility of the underlying turbulent velocity field and also the irreversibility mechanism arising from the non-local contribution to their velocity dynamics. Concerning the FIT and BIT mean-square dispersion of fluid particles, our DNS data shows that the dispersion is accurately described for small-times by the ballistic law for all separations. For finite times in the dissipation range we do not observe evidence of the simple exponential law suggested by Batchelor. However, it is possible that the initial particle separations were not small enough to observe the exponential law before the pair left the dissipation range. We also observe clear RO scaling for initial separations lying between 3 and 4 $\eta$, and the FIT and BIT Richardson’s constants were found to be in excellent agreement with experimental data. For smaller separations the contaminating effect of the small scale separation behavior means that the RO law is never attained for the time spans we have data, although it does appear to be approaching this asymptotically. For larger initial separations the limited scale separation in our DNS means that the RO law is not reached by the time the fluid particle pairs are at integral scale separations. We developed theoretical explanations and predictions for the BIT mean-square dispersion of inertial particles. The small-time theory, which in the dissipation range describes the dispersion for times $\leq\max[St\tau_\eta,\tau_\eta]$ agrees very well with DNS data, capturing the effects of changes in both $r^0$ and $St$. We also showed that in the small $St$ regime, the effects of preferential sampling causes the inertial particles to separate slower than fluid particles FIT, but not BIT. The long-time theory, valid for times $\gg\max[St\tau_\eta,\tau_\eta]$, is essentially based upon an expansion around the fluid particle RO $\mathcal{T}^3$ law. It predicts that the inertial particle mean-square separation approaches a RO law at a rate $\propto\mathcal{T}^{-1}\log[\mathcal{T}/\hat{\mathcal{T}}]$ and that the particle inertia causes the dispersion to be greater than that for fluid particles, in contrast to the FIT case in [@bec10b] where the inertia causes the particles to separate more slowly than the fluid particles in the long-time limit. The DNS data confirms the predictions of the long-time theory provided that at these times the local Stokes number is small. In our DNS this is only satisfied for $St\leq\mathcal{O}(1)$ because of the moderate $Re_\lambda$ of the flow. However, this condition would be satisfied for arbitrary $St$ at sufficiently long-times in the limit ${Re_{\lambda}\to\infty}$. The research presented in this paper will be of use for understanding mixing processes of inertial particles in turbulence, which is connected to BIT and not FIT dispersion. We have shown how dramatically the BIT and FIT dispersion can differ for inertial particles, highlighting how inaccurate it is to approximate them as being equivalent in mixing models for inertial particles. The work is also important for the development of the theory presented in [@pan10] for the relative velocities of inertial particles in isotropic turbulence. In [@pan10] they approximated the BIT mean square separation by its FIT counterpart since no theory or data was available to inform them of the BIT behavior. In [@bragg14c] we argued that this approximation is responsible for some error in their theory predictions. In future work we intend to develop the theory presented in [@pan10] by using the BIT closures developed in this paper. Acknowledgements {#acknowledgements .unnumbered} ================ The work was supported by the National Science Foundation through CBET grant 0967349, and through a graduate research fellowship to PJI. Additional funding was provided by Cornell University. Computational simulations were performed on Yellowstone (ark:/85065/d7wd3xhc) at the U. S. National Center for Atmospheric Research [@yellowstone] under grants ACOR00001 and P35091057.

--- abstract: 'A modified realisability interpretation of infinitary logic is formalised and proved sound in constructive type theory (CTT). The logic considered subsumes first order logic. The interpretation makes it possible to extract programs with simplified types and to incorporate and reason about them in CTT.' address: 'Department of Mathematics, Uppsala University' author: - Erik Palmgren bibliography: - 'ModBiblio.bib' nocite: '[@BBSS:PfThyWo]' title: | Internalising modified realisability\ in constructive type theory --- Modified realisability ====================== Modified realisability interpretation is a well-known method for giving constructive interpretation of some intuitionistic logical system into a simple type structure [@Troelstra:Meta]. The method is used, for instance, in Minlog and Coq for extracting programs from proofs (cf. [@Schwicht:MLCF] and [@LetouzeyThesis]). These programs are to a large extent free from the computationally irrelevant parts that might be present in programs arising from direct interpretations into constructive type theory. The realisability interpretation requires a separate proof of correctness, which is usually left unformalised. In this note we present a completely formalised modified realisability interpretation carried out in the proof support system Agda [@CCoquand:Agda]. We shall here use what is called [*modified realisability with truth*]{} which has the property that anything realised is also true in the system (Theorem \[corr\]). One difference from usual interpretations as in Minlog is that the logic interpreted goes beyond first order logic: it is a (constructively) infinitary logic, which arises naturally from the type-theoretic notion of universe. Our extension to infinitary logic seems to be a novel result. Agda is based on Martin-Löf constructive type theory [@ML:AITT] with an infinite hierarchy of universes $\#0={\sf Set}$, $\#1={\sf Type}$, $\#2={\sf Kind}$, $\#3$, …. Each of these universes is closed under the formation of generalised inductive data types. We define in Agda an inductive type ${\sf SP}$ of propositions, so called [*simple propositions*]{}, by induction: for each small type $A$ (i.e. a member of ${\sf Set}$) an atomic proposition ${\sf atom}(A):{\sf SP}$ is introduced; ${\sf SP}$ contains $\bot$ and is closed under propositional connectives ($\land$, $\lor$, $\to$) and for any small type $A$ and any propositional function $P:A \to {\sf SP}$ the quantified propositions $\forall(A,P)$ and $\exists(A,P)$ belong to ${\sf SP}$. There is an obvious homomorphic embedding ${\sf Tp}$ of ${\sf SP}$ into the small types defined by ${\sf Tp}(\bot) =\emptyset$, ${\sf Tp}({\sf atom}(A)) = A$, ${\sf Tp}(P \lor Q) = {\sf Tp}(P) + {\sf Tp}(Q)$, ${\sf Tp}(P \land Q) = {\sf Tp}(P) \times {\sf Tp}(Q)$, ${\sf Tp}(P \to Q) = {\sf Tp}(P) \to {\sf Tp}(Q)$, ${\sf Tp}(\forall(A,P)) = (\Pi x:A){\sf Tp}(P(x))$ and ${\sf Tp}(\exists(A,P)) = (\Sigma x:A){\sf Tp}(P(x))$. We shall sometimes write $(\forall x:A)P(x)$ for $\forall(A,P)$ etc. The simple propositions may be realised by terms from a simplified type structure. All atomic propositions will be realised by the unique element ${\bf elt}$ of the unit type ${\sf Un}$. Define another homomorphism ${\sf Cr}$ (for [*crude type*]{}) from ${\sf SP}$ to small types by letting $$\begin{aligned} {\sf Cr}(\bot) &=& {\sf Un} \\ {\sf Cr}({\sf atom}(A)) &=& {\sf Un} \\ {\sf Cr}(P \land Q) &=& {\sf Cr}(P) \times {\sf Cr}(Q) \\ {\sf Cr}(P \lor Q) &=& {\sf Cr}(P) + {\sf Cr}(Q) \\ {\sf Cr}(P \to Q) &=& {\sf Cr}(P) \to {\sf Cr}(Q) \\ {\sf Cr}(\forall(A,P)) &=& (\Pi x:A){\sf Cr}(P(x)) \\ {\sf Cr}(\exists(A,P)) &=& (\Sigma x:A){\sf Cr}(P(x)). \\\end{aligned}$$ The only difference from ${\sf Tp}$ is thus in the translation of absurdity and atoms. We note that a crude type may still be a dependent type, if the simple proposition is truly infinitary. For example, this is the case with ${\sf Cr}(\exists(A,P))$, if $A=N$ and $P(0) = \top$, $P(S(n)) = Q(n) \land P(n)$. Another variant of the crude type map ${\sf Cr}'$ will be employed in Theorem \[th15\] below, which is defined as ${\sf Cr}$, except that $${\sf Cr}'(\exists(A,P)) = {\sf Un} + (\Sigma x:A){\sf Cr}'(P(x)).$$ The unit type appearing in the disjoint sum ensures that the type is never empty, which is crucial for interpreting the full absurdity axiom. The modified realisability ${\sf MR}(S,r)$ of a simple proposition $S:{\sf SP}$ by an element of crude type $r:{\sf Cr}(S)$ is defined as a small proposition (or small type) by the following recursion on $S$. (We use the identification of propositions and types for small types, so that $\land$ and $\lor$ are used interchangeably with $\times$ and $+$, respectively.) $$\begin{aligned} {\sf MR}(\bot,r) &=& \bot \\ {\sf MR}({\sf atom}(P),r) &=& P \\ {\sf MR}(A\land B,r) &=& {\sf MR}(A,r.1) \land {\sf MR}(B,r.2) \\ {\sf MR}(A\lor B,{\sf inl}(s)) &=& {\sf MR}(A,s) \\ {\sf MR}(A\lor B,{\sf inr}(t)) &=& {\sf MR}(B,t) \\ {\sf MR}(A\to B,r) &=& ({\sf Tp}(A) \to {\sf Tp}(B))\\ & & \land\; (\Pi s:{\sf Cr}(A))({\sf MR}(A,s) \to {\sf MR}(B,r(s))) \\ {\sf MR}(\forall(A,P),r) &=& (\Pi x:A){\sf MR}(P(x),r(x)) \\ {\sf MR}(\exists(A,P),r) &=& {\sf MR}(P(r.1),r.2). \\\end{aligned}$$ Here $r.1$ and $r.2$ denote the first and second projections. The above constructions work in many different type-theoretic settings. What is needed is a type universe $U$ closed under $\Pi$, $\Sigma$, $+$ and containing basic types ${\sf Un}$ and $\emptyset$. Moreover the inductive construction ${\sf SP}_U$ is should be made relative to $U$ instead of ${\sf Set}$. Then $${\sf Tp}_U : {\sf SP}_U \to U \qquad {\sf Cr}_U : {\sf SP}_U \to U$$ are defined by recursion on ${\sf SP}_U$ similarly to the above, and so is $${\sf MR}_U : (\Pi s :{\sf SP}_U)({\sf Cr}_U(s) \to U).$$ The following correctness, or conservativity, result states that each simple proposition, which is realised, is also true in the standard interpretation. \[corr\] For any $S:{\sf SP}$ and $r:{\sf Cr}(S)$, if ${\sf MR}(S,r)$ then ${\sf Tp}(S)$. The proof goes by induction on $S$. For $S=\bot$ or $S={\sf atom}(A)$ the result is immediate. For $S=A\to B$ we took care to define realisability so that this is direct as well. Here are two examples of the inductive step. Suppose ${\sf MR}(A\lor B,r)$. If $r={\sf inl}(s)$, then ${\sf MR}(A,s)$ is true. By the inductive hypothesis, we get ${\sf Tp}(A)$ and hence also ${\sf Tp}(A \lor B)$. The argument for $r={\sf inr}(t)$ is similar. Assume ${\sf MR}(\forall(A,P),r)$. Let $a \in A$. Then ${\sf MR}(P(a),r(a))$, and so by the inductive hypothesis ${\sf Tp}(P(a))$. Since $a$ was arbitrary we have actually ${\sf Tp}(\forall(A,P))$. As a corollary there is an extraction theorem for $\forall\exists$-formulae: \[AE\] For small types $A$ and $B$ and a simple proposition $P(x,y)$ where $x:A$ and $y:B$, let $$S =(\forall x:A) (\exists y: B)P(x,y).$$ If ${\sf MR}(S, r)$ for some $r$, then there is some $f:A \to B$ such that ${\sf Tp}(P(x,f(x)))$ for all $x:A$. Thereby the program $f$ extracted also satisfies its specification ${\sf Tp}(P(x,f(x)))$ within type theory. For $P(x,y) = {\sf atom}(R(x,y))$ this is equivalent to $R(x,f(x))$. Note the difference in the $\forall$-case from usual interpretations, which go from theories to theories [@Troelstra:Meta]. It is not required that ${\sf Tp}(\Pi(A,P))$ is added to the condition, since this follows from the correctness theorem in the present internalised version. We present an intuitionistic infinitary propositional logic $IPC^-_{\infty}$ in type theory in which quantifiers are understood as infinitary versions of conjunction and disjunction. The system has a restriction on the absurdity axiom to atomic formulae. $$\small \begin{array}{cc} A\vdash A \qquad &\qquad \vcenter{\infer{A\vdash C}{A \vdash B & B \vdash C}} \\ [2ex] A \vdash {\sf atom}(P), \rlap{\ for any inhabited $P$}\qquad &\qquad \\ [2ex] A \land B \vdash A \qquad A \land B \vdash B\qquad &\qquad \vcenter{\infer{C\vdash A\land B}{C\vdash A & C\vdash B}} \\ [2ex] \bot \vdash {\sf atom}(P) \qquad &\qquad \\ [2ex] A \vdash A \lor B \qquad B \vdash A \lor B \qquad &\qquad\vcenter{\infer{A \lor B \vdash C}{A \vdash C & B \vdash C}} \\ [2.8ex] \vcenter{\infer{A \vdash B\to C}{A \land B \vdash C}} \qquad &\qquad\vcenter{\infer{A \land B \vdash C}{A \vdash B\to C}} \\ [2.8ex] \vcenter{\infer{A\vdash \forall(S,P)}{A \vdash P(t) \quad (t:S)}}\qquad &\qquad\vcenter{\infer{A\vdash P(t)}{A\vdash \forall(S,P) & t:S}} \\ [2.8ex] \vcenter{\infer{\exists(S,P) \vdash A}{P(t) \vdash A \quad (t:S)}}\qquad &\qquad\vcenter{\infer{P(t) \vdash A}{\exists(S,P) \vdash A \quad t:S}}\\ \end{array}$$ Note in particular that the existential quantifier is of the weak kind, as in first order logic. For $S=\emptyset$ each $\exists(S,P)$ works as absurdity constant. However, if we wish to avoid empty sets as types of realisers, the restricted absurdity axiom $\bot \vdash {\sf atom(P)}$ should be used. The full absurdity rule can be derived from the restricted one, for those propositions which do not include quantification over empty sets. By this procedure we can in principle extract simply typed programs as in Minlog. We say that a sequent $A \vdash B$ is [*${\sf MR}$-realised,*]{} if there is some $r$ such that ${\sf MR}(A\to B,r)$ is true. A rule is [*realised*]{} if whenever all the sequents above the rule bar are realised, then so is the sequent below the bar. \[th14\] The axioms and rules of the system $IPC^-_\infty$ are ${\sf MR}$-realised. To strengthen the weak absurdity axiom to the full axiom $$\bot \vdash A$$ where $A:{\sf SP}$ may be arbitrary, we use the crude type map ${\sf Cr}'$ instead and introduce ${\sf MR}'$. This is defined recursively as ${\sf MR}$ apart from the case for the existential quantifier: $$\begin{aligned} {\sf MR}'(\exists(S,P),{\sf inl}(s)) &=& \bot\\ {\sf MR}'(\exists(S,P),{\sf inr}(t)) &=& {\sf MR}'(P(t.1),t.2). \\\end{aligned}$$ Theorem \[corr\] and Corollary \[AE\] now go through with ${\sf MR}'$ and ${\sf Cr}'$ in place of ${\sf MR}$ and ${\sf Cr}$. The proof of soundness of the logical rules and axioms is similar as for Theorem \[th14\], with the exception for the verification of the absurdity rule, and the left existential rule. This requires a special device. Namely a function which to each $P:{\sf SP}$ assigns an element, called ${\sf element}(P)$, of ${\sf Cr}'(P)$ is necessary. This function is defined straightforwardly by recursion on $P$. Some key clauses are $$\begin{aligned} {\sf element}(\exists(A,P)) &=& {\sf inl}({\bf elt}) \\ {\sf element}(\forall(A,P)) &=& \lambda x.{\sf element}(P(x)) \\ {\sf element}(A\lor B) &=& {\sf inl}({\sf element}(A)). \\\end{aligned}$$ Observe that no such element need to exist when employing the first definition of ${\sf Cr}$, e.g. in the case ${\sf Cr}(\exists(\emptyset,P)) = (\Sigma x:\emptyset){\sf Cr}(P(x))$. \[th15\] The axioms and rules of the full system $IPC_\infty$ ($IPC^-_\infty$ and the full absurdity axiom) are ${\sf MR}'$-realised. We mention some useful mathematical axioms that are realisable: For each propositional function $P:{\mathbb N} \to {\sf SP}$ the induction scheme $$P(0) \land (\forall x:{\mathbb N})[P(x) \to P(S(x))]\to (\forall x:{\mathbb N})P(x)$$ is both ${\sf MR}$-realised and ${\sf MR}'$-realised. For any binary propositional function $P:A \times B \to {\sf SP}$ the type-theoretic choice principle $$(\forall x:A)(\exists y:B)P(x,y) \to (\exists g:A\to B)(\forall x:A)P(x,g(x))$$ is ${\sf MR}$-realisable. In case $B$ is inhabited, the principle is ${\sf MR}'$-realisable as well. The non-trivial part is to prove the second statement. Suppose $b_0:B$ and $r:{\sf Cr}'(S)$ and $p: {\sf MR}'(S,r)$, where $S = (\forall x:A)(\exists y:B)\,P(x,y)$. Define an auxiliary operation $f(x,w):(\Sigma y:B){\sf Cr}'(P(x,y))$ where $x:A$ and $w: {\sf Cr}'((\exists y:B)P(x,y))$, by cases $$\begin{aligned} f(x, {\sf inl}(u)) &=& \langle b_0, {\sf element}(P(x,b_0)) \rangle \\ f(x, {\sf inr}(y)) &=& y.\end{aligned}$$ The realiser $k$ for the implication is now given by $$k(r) = \langle \lambda x. f(x,r(x)).1, \lambda x. f(x,r(x)).2 \rangle$$ To prove it is a realiser, use $\bot$-elimination for the case $r(x) = {\sf inl}(u)$. The following result is often useful to verify realisability. If the ${\sf Tp}$-translation of the proposition $$(\forall x_1:A_1)\cdots (\forall x_n:A_n)[Q(x_1,\ldots,x_n) \to P(x_1,\ldots,x_n)]$$ is true and $P$ is atomic or $\bot$, then the proposition is ${\sf MR}$-realised as well as ${\sf MR}'$-realised. The realising function is trivial for such a proposition: $(\lambda x_1)\cdots (\lambda x_n)(\lambda r){\bf elt}$. Theorem \[corr\], a special property of modified realisability with truth, is necessary here. Many stronger “transfer principles” are possible to establish. See [@BBS:Refined] for further results and references. An Example ========== We test the formalisation and extraction procedure on a simple example, which is due to Berger and Schwichtenberg. The extracted function computes Fibonacci numbers efficiently by “memoization.” A binary predicate $G$ on natural numbers is given. From the axioms - $G(0,0)$ - $G(1,1)$ - $(\forall m, k, \ell)[G(m,k) \land G(S(m),\ell) \to G(S(S(m)),k+\ell)].$ one derives by induction and intuitionistic logic the proposition - $(\forall x)(\exists k,\ell)G(x,k) \land G(S(x),\ell)$. Thus there is some realiser $f$ so that $${\sf MR}({\rm Ax1}\, \&\, {\rm Ax2}\, \&\, {\rm Ax3} \vdash {\rm P},f).$$ The extracted program $p$ (which is [fib\_prog]{} in the Appendix) for computing the Fibonacci sequence is then given by $$p(x) = f(nc,x).1$$ where $nc$ ([nocontent]{} in the Appendix) is the trivial realiser for ${\rm Ax1}\, \&\, {\rm Ax2}\, \&\, {\rm Ax3}$. After a normalisation process one gets the program: p x = (case x of { (zero) -> t; (succ x') -> h x' g (rec (\(z::Nat) -> C) x' t (\(x''::Nat) -> \(y::C) -> h x'' g y));}).1 where C = Sigma Nat (\(k::Nat) -> Sigma Nat (\(l::Nat) -> Unit)) h v p q = <q.2.1; <case q.2.1 of {(zero) -> q.1; (succ u) -> succ (q.1 + u); } ;<q.2.2.2; e>>> t = <zero; <succ zero; <e;e>>> g = \(x,y,z::Nat) -> \(h,j::Unit) -> e e = elt@_ Note that all truly dependent types have disappeared. The type $C$ is really the type ${\sf N}\times ({\sf N} \times {\sf Un})$. The normalised program has been computed using the partial normalisation procedure of Agda on selected subexpressions, and was thus not completely automatic. We also introduced the abbreviations $C, h,t,g,e$ by hand. Some syntactical sugar for lambda expressions and pairs is added. The formalisation ================= The formalisation have been carried out in Agda/IAgda (version 2003-08-09) with the aid of the graphical user interface Alfa. The relevant files are available at the URL `www.math.uu.se/~palmgren/modif`

--- abstract: 'The Kassiopeia software package was originally developed to simulate electromagnetic fields and charged particle trajectories for neutrino mass measurement experiments. Recent additions to Kassiopeia also allow it to simulate neutral particle trajectories in magnetic fields based on their magnetic moments. Two different methods were implemented: an exact method that can work for arbitrary fields and an adiabatic method that is limited to slowly-varying fields but is much faster for large precession frequencies. Additional interactions to simulate reflection of ultracold neutrons from material walls and to allow spin-flip pulses were also added. These tools were used to simulate neutron precession in the Paul Scherrer Institute’s neutron electric dipole moment experiment and predict the values of the longitudinal and transverse relaxation times as well as the trapping lifetime. All three parameters are found to closely match the experimentally determined values when simulated with both the exact and adiabatic methods, confirming that Kassiopeia is able to accurately simulate neutral particles. This opens the door for future uses of Kassiopeia to prototype the next generation of atomic traps and ultracold neutron experiments.' address: '$^1$Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, MA 02139, United States of America' author: - 'Z Bogorad$^1$, P M Mohanmurthy$^1$, and J A Formaggio$^1$' bibliography: - 'bibliography.bib' title: Ultracold Neutron Storage Simulation Using the Kassiopeia Software Package --- [Keywords: *simulation, software, particle tracking, ultracold neutrons*]{} \[sec:introduction\]Introduction ================================ At sufficiently low energies, neutrons can be reflected from material walls by the coherent strong interaction [@ZeldovichUCN; @SteyerlUCN]. Such neutrons can be stored in material traps, and are termed “ultracold neutrons" (UCNs). The ability of UCNs to be stored for prolonged periods makes them of interest in several experimental areas. They can be used to measure the neutron lifetime [@NeutronLifetimeBeam; @NeutronLifetimeBottle], which continues to be of interest due to the remaining discrepancy between different experimental measurements [@NeutronLifetimeReview]. UCNs can also be used to constrain the neutron electric dipole moment, which is related to the strong CP problem and can constrain beyond the Standard Model physics [@StrongCPReview; @NuclearAxionSearch; @NuclearLorentzConstraints]. Currently, the best upper bound on the neutron electric dipole moment comes from measurements on ultracold neutrons, described in [@BestnEDMLimit; @RevisednEDMLimit]. Several software packages have previously been used to simulate ultracold neutrons, including Geant4 [@Geant4], PENTrack [@PENTrack], STARucn [@STARucn] and MCUCN [@MCUCN]. Some comparisons between these are made in [@PENTrack; @UCNSimulationComparison]. In this paper, we demonstrate the ability of Kassiopeia [@Kassiopeia], a software package originally developed for neutrino experiments, to accurately simulate ultracold neutron storage. In Section \[sec:old\_simulation\], we describe some general features of Kassiopeia, not specific to UCNs. We then describe neutral particle tracking features that have been added to Kassiopeia in order to simulate UCNs in Section \[sec:spin\_simulation\]. Next, we describe a specific experiment that we simulated and the parameters of our simulation in Section \[sec:UCN\_storage\], and describe our results in Section \[sec:results\]. Finally, in Section \[sec:applications\] we consider possible future applications of Kassiopeia’s neutral particle tracking capabilities. \[sec:old\_simulation\]Simulation in Kassiopeia =============================================== Kassiopeia was created as a combined field solver and particle tracker for the KATRIN collaboration [@Kassiopeia]. It is developed in `C++` using a modular design that allows it to be adapted to various applications using different field solvers, equations of motion and interactions. The standard distribution, available at <https://github.com/KATRIN-Experiment/Kassiopeia>, implements a variety of these features already, and Kassiopeia is designed for easy development of additional modules for future applications. Individual simulations can then be configured using an Extensible Markup Language (XML) file. Kassiopeia has several implemented methods to calculate the electromagnetic fields associated with a particular current geometry. Magnetic fields can be calculated using zonal harmonic expansions [@CoronaThesis] or the fast multipole method [@BarrettThesis], while electric fields can be computed using the boundary element method [@CoronaThesis]. Both types of fields can also be pre-calculated and imported, as is done in this work. Kassiopeia also has several modules, called “trajectories", for solving particles’ equations of motion. Different trajectories include different terms in the equations of motion and use different representations of particle states. Early versions of Kassiopeia included two main trajectory modules: an exact trajectory module, which describes particles via their position and momentum in laboratory coordinates and is limited only by the integration step size, and an adiabatic trajectory module, which calculates and records particles’ position and motion around the magnetic field lines that they adiabatically follow at sufficiently low energies. Both of these modules implemented equations of motion that assumed a non-zero particle charge, which prevented Kassiopeia from simulating other particles of interest, such as neutrons or neutral atoms. To remedy this, two new trajectory modules were added to Kassiopeia: an exact spin trajectory and an adiabatic spin trajectory, described below. Both modules describe particles’ position and momentum in laboratory coordinates, as in the original exact trajectory, but then add additional terms based on particles’ magnetic moments, determined by their spins. These additional terms are potentially the only non-zero terms in the equations of motion for neutral particles. \[sec:spin\_simulation\]Spin Features in Kassiopeia =================================================== At each integration step of a trajectory calculation, two parameters of the particle’s evolution must be calculated: the time derivative of the spin, and the force on the particle as a function of the spin and surrounding fields. We therefore next describe how these are calculated for the exact spin trajectory. The relativistic generalization of a classical spin is the four-vector $S = (S_0, \bi{S})$ given by, for a particle with classical spin $\bi{s}$ and velocity $\bi{v}=\bbeta c$ and Lorentz factor $\Gamma$ (to avoid confusion with the gyromagnetic ratio $\gamma$) [@Jackson], $$\begin{aligned} S_0 &=& \Gamma\bbeta\cdot\bi{s} \\ \bi{S} &=& \bi{s}+\frac{\Gamma^2}{\Gamma+1}(\bbeta\cdot\bi{s})\bbeta.\end{aligned}$$ The equations of motion for a relativistic spin with 4-velocity $U^\alpha$ in fields $F^{\mu\nu}$, using the particle’s proper time $\tau$, are given by the BMT equation: $$\frac{\rmd S^\alpha}{\rmd \tau} = \frac{\gamma}{c}\left[F^{\alpha\beta}S_\beta+\frac{1}{c^2}U^\alpha(S_\lambda F^{\lambda\mu}U_\mu)\right]-\frac{1}{c^2}U^\alpha\left(S_\lambda\frac{\rmd U^\lambda}{\rmd \tau}\right).$$ The spin-dependent contribution to the particle’s motion is given by the magnetic dipole force term, which is implemented nonrelativistically: $$\mathbf{F} = \nabla(\gamma\mathbf{s}\cdot\mathbf{B})$$ This exact spin tracking method is completely general, except for the assumption that relativistic corrections to the magnetic dipole force term are irrelevant. However, the step size for numerically integrating these equations of motion is limited by the precession rate of the spins, $\gamma|\mathbf{B}|$. This is typically fast relative to other timescales of a particle’s motion, which can make the exact spin tracking method impractically slow for experiments with long timescales. Given the high computational cost of simulating tracks with the exact spin trajectory, an adiabatic spin trajectory was also implemented. The adiabatic spin trajectory records particle spins using two variables instead of four: an aligned spin $m$ that gives the component of the spin along the magnetic field at the particle’s position, and a spin angle $\phi$ that gives the orientation of the spin around that field. This spin angle is defined with respect to two unit vectors orthogonal to the local magnetic field, $\bi{e_1}$ and $\bi{e_2}$, defined, for a magnetic field $\bi{B} = (B_x, B_y, B_z)$ in laboratory coordinates, as: $$\begin{aligned} \bi{e_1} &:=& (B_z,0,-B_x)/\sqrt{B_x^2+B_z^2} \\ \bi{e_2} &:=& \bi{B}\times \bi{e_1}/|\bi{B}|\end{aligned}$$ Given these axes, $\phi$ gives the angle of the projection of the spin into the plane perpendicular to $\bi{B}$ away from $\bi{e_1}$, in the direction of $\bi{e_2}$: $$\bi{s} = (s^2-m^2)^{1/2}(\bi{e_1}\cos\phi + \bi{e_2}\sin\phi) + m\bi{B}/|\bi{B}| \label{eq:local_coords}$$ The relationship between these vectors is illustrated in . ![The vectors used to define the adiabatic coordinates $m$ and $\phi$ from the spin vector $\bi{s}$ in .[]{data-label="fig:VectorPlot"}](VectorPlot.png){width="0.4\linewidth"} These two parameters, $m$ and $\phi$, along with the position $x$, then give the adiabatic spin equations of motion (excluding contributions independent of spin) [@AdiabaticSpinEoM]: $$\begin{aligned} \frac{M}{\hbar}\ddot{\bi{x}} &=& \gamma m\nabla |\bi{B}| + \gamma |\bi{B}|(s^2-m^2)^{1/2}\nabla \bi{b}\cdot \bi{c} \\ \frac{\dot{m}}{s} &=& (s^2-m^2)^{1/2}(\dot{\bi{x}}\cdot\nabla \bi{b}\cdot \bi{c}) \label{adiabatic_mdot} \\ \dot{\phi} &=& -\gamma|\bi{B}|-\dot{\bi{x}}\cdot \bi{A}-\frac{sm}{(s^2-m^2)^{1/2}}(\dot{\bi{x}}\cdot\nabla \bi{b}\cdot \bi{a})\end{aligned}$$ where $x$ is position, $\gamma$ is the particle’s gyromagnetic ratio, $M$ is its mass, $s$ is the magnitude of its total spin, and we have defined $$\begin{aligned} \bi{a} &=& -\bi{e_1}\cos\phi+\bi{e_2}\sin\phi \\ \bi{b} &=& \bi{B}/|\bi{B}| \\ \bi{c} &=& \bi{e_1}\cos\phi+\bi{e_2}\sin\phi \\ \bi{A} &=& \nabla\bi{e_1}\cdot\bi{e_2}\end{aligned}$$ for convenience. This adiabatic spin trajectory is accurate when particles are non-relativistic and spin precession is fast compared to the rate at which the field at the particle’s location changes. Assuming the spin precession rate is dominated by the first term $\gamma|\bi{B}|$, this requires $\gamma|\bi{B}| \gg |\dot{\bi{x}}\cdot\nabla\bi{B}|/|\bi{B}|$, which should hold for all applications discussed in this work. During a single integration time step, the exact trajectory requires the spin to rotate by much less than a radian, so the left-hand side of that inequality approximately sets the maximum step size for exact spin integration. The adiabatic trajectory can handle large rotations per time step but assumes a constant field during the step, so the right-hand side limits the step size for adiabatic spin integration. In this limit, therefore, the adiabatic spin trajectory can be much faster. A comparison of the outputs from the adiabatic and exact trajectories is shown in . Note that the exact trajectory becomes increasingly inaccurate as the integration time approaches the precession period $33\ \mu$s, even for $10$ ms tracks. ![A comparison of neutron spin angles around the local magnetic field (in the coordinate system defined by ), including an adiabatic trajectory with an integration time step of $100\ \mu$s and exact trajectories with time steps of $5$, $10$, and $20\ \mu$s. These were run in a version of the neutron precession chamber described later in this work, but with the magnetic field scaled up by a factor of $1000$, resulting in a precession frequency of approximately $30$ MHz. Note that the plot necessarily suffers from aliasing, but sampling was done simultaneously so that points at the same time can be compared.[]{data-label="fig:ComparisonPlot"}](ComparisonPlot.png){width="0.8\linewidth"} There may, however, be other limitations on the step size that prevent this from being the case. In particular, for the ultracold neutron tracking described in this work, weak magnetic fields mean that the limiting factor on the time step is error in position, not spin, so adiabatic and exact spin tracking require similar computation time. \[sec:UCN\_storage\]Ultracold Neutron Storage ============================================= Having established that the exact and adiabatic spin tracking methods agree under appropriate conditions, we next consider how well Kassiopeia can reproduce experimental data from the Paul Scherrer Institute’s ultracold neutron precession chamber [@PSIResults]. The PSI neutron precession chamber is a cylinder of radius 23.5 cm and height 12 cm. The interior rounded surface is primarily deuterated polystyrene (dPS), with small windows of deuterated polyethylene (dPE), while the two flat ends of the cylinder (the electrodes) are covered in diamond-like carbon (DLC). For purposes of our simulation, the entire surface was assumed to be dPS, since it is the primary surface cover and any changes associated with including dPE and DLC are likely to be smaller than the uncertainties in the values described below. UCN reflections from surfaces are described by three features: the reflection probability, the depolarization probability, and the distribution of outgoing angles for reflected neutrons. The first two are described in [@UCNReflectionThesis], while the third is described in [@UCNAngles]. We will summarize these and discuss our estimates of the associated parameters. The UCN reflection probability at an energy $E$ and angle to the surface normal $\theta$ is given by $$P(R) = 1 - \eta\left(\frac{E_\perp}{V_f-E_\perp}\right)^{1/2} \label{p_reflection}$$ where $E_\perp = E\cos^2\theta$ is the component of energy from normal momentum, $V_f$ is the real part of the surface’s optical potential, and $\eta$ is another constant of the surface. For dPS, these values are known only approximately, with $\eta \sim (1-3)\times10^{-4}$ and $V_f \sim 1.6\times10^{-7}$ eV [@UCNReflectionThesis], while a range of values for DLC were found in [@DLCMeasurement]. The UCN depolarization probability is the simplest to describe, and is given by a constant $\beta$. For dPS, this is known very roughly as $\beta \sim 10^{-6}-10^{-5}$ [@UCNReflectionThesis], and again a range of values for DLC are given in [@DLCMeasurement]. Detailed theoretical treatments of the distribution of outgoing angles for reflected UCNs can be found in [@UCNAngles; @UCNReflectionPaper]. Since the behavior of UCNs in the precession chamber should not depend significantly on the exact angular distribution, we derive an approximate form. We start from the microroughness model result which, up to a normalization constant, can be modeled as: $$p(\theta_f,\Phi_f) \propto |S(\theta_f)|^2\exp\left[-\frac{(wk)^2}{2}(\sin^2\theta_i+\sin^2\theta_f-2\sin\theta_i\theta_f\cos\Phi_f)\right]$$ Here, $\theta_i$ and $\theta_f$ are the incident and outgoing angles to the normal, $\phi_f$ is the change in direction around the normal, $w$ is the surface roughness, $k$ is the neutron wave vector, and $$\begin{aligned} |S(\theta_f)| &=& \left|\frac{2\cos\theta_f}{\cos\theta_f+\sqrt{\cos^2\theta_f-k_c^2/k^2}}\right| \\ &=& 2k\cos\theta_f/k_c\end{aligned}$$ where $k_c = \sqrt{2mV_f/\hbar}$ is the wave number corresponding to the optical potential and we consider only neutrons with $k<k_c$, since higher-energy neutrons quickly escape the precession chamber. Then $$\begin{aligned} \fl p(\theta_f,\Phi_f)&\propto& \cos^2\theta_f\exp\left[-\frac{1}{2}(wk)^2(\sin\theta_i-\sin\theta_f)^2+(wk)^2\sin^2\theta_i(\cos\Phi_f-1)\right] \\ %\fl \approx (\cos^2\theta_i-\cos\theta_i\sin\theta_i(\theta_f-\theta_i))\exp\left[-\frac{1}{2}(wk)^2\cos^2\theta_i(\theta_f-\theta_i)^2-\frac{(wk)^2}{2}\sin^2\theta_i\Phi_f^2\right] \\ \fl &\approx& (1-\tan\theta_i\ \Delta\theta)\exp\left[-\frac{(wk)^2}{2}\cos^2\theta_i\Delta\theta^2-\frac{(wk)^2}{2}\sin^2\theta_i\Delta\theta^2\right]\end{aligned}$$ where we defined $\Delta\theta = \theta_f-\theta_i$. This gives a simple approximate distribution of outgoing angles that can be efficiently sampled. Note that this derivation assumes that the change in incident and outgoing angles is small. This is a rough approximation in our case, but we found that our simulations were insensitive to the exact parameters of the reflected distribution, so this should not significantly affect our results. For our simulations, we used $\beta = 6\times 10^{-6}$, $\eta = 1\times 10^{-4}$, $V_f = 150$ neV, and $w = 30$ nm, which were consistent with the available estimated ranges of each value (see [@dPSCorrelationLength; @DLCCorrelationLength] for measurements of $w$). The first two parameters were tuned so as to be most consistent with PSI’s data. In the future, it might be possible to use our simulations as a means of actually determining these values, but this was beyond the scope of this work. For each test simulation, the precession chamber’s interior had an essentially constant magnetic field of approximately $1\ \mu$T along the cylinder radius, with $\mathcal{O}$(nT) deviations along each coordinate axis. In PSI’s experiment, the apparatus would also contain an electric field in either axial direction, but this was not included in our simulation since electric effects are below the sensitivity of the experiment [@BestnEDMLimit]. The axial magnetic fields were different (and in particular were in opposite directions) for the two electric field directions, and we received a set of magnetic field measurements of each field configuration [@PSIFields] and then used a combination of natural neighbors interpolation (to reach a rectilinear lattice) and cubic interpolation (within lattice cells) to estimate the magnetic field everywhere. During a particular run of our simulation, a number of neutrons would be added to just above the bottom of the precession chamber. While neutrons may not actually all start near the bottom surface, their motion around the trap is sufficiently fast relative to the simulation times that this should not affect the results significantly. Neutron kinetic energies were generated from a Gaussian distribution of mean $150$ neV and standard deviation $50$ neV. This is a higher energy distribution than the energy distribution estimated by PSI [@PSIEnergies]. It may be possible to match the measured energy distribution more closely by varying the parameters described above, but this was beyond the scope of this work. During a given run, all neutrons would also begin with the same spin, which varied by experiment as described below. Three example tracks inside of the simulated PSI precession chamber are shown in . ![A VTK image of three neutrons tracked for 1 second inside of the PSI precession chamber. Colors correspond to kinetic energies, with higher energies at the blue end of the spectrum. Note the approximately specular reflections and the slight curvature in the tracks induced by gravity.[]{data-label="fig:ExampleTracks"}](ExampleTracks.png){width="0.8\linewidth"} \[sec:results\]Results and Discussion ===================================== We compared our simulation’s results with three previously measured parameters of PSI’s precession chamber: the longitudinal relaxation time T1 [@PSI_T1], the transverse relaxation time T2 [@PSI_T2], and the effective trap lifetime $\tau$ including neutron decay and wall losses [@PSI_T1]. A total of 17280 neutrons were simulated up to maximum times of 500 seconds. This was done using 4 parallel `c4.8xlarge` instances of Amazon Elastic Compute Cloud (EC2), or 144 total cores. Kassiopeia does not directly support parallelization of particle tracking, so we instead ran 144 separate simulations of 120 neutrons each and then combined the outputs. This took approximately 2500 CPU-hours, though further optimization is likely possible. The Longitudinal Relaxation Time T1 ----------------------------------- The longitudinal relaxation time, T1, is the time constant of the decay of average polarization for neutrons with initial spins that are either aligned or anti-aligned relative to the local magnetic field. In particular, it is the time constant for decay towards the thermal distribution of aligned spin. It should be noted that, even for ultracold neutrons, the $\mu$T-order magnetic fields in the precession chamber do not significantly affect neutron energies since $\mu_N|\bi{B}|$ is far less than the kinetic energy, so the ultimate thermal evolution of neutrons in the chamber is towards an essentially uniform distribution of spins. On a microscopic level, longitudinal relaxation is associated with inhomogeneities of the magnetic field, and with depolarization due to reflections. Since the adiabatic equation of motion for fully aligned or anti-aligned spins always gives a zero derivative of the aligned spin, our simulations instead started spins at $1^\circ$ or $179^\circ$ to the magnetic field, though we found that any angle up to a few degrees led to indistinguishable results, as the dominant contribution to T1 came from reflection depolarization. We simulated 4320 neutrons for 500 seconds using both the adiabatic and exact trajectories with half initially aligned and half initially anti-aligned, and compared our average polarization as a function of time to that obtained by PSI. Our results can be seen in , where a relative polarization of 1 indicates an average polarization equal to the initial value ($\pm$1 in simulations, but smaller magnitudes in PSI’s experiments). ![A comparison of the fractional remaining longitudinal polarization measured at PSI with the results of our adiabatic and exact simulations, including 1$\sigma$ confidence intervals for the simulations and 1$\sigma$ error bars for the experimental data. Here, a $0.95$ relative polarization corresponds to an average spin along the z-axis equal to $0.95$ times the initial average spin along the z-axis. Each data set averages over the two field configurations and over initially aligned and anti-aligned neutrons. We see good agreement over the entire period; the deviation of the exact data at long times is not statistically significant.[]{data-label="fig:T1_Plot"}](T1PlotFinal.png){width="0.8\linewidth"} As shows, both the adiabatic and exact simulations were able to accurately reproduce the measured longitudinal depolarization rate. Both simulations are within 1$\sigma$ of the experimental data (using the combined error) until around 350 seconds, and the adiabatic simulations continue to closely match the experiment for the entire run. The apparent discrepancy between the experimental data and the exact trajectory at high times is not statistically significant (see ). The Transverse Relaxation Time T2 --------------------------------- The transverse relaxation time, T2, is the time constant of the decay of average polarization for neutrons with initial spins that are aligned along an axis perpendicular to the local magnetic field. The dominant contribution to transverse relaxation of neutrons in the precession chamber is the variation in the magnetic field over the trap volume. Note that spin-spin interactions, which may be the dominant contribution to T2 in other contexts, should not be significant for neutrons and cannot be implemented in Kassiopeia, which tracks particles one at a time. ![A comparison of the fractional remaining transverse polarization measured at PSI with the results of our adiabatic and exact simulations, including 1$\sigma$ confidence intervals for the simulations and 1$\sigma$ error bars for the experimental data. Here, a $0.95$ relative polarization corresponds to an average spin along the axis of average precession equal to $0.95$ times the initial average spin along the initial polarization axis. Each data set averages over the two field configurations and over two initial spin directions. We see good agreement over the entire period; the deviation of the exact data at long times is not statistically significant.[]{data-label="fig:T2_Plot"}](T2PlotFinal.png){width="0.8\linewidth"} In PSI’s T2 measurements, neutrons began with polarizations aligned or anti-aligned with the magnetic field and then T2 was measured between two $\pi/2$ pulses that rotated them into and out of the perpendicular plane. While ideal spin rotating pulses are implemented in Kassiopeia, they were not used for this simulation since we were interested only in the behavior occurring within the field-perpendicular plane. We simulated 4320 neutrons for 500 seconds using both trajectory types and two different initial spin directions perpendicular to the magnetic field. We then compared our average polarization as a function of time to that obtained by PSI. Our results can be seen in , with relative polarization defined as for T1, but along the axis corresponding to average precession over that time instead of along the z-axis. As shows, both simulations were also able to accurately reproduce the measured transverse depolarization rate. The adiabatic simulation matches all of the experimental values, however the exact simulation results appear to deviate from the data at long times. However, this discrepancy is again not statistically significant (see ), as our simulation errors are highly correlated: neutrons that deviate significantly from the average polarization (due to depolarization on reflection, for example) are likely to remain in the trap for some time, lowering the average polarization at those later times as well. The Effective Trap Lifetime $\tau$ ---------------------------------- Since neutron reflection from material surfaces is not guaranteed, especially at higher energies (see ), the effective lifetime of neutrons in the precession chamber is significantly lower than their decay lifetime. Neutrons may also be lost due to holes or other imperfections in a physical trap, which should be included in the estimated value of $\eta$. ![A comparison of the fraction of neutrons remaining in the trap measured at PSI with the results of our adiabatic and exact simulations, including 1$\sigma$ confidence intervals for the simulations and 1$\sigma$ error bars for the experimental data. Averages for each data set are taken as in and . Good agreement is seen along the entire PSI data set. Note that the exact data uses a shifted y-axis for visual clarity. The experimental values are arbitrarily normalized as the initial neutron count is not known.[]{data-label="fig:LT_Plot"}](LTPlotFinal.png){width="0.8\linewidth"} Lifetimes may be extracted from both the T1 and T2 simulations, since the lifetime should be nearly independent of the neutrons’ spins in a $\mu$T field. The results from both sets of simulations, as well as from PSI’s experiments, are shown in once again shows good agreement between each of the four simulations and the experimental data. The adiabatic T1 simulation deviates significantly from the experimental data (see ), but the lifetime is strongly dependent on the chosen value of $\eta$ and the initial energy distribution–neither of which is precisely known for the PSI trap–so this discrepancy is not practically significant. Summary of Results ------------------ A summary of our simulated data’s agreement to the PSI experimental data is presented in , showing generally good agreement. The apparent overestimation of errors is expected to be the result of correlations between our errors at different times. ------------ -------------- ------ -------------- ------ Simulation $\chi^2$/ndf $p$ $\chi^2$/ndf $p$ T1 Ad 0.22 1.00 2.43 0.02 T1 Ex 0.66 0.84 0.53 0.90 T2 Ad 0.13 1.00 0.56 0.88 T2 Ex 1.29 0.29 0.85 0.66 ------------ -------------- ------ -------------- ------ : The results of $\chi^2$ testing of our results against the PSI experimental data, including left-sided p-values. Note that ndf=11 for the T1 and T2 measurements and ndf=9 for the lifetime measurements. These values suggest that we are over-estimating our errors, likely due to our results being strongly correlated. The T1 adiabatic lifetime data’s disagreement with the experimental data may indicate that our assumed value of $\eta$ was not correct.[]{data-label="tab:Chi_Squared"} This demonstrates that Kassiopeia can reproduce the effective lifetime as well as both decoherence times for neutrons within the PSI precession chamber, confirming that it can accurately simulate neutral particles. \[sec:applications\]Other Applications ====================================== Having established that Kassiopeia can accurately simulate neutral particles’ motion in at least one context, we next consider a few potential future applications of these features. The application most directly stemming from the simulations in this paper is simulated prototyping of future ultracold neutron experiments. In this work, we simulated neutron precession using a previously measured magnetic field, but Kassiopeia also includes extensive field calculation methods, which can be used to calculate the magnetic fields inside of a trap based on its design [@CoronaThesis; @BarrettThesis]. Kassiopeia could therefore be used to estimate the effective lifetime and relaxation times of future ultracold neutron experiments. Another potential application is simulating tritium storage methods in the Project 8 neutrino mass experiment [@Project8]. Currently, atomic tritium is stored using a purely magnetic trap. We have previously simulated versions of this trap in order to obtain estimates of necessary currents and effective lifetimes, but development of the final trap configuration is ongoing and further simulation work could be done in this area. Furthermore, one current proposal is to use a combined magneto-gravitational trap instead, which has not been simulated as of writing. Neutral particle simulation can also be applied to atomic experiments. Experiments such as ACME [@ACME] and others such as [@YbFeEDM] use molecular beams to measure the electric dipole moment. Molecules are prepared in particular spin states and then precess through constant electric and magnetic fields; a dependence of precession on the electric field can indicate an electric dipole moment. Kassiopeia may be able to simulate molecular evolution within these fields to aid with estimating systematic errors. The authors would like to thank Mathieu Guigue for some helpful discussion. We would also like to acknowledge the KATRIN and Project 8 collaborations as well as the Paul Scherrer Institute’s nEDM collaboration for their support of this project. This work was supported by the MIT Undergraduate Research Opportunities Program, by the Department of Energy Office of Nuclear Physics under award number DE-SC0011091, and by SERI-FCS under award number 2015.0594. Other allied works were supported by Sigma Xi under grant number G2017100190747806. References {#references .unnumbered} ==========

--- abstract: 'Polarization-adjusted convolutional (PAC) codes are special concatenated codes in which we employ a one-to-one convolutional transform as a pre-coding step before the polar transform. In this scheme, the polar transform (as a mapper) and the successive cancellation process (as a demapper) present a synthetic vector channel to the convolutional transformation. The numerical results show that this concatenation improves the Hamming distance properties of polar codes. In this work, we implement the parallel list Viterbi algorithm (LVA) and show how the error correction performance moves from the poor performance of the Viterbi algorithm (VA) to the superior performance of list decoding by changing the constraint length, list size, and the sorting strategy (local sorting and global sorting) in the LVA. Also, we analyze the latency of the local sorting of the paths in LVA relative to the global sorting in the list decoding and the trade-off between the sorting latency and the error correction performance.' author: - '[^1]' title: List Viterbi Decoding of PAC Codes --- Polarization-adjusted convolutional codes, polar codes, Viterbi decoding, list decoding, path metric sorting. Introduction {#sec:intro} ============ Polar codes proposed by Arikan in [@arikan] are the first class of channel codes with an explicit construction that was proven to achieve the symmetric (Shannon) capacity of a binary-input discrete memoryless channel (BI-DMC) using a low-complexity successive cancellation (SC) decoder (SCD). Nevertheless, the error correction performance of finite-length polar codes under SCD is not satisfactory due to the existence of partially polarized channels. To address this issue, SC list decoding (SCLD or in short LD) was proposed in [@tal]. Recently in [@arikan2], Arikan proposed a concatenation of a convolutional transform with the polarization transform [@arikan] in which a message is first encoded using a convolutional transform and then transmitted over polarized synthetic channels as shown in Fig. \[fig:PAC\_scheme\]. These codes are called “polarization-adjusted convolutional (PAC) codes". It was shown in [@li2] that a properly designed pre-transformation such as convolutional transform can improve the distance properties of polar codes. ![ PAC Coding Scheme[]{data-label="fig:PAC_scheme"}](coding_diag.eps){width="1\columnwidth"} In [@rowshan-pac1], we studied the implementation of tree search algorithms including the conventional list decoding and complexity-efficient Fano decoding for PAC codes. However, due to the convolutional pre-transformation, PAC codes can also be easily encoded and decoded based on the trellis by employing the Viterbi algorithm (VA) [@viterbi; @forney] and the extended/list VA [@hashimoto] as a decoder. The basic Viterbi algorithm was employed in [@arikan_viterbi] as an ML decoder for short polar codes in a comparison with Reed-Muller (RM) codes. In this paper, we illustrate the implementation of the parallel list Viterbi algorithm (LVA) [@hashimoto; @seshadri] for PAC codes, and analyse the impact of list size and constraint length on the error correction performance. We also analyze the latency of the path sorting at each state on the trellis relative to global sorting in tree-based list decoding. Preliminaries {#sec:prelim} ============= Polarization-adjusted convolutional (PAC) codes are denoted by [*PAC*]{}$(N,K,\mathcal{A},\mathbf{g})$, where $N = 2^n$ is the length of the PAC code. A rate profiler first maps the $K$ information bits to $N$ bits. Then, the convolutional transform (with polynomial coefficients vector $\mathbf{g}$) scrambles the resulting $N$ bits before feeding them to the classical polar transform (Fig. \[fig:PAC\_scheme\]). The information bits $\mathbf{d}=[d_0,d_1,...,d_{K-1}]$ are interspersed with $N-K$ zeros and mapped to the vector $\mathbf{v}=[v_0,v_1,...,v_{N-1}]$ using a rate-profile which defines the code construction. The rate-profile is defined by the index set $\mathcal{A}\subseteq \{0,\ldots, N-1\}$, where the information bits appear in $\mathbf{v}$. This set can be defined as the indices of sub-channels in the polarized vector channel with high reliability. These sub-channels are called [*good channels*]{}. The bit values in the remaining positions $\mathcal{A}^c$ in $\mathbf{v}$ are set to 0. The input vector $\mathbf{v}$ is transformed to vector $\mathbf{u}=[u_0,\ldots,u_{N-1}]$ as $u_i = \sum_{j=0}^m g_j v_{i-j}$ using the binary generator polynomial of degree $m$, with coefficients $\mathbf{g}=[g_0,\ldots,g_m]$. The convolutional transform combines $m$ previous input bits stored in a shift register with the current input bit $v_i$ to calculate $u_i$ (see subroutine [*conv1bEnc*]{} in Algorithm \[alg:LVA\]). The parameter $m+1$, in bits, is called the [*constraint length*]{} of the convolutional code. As a result of this pre-transformation, $u_i$ for $i\in\mathcal{A}^c$ are no longer frozen as in polar codes. Note that this convolutional transformation is one-to-one, therefore the output vector $\mathbf{u}$ is not a traditional convolutional codeword. The rate-profiling process performed before the convolutional transformation creates the redundancy by inserting $N-K$ zeros in the length-$K$ input sequence $\mathbf{d}$. Finally, as shown in Fig. \[fig:PAC\_scheme\], vector $\mathbf{u}$ is mapped to vector $\mathbf{x}$ ($\mathbf{x}=\mathbf{u}\mathbf{P}_n$) by the polar transform $\mathbf{P}_n=\mathbf{P}^{\otimes n}$ defined as the $n$-th Kronecker power of $\mathbf{P} = {\footnotesize \begin{bmatrix} 1 & 0 \\ 1 & 1 \end{bmatrix} }$. The $\mathbf{x}$ vector is transmitted through a noisy channel and received as the vector $\mathbf{y}$. The channel log-likelihood ratios (channel LLRs) computed based on the received signals $\mathbf{y}$ by $\lambda_n^t= \ln\frac{P(Y_t=y_i|X_t=+1)}{P(Y_i=y_t|X_t=-1)}=\frac{2}{\sigma^2}y_t$. The outputs of demapping by successive cancellation process are denoted by $\lambda_0^{0,N-1}$ which are simply shown by $\lambda$ in Fig. \[fig:PAC\_scheme\]. Note that the subscript $n$ and $0$ in $\lambda_0^{n,N-1}$ and $\lambda_0^{0,N-1}$ denote respectively the first and the last stages of the SC factor graph shown in Fig. 1 of [@rowshan-stepped]. In the next section, we describe the decoding process and define the path metric. List Viterbi Decoding {#sec:LVA} ===================== The Viterbi algorithm [@viterbi] is the most popular decoding procedure for convolutional codes (CCs), which is based on their [*trellis diagram*]{} graphical representation [@forney]. A trellis is a directed graph where the nodes represent the encoder state. The branch sequences on the trellis are generated by a finite state machine with inputs $\mathbf{v}$ and states $\mathcal{S}=\{s_1,...,s_{2^m}\}$ and the code is called the trellis code. The Viterbi algorithm traverses the trellis from left to right, finding the maximum likelihood transmitted sequence estimate, when reaching the last stage $t=N-1$. PAC codes can be encoded and decoded on the trellis. The trellis used for PAC codes is an irregular trellis which is shown in Fig. \[fig:irregular\_trellis1\] and \[fig:irregular\_trellis2\]. As shown, when there is a sub-sequence of at least $m$ zeros in the input $\mathbf{v}$, the current states of all the paths on the trellis transit toward all-zero state. ![The truncated trellis for PAC codes. Since $v_t=0$ for $t\in \mathcal{A}^c$, the path does not split. The dashed-line arrows represent the input 0 and the solid-line arrows represent the input 1 to the convolutional transform.[]{data-label="fig:irregular_trellis1"}](irregular_trellis1.eps){width="1\columnwidth"} In convolutional coding, there are three methods to obtain the finite code sequences: (1) code truncation where the encoder stops after a finite block-length, $N$, and the code sequence is truncated. This method leads to a substantial degradation of error protection, because the last encoded information bits influence a small number of code bits. (2) code termination where we add some tail bits to the code sequence in order to ensure a predefined end state (usually, the all-zero state) of the encoder, which leads to low error probabilities for the last bits, (3) tail-biting where we choose a starting state that ensures the starting and ending states are the same (this state value does not necessarily have to be the all-zero state). This scheme avoids the rate loss incurred by zero-tail termination at the expense of a more complex decoder. For encoding PAC codes, we use the code truncation, thus we do not add any tail bits. This will not degrade the error protection of last bits because the last encoded bits are transmitted over the high-reliability sub-channels in the polar transform. ![The irregularity of the trellis where $v_t=0$ for $t\in \mathcal{A}^c$ or $t=[i+1,...,j]$ for $j>i$. The paths from $t=i+1$ to $t=j$ are not pruned. []{data-label="fig:irregular_trellis2"}](irregular_trellis2.eps){width="1\columnwidth"} The fundamental idea behind the Viterbi decoding is as follows. A coded sequence $\mathbf{u}$, the output of the convolutional transform in Fig. \[fig:PAC\_scheme\], corresponds to a path through the trellis. Due to the noise in the channel, the received vector $\mathbf{y}$ after demapping may not correspond exactly to a path on the trellis. The decoder finds a path through the trellis which has the highest probability to be the transmitted sequence $\mathbf{u}$ over the polarized vector channel. The probability to be [*[maximized]{}*]{} is $$\label{eq:pac_metric1} \begin{multlined} P(\hat{\mathbf{u}}|\mathbf{y})=\prod_{t=0}^{N-1} P(\hat{u}_t|\hat{u}_0^{t-1},y_0^{N-1}) \end{multlined}$$ In practice, it is convenient to deal with the logarithm of (\[eq:pac\_metric1\]) to use an additive metric. Consider now a partial sequence $\hat{u}_0^{t-1}=[\hat{u}_0, \hat{u}_1,\ldots, \hat{u}_{t-1}]$ at the output of the convolutional transform. This sequence determines a path, or a sequence of states, through the trellis for the code. Let $M_{t-1}(s^\prime)=-\sum_{i=0}^{t-1}\log P(\hat{u_i}|\hat{u}_0^{i-1},y_0^{N-1})$ denote the [*path metric*]{} for the sequence $\hat{u}_0^{t-1}$ terminating in state $s^\prime$. We seek to minimize the path metric for the entire codewords ($t=N-1$) to maximize the probability in (\[eq:pac\_metric1\]). Now let the sequence $\hat{u}_0^{t}$ be obtained by appending $\hat{u}_t$ to $\hat{u}_0^{t-1}$ and suppose $\hat{u}_t$ is such that the state at time $t + 1$ is $s$. The path metric for this longer sequence is $$\begin{aligned} M_t(s)&=&-\sum_{i=0}^{t}\log P(\hat{u}_i|\hat{u}_0^{i-1},y_0^{N-1}) \label{eq:pac_metric2}\\ &=&M_{t-1}(s^\prime)+\mu_t(s^\prime,s) \label{eq:pac_metric3}\end{aligned}$$ where $\mu_t(s^\prime,s)=-\log P(\hat{u}_t|\hat{u}_0^{t-1},y_0^{N-1})$ denotes the [*branch metric*]{} for the trellis transition from state $s'$ at time $t$ to state $s$ at time $t + 1$. The path metric along a path to state $s$ at time $t$ is obtained by adding the path metric to the state $s^\prime$ at time $t - 1$ to the branch metric for an input that moves the encoder from state $s^\prime$ to state $s$. If there is no such input, i.e., $s^\prime$ and $s$ are not connected on the trellis, then the branch metric is considered $\infty$. To simplify the arithmetic operation, we can define $\mu_t$ based on $\lambda_0^t(s^\prime,s)$ or simply $\lambda_0^t$. $$\label{eq:pac_metric4} \begin{multlined} \mu_t(s^\prime,s) = -\log P(\hat{u}_t|\hat{u}_0^{t-1},y_0^{N-1})\\=-\log\left(\frac{e^{(1-\hat{u}_t)\lambda_0^t}}{e^{\lambda_0^t}+1}\right) =\log\left(1+e^{-(1-2\hat{u}_t)\lambda_0^t}\right) \end{multlined}$$ where the last equality holds only for $\hat{u}_t=\hat{u}_t(s^\prime,s)=$ 0 and1. Now, for the value of $\hat{u}_t$ that equals $h(\lambda_0^t)$, $$\label{eq:sc_hard_decision} h(\lambda_0^t) = \begin{dcases*} 0 & $\lambda_0^t>0$,\\ 1 & otherwise\\ \end{dcases*}$$ the term $e^{-(1-2\hat{u}_t)\lambda_0^t}=e^{-|\lambda_0^t|}$ is small and hence $\log(1+~e^{-|\lambda_0^t|}) \approx 0$. Otherwise, we can approximate $\log(1+e^{|\lambda_0^t|})\approx |\lambda_0^t|$. Thus $$\label{eq:pm_func} \mu_t(s^\prime,s)=\mu_t(\lambda_0^t,\hat{u}_t)\! \approx \! \begin{dcases*} 0 & if $\hat{u}_t = h(\lambda_0^t)$\\ |\lambda^t_0| & otherwise \\ \end{dcases*}$$ It turns out that this branch metric is equivalent to the one suggested for the list decoding of polar codes in [@yuan; @balatsoukas] and PAC codes in [@rowshan-pac1]. When paths merge at state $s$, we need to select one of them in order to extend it at the next time step. Suppose $M_{t-1}(s^\prime_0)$ and $M_{t-1}(s^\prime_1)$ are the path metrics of the paths ending at states $s^\prime_0,s^\prime_1\in \{0,1,...,2^m-1\}$ at time $t$. Suppose further that both of these states are connected to state $s$ at time $t + 1$, as illustrated in Fig. \[fig:paths\_merge\]. ![ Merging two paths at state $s$[]{data-label="fig:paths_merge"}](paths_merge.eps){width="0.8\columnwidth"} According to the Bellman’s principle of optimality [@belman], to obtain the maximum likelihood (ML) path through the trellis, the path to any state at each time step must be locally an ML path. This is the governing principle of the Viterbi algorithm. Thus, when the two or more paths merge, the path with the smallest path metric is retained (the [*survivor path*]{} or in short the [*survivor*]{}) and the other path is eliminated from further consideration. This defines the [*add-compare-select*]{} step of the Viterbi algorithm $$\label{eq:pac_metric5} \begin{multlined} M_t(s) = \min\{ M_{t-1}(s^\prime_0) + \mu_t(s^\prime_0,s),\\ M_{t-1}(s^\prime_1) + \mu_t(s^\prime_1,s) \} \end{multlined}$$ Note that the initial path metrics are $M_0(0) = 0$ and $M_0(s^\prime) = \infty$ for $s^\prime = 1,2, . . . , 2^m - 1$. In [@hashimoto], the conventional Viterbi algorithm was generalized to list-type VA where instead of one path, the $L$ paths with smallest metric are selected and extended at time $t$. Hence, (\[eq:pac\_metric5\]) is generalized as $$\label{eq:pac_metric_lva} \begin{multlined} M_t(s,k) = \min_{\substack{1\leq l\leq L\\s^\prime}}^{(k)}\{ M_{t-1}(s^\prime,l) + \mu_t(s^\prime,s) \} \end{multlined}$$ where $\min^{(k)}$ denotes the $k$-th smallest value ($1\leq k\leq L$). From (\[eq:pac\_metric\_lva\]), one can observe some similarity between list decoding of PAC codes and list Viterbi algorithm (LVA). The main difference is that in the LVA, the paths are sorted locally at each state, while in list decoding all the paths are sorted globally and then half of them are discarded. Algorithm \[alg:LVA\] illustrates the list Viterbi algorithm. In the beginning, there is a single path in the list. When the index of the current bit is in the set $\mathcal{A}^c$, the decoder knows its value, usually $v_t=0$, and therefore it is encoded into $u_t$ based on the current memory state $S$ and the generator polynomial $\mathbf{g}$ in line 8. Then, using the decision LLR $\lambda_0^t$ obtained in line 6, the corresponding path metric is calculated using subroutine $calcM$. Note that in the algorithm \[alg:LVA\], instead of $M_t(s,k)$ in \[eq:pac\_metric\_lva\], we use $M_t(k)$. Although the metric is calculated in lines 9 and 26-27 regardless of the current state of each corresponding path, when we sort the paths in line 16, we consider their current states. Eventually, the decoded value $u_t$ is fed back into SC process in line 10 to calculate the partial sums. On the other hand, if the index of the current bit is in the set $\mathcal{A}$ (see lines 12-17), there are two options for the value of $v_t$, i.e., 0 and 1, to be considered in line 23. For each option of 0 and 1, the aforementioned process for $t\in \mathcal{A}^c$ including convolutional encoding, and calculating the path metric is performed and then the two encoded values $u_t=0$ and $1$ are fed back into SCprocess to update the partial sums $\beta_\pi$. The vector $\mathbf{\lambda}[\pi]$ as the input argument of the subroutine [*updateLLRs*]{} constitutes the $N-1$ intermediate LLR values of path $\pi$. The subroutine updateLLRs updates all the intermediate LLRs and gives $\lambda_0^t[\pi]$. Similarly, the vector $\beta_\pi$ constitutes the $N-1$ intermediate partial sums of path $\pi$ which is needed to compute the intermediate LLRs. The partial sums are updated after decoding each bit by the subroutine [*updatePartialSums*]{}. The subroutines [*updateLLRs*]{}, [*updatePartialSums*]{}, and [*prunePaths*]{} in Algorithm \[alg:LVA\] are identical to the ones used in SCL decoding of polar codes. \[alg:LVA\] $\Pi \gets \{1\}$ $m \gets |\mathbf{g}|-1$ $\hat{v}_0^{N-1}[1:|\Pi|] \gets$ sort($\hat{v}_0^{N-1}[1:|\Pi|]$) // [in ascending order]{} $\mathbf{\hat{d}} \gets$ extractData($\hat{v}_0^{N-1}[0]$, $\mathcal{A}$) // [inverse of rate-profiling]{} $\mathbf{\hat{d}}$; Generalization of List Viterbi Algorithm {#sec:universality} ======================================== Successive Cancellation List Viterbi algorithm (SC-LVA or in short LVA) for decoding of PAC codes can be considered a generalized decoder for PAC codes in a sense that it can be converted to SC decoding, SC list decoding and Viterbi decoding by changing the parameters of the algorithm. In terms of sorting strategy for the path metrics at each time step, there are two strategies to consider: - [*global sorting*]{} of all the paths regardless of their current states. In this case, LVA will not have a fixed number of survivors for each state (or at each node on the trellis) and the decoding reduces to SC list decoding (LD) of PAC codes. In this case, the performance improves by increasing the list size, $L$. A special case of list decoding is SC decoding when $L=1$. - [*local sorting*]{} of the paths with the same current state (the paths connected to the same node on the trellis). This is the conventional LVA for PAC codes described in section \[sec:LVA\]. In this case, by increasing either the list size $L$ or the number of states $|\mathcal{S}|$, while keeping the other parameter constant, the performance improves. However, if we keep the product of $L\cdot|\mathcal{S}|$ constant, an increase in $L$ improves the performance. Note that in this case, if $|\mathcal{S}|$ becomes two small such as $|\mathcal{S}|=2$, the convolution has a limited span and results in a degradation in FER performance as we will see in Section \[sec:results\]. Needless to mention that if we increase $L\cdot|\mathcal{S}|$, the performance improves. We note that the PAC code changes by changing $|\mathcal{S}|$, since we are using a different $\mathbf{g}$. Since it was observed that the FER performance of PAc codes is not significantly affected by the change of $\mathbf{g}$, we can vary this parameter and the local list size and observe the tradeoffs of the different decoders. Additionally, when we choose only one path at each state ($L=1$), LVA is converted to a standard Viterbi algorithm (VA) for PAC codes, which was described in Section \[sec:LVA\]. In this case, as the number of states, $|\mathcal{S}|$, on the trellis increases, the performance improves. Also note that PAC coding with $\mathbf{g}=[1]$ or $m=0$ is equivalent to polar coding simply because there is no pre-transformation or pre-coding in this case. Sorting Latency {#sec:sorting} =============== As discussed in the previous section, the error correction performance of the decoding changes with the sorting strategy as well as the list size and the number of states. Now, let us analyse the sorting complexity in list decoding and list Viterbi decoding. Suppose the total number of survivor paths is the same in LD and LVA, i.e., $L_{LD}=L_{LVA}\cdot 2^m$. As we will observe in the next section, in the condition of the same number of survivors, LD slightly outperforms LVA due to the global sorting strategy. However, in case of parallelism which is popular in the hardware design, the local sorting in LVA can improve the latency significantly. ![The reduced bitonic sorting network for LVA with $L=4$. The order of $L$ smallest path metrics is not needed.[]{data-label="fig:sorting_network"}](sorting_network.eps){width="1\columnwidth"} Let us consider a bitonic sorter [@batcher] with $1+\log L$ super-stages that can sort $2L$ path metrics shown in Fig. \[fig:sorting\_network\]. At each super-stage with index $\psi\in\{1,...,1+\log L\}$, there are $\psi$ stages (i.e., the number of stages at each super-stage equals the index ($\psi$) of the corresponding super-stage, see the top and the bottom of Fig. \[fig:sorting\_network\]), each including $L$ pairs of a component (shown by vertical connections in Fig. \[fig:sorting\_network\]) consists of a comparator and 2-to-2 multiplexer, which work in parallel. This sorter was used for list decoding of polar codes in [@lin] and later improved in [@balatsoukas-sorter]. The length of the critical path of the sorter is determined by the total number of stages which is computed based on the sum of the arithmetic progression as follows: $$\label{eq:total_stages} \Psi_{LD} = \sum_{\psi=1}^{1+\log_2 L} \psi=\frac{1}{2}(1+\log_2 L)(2+\log L)$$ From (\[eq:total\_stages\]), one can see the impact of the list size, $L$, on the latency of the sorter and consequently the whole decoder. The pruned bitonic sorter suggested in [@balatsoukas-sorter] removes one stage out of $\Psi_{LD}$ stages, which is not significant in the case of large $L$, although the pruned network reduces the silicon area in hardware implementation. An efficient solution for a significant reduction in the latency is to employ list VA where the sorting is performed locally at each state. Therefore, the parameter $L$ in (\[eq:total\_stages\]) is divided by the number of states. It turns out the the order of the sorted metric in LVA is not needed unlike in the pruned bitonic sorter where the pruning is performed based on our prior knowledge about the order and the relations between adjacent metric before and after the tree extension. Hence, we can remove the last $log_2L$ stages in the last super-stage As a result, the total number of stages is: $$\label{eq:total_stages_lva} \Psi_{LVA} = \frac{1}{2}\left(1+\log_2 \frac{L}{2^{m}}\right)\left(2+\log \frac{L}{2^{m}}\right)-\log_2L$$ Thus, list VA results in a significant reduction in the latency of the decoding. For instance, for list decoding of PAC(256,128) with $m=6$ and $L=128$ which has 128 survivors at each decoding stage, the total number of sorting stages throughout decoding is $K \Psi_{LD}=128\times 36=4608$. However, in decoding of the same code under list VA with $m=4$ and $L=128/2^4=8$ which has 32 survivors at each decoding stage, $K\Psi_{LVA}=128\times (10-3)=896$, which is 80% smaller than its counterpart. Note that this reduction comes at the cost of a slight degradation in the FER performance. In a software implementation, the sorting algorithms such as Heapsort and Mergesort cannot perform better than $O(2L\log (2L))$ in terms of time complexity. By employing LVA, the time complexity reduces to $O(2^m 2L/2^m\log (2L/2^m))=O(2L\log (2L/2^m))$. {width="1\columnwidth"} \[fig:FER1\] {width="1\columnwidth"} \[fig:FER2\] Numerical Results {#sec:results} ================= In this Section, the error correction performance of list Viterbi algorithm for PAC(128,64) on the trellis with different setups is illustrated and analyzed. The RM rate-profile [@rowshan-pac1] and the generator polynomials 0o3, 0o7, 0o17, 0o33, 0o73, and 0o133 ($m=1,\ldots, 6$), are used for convolutional transform (pre-coding) for the results shown in Fig. \[fig:FER1\] corresponding to a number of states $|\mathcal{S}|= 2, 4, 8, 16, 32$ and 64 (for LD), respectively. Also, for the results shown in Fig. \[fig:FER2\], the generator polynomials 0o133, 0o73 and 0o733 are used for convolutional transform with a number of states $|\mathcal{S}|=64$ (for LD), 32 and 256, respectively. The codewords are modulated based on BPSK and transmitted over the AWGN channel. Fig. \[fig:FER1\] compares the FER performance under LVA with various list sizes $L$, while the total number of survivor paths at each time step $t$ remains constant (32 survivors). As can be seen, the performance improves as $L$ increases. Fig. \[fig:FER2\] shows that as the total number of survivors increases, the gap between the performance of LD, VA and LVA decreases. This makes LVA a better candidate when employing a very large list size, given latency advantage shown in Section \[sec:sorting\]. Conversely, when list size is small, the performance of LVA with list size $L$ is close to the performance of LD with list size $L/2$ as it is shown in Fig. \[fig:FER2\] for LVA with $L=32$ and LD with $L=16$. Conclusion ========== In this paper, we investigate the implementation of the list Viterbi decoding for PAC codes. We show that LVA can be considered a general decoding scheme, which can transition from list decoding to Viterbi algorithm decoding by changing the number of states and the local list size. The results show that as the local list size increases, the performance improves. This implies that in the local sorting of the paths, the probability of discarding the correct path is higher than the global sorting in list decoding. On the other hand, the local sorting has the advantage of a significantly lower latency than global sorting. Therefore, depending on the application, we can trade latency for performance, specially when the list size is large. [9]{} E. Arikan, “Channel polarization: A method for constructing capacity-achieving codes for symmetric binary-input memoryless channels," [*IEEE Trans. Inf. Theory,*]{} vol. 55, no. 7, pp. 3051-3073, Jul. 2009. I. Tal and A. Vardy, “List decoding of polar codes," [*IEEE Int. Symp. on Information Theory,*]{} St. Petersburg, Russia, Jul. 2011, pp. 1–5. E. Arikan, “From sequential decoding to channel polarization and back again," arXiv preprint arXiv:1908.09594 (2019). B. Li, H. Zhang, J. Gu, “On Pre-transformed Polar Codes," arXiv preprint arXiv:1912.06359 (2019). M. Rowshan, A. Burg and E. Viterbo, “Polarization-adjusted Convolutional (PAC) Codes: Fano Decoding vs List Decoding," arXiv preprint arXiv:2002.06805 (2020). A. Viterbi, “Error bounds for convolutional codes and an asymptotically optimum decoding algorithm," in [*IEEE Transactions on Information Theory*]{}, vol. 13, no. 2, pp. 260-269, April 1967. G.D. Forney, “The Viterbi Algorithm," [*Proc. of the IEEE*]{}, Vol. 61, No. 3, pp. 268-278, Mar. 1973. E. Arıkan, H. Kim, G. Markarian, U. Ozgur and E. Poyraz, “Performance of short polar codes under ML decoding," in [*Proc. ICT-Mobile Summit Conf.*]{}, Santander, Spain, 2009, pp. 1-6. T. Hashimoto, “A list-type reduced-constraint generalization of the Viterbi algorithm," in [*IEEE Transactions on Information Theory*]{}, vol. 33, no. 6, pp. 866-876, November 1987. N. Seshadri and C.-E. W. Sundberg, “List Viterbi decoding algorithms with applications," in [*IEEE Transactions on Communications*]{}, vol. 42, no. 234, pp. 313-323, Feb-Apr 1994. M. Rowshan and E. Viterbo, “Stepped List Decoding for Polar Codes," 2018 [*IEEE 10th International Symposium on Turbo Codes & Iterative Information Processing (ISTC)*]{}, Hong Kong, Hong Kong, 2018, pp. 1-5. B. Yuan and K. K. Parhi, “Successive cancellation list polar decoder using log-likelihood ratios," [*2014 48th Asilomar Conference on Signals, Systems and Computers*]{}, Pacific Grove, CA, 2014, pp. 548-552. A. Balatsoukas-Stimming, M. Bastani Parizi, and A. Burg, “LLR-based successive cancellation list decoding of polar codes," [*IEEE Trans. Signal Processing,*]{} vol. 63, no. 19, pp. 5165-5179, Oct 2015. R. E. Bellman and S. E. Dreyfus, “Applied Dynamic Programming", Princeton University Press, Princeton, NJ, 1962. K. E. Batcher, “Sorting networks and their applications," in [*Proc. AFIPS Spring Joint Comput. Conf.*]{}, vol. 32, 1968, pp. 307-314. J. Lin and Z. Yan, “Efficient list decoder architecture for polar codes," in [*Proc. IEEE Int. Symp. on Circuits and Systems (ISCAS)*]{}, Jun. 2014, pp. 1022–1025. A. Balatsoukas-Stimming, M. Bastani Parizi and A. Burg, “On metric sorting for successive cancellation list decoding of polar codes," [*2015 IEEE International Symposium on Circuits and Systems (ISCAS)*]{}, Lisbon, 2015, pp. 1993-1996. [^1]: M. Rowshan and E. Viterbo are with the Department of Electrical and Computer Systems Engineering (ECSE), Monash University, Melbourne, VIC3800, Australia. E-mail: {mohammad.rowshan, emanuele.viterbo}@monash.edu. These authors’ work was supported by the Australian Research Council under Discovery Project ARC DP160100528.

--- abstract: 'We introduce a natural nondegeneracy condition for Poisson structures, called holonomicity, which is closely related to the notion of a log symplectic form. Holonomic Poisson manifolds are privileged by the fact that their deformation spaces are as finite-dimensional as one could ever hope: the corresponding derived deformation complex is a perverse sheaf. We develop some basic structural features of these manifolds, highlighting the role played by the divergence of Hamiltonian vector fields. As an application, we establish the deformation-invariance of certain families of Poisson manifolds defined by Feigin and Odesskii, along with the “elliptic algebras” that quantize them.' author: - 'Brent Pym[^1]' - 'Travis Schedler[^2]' bibliography: - 'holonomic-poisson.bib' title: Holonomic Poisson manifolds and deformations of elliptic algebras --- Introduction ============ Holonomic Poisson manifolds =========================== Holonomicity and the modular vector field ========================================= Perverse sheaves and applications {#sec:compfact} ================================= [^1]: School of Mathematics, University of Edinburgh, <brent.pym@ed.ac.uk> [^2]: Department of Mathematics, Imperial College London, <t.schedler@imperial.ac.uk>

--- abstract: 'To calculate realistic models of objects with Ni in their atmospheres, accurate atomic data for the relevant ionization stages needs to be included in model atmosphere calculations. In the context of white dwarf stars, we investigate the effect of changing the Ni [iv]{}-[vi]{} bound-bound and bound-free atomic data has on model atmosphere calculations. Models including PICS calculated with [autostructure]{} show significant flux attenuation of up to $\sim 80$% shortward of 180Å in the EUV region compared to a model using hydrogenic PICS. Comparatively, models including a larger set of Ni transitions left the EUV, UV, and optical continua unaffected. We use models calculated with permutations of this atomic data to test for potential changes to measured metal abundances of the hot DA white dwarf G191-B2B. Models including [autostructure]{} PICS were found to change the abundances of N and O by as much as $\sim 22$% compared to models using hydrogenic PICS, but heavier species were relatively unaffected. Models including [autostructure]{} PICS caused the abundances of N/O [iv]{} and [v]{} to diverge. This is because the increased opacity in the [autostructure]{} PICS model causes these charge states to form higher in the atmosphere, moreso for N/O [v]{}. Models using an extended line list caused significant changes to the Ni [iv]{}-[v]{} abundances. While both PICS and an extended line list cause changes in both synthetic spectra and measured abundances, the biggest changes are caused by using [autostructure]{} PICS for Ni.' author: - | S. P. Preval$^{1,2}$[^1], M. A. Barstow$^{1}$, N. R. Badnell$^{2}$, I. Hubeny$^{3}$, J. B. Holberg$^{4}$\ $^{1}$Department of Physics and Astronomy, University of Leicester, University Road, Leicester, LE1 7RH\ $^{2}$Department of Physics, University of Strathclyde, Glasgow, G4 0NG\ $^{3}$Steward Observatory, University of Arizona, Tucson, AZ 85721, USA\ $^{4}$Lunar and Planetary Laboratory, Sonett Space Sciences Building, University of Arizona, Tucson, AZ 85721, USA\ bibliography: - 'simonpreval.bib' date: 'Accepted XXX. Received YYY; in original form ZZZ' title: 'Hot DA white dwarf model atmosphere calculations: Including improved Ni PI cross sections' --- \[firstpage\] white dwarfs, opacity, atomic data, stellar atmospheres. Introduction ============ The presence of metals in white dwarf (WD) atmospheres can have dramatic effects on both the structure of the atmosphere, and the observed effective temperature ($T_{\mathrm{eff}}$). For example, these effects have been demonstrated convincingly by [@barstow98a]. The authors determined the $T_{\mathrm{eff}}$ and surface gravity (log $g$) of several hot DA white dwarfs using a set of model atmosphere grids, which were either pure H & He, or heavy metal-polluted. It was found that the $T_{\mathrm{eff}}$ determined using the pure H model grid was $\approx{4000-7000}$K higher than if a heavy metal-polluted model grid were used. Conversely, there was little to no difference in the measured log $g$ when using either model grid. It can be inferred that the completeness of the atomic data supplied in model calculations might have a significant effect on the measured $T_{\mathrm{eff}}$. A study by [@chayer95a] considered the effects of radiative levitation on the observed atmospheric metal abundances at different $T_{\mathrm{eff}}$ and log $g$. In addition, these calculations were done using Fe data sets of varying line content. It was found that the number of transitions included in the calculation greatly affected the expected Fe abundance in the atmosphere (cf. @chayer95a, Figure 11). This result implies that the macroscopic quantities determined in a white dwarf, such as metal abundance, are extremely sensitive to the input physics used to calculate the model grids. Therefore, this means that any atomic data that is supplied to the calculation must be as complete and accurate as possible in order to calculate the most representative model. While the [@chayer95a] study considered only the variations in observed Fe abundance, it is reasonable to assume that the set of atomic data supplied may also have an impact on the $T_{\mathrm{eff}}$ and log $g$ measured. Studies of white dwarf metal abundances, such as those of [@barstow98a; @vennes01a; @preval13a], used model atmospheres incorporating the atomic data of [@kurucz92a] (Ku92 hereafter) in conjunction with photoionization cross-section (PICS) data from the Opacity Project (OP) for Fe, and approximate hydrogenic PICS for Ni. A more comprehensive dataset has since been calculated by [@kurucz11a] (Ku11 hereafter), containing a factor $\sim{10}$ more transitions and energy levels for Fe and Ni [iv-vi]{} than its predecessor. In Table \[table:lines\] we give a comparison of the number of Fe and Ni [iv-vii]{} transitions available between Ku92 and Ku11. In both the Ku92 and Ku11 datasets, the energy levels etc were calculated using the Cowan Code [@cowanbook1981]. Based on the work discussed above, it is prudent to explore the differences between model atmospheres calculated using the Ku92 and Ku11 transition data, and the effect this has on measurements made using such models. Ion No of lines 1992 No of lines 2011 ------------ ------------------ ------------------ Fe [iv]{} 1,776,984 14,617,228 Fe [v]{} 1,008,385 7,785,320 Fe [vi]{} 475,750 9,072,714 Fe [vii]{} 90,250 2,916,992 Ni [iv]{} 1,918,070 15,152,636 Ni [v]{} 1,971,819 15,622,452 Ni [vi]{} 2,211,919 17,971,672 Ni [vii]{} 967,466 28,328,012 Total 10,420,643 111,467,026 : The number of lines present in the Ku92 and Ku11 atomic databases for Fe and Ni [iv]{}-[vii]{}. \[table:lines\] Ni [v]{} absorption features were first discovered in the hot DA white dwarfs G191-B2B and REJ2214-492 by [@holberg94a] using high dispersion UV spectra from the International Ultraviolet Explorer (IUE). The authors derived Ni abundances of $\sim 1\times{10}^{-6}$ and $\sim 3\times{10}^{-6}$ as a fraction of H for G191-B2B and REJ2214-492 respectively. Compared to their measurements for Fe of $\sim 3\times{10}^{-5}$ and $\sim 1\times{10}^{-4}$ for G191-B2B and REJ2214-492 respectively, Ni is $\sim 3$% of the Fe abundance for these stars. [@werner94a] also found Ni in two other hot DA white dwarfs, namely Feige 24, and RE0623-377, measuring Ni abundances of $1-5\times{10}^{-6}$ and $1-5\times{10}^{-5}$, respectively. More recently, [@preval13a] and [@rauch13a] have measured Ni abundances for G191-B2B. Using Ni [iv]{} and [v]{}, [@preval13a] found abundances of $3.24\times{10}^{-7}$ and $1.01\times{10}^{-6}$ respectively, while [@rauch13a] measured a single Ni abundance of $6\times{10}^{-7}$. To date, there has been no attempt to include representative PICS for Ni in white dwarf model atmosphere calculations. It is for this reason that we choose to focus on Ni, and examine the effects of both including new PI cross sections, and transition data. We structure our paper as follows; We first describe the atomic data calculations required to generate PICS for Ni [iv-vi]{}. Next, we describe the model atmosphere calculations performed, and the tests we conducted. We then discuss the results. Finally, we state our conclusions. Atomic data calculations ======================== The PICS data presented in this paper were calculated using the distorted wave atomic collision package [autostructure]{} [@badnell86a; @badnell97a; @badnell11a]. [autostructure]{} is an atomic structure package that can model several aspects of an arbitrary atom/ion a priori, including energy levels, transition oscillator strengths, electron-impact excitation cross sections, PICS, and many others. [autostructure]{} is supplied with a set of configurations describing the number of electrons and the quantum numbers occupied for a given atom or ion. The wavefunction $P_{nl}$ for a particular configuration is obtained by solving the one particle Schrodinger equation $$\left[\frac{d^2}{dr^2}-\frac{l(l+1)}{r^2}+2V_{\mathrm{Eff}}(r)+E_{nl}\right]P_{nl}=0,$$ where $n$ and $l$ are the principle and orbital angular momentum quantum numbers respectively, and $V_{\mathrm{Eff}}$ is an effective potential as described by [@eissner69a], which accounts for the presence of other electrons. Scaling parameters $\lambda_{nl}$ can be included to scale the radial coordinate, which are related to the effective charge ‘seen’ by a particular valence electron, and typically has a value close to unity. The $\lambda_{nl}$ can be varied according to the task. For an example, the parameters may be varied to minimise an energy functional, or they may be varied such that the difference between the calculated energy levels and a set of observed energy levels is minimised. Three coupling schemes are available in calculating the wavefunctions, dependent upon the resolution required, and the type of problem being considered. These are LS coupling (term resolved), intermediate coupling (IC, level resolved), or configuration average (CA, configuration resolved). We use IC with the aim of reproducing the energy levels from Ku11 as closely as possible. The energy level structure is first determined by running [autostructure]{} with the configurations used by Ku11, listed in Table \[table:asconfig\]. With this information, 17 $\lambda_{nl}$ are then specified for orbitals 1s to 6s set initially to unity. The 1s parameter is fixed as it does not converge in the presence of relativistic corrections. The parameters were then varied to give as close agreement to the Ku11 data as possible. The result of parameter variation is given in Table \[table:asparm\]. Ion Configurations $N_\mathrm{config}$ Levels Lines ----------- -------------------------------------------------------------------------------------------------------- --------------------- -------- ------------ Ni [iv]{} $\mathbf{3d^7}$, $3d^6 4s$, $3d^6 5s$,$3d^6 6s$, $3d^6 7s$, $3d^6 8s$ 85 37,860 32,416,571 $3d^6 9s $, $3d^5 4s^2 $, $3d^5 4s5s $, $3d^5 4s6s $, $3d^5 4s7s $, $3d^5 4s8s $ $3d^5 4s 9s $, $3d^4 4s^2 5s$, $3d^6 4d $, $3d^6 5d $, $3d^6 6d $, $3d^6 7d $ $3d^6 8d $, $3d^6 9d $, $3d^5 4s4d $, $3d^5 4s5d $, $3d^5 4s6d $, $3d^5 4s7d $ $3d^5 4s 8d $, $3d^5 4s 9d $, $3d^4 4s^2 4d$, $3d^5 4p^2 $, $3d^6 5g $, $3d^6 6g $ $3d^6 7g $, $3d^6 8g $, $3d^6 9g $, $3d^5 4s 5g $, $3d^5 4s 6g $, $3d^5 4s 7g $ $3d^5 4s 8g $, $3d^5 4s 9g $, $3d^6 7i $, $3d^6 8i $, $3d^6 9i $, $3d^5 4s 7i $ $3d^5 4s 8i $, $3d^5 4s 9i$ $3d^6 4p $, $3d^6 5p $, $3d^6 6p $, $3d^6 7p $, $3d^6 8p $, $3d^6 9p $ $3d^5 4s 4p $, $3d^5 4s 5p $, $3d^5 4s 6p $, $3d^5 4s 7p $, $3d^5 4s 8p $, $3d^5 4s 9p $ $3d^4 4s^2 4p $, $3d^6 4f $, $3d^6 5f $, $3d^6 6f $, $3d^6 7f $, $3d^6 8f $ $3d^6 9f $, $3d^5 4s 4f $, $3d^5 4s 5f $, $3d^5 4s 6f $, $3d^5 4s 7f $, $3d^5 4s 8f $ $3d^5 4s 9f $, $3d^4 4s^2 4f$, $3d^6 6h $, $3d^6 7h $, $3d^6 8h $, $3d^6 9h $ $3d^5 4s 6h $, $3d^5 4s 7h $, $3d^5 4s 8h $, $3d^5 4s 9h $, $3d^6 8k $, $3d^6 9k $ $3d^5 4s 8k $, $3d^5 4s 9k $, $3p^5 3d^8$ Ni [v]{} $\mathbf{3d^6}$, $3d^5 4d$, $3d^5 5d$, $3d^5 6d$, $3d^5 7d $, $3d^5 8d $ 87 37,446 34,066,259 $3d^5 9d $, $3d^5 10d $, $3d^4 4s4d $, $3d^4 4s5d $, $3d^4 4s6d $, $3d^4 4s7d $ $3d^4 4s 8d $, $3d^4 4s9d $, $3d^4 4s10d $, $3d^5 4s $, $3d^5 5s $, $3d^5 6s $ $3d^5 7s $, $3d^5 8s $, $3d^5 9s $, $3d^5 10s $, $3d^4 4s^2 $, $3d^4 4s5s $ $3d^4 4s 6s $, $3d^4 4s7s $, $3d^4 4s8s $, $3d^4 4s9s $, $3d^4 4s10s $, $3d^5 5g $ $3d^5 6g $, $3d^5 7g $, $3d^5 8g $, $3d^5 9g $, $3d^4 4s5g $, $3d^4 4s6g $ $3d^4 4s 7g $, $3d^4 4s8g $, $3d^4 4s9g $, $3d^5 7i $, $3d^5 8i $, $3d^5 9i $ $3d^4 4s 7i $, $3d^4 4s8i $, $3d^4 4s9i $, $3d^4 4p^2$ $3d^5 4p $, $3d^5 5p $, $3d^5 6p $, $3d^5 7p $, $3d^5 8p $, $3d^5 9p $ $3d^5 10p $, $3d^4 4s4p $, $3d^4 4s5p $, $3d^4 4s6p $, $3d^4 4s7p $, $3d^4 4s8p $ $3d^4 4s 9p $, $3d^4 4s10p $, $3d^3 4s^2 4p$, $3d^5 4f $, $3d^5 5f $, $3d^5 6f $ $3d^5 7f $, $3d^5 8f $, $3d^5 9f $, $3d^5 10f $, $3d^4 4s4f $, $3d^4 4s5f $ $3d^4 4s 6f $, $3d^4 4s7f $, $3d^4 4s8f $, $3d^4 4s9f $, $3d^5 6h $, $3d^5 7h $ $3d^5 8h $, $3d^5 9h $, $3d^4 4s6h $, $3d^4 4s7h $, $3d^4 4s8h $, $3d^4 4s9h $ $3d^5 8k $, $3d^5 9k $, $3d^4 4s8k $, $3d^4 4s9k $, $3p^5 3d^7$ Ni [vi]{} $\mathbf{3d^5}$, $3d^4 4d$, $3d^4 5d$, $3d^4 6d$, $3d^4 7d$, $3d^4 8d$ 122 29,366 42,412,822 $3d^4 9d $, $3d^4 10d $, $3d^3 4s 4d $, $3d^3 4s5d $, $3d^3 4s6d $, $3d^3 4s7d $ $3d^3 4s 8d $, $3d^3 4s 9d $, $3d^3 4s 10d $, $3d^2 4s^2 4d $, $3d^2 4s^2 5d $, $3d^2 4s^2 6d $ $3d^2 4s^2 7d $, $3d^2 4s^2 8d $, $3d^2 4s^2 9d $, $3d^2 4s^2 10d $, $3d^4 4s $, $3d^4 5s $ $3d^4 6s $, $3d^4 7s $, $3d^4 8s $, $3d^4 9s $, $3d^4 10s $, $3d^3 4s^2 $ $3d^3 4s 5s $, $3d^3 4s 6s $, $3d^3 4s 7s $, $3d^3 4s8s $, $3d^3 4s9s $, $3d^3 4s10s $ $3d^2 4s^2 5s $, $3d^2 4s^2 6s $, $3d^2 4s^2 7s $, $3d^2 4s^2 8s$, $3d^2 4s^2 9s$, $3d^2 4s^2 10s$ $3d^4 5g $, $3d^4 6g $, $3d^4 7g $, $3d^4 8g $, $3d^4 9g $, $3d^4 10g $ $3d^3 4s 5g $, $3d^3 4s 6g $, $3d^3 4s 7g $, $3d^3 4s8g $, $3d^3 4s9g $, $3d^3 4s10g $ $3d^4 7i $, $3d^4 8i $, $3d^4 9i $, $3d^3 4s7i $, $3d^3 4s8i $, $3d^3 4s9i $, $3d^3 4p^2$ $3d^4 4p $, $3d^4 5p $, $3d^4 6p $, $3d^4 7p $, $3d^4 8p $, $3d^4 9p $ $3d^4 10p $, $3d^4 11p $, $3d^3 4s 4p $, $3d^3 4s5p $, $3d^3 4s6p $, $3d^3 4s7p $ $3d^3 4s 8p $, $3d^3 4s 9p $, $3d^3 4s 10p $, $3d^3 4s11p $, $3d^2 4s^2 4p $, $3d^2 4s^2 5p $ $3d^2 4s^2 6p$, $3d^2 4s^2 7p$, $3d^2 4s^2 8p $, $3d^2 4s^2 9p $, $3d^2 4s^2 10p $, $3d^2 4s^2 11p$ $3d^4 4f $, $3d^4 5f $, $3d^4 6f $, $3d^4 7f $, $3d^4 8f $, $3d^4 9f $ $3d^4 10f $, $3d^4 11f $, $3d^3 4s 4f $, $3d^3 4s5f $, $3d^3 4s6f $, $3d^3 4s7f $ $3d^3 4s 8f $, $3d^3 4s 9f $, $3d^3 4s 10f $, $3d^3 4s11f $, $3d^2 4s^2 4f $, $3d^2 4s^2 5f $ $3d^2 4s^2 6f $, $3d^2 4s^2 7f $, $3d^2 4s^2 8f$, $3d^2 4s^2 9f$, $3d^2 4s^2 10f $, $3d^2 4s^2 11f$ $3d^4 6h $, $3d^4 7h $, $3d^4 8h $, $3d^4 9h $, $3d^3 4s6h $, $3d^3 4s7h $ $3d^3 4s 8h $, $3d^3 4s 9h $, $3d^4 8k $, $3d^4 9k $, $3d^3 4s8k $, $3d^3 4s9k$, $3p^5 3d^6$ \[table:asconfig\] Orbital Ni [iv]{} Ni [v]{} Ni [vi]{} --------- ----------- ---------- ----------- $2s$ 1.31391 1.31355 1.31487 $2p$ 1.12294 1.12144 1.11996 $3s$ 1.09887 1.11219 1.12750 $3p$ 1.05876 1.07206 1.08779 $3d$ 1.06828 1.09148 1.10379 $4s$ 1.13492 1.15167 1.19770 $4p$ 0.91520 0.90669 0.92207 $4d$ 1.40156 1.48359 1.49036 $4f$ 1.08824 1.03771 1.10538 $5s$ 1.03946 1.03552 1.25294 $5p$ 0.99302 0.96743 1.00439 $5d$ 1.09409 1.08876 1.15671 $5f$ 1.06373 1.01089 1.01693 $5g$ 1.37199 1.17219 1.83712 $6s$ 1.02002 1.01140 1.05828 : Calculated IC scaling parameters from [autostructure]{}. \[table:asparm\] By comparing Table \[table:lines\] with Table \[table:asconfig\], it can be seen that the latter has more transitions than the former. This is because the transitions listed in Ku11 are limited by their strength. Any observed/well known transitions with log $gf<-9.99$, or predicted transitions with a log $gf<-7.5$ were omitted. In addition, the Ku11 database omitted radiative transitions between two autoionizing levels. We do the same. After calculating a set of scaling parameters for a particular ion, the accompanying PICS can be obtained. In order to see what potential effect (if any) replacing the hydrogenic PICS with more realistic data would have, we limited our calculations to considered direct photoionization (PI) only, neglecting resonances from photoexcitation/autoionization. The final photoionized configurations used in calculating the PICS were constructed by removing the outer most electron from each configuration in Table \[table:asconfigpi\]. These PI configurations are listed in Table \[table:asconfigpi\]. The PICS are evaluated for a table of 50 logarithmically spaced ejected electron energies spanning 0 to 100 Ryd. The PICS in the ejected electron energy frame are then linearly interpolated to the incident photon energy frame using two point interpolation. Ion Configurations ----------- -------------------------------------------------------- Ni [iv]{} $3p^5 3d^7$, $3d^6$, $3d^5 4s$, $3d^5 4p$, $3d^4 4s^2$ Ni [v]{} $3p^5 3d^6$, $3d^5$, $3d^4 4s$, $3d^4 4p$, $3d^3 4s^2$ Ni [vi]{} $3d^4$, $3p^5 3d^5$, $3d^3 4s$, $3d^3 4p$, $3d^2 4s^2$ : Final photoionized configurations used in the [autostructure]{} calculations. \[table:asconfigpi\] [autostructure]{} employs the distorted wave method, which is an approximation. The calculations performed for the OP were done using an R-Matrix approach, which can potentially give the most accurate result in calculating PICS. Furthermore, the R-Matrix method automatically includes resonances arising from photoexcitation/autoionization, and interference between these two processes. In [autostructure]{}, when the resonances are included separately, interference effects are neglected. The downside to using an R-Matrix calculation is the length of time and computer resources required to perform the calculation. With regards to accuracy, [@seaton2004a] calculated term-resolved PI calculations for Fe [viii]{} to Fe [xiii]{} using [autostructure]{}. The authors then replaced data calculated for the OP, which consisted of R-Matrix plus [superstructure]{} [@eissner69a] data for these ions, and re-evalulated the Rosseland Means for a solar mixture. The Rosseland means calculated using the [autostructure]{} data were found to be close to those calculated with the OP data. This implies that the distorted wave method is a good indicator of the potential effects of including new data. Therefore, it is instructive to perform such calculations before committing to a large R-Matrix calculation. Stellar atmosphere calculations {#sec:stelatcal} =============================== All model atmospheres in this work were calculated using the non-local thermodynamic equilibrium (NLTE) stellar atmospheres code [tlusty]{} [@hubeny88a; @hubeny95a], version 201. The models were then synthesised using [synspec]{} [@hubeny11a]. [tlusty]{} benefits from the hybrid CL/ALI method, which combines the Complete Linearisation (CL) and Accelerated Lambda Iteration (ALI) methods in order to accelerate the rate of convergence of a model. [tlusty]{} has two methods in which to treat heavy metal opacity, namely opacity distribution functions (ODF), and opacity sampling (OS). Nominally, OS is a Monte-Carlo sampling method. However, in the limit of high resolution, OS is an exact method for accounting for opacity. We use OS, and specify a resolution of 5 fiducial[^2] Doppler widths. As the Ku11 data contains more energy levels than Ku92, we constructed new model ions for Ni [iv-vi]{} according to the prescription of [@anderson89a], however, superlevels were calculated using energies of either even or odd parity. If we created superlevels with a mixture of even or odd levels, then we would have to consider transitions between levels within the superlevel (cf. @hubeny95a). The Ni [iv-vi]{} ions have 73, 90, and 75 superlevels respectively. The PICS for each superlevel $\bar{\sigma}_{\mathrm{PI}}(E_{\gamma})$ as a function of photon energy $E_{\gamma}$ were calculated as an average of the PICS for each individual level $\sigma_{i}(E)$ used to create the superlevel, weighted by the statistical weights of each level $g_{i}$. This can be written as $$\bar{\sigma}_{\mathrm{PI}}(E_{\gamma})=\frac{\sum_{i=1}^{N}{\sigma_{i}(E_{\gamma})g_{i}}}{\sum_{i=1}^{N}{g_{i}}},$$ where $N$ is the number of levels used to form the superlevel. We refer to these summed PICS as supercross-sections hereafter. In Figure \[fig:xsects\] we have plotted the total supercross-sections for Ni [iv-vi]{} for both the hydrogenic and autostructure cases. Prior to summation, each supercross-section was multiplied by a boltzmann constant. It can be seen that the total supercross-sections calculated in [autostructure]{} are far larger than their hydrogenic counterparts. ![Plot of summed supercross-sections for Ni [iv-vi]{} calculated using autostructure (solid red) and a hydrogenic approximation (dotted blue). Prior to addition, each supercross-section was weight by a boltzmann factor and the statistical weight of the superlevel concerned.[]{data-label="fig:xsects"}](xsectout.eps){width="80mm"} As there are two different Ni line lists (Ku92 and Ku11), and two different sets of PICS (hydrogenic and those calculated with [autostructure]{}), there are four different combinations that we can test. Therefore, we calculated four different models in NLTE. We refer to these as Models 1, 2, 3, and 4. In Model 1, we use the Ku92 transitions and hydrogenic PICS. In Model 2, we use the Ku11 transitions and hydrogenic PICS. In Model 3, we use the Ku92 transitions and [autostructure]{} PICS. Finally, in Model 4, we use the Ku11 transitions and [autostructure]{} PICS. In all four cases, we based the models on a G191-B2B like atmosphere, and calculated the models with $T_{\mathrm{eff}}=52500$K, log $g=7.53$ [@barstow03b], and metal abundances listed in Table \[table:atmosabun\]. These abundances were taken from [@preval13a], measured in their analysis of the hot DA white dwarf G191-B2B. In the case where there was more than one ionization stage considered, we used the abundances with the smallest uncertainty. Listed in Table \[table:opacions\] are the model ions used in [tlusty]{}, along with the number of superlevels included. Metal Abundance X/H ------- ----------------------- He $1.00\times{10}^{-5}$ C $1.72\times{10}^{-7}$ N $2.16\times{10}^{-7}$ O $4.12\times{10}^{-7}$ Al $1.60\times{10}^{-7}$ Si $3.68\times{10}^{-7}$ P $1.64\times{10}^{-8}$ S $1.71\times{10}^{-7}$ Fe $1.83\times{10}^{-6}$ Ni $1.01\times{10}^{-6}$ : Metal abundances used in calculating the four model atmospheres described in text as a fraction of H. These abundances originate from [@preval13a], where the values with the lowest statistical uncertainty were used. \[table:atmosabun\] Ion Nsuperlevels -------------- -------------- H [i]{} 9 H [ii]{}\* 1 He [i]{} 24 He [ii]{} 20 He [iii]{}\* 1 C [iii]{} 23 C [iv]{} 41 C [v]{}\* 1 N [iii]{} 32 N [iv]{} 23 N [v]{} 16 N [vi]{}\* 1 O [iv]{} 39 O [v]{} 40 O [vi]{} 20 O [vii]{}\* 1 Al [iii]{} 23 Al [iv]{}\* 1 Si [iii]{} 30 Si [iv]{} 23 Si [v]{}\* 1 P [iv]{} 14 P [v]{} 17 P [vi]{}\* 1 S [iv]{} 15 S [v]{} 12 S [vi]{} 16 S [vii]{}\* 1 Fe [iv]{} 43 Fe [v]{} 42 Fe [vi]{} 32 Fe [vii]{}\* 1 Ni [iv]{} 38 (73) Ni [v]{} 48 (90) Ni [vi]{} 42 (75) Ni [vii]{}\* 1 : List of model ions and the number of levels used in model atmosphere calculations described in the text. Ions marked with \* were treated approximately as single level ions by [tlusty]{}. For Ni, the number of levels outside and inside the brackets correspond to the number of levels for the Ku92 and Ku11 model ions, respectively. \[table:opacions\] Spectral Energy Distribution ---------------------------- For this comparison, we considered the differences between the spectral energy distributions (SED) of each model. For each model, we synthesize three spectra covering the extreme ultraviolet (EUV), the ultraviolet (UV), and the optical regions. We then calculated the residual between models 1 and 2, 1 and 3, and 1 and 4 using the equation $$\label{eq:res} \mathrm{Residual}=\frac{F_{\mathrm{i}}-F_{\mathrm{1}}}{F_{\mathrm{1}}},$$ where $F_{\mathrm{1}}$ is the flux for model 1, and $F_{\mathrm{i}}$ is the flux for model 2, 3, or 4. Abundance variations -------------------- For this comparison, we wanted to examine the differences between abundances measured for G191-B2B when using each of the four models described above. Using a similar method to [@preval13a], we measured the abundances for G191-B2B using all four of the models described above. The observational data for G191-B2B consists of three high S/N spectra constructed by co-adding multiple datasets. The first spectrum uses data from the Far Ultraviolet Spectroscopic Explorer (FUSE) spanning 910-1185Å, and the other two use data from the Space Telescope Imaging Spectrometer (STIS) aboard the Hubble Space Telescope (HST) spanning 1160-1680Å and 1625-3145Å, respectively. A full list of the data sets used, and the coaddition procedure, is given in detail in [@preval13a]. The model grids for each metal was constructed by using [synspec]{}. [synspec]{} takes a starting model converged assuming NLTE, and is able to calculate a spectrum for smaller or larger metal abundances by stepping away in LTE. We used the X-ray spectral package [xspec]{} [@arnaud96a] to measure the abundances. [xspec]{} takes a grid of models and observational data and interpolates between these models using a chi square ($\chi^2$) minimisation procedure. [xspec]{} is unable to use observational data with a large number of data points. To remedy this, we isolate individual absorption features for various ions and then use [xspec]{} to measure the abundances. A full list of the absorption features used and the sections of spectrum extracted is given in Table 9 in [@preval13a]. In addition to this list, we also include measurements of the O [v]{} abundance using the excited transition with wavelength 1371.296Å. Results and discussion. ======================= Here we discuss the results obtained from the four models calculated using permutations of the Ku92 and Ku11 atomic data, and the hydrogenic and [autostructure]{} cross-section data. Models 1, 2, 3, and 4 used Ku92/hydrogenic, Ku11/hydrogenic, Ku92/[autostructure]{}, and Ku11/[autostructure]{} respectively. SED variations -------------- In this subsection we discuss the differences between spectra synthesised for the four models described above. ### EUV Of all the spectral regions, the EUV undergoes the most dramatic changes. However, the EUV region appears to be relatively insensitive to whether Ku92 or Ku11 is used. In Figure \[fig:g191euvcomp1\] we have plotted the EUV region for models 1 and 2, along with the residual between these two models as defined in text. Below 200Å the flux of model 2 appears to increase as wavelength decreases, being $\sim 15$% larger by 50Å. This may be due to how the superlevels are partitioned in the model calculation rather than a decrease in opacity from the Ku11 line list. In Figure \[fig:g191euvcomp2\] we plot the same region, but with models 1 and 3. Significant changes occur below 180Å, with the flux of model 3 being greatly attenuated, reaching a maximum of $\sim 80$% with respect to model 1. This is indicative of a larger opacity due to the [autostructure]{} PICS for Ni. Model 4 showed a combination of effects from models 2 and 3. {width="140mm"} {width="140mm"} ### UV In the case of the UV region, not a lot changes when using the [autostructure]{} PICS. In Figure \[fig:g191uvcomp1\] we have plotted synthetic spectra for Models 1 and 3 in the UV region. It can be seen that changes are limited to absorption features only, with the vast majority only changing depth by $\sim 3$%. The obvious exception to this is the N [v]{} doublet near 1240Å, where the depth has changed by $\sim 5-6$%. In Figure \[fig:g191uvcomp2\] we have now plotted models 1 and 2. Again, changes are limited to absorption features, but these are now far more pronounced, with depth changes of up to and beyond $10$%. These features can be attributed to Ni, and a few lighter metals, the abundances of which we discuss later. Again, Model 4 showed a combination of the changes seen in models 2 and 3. {width="140mm"} {width="140mm"} ### Optical Very little to no change occurs in the optical region, regardless of line list or PICS used to calculate the models. In Figures \[fig:g191optcomp1\] and \[fig:g191optcomp2\] we plot the synthetic spectra of models 1 and 2, and 1 and 3 in the optical region respectively. In both cases, changes to both the continuum flux and H-balmer lines can be seen, but these are restricted to $<0.1$%. The same also occurs for model 4. Because these changes are so small, it is highly unlikely that measurements made using these models would be significantly different. {width="140mm"} {width="140mm"} Abundance measurements ---------------------- In Table \[table:newobsabun\] we list the abundance measurements made using the various models. We have also given the abundance differences between models 1 and 2, 1 and 3, and 1 and 4. Seven ions were found to have statistically significant abundance differences dependent on which model was used, namely N [v]{}, O [iv]{}, O [v]{}, Fe [iv]{}, Fe [v]{}, Ni [iv]{}, and Ni [v]{}. ### For N [v]{}, significant changes are seen in models 2, 3, and 4. Using model 1, we measured the abundance of N [v]{} to be $1.65^{+0.02}_{-0.02}\times{10}^{-7}$, whereas for models 2, 3, and 4, we find $1.77^{+0.02}_{-0.02}\times{10}^{-7}$, $1.87^{+0.02}_{-0.02}\times{10}^{-7}$, and $1.99^{+0.02}_{-0.02}\times{10}^{-7}$ respectively. The abundances measured using models 2, 3, and 4 correspond to increases from model 1 of $\sim 7$%, $\sim 13$%, and $\sim 20$% respectively. This suggests that both the number of transitions and the PICS used cause significant changes to the abundance. ### In the case of O [iv]{}, significant changes in the abundances are seen when using the [autostructure]{} PICS in models 3 and 4. For O [iv]{}, we measured the abundance to be $4.63^{+0.12}_{-0.12}\times{10}^{-7}$ using model 1. For models 3 and 4, we measured abundances of $4.38^{+0.12}_{-0.12}\times{10}^{-7}$ and $4.31^{+0.12}_{-0.12}\times{10}^{-7}$ respectively, corresponding to decreases of $\sim 5$% and $\sim 7$% respectively. A similar case occurs for O [v]{}, where statistically significant differences occur for models 3 and 4. Using model 1, we measured an abundance of $1.47^{+0.07}_{-0.07}\times{10}^{-6}$, whereas for models 3 and 4 we measured abundances of $1.69^{+0.08}_{-0.08}\times{10}^{-6}$ and $1.79^{+0.08}_{-0.08}\times{10}^{-6}$ respectively. This corresponds to an increase of $\sim 15$% and $\sim 22$% for models 3 and 4 over model 1 respectively. These results suggest that the largest changes to the O [iv]{}-[v]{} may be cause by the PICS rather than the number of transitions included in the line list. ### The Fe [iv]{}-[v]{} abundances appear relatively insensitive to changes in the line list used and the PICS. A statistically significant difference was only observed between abundances measured using models 1 and 4. For Fe [iv]{} we measured an abundance of $2.05^{+0.04}_{-0.04}\times{10}^{-6}$ for model 1, and $1.98^{+0.04}_{-0.04}\times{10}^{-6}$ for model 4. This is a $\sim 3$% decrease from model 1. For Fe [v]{}, we measured abundances of $5.37^{+0.08}_{-0.08}\times{10}^{-6}$ and $5.20^{+0.07}_{-0.07}\times{10}^{-6}$ for models 1 and 4 respectively, corresponding to a $\sim 3$% decrease. These changes appear to suggest that a combination of both the number of transitions and PICS is required to change the abundance. We discuss this further below. ### Interestingly, statistically significant changes compared to model 1 are seen in the Ni [iv]{} abundance when models 2 and 3 are used, but not for model 4. For model 1, we measured the abundance to be $3.00^{+0.09}_{-0.04}\times{10}^{-7}$. For models 2 and 3, we measure the abundances to be $2.81^{+0.05}_{-0.05}\times{10}^{-7}$ and $3.32^{+0.10}_{-0.10}\times{10}^{-7}$ respectively. This corresponds to a decrease of $\sim 6$% for model 2, and an increase of $\sim 11$% for model 3. In the case of model 4, an abundance of $2.97^{+0.12}_{-0.05}\times{10}^{-7}$ was measured. The Ni [v]{} abundance appears to depend strongly upon whether Ku92 or Ku11 atomic data was included. In model 1, an abundance of $9.88^{+0.28}_{-0.21}\times{10}^{-7}$ was measured. For model 2, we measured the abundance to be $1.22^{+0.04}_{-0.04}\times{10}^{-6}$, being a $\sim 23$% increase from model 1. For model 3, only a very small difference was noted. We measured the abundance to be $9.81^{+0.23}_{-0.21}\times{10}^{-6}$, which is a $<1$% decrease. In model 4, we again see a large increase in the abundance from model 1, measuring $1.20^{+0.04}_{-0.04}\times{10}^{-6}$, which is a $\sim 21$% increase. ### Ionization fraction agreement In addition to the differences noted above, abundances for different ionization stages of N and O were found to diverge depending upon the PICS or atomic data used. In the case of N, the difference between the N [iv]{} and N [v]{} abundances for model 1 is $0.02^{+0.22}_{-0.22}\times{10}^{-7}$, whereas for models 2, 3, and 4 the differences are $-0.12^{+0.22}_{-0.22}\times{10}^{-7}$, $-0.25^{+0.22}_{-0.22}\times{10}^{-7}$, and $-0.39^{+0.22}_{-0.22}\times{10}^{-7}$ respectively. For O, the difference between the O [iv]{} and O [v]{} abundances for model 1 is $-1.01^{+0.07}_{-0.07}\times{10}^{-6}$, whereas for models 2, 3, and 4 the differences are $-1.11^{+0.07}_{-0.07}\times{10}^{-6}$, $-1.25^{+0.08}_{-0.08}\times{10}^{-6}$, and $-1.36^{+0.08}_{-0.08}\times{10}^{-6}$ respectively. The reason for this can be seen upon inspection of the ionization fractions for N and O. In Figures \[fig:ionnitro\] and \[fig:ionoxy\] we have plotted the ionization fractions for N and O against column mass for models 1 and 3. In both cases it can be seen that the model 3 ionization fractions have been shifted to smaller column masses. In the case of O, this effect is far more pronounced. In addition, it can also be seen that the shift to smaller column masses is larger for N/O [v]{} than it is for N/O [iv]{}. Therefore, this explains why the abundance measurements diverge. Notwithstanding changes to the atomic data and PICS, the overall agreement between abundances measured for different ionization stages of particular species is generally poor. For this work, we adopted $T_{\mathrm{eff}}=52,500$K and log $g=7.53$ as measured by [@barstow03b] for G191-B2B for models 1 to 4. These values were used for consistency with the work described by [@preval13a]. Since then, measurements of $T_{\mathrm{eff}}$ and log $g$ for G191-B2B have been revised upward by [@rauch13a] to 60,000K and 7.60 respectively. The agreement between abundances measured for different ionization stages is a sensitive function of $T_{\mathrm{eff}}$, atmospheric composition, and to a lesser extent (sans H) log $g$. Our work was not focused on finding the best combination of $T_{\mathrm{eff}}$, log $g$, and atmospheric composition, but instead focused on whether a change, if any, occured to the measured abundances when altering the atomic data and PICS. {width="140mm"} {width="140mm"} ### Abundance differences Interestingly, the difference between abundances measured using models 1 and 4 can be related to the differences between abundances measured using models 1 and 2, and 1 and 3. For example, the difference between the N [v]{} abundances measured using models 1 and 2 is $-0.12^{+0.02}_{-0.02}\times{10}^{-7}$, whereas for models 1 and 3, it is $-0.23^{+0.02}_{-0.02}\times{10}^{-7}$. If we add these differences from the abundance found in model 1, we obtain a total of $-0.35^{+0.03}_{-0.03}\times{10}^{-7}$. It is for this reason that we include an extra column in Table \[table:newobsabun\], where the differences between abundances measured using models 1 and 2, and 1 and 3, are summed together. It can be seen that in all cases, the sum of these is equal to (within the uncertainties) the 1-4 column. This is easily explained in terms of the opacity. Recall that the total opacity in a stellar atmosphere is just a linear sum of each individual contribution. In this case, it is the bound-free (PICS) and the bound-bound (Ku92 or Ku11) that is being added. Model 2 and Model 3 use the Ku11/Hydrogenic PICS and the Ku92/[autostructure]{} PICS respectively. Given that Model 1 uses the Ku92/Hydrogenic PICS, substracting the abundance found in Model 2 from Model 1 shows the effect of including more Ni transitions. Likewise, substracting the abundance found in Model 3 from Model 1 shows the effect of including more realistic PICS. Therefore, adding these two differences together will give the combination of these two effects. This explains why statistically significant differences were observed only when using model 4 for Fe [iv]{}-[v]{}, in that the effects of both the line list and the PICS combine. Ion Model 1 Model 2 Model 3 Model 4 1 - 2 1 - 3 1 - 4 $\Sigma$ ------------ ---------------------------- ---------------------------- ---------------------------- ---------------------------- -------------------------------------- -------------------------------------- -------------------------------------- ----------------------------- C [iii]{} $1.83^{+0.03}_{-0.03}[-7]$ $1.83^{+0.03}_{-0.03}[-7]$ $1.83^{+0.03}_{-0.03}[-7]$ $1.83^{+0.03}_{-0.03}[-7]$ $ 0.00^{+0.04}_{-0.04}[-7]$ $ 0.00^{+0.04}_{-0.04}[-7]$ $ 0.00^{+0.04}_{-0.04}[-7]$ $ 0.00^{+0.06}_{-0.06}[-7]$ C [iv]{} $3.00^{+0.10}_{-0.14}[-7]$ $3.00^{+0.10}_{-0.17}[-7]$ $3.00^{+0.10}_{-0.17}[-7]$ $3.00^{+0.09}_{-0.20}[-7]$ $ 0.00^{+0.14}_{-0.22}[-7]$ $ 0.00^{+0.15}_{-0.22}[-7]$ $ 0.00^{+0.14}_{-0.24}[-7]$ $ 0.00^{+0.21}_{-0.31}[-7]$ N [iv]{} $1.67^{+0.22}_{-0.22}[-7]$ $1.65^{+0.22}_{-0.22}[-7]$ $1.62^{+0.22}_{-0.22}[-7]$ $1.60^{+0.22}_{-0.22}[-7]$ $ 0.02^{+0.31}_{-0.31}[-7]$ $ 0.05^{+0.31}_{-0.31}[-7]$ $ 0.07^{+0.31}_{-0.31}[-7]$ $ 0.07^{+0.44}_{-0.44}[-7]$ N [v]{} $1.65^{+0.02}_{-0.02}[-7]$ $1.77^{+0.02}_{-0.02}[-7]$ $1.87^{+0.02}_{-0.02}[-7]$ $1.99^{+0.02}_{-0.02}[-7]$ $\mathit{-0.12^{+0.02}_{-0.02}[-7]}$ $\mathit{-0.23^{+0.02}_{-0.02}[-7]}$ $\mathit{-0.35^{+0.02}_{-0.02}[-7]}$ $-0.35^{+0.03}_{-0.03}[-7]$ O [iv]{} $4.63^{+0.12}_{-0.12}[-7]$ $4.54^{+0.12}_{-0.12}[-7]$ $4.38^{+0.12}_{-0.12}[-7]$ $4.31^{+0.12}_{-0.12}[-7]$ $ 0.09^{+0.17}_{-0.17}[-7]$ $\mathit{ 0.25^{+0.17}_{-0.17}[-7]}$ $\mathit{ 0.32^{+0.17}_{-0.17}[-7]}$ $ 0.34^{+0.24}_{-0.24}[-7]$ O [v]{} $1.47^{+0.07}_{-0.07}[-6]$ $1.56^{+0.07}_{-0.07}[-6]$ $1.69^{+0.08}_{-0.08}[-6]$ $1.79^{+0.08}_{-0.08}[-6]$ $-0.09^{+0.10}_{-0.10}[-6]$ $\mathit{-0.23^{+0.10}_{-0.10}[-6]}$ $\mathit{-0.32^{+0.11}_{-0.11}[-6]}$ $-0.32^{+0.14}_{-0.14}[-6]$ Al [iii]{} $1.62^{+0.09}_{-0.09}[-7]$ $1.63^{+0.10}_{-0.10}[-7]$ $1.63^{+0.10}_{-0.10}[-7]$ $1.63^{+0.10}_{-0.10}[-7]$ $-0.01^{+0.13}_{-0.13}[-7]$ $-0.01^{+0.13}_{-0.13}[-7]$ $-0.01^{+0.13}_{-0.13}[-7]$ $-0.02^{+0.18}_{-0.18}[-7]$ Si [iii]{} $2.90^{+0.39}_{-0.23}[-7]$ $2.89^{+0.38}_{-0.23}[-7]$ $2.92^{+0.41}_{-0.23}[-7]$ $2.92^{+0.41}_{-0.23}[-7]$ $ 0.00^{+0.54}_{-0.32}[-7]$ $-0.02^{+0.57}_{-0.33}[-7]$ $-0.02^{+0.56}_{-0.33}[-7]$ $-0.02^{+0.79}_{-0.46}[-7]$ Si [iv]{} $3.32^{+0.20}_{-0.20}[-7]$ $3.33^{+0.20}_{-0.20}[-7]$ $3.34^{+0.20}_{-0.20}[-7]$ $3.35^{+0.20}_{-0.20}[-7]$ $-0.01^{+0.28}_{-0.28}[-7]$ $-0.02^{+0.28}_{-0.28}[-7]$ $-0.03^{+0.28}_{-0.28}[-7]$ $-0.03^{+0.40}_{-0.40}[-7]$ P [iv]{} $1.34^{+0.20}_{-0.20}[-7]$ $1.34^{+0.20}_{-0.20}[-7]$ $1.30^{+0.19}_{-0.20}[-7]$ $1.30^{+0.19}_{-0.20}[-7]$ $ 0.00^{+0.28}_{-0.28}[-7]$ $ 0.03^{+0.28}_{-0.28}[-7]$ $ 0.04^{+0.28}_{-0.28}[-7]$ $ 0.03^{+0.40}_{-0.40}[-7]$ P [v]{} $1.91^{+0.03}_{-0.03}[-8]$ $1.91^{+0.03}_{-0.03}[-8]$ $1.91^{+0.03}_{-0.03}[-8]$ $1.90^{+0.03}_{-0.03}[-8]$ $ 0.01^{+0.04}_{-0.04}[-8]$ $ 0.00^{+0.04}_{-0.04}[-8]$ $ 0.01^{+0.04}_{-0.04}[-8]$ $ 0.01^{+0.06}_{-0.06}[-8]$ S [iv]{} $2.01^{+0.03}_{-0.03}[-7]$ $1.99^{+0.03}_{-0.03}[-7]$ $1.99^{+0.03}_{-0.03}[-7]$ $1.96^{+0.03}_{-0.03}[-7]$ $ 0.03^{+0.05}_{-0.05}[-7]$ $ 0.02^{+0.05}_{-0.05}[-7]$ $ 0.05^{+0.05}_{-0.05}[-7]$ $ 0.05^{+0.07}_{-0.07}[-7]$ S [vi]{} $7.55^{+0.20}_{-0.20}[-8]$ $7.64^{+0.20}_{-0.20}[-8]$ $7.51^{+0.20}_{-0.20}[-8]$ $7.59^{+0.20}_{-0.20}[-8]$ $-0.08^{+0.28}_{-0.28}[-8]$ $ 0.05^{+0.28}_{-0.28}[-8]$ $-0.03^{+0.28}_{-0.28}[-8]$ $-0.03^{+0.40}_{-0.40}[-8]$ Fe [iv]{} $2.05^{+0.04}_{-0.04}[-6]$ $2.01^{+0.04}_{-0.04}[-6]$ $2.02^{+0.04}_{-0.04}[-6]$ $1.98^{+0.04}_{-0.04}[-6]$ $ 0.04^{+0.05}_{-0.05}[-6]$ $ 0.03^{+0.05}_{-0.05}[-6]$ $\mathit{ 0.07^{+0.05}_{-0.05}[-6]}$ $ 0.07^{+0.07}_{-0.07}[-6]$ Fe [v]{} $5.37^{+0.08}_{-0.08}[-6]$ $5.31^{+0.07}_{-0.07}[-6]$ $5.27^{+0.07}_{-0.07}[-6]$ $5.20^{+0.07}_{-0.07}[-6]$ $ 0.06^{+0.11}_{-0.11}[-6]$ $ 0.10^{+0.11}_{-0.11}[-6]$ $\mathit{ 0.17^{+0.11}_{-0.11}[-6]}$ $ 0.16^{+0.16}_{-0.16}[-6]$ Ni [iv]{} $3.00^{+0.09}_{-0.04}[-7]$ $2.81^{+0.05}_{-0.05}[-7]$ $3.32^{+0.10}_{-0.10}[-7]$ $2.97^{+0.12}_{-0.05}[-7]$ $\mathit{ 0.19^{+0.10}_{-0.06}[-7]}$ $\mathit{-0.32^{+0.13}_{-0.10}[-7]}$ $ 0.03^{+0.15}_{-0.07}[-7]$ $-0.13^{+0.16}_{-0.12}[-7]$ Ni [v]{} $9.88^{+0.28}_{-0.21}[-7]$ $1.22^{+0.04}_{-0.04}[-6]$ $9.81^{+0.23}_{-0.21}[-7]$ $1.20^{+0.04}_{-0.04}[-6]$ $\mathit{-2.30^{+0.29}_{-0.22}[-7]}$ $ 0.06^{+0.37}_{-0.30}[-7]$ $\mathit{-2.12^{+0.29}_{-0.22}[-7]}$ $-2.24^{+0.47}_{-0.37}[-7]$ \[table:newobsabun\] General discussion ------------------ Ideally, any calculation should be as accurate as possible including the most up-to-date data available. However, this also needs to be balanced in terms of time constraints, and the task at hand. We have seen that in the EUV, the choice of using either Ku92 or Ku11 is irrelevant as the change is very small. The shape of the continuum, however, is very sensitive to the PICS used. The downside to using the larger line list from Ku11 increases the calculation time significantly. For example, Model 1 took $\sim 17500$ seconds (292 minutes) to converge, whereas Model 2 took $\sim 37000$ seconds (617 minutes). This is because the Ku11 data has more energy levels, and is hence split into a larger number of superlevels than for Ku92 data. From a wider perspective, the PICS calculated using [autostructure]{} caused the most changes, in that the EUV continuum was severly attenuated, and abundances for N and O were changed. When using the Ku11 line list in model atmospheres, abundances for Ni changed significantly while the continua for various spectral regions were left relatively unchanged. Given that a calculation with Ku11 data takes twice as long to do than with Ku92, an abundance change in only Ni [iv]{}-[v]{} is relatively little payoff compared to the physics we can learn from changing the PICS. The way forward in improving the quality of future model atmosphere calculations is clear; effort should be focused on improving the PICS data for ions where it exists, as well as filling in gaps where it is required (in this case, for Ni). This piece of work has been a proof-of-concept endeavour. While we have shown that replacing hydrogenic cross-section data with more realistic calculations has a significant effect on synthesised spectra and measurements, we have only considered direct PI. If a direct PI-only calculation has this large an effect, then it stands to reason that a full calculation including photoexcitation/autoionization resonances will cause a greater effect. The applications of this work is not limited to white dwarf stars. This data can be used in stellar atmosphere models for objects of any kind, and any temperature range. We chose to demonstrate the effects of our calculations on a hot DA white dwarf star as calculations for these objects are relatively simple. At this temperature regime, we do not need to worry about the effects of 3D modelling, convection etc. Future work ----------- As mentioned in our discussion, this work has been a proof-of-concept. The next step is to extend our PICS calculations to include other ions of Ni. Once this is done, we plan to include the omitted resonances to our calculations, and re-examine the effect including this data has on model spectra and measurements. The present PICS data was calculated using a distorted wave approximation. A potentially more accurate calculation can be achieved using the R-matrix method as in the Opacity Project. Therefore, we aim to do some test calculations to compare Ni PICS using both the R-matrix or distorted wave approximation. The EUV spectra of hot metal polluted white dwarfs has historically been difficult to model (cf. @lanz96a [@barstow98a]), the key to which may be the input atomic physics. Therefore, we will also consider the quality of fits to the EUV spectra of several metal polluted white dwarf stars. Conclusion ========== We have presented our PICS calculations of Ni [iv]{}-[vi]{} using the distorted wave code [autostructure]{}. We investigated the effect of using two different line lists (Ku92 and Ku11) and two different sets of PICS (hydrogenic and [autostructure]{}) on synthesized spectra and abundance measurements based on the hot DA white dwarf G191-B2B. This investigation was done by calculating four models (labelled 1, 2, 3, and 4) with permutations of the Ni line list and PICS used. Model 1 used Ku92 line list/hydrogenic PICS, Model 2 used Ku11 line list/hydrogenic PICS, Model 1 used Ku92 line list/[autostructure]{} PICS, and model 4 used Ku11 line list/[autostructure]{} PICS. We synthesised model spectra for each of the four models in the EUV, UV, and optical regions. In the EUV, model 3 showed large attentuation shortward of 180Å of up to $\sim 80$% relative to model 1, whereas model 2 was relatively unchanged. In the UV, the continuum was unchanged in models 2 and 3. However, in model 2, the Ni absorption feature depths changed significant, increasing in depth by up to $\sim 10$%. Absorption features in model 3 were relatively unchanged, with depth changes of $\sim 3$% across the spectrum. In the optical, changes in flux were so small ($<0.1$%) across models that these are unlikely to be observed, nor would it be possible to differentiate between them. Model 4 was not plotted in the EUV, UV, or optical as the resultant spectrum was just a combination of the effects observed in models 2 and 3. We measured metal abundances for G191-B2B using all four models. This was to see if there were any differences in the metal abundances measured when changing the PICS or atomic data included in the model calculation. Statistically significant (consistent with non-zero difference compared to model 1) abundance changes were observed in N [v]{} over all models, and O [iv]{}-[v]{} when using models 3 and 4. This suggests that the N abundances are sensitive to both the line list and PICS used, while the O abundances were only sensitive to the PICS. The Fe [iv]{}-[v]{} abundances only changed by a statistically significant amount for model 4, implying a combination of the line list and cross-section caused the change. The Ni [iv]{}-[v]{} abundances changed by a statistically significant amount for models 2 and 4, implying the line list caused the difference. Interestingly, for each metal abundance, the difference between measurements made using models 1 and 4 could be found by summing the differences between measurements made using models 1 and 2, and models 1 and 3. This is in keeping with the assumption that predicted radiation for small variations of the opacity sources scales roughly linearly with the opacity. In addition, we found the abundances of N/O [iv]{} and [v]{} diverged depending on the PICS used. A comparison of the ionization fractions calculated using models 1 and 3 showed that the charge states for model 3 formed higher in the atmosphere than the charge states for model 1. Furthermore, N/O [v]{} experiences a larger change with respect to depth formation than N/O [iv]{}, explaining the divergence of the abundance measurements. Our work has demonstrated that, even with a limited calculation, the Ni PICS have made a significant difference to the synthetic spectra, and by extension what is measured from observational data. Comparatively, an extended line list such as Ku11 offers little benefit given the extended computational time required to converge a model atmosphere, and the small pay off (changed Ni abundance). It is our opinion that future atomic data calculations for stellar atmosphere models should not necessarily focus on how big the line list is, but the quality of the PICS. Acknowledgments {#acknowledgments .unnumbered} =============== We gratefully acknowledge the time and effort expended by Robert Kurucz in assisting with this project. We also thank Simon Jeffery and Nigel Bannister for helpful discussions. SPP and MAB acknowledge the support of an STFC student grant. \[lastpage\] [^1]: E-mail: simon.preval@strath.ac.uk [^2]: the Doppler width for Fe absorption features at $T_{\mathrm{eff}}$.

--- abstract: 'We discuss the problem of coexistence of genuine quantum multipartite correlations and classical multipartite correlations. We introduce a postulate which any measure of genuine multipartite classical correlations should satisfy. We show that covariance does not satisfy this postulate. Finally we propose a definition of genuine multipartite correlations and illustrate it with examples.' author: - Andrzej Grudka - 'Micha[ł]{} Horodecki' - 'Pawe[ł]{} Horodecki' - Ryszard Horodecki title: Note on genuine multipartite classical correlations --- One of the most important problems in quantum information theory is the problem of quantifying correlations. Henderson and Vedral raised the problem of separating total correlations in a bipartite state into a quantum and a classical part. They also proposed a measure of purely classical bipartite correlations [@Henderson1]. It was shown [@Hayden1] that there exist bipartite states which have almost maximal entanglement of formation and almost no mutual information and hence, almost no classical correlations. In a series of papers a thermodynamical approach to quantifying correlations was developed [@Oppenheim1; @Horodecki3]. It is well known that bits of information can be used to extract work from the heat bath. If we have a bipartite quantum state we can ask how much work we can extract from the heat bath under different classes of operations. In particular quantum information deficit was defined. It is a difference between globally and locally (with the use of local operations and classical communication) extractable work from the heat bath. Recently it was shown that if we only allow sending a particle from one party to another through a dephasing channel (i.e. classical channel) then there exist quantum states for which quantum information deficit is equal to quantum mutual information [@Pankowski1]. Hence, only quantum correlations can be used to extract work from the heat bath. Kaszlikowski *et al.* raised the problem of coexistence of quantum and classical correlations in multipartite systems [@Kaszlikowski]. They constructed a class of $n$-qubit states for which $n$-party covariance defined as $\text{Cov}(X_{1},... X_{n})=\langle (X_{1}-\langle X_{1}\rangle... (X_{n}-\langle X_{n}\rangle) \rangle$ is zero for all choices of local observables $X_{i}$ and the state is entangled across any bipartite cut. In this note we expose the weakness of $n$-party covariance as an indicator of $n$-party correlations and show that the class of states considered by Kaszlikowski *et al.* is classically correlated across any bipartite cut. Let us begin with an example of a tripartite classically correlated state for which $3$-party covariance vanishes. We take as our example the following state $$\varrho=\frac{1}{2}(|000\rangle \langle 000|+|111\rangle \langle 111|)$$ One can easily check that three-party covariance vanishes for all local observables (i.e. Pauli operators). However, the state has tripartite correlations. On the other hand if we take a similar but fourpartite state, i.e. $$\varrho=\frac{1}{2}(|0000\rangle \langle 0000|+|1111\rangle \langle 1111|)$$ and we choose as local observables $\sigma_{z,i}$ then four-party covariance equals to one. It is clear that for equal mixture of two pure states the value of four-party covariance depends on the difference in the number of $|1\rangle$’s in two terms. We now give a postulate which each measure of multipartite correlations should satisfy. [**Postulate.**]{} *If the $n$-partite state does not have genuine $n$-partite correlations according to some measure, then after adding $k$ ancillas in a product state (each ancilla can be given to one party only), performing by each party local operations on his particle and his ancillas, and distributing ancillas to $k$ new parties, the $n+k$-partite state does not have genuine $n+k$-partite correlations according to the same measure.* Note that the four-partite state can be obtained from three-partite state by performing $\text{CNOT}$ gate with a control qubit from the three-partite state and an additional ancilla prepared in the state $|0\rangle$.The party who performed $\text{CNOT}$ gate sends his ancilla to the fourth party. We observe that for three particle state covariance is equal to $0$ and for four particle state (obtained with the previously mentioned operations) covariance is equal to $1$. Hence, covariance does not satisfy our postulate. We also propose a definition of multipartite classical correlations (see also: [@Zhou] ). Let us assume that we have an $n$-particle state. We divide the particles into two disjoint sets $A$ and $B$. Next we perform local measurement on each particle in the set $A$ and each particle in the set $B$., i.e., we do not perform collective measurements even on qubits which are in the same set. [**Definition 1.**]{} *We say that multipartite classical state, i.e. embedded classical distribution has classical correlations across bipartite cut $A:B$ if the probability distribution does not factorize. We say that multipartite classical state has genuine multipartite classical correlations if it has classical correlations across any bipartite cut.* Note by the way that if the classical distribution is product under any $1:(n-1)$ bipartite cut then it is completely product, i.e., has no classical correlations at all. [**Definition 2.** ]{} *We say that multipartite quantum state has classical correlations across bipartite cut $A:B$ if there exist local single partite measurements such that the outcomes of measurements on particles in the set $A$ are classically correlated with outcomes of measurement on particles in the set $B$ in the sense of Definition 1. We say that the multipartite quantum state has genuine multipartite classical correlations if it has classical correlations across any bipartite cut.* [**Observation.**]{} Below we show that genuine multipartite classical correlations can take different forms. One could define genuine multipartite classical correlations if there are classical correlations between any pair of particles. On the other hand one could define genuine multipartite classical correlations if the state of $n$ particles is classically correlated across any bipartite cut and the state of any $n-1$ particles is not classically correlated across any bipartite cut. Obviously, for $n \geq 3$ these two types of classical correlations cannot coexist. However Definition 2 covers all types of genuine multipartite classical correlations. Here are examples of states having those opposite types of genuine multipartite classical correlations. Let us consider the $n$-partite state: $$\varrho=\frac{1}{2}\sum_{i=0}^{1}|ii...i\rangle \langle ii...i|$$ For this state mutual information is equal to $1$ for any bipartite cut. According to Definition 2 this state has genuine multipartite classical correlations. Moreover there are maximal correlations between any pair of particles, i.e. if we trace $n-2$ particles, then $I(A:B)=1$ for the remaining two particles. Let us also consider another $n$-partite state: $$\varrho=\frac{2}{n}\sum_{i_{1}i_{2}...i_{n}=0: \text{even parity}}^{1}|i_{1}i_{2}...i_{n}\rangle \langle i_{1}i_{2}...i_{n}|$$ For this state mutual information is also equal to $1$ for any bipartite cut. According to Definition 2 this state has also genuine multipartite classical correlations. However there are no classical correlations between any $n-1$ particles, i.e. if we trace one particle, then the state is product for the remaining $n-1$ particles. Similar postulate and definitions can be applied in entanglement theory for pure states. Analogous examples would be then W and GHZ states. Both have genuine multipartite entanglement, yet the types are opposite. E.g. for three particles, if we trace one qubit from W state, then the remaining two particles are entangled. If we trace one qubit from GHZ state, then the remaining two particles are not entangled. For mixed states, the definitions would be lifted by convexity, i.e. a mixed state is called genuinely multipartite entangled if it cannot be represented as a mixture of states that are all not genuinely multiparticle entangled. Further in this paper we shall analyze in more details the class of states introduced by Kaszlikowski *et al.* [@Kaszlikowski] in light of the above discussion. These are $n$-partite ($n$ is odd) states of the form: $$\varrho=\frac{1}{2}(|W \rangle \langle W|+|W'\rangle \langle W'|)$$ where $$|W\rangle=\frac{1}{\sqrt{n}}(|00...01\rangle + |00...10\rangle + ... |10...00\rangle)$$ and $$|\overline{W}\rangle=\frac{1}{\sqrt{n}}(|11...10\rangle + |11...01\rangle + ... |01...11\rangle)$$ One can check that $n$-party covariance vanishes for all choices of local observables. Let us choose as local observables $\sigma_{z,i}$ then $n$-party covariance vanishes because $W$-state contains one $|1\rangle$ and $\overline{W}$-state contains even number of $|1\rangle$. We see simliarity with the before mentioned classically correlated states. Henderson and Vedral gave properties which should be satisfied by a measure of classical correlations $C$: 1\. $C=0$ for $\varrho_{AB}=\varrho_{A} \otimes \varrho_{B}$ 2\. $C$ is invariant under local unitary transformations 3\. $C$ is non-increasing under local operations 4\. $C=S(\varrho_{A})=S(\varrho_{B})$ for pure states. They also suggested the following measure of classical correlations: $$C_{B}(\varrho_{AB})=\text{max}_{B_{i}^{\dagger}B_{i}} S(\varrho_{A})-\sum_{i}p_{i}S(\varrho_{A}^{i}),$$ where $B_{i}^{\dagger}B_{i}$ is POVM performed on the subsystem $B$ and $\varrho_{A}^{i}=\text{Tr}_{B}(B_{i}\varrho_{AB}B_{i}^{\dagger})/\text{Tr}_{AB}(B_{i}\varrho_{AB}B_{i}^{\dagger})$ is postmeasurement state of the subsystem $A$ after obtaining an outcome $i$. We now show that the state under consideration is classically correlated across any bipartite cut. We daphase each qubit in the basis $\{ |0\rangle, |1\rangle \}$. Due to property 3 dephasing does not increase classical correlations. After dephasing the state of the system is: $$\varrho=\frac{1}{2n}(|00...01\rangle \langle 00...01| +....+ |10...00\rangle \langle 10...00|+ |11...10\rangle \langle 11...10| +....+ |01...11\rangle \langle 01...11|$$ One can check that Henderson-Vedral measure of classical correlations of the dephased state is equal to quantum mutual information $I(A:B)=S(\varrho_{A})+S(\varrho_{B})-S(\varrho_{AB})$. It is also important that one can attain maximum by performing local von Neumann measurements in the basis $\{ |0\rangle, |1\rangle \}$ on each qubit in the subset $B$. The von Neumann entropy of the density matrix is $S(\varrho)=\log(2n)$. Let us now calculate the reduced density matrix of the subset of $k$ particles $\varrho_{k}$. Because of high symmetry of the state the reduced density matrix depends only on the number of particles in the subset and does not depend on the particular choice of particles. A straightforward calculation gives $$\begin{aligned} \varrho_{k}=\frac{n-k}{2n}|00...00\rangle \langle 00...00|+\frac{1}{2n}|00...01\rangle \langle 00...01|+...+\frac{1}{2n}|10...00\rangle \langle 10...00|+\nonumber\\ +\frac{n-k}{2n}|11...11\rangle \langle 11...11|+\frac{1}{2n}|11...10\rangle \langle 11...10|+...+\frac{1}{2n}|01...11\rangle \langle 01...11|\end{aligned}$$ The von Neumann entropy of the reduced density matrix for $k \geq 3$ is: $$S(\varrho_{k})=1+H(\frac{k}{n})+\frac{k}{n}\log k,$$ where $H(x)=-x\log(x)-(1-x)\log(1-x)$ is binary entropy. For $k=1$ and $k=2$ we have $S(\varrho_{1})=1$ and $S(\varrho_{2})=1+H(\frac{2}{n})$ respectively. Now for any bipartite cut (with $k$ particles in the first set and $n-k$ particles in the second set) we have the following expression for quantum mutual information of the dephased state for $n \geq 5$ and $3 \leq k \leq n-3$ : $$I(A:B)=S(\varrho_{k})+S(\varrho_{n-k})-S(\varrho)=1+H(\frac{k}{n})$$ For $k=1$ or $k=n-1$ and $k=2$ or $k=n-2$ we have: $$I(A:B)=1$$ and $$I(A:B)=H(\frac{2}{n})+\frac{n-2}{n}$$ respectively. For $n=3$ and $k=1$ or $k=2$ we obtain: $$I(A:B)=\frac{1}{3}.$$ We see that quantum mutual information is nonzero for any odd $n$ and any $k$, i.e., there are classical correlations for any bipartite cut. It is also interesting to calculate quantum mutual information between any pair of particles of the dephased state: $$I(A:B)=S(\varrho_{1})+S(\varrho_{1})-S(\varrho_{2})=1-H(\frac{2}{n})$$ Hence for any odd $n$ there are classical correlations between any pair of particles. Let us now show that if we define genuine multipartite classical correlations according to Definiition 2 then a state which is non-product across any cut is classically correlated. [**Lemma.**]{} *Any $n$-qubit state which is non-product across any cut has genuine multipartite classical correlations.* *Proof.* We show that the state which does not have classical multipartite correlations across some bipartite cut $A:B$ has to be product across this cut. We divide qubits of an $n$-qubit state into two disjoint sets $A$ and $B$. We perform on each qubit information complete measurement given by six POVM elements $\{(I \pm \sigma_{i})/6: i=x,y,z\}$. Thus the measurement on the whole system is also information complete. If the outcomes of measurements on particles in the set $A$ are not correlated with outcomes of measurements on particles in the set $B$ then by definition the probability distribution has to factorize across cut $A:B$. However the probability distribution of outcomes of information complete measurement specifies density matrix. We obtain that the density matrix of the whole system has to be product across cut $A:B$. In conclusion we have proposed a postulate which should be satisfied by each measure of genuine multipartite classical correlations and we have shown that covariance does not satisfy it. We have also proposed a definition of genuine multipartite classical correlations and discussed different types of such correlations. We believe that our postulate will lead to deeper understanding of nature of multipartite correlations. We acknowledge discussion with D. Kaszlikowski, M. Piani, A. Sen(De), U. Sen, V. Vedral, and A. Winter. This work was supported by the European Commission through the Integrated Project FET/QIPC “SCALA”. [7]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , ****, (). , , , ****, (). , , , , ****, (). , , , , , , , ****, (). , (). , , , , , (). , , , , ****, ().

--- abstract: 'The light curve of an eclipsing system shows anomalies whenever the eclipsing body passes in front of active regions on the eclipsed star. In some cases, the pattern of anomalies can be used to determine the obliquity $\Psi$ of the eclipsed star. Here we present a method for detecting and analyzing these patterns, based on a statistical test for correlations between the anomalies observed in a sequence of eclipses. Compared to previous methods, ours makes fewer assumptions and is easier to automate. We apply it to a sample of 64 stars with transiting planets and 24 eclipsing binaries for which precise space-based data are available, and for which there was either some indication of flux anomalies or a previously reported obliquity measurement. We were able to determine obliquities for ten stars with hot Jupiters. In particular we found $\Psi \lesssim$10$^\circ$ for Kepler-45, which is only the second M dwarf with a measured obliquity. The other 8 cases are G and K stars with low obliquities. Among the eclipsing binaries, we were able to determine obliquities in 8 cases, all of which are consistent with zero. Our results also reveal some common patterns of stellar activity for magnetically active G and K stars, including persistently active longitudes.' author: - 'Fei Dai, Joshua N. Winn, Zachory Berta-Thompson, Roberto Sanchis-Ojeda, Simon Albrecht' bibliography: - 'manuscript.bib' title: 'Stellar obliquities and magnetic activities of Planet-Hosting Stars and Eclipsing Binaries based on Transit Chord Correlation' --- Introduction ============ The Solar System is nearly coplanar, with the Sun’s equator and the planetary orbits aligned to within a few degrees. This makes sense because the Sun and the planets likely formed from a flattened disk of material with a well-defined sense of rotation. However, the discovery of planets with orbits nearly perpendicular to the star’s equator [e.g., HAT-P-11b, @Winn2010], or nearly retrograde [e.g., WASP-17b, @Triaud2010] show that planet formation and evolution can be more complicated than in this simple picture. There is hope that the distribution of obliquities, and its dependence on stellar and planetary properties, will help to elucidate the formation history and orbital evolution of planets [as recently reviewed by @WinnFabrycky2015; @Triaud2017]. Many methods are available for measuring stellar obliquities: the Rossiter-McLaughlin effect [@Queloz2000]; the $v \rm{sin} i_\star$ method [@Schlaufman2010; @Winn2017vsini]; the photometric variability method [@Mazeh+2015]; the gravity-darkening method [@Barnes2009; @Masuda2015]; the asteroseismic method [@Chaplin2013; @Huber2013; @VanEylen2014]; and the method based on spot-crossing anomalies [@Sanchis-Ojeda2011WASP; @Desert2011; @Mazeh2015; @Holczer2015], the subject of this paper. When a transiting planet is blocking a dark starspot, the loss of light is smaller than when it is blocking an unspotted portion of the photosphere. This produces a positive glitch in the light curve. Similarly, transits over bright regions (plages and faculae) produce negative glitches. In some cases the stellar obliquity can be deduced from the pattern of recurrence of spot-crossing anomalies in a sequence of transits. As a simple example, consider a low-obliquity star with long-lived starspots and a rotation period $P_{\rm rot}$ much longer than the planet’s orbital period $P_{\rm orb}$. In such a case, the anomalies that are seen during one transit will recur in the next transit, but shifted forward in time because the active regions have advanced across the the visible hemisphere of the star due to stellar rotation. Conversely, if the stellar obliquity is high, the rotation of the star carries the active regions outside of the “transit chord”, the narrow strip on the stellar disk that is traversed by the transiting planet. In this case, the anomalies produced by a given spot will not recur in subsequent transits. Unlike some of the other methods for measuring obliquities, the spot-crossing method does not require high-resolution spectroscopy, and can therefore be applied to relatively faint stars. However, the spot-crossing method does require a large number of consecutive light curves with a high signal-to-noise ratio (SNR) to allow spot-crossing events and their recurrence to be detected. As a result, it has only been applied to about a dozen systems. Most of these were discovered by the space missions [*CoRoT*]{} and [*Kepler*]{}. The method of analysis has generally relied on the identification of discrete anomalies, often through visual inspection, a laborious and somewhat subjective procedure. Most of the modeling has been performed with idealized assumptions such as circular spots of uniform intensity. The motivation for our work was to develop a more objective method that does not rely on explicit spot modeling, and can be more easily applied to a large sample of systems. Instead of assuming that the active regions are discrete dark and bright spots, we treat the transit light curve as a measure of the intensity distribution of the stellar photosphere along the transit chord. We do not model the intensity distribution with discrete spots, although we still must hope that the active regions persist and are nearly stationary in the rotating frame of the star for at least a few orbital periods. Given the values of $P_{\rm rot}$ and $P_{\rm orb}$, we can calculate the angle by which the active regions should advance in between transits if the obliquity is low. Then we can seek evidence for correlations — with the appropriate lag — between the anomalies observed in a series of transits. A significant correlation implies a low obliquity. We can search for evidence of retrograde rotation in a similar way. For convenience we call this the Transit Chord Correlation (TCC) method, although we do not claim it is a completely new concept. It is closely related to eclipse mapping [@Horne1985], which has long been used to probe the brightness distribution of stars and accretion disks. Eclipse mapping has also been applied to a couple of stars with transiting planets [@Huber2010corot; @Scandariato2017]. The main difference is that the previous investigators assumed zero obliquity and sought to determine the brightness distribution across the transit chord, while we use the method to try and determine the obliquity. Similar in spirit to our method is the one described by @Mazeh2015 and @Holczer2015. They were able to distinguish prograde and retrograde motion in a few [*Kepler*]{} systems by searching for a significant correlation between the observed transit timing variations (TTV) and the time derivative of the stellar flux immediately before and after transits. Stars with prograde rotation should show a correlation if a single active region is responsible for both the apparent TTV and the out-of-transit variation, while retrograde stars should show an anti-correlation. Two advantages of this method are its simplicity and the freedom from the requirement that the active regions have a lifetime of a few orbital periods. One problem is that the out-of-transit variations are the net effect of all the active regions on the star, and are not necessarily dominated by the region that is responsible for the transit anomalies. Our TCC method does not rely on the out-of-transit variations, and is also able to provide higher precision in the obliquity determination. In this paper, we explain the TCC method, validate it through application to systems for which the stellar obliquity has been measured by independent methods, and apply it to all the transiting planets for which the method is currently feasible. We also apply the TCC method to a sample of eclipsing binaries drawn from the [ *Kepler*]{} survey. Section 2 describes the TCC method in greater detail. Section 3 describes the target selection and light curve preparation and analysis. Section 4 presents the results for the obliquities, as well as some interesting features we noticed in the pattern and time evolution of active regions. Section 5 discusses these results in the context of what was previously known about both stellar obliquities and stellar activity. Method {#Method} ====== Conceptual illustration ----------------------- Figure \[sky\_model\] illustrates the concept of the TCC method. This figure shows the face of a star and the corresponding light curves of three consecutive transits. On the left are the light curves for a well-aligned star ($\Psi \approx 0^{\circ}$) with a random pattern of active regions that slowly rotates across the visible hemisphere. The anomalies are seen to move progressively to the right along the time axis. We isolate the anomalies by subtracting the best-fitting transit model. Then we transform the time coordinate into the longitude $\Phi$ of the star that is being crossed by the planet, using the equation $$\label{eqn:longitude} \Phi(t) = \sin^{-1} \left[ \frac{x(t)}{\sqrt{1-y(t)^2}} \right] - \frac{2\pi(t-t_{\rm ref})}{P_{\text{rot}}}$$ where $x(t)$ and $y(t)$ are the sky-plane coordinates of the planet in the system defined by @Winn2010b, in which the origin is at the center of the stellar disk and the $x$ direction is aligned with the planet’s trajectory. The reference time $t_{\rm ref}$ is arbitrary; we adopt the usual [*Kepler*]{} convention of BJD2454833. Note that $\Phi$ is the longitude of the star in its own rotating frame. Thus, active regions maintain a constant value of $\Phi$ even as they rotate across the star’s visible hemisphere, as long as they do not evolve or migrate significantly on the timescale of the orbital period of the planet (typically 1–10 days, for the planets considered in this work). We also note that this transformation assumes that $P_{\rm rot}$ is a constant and thereby neglects the effects of latitudinal differential rotation. When the star has a low obliquity, $P_{\rm rot}$ should be understood as the rotation period at the latitude of the transit chord. Now that we have calculated the stellar longitude that is being blocked by the planet as a function of time, we can plot the light-curve residuals as a function of stellar longitude. This is shown in the bottom panels of Figure \[sky\_model\]. The three patterns of residuals are seen to be very similar, as expected, since we have assumed that the active regions are stationary in the star’s reference frame. The slight change of the observed pattern between transits are caused by geometric foreshortening and limb darkening of the active regions. In this well-aligned case, we would observe strong correlations between the residuals of successive transits. The right side of Figure \[sky\_model\] shows similar illustrations for a star that has an obliquity $\Psi$ of 45$^\circ$. In this case, the active regions rotate across the transit chord, rather than along the chord. When a naive observer transforms time into stellar longitude under the assumption $\Psi \approx 0^\circ$, the residuals show no correlation. This is because the anomalies that are seen in successive transits are produced by different active regions. In this case $\Phi$, as calculated by Equation \[eqn:longitude\], is not really the stellar longitude. To be more general we will refer to $\Phi$ as the [*transit longitude*]{} rather than the stellar longitude; it corresponds to the stellar longitude only when the star has a low obliquity. The TCC (defined below) is a statistic that quantifies the degree of correlation between the residuals of successive transits as a function of transit longitude. If the obliquity is low and the SNR is high enough, the TCC should be high when the correct stellar rotation period $P_{\rm rot}$ is used to calculated $\Phi$ in Equation \[eqn:longitude\]. The high TCC indicates there is a pattern of active regions beneath the transit chord that is stationary in stellar longitude, which can only happen for a low obliquity. A similar test for perfectly retrograde obliquities can be performed by switching the sign of the second term in Equation \[eqn:longitude\]. Light curve fitting ------------------- Figure \[flow\_chart\] summarizes the TCC method with a flow chart. We begin with the observed light curve. We isolate the individual transits, retaining only the segments of data spanning twice the transit duration and centered on each transit midpoint. The transit data are used for the TCC computation. The rest of the data are used only to estimate the stellar rotation period. For this purpose we look for the strongest peak in the Lomb-Scargle periodogram of the light curve, and designate this as $P_{\rm phot}$, the photometric rotation period, with an uncertainty equal to the full width at half maximum of the peak. We also estimate the stellar rotation period based on the auto-correlation function [@McQuillan2013; @McQuillan2014], and consider $P_{\rm phot}$ to be securely measured when the results from the periodogram and the auto-correlation function are in agreement. We then fit a simple transit model assuming no active regions, and compute the residuals between the data and the best-fitting model. For the model we use the [Batman]{} code [@Kreidberg2015], with the following parameters: the orbital period $P_{\rm orb}$; the planet-to-star radius ratio ($R_p/R_\star$); the scaled orbital distance ($a/R_\star$); the impact parameter ($b \equiv a\cos i/R_\star$); the quadratic limb-darkening coefficients ($u_1$ and $u_2$), the transit midpoint ($t_0$); and two coefficients of a quadratic function of time to account for the longer-timescale stellar flux variation near the time of the transit. We adopt the usual $\chi^2$ likelihood function and obtained the maximum-likelihood solution using the Levenberg-Marquardt algorithm as implemented in the [Python]{} package [lmfit]{} [@lmfit]. After each transit is fitted individually we “rectify” the light curves, dividing the data by the best-fitting quadratic function of time. Then we fit all the rectified light curves together with a reduced set of parameters: $P_{\rm orb}$, $R_p/R_\star$, $a/R_\star$, $b$, $t_0$, $u_1$ and $u_2$. Next we allow for the possibility of TTVs, and for transit depth variations caused by untransited active regions. When the untransited portions of the star are relatively faint, the loss of light due to the planet increases, and vice versa. We account for these effects by fitting the individual light curves again but with only two adjustable parameters: the time of transit ($t_0$), and an additive constant ($\Delta F$) to account for the overall changes in flux of the star. We hold the other parameters fixed at the values determined in the preceding joint fit of all the light curves. The model for the observed transit light curve is then $$F_{\text{calc}} = \frac{F_{\text{tra}} + \Delta F} {1 + \Delta F},$$ where $F_{\text{tra}}$ is the transit model calculated by [Batman]{} and the denominator sets $F_{\text{calc}} = 1$ outside of transits, the same normalization that was adopted for the data. To prevent overfitting we require $|\Delta F|$ to be smaller than the observed peak-to-peak variation of the relative flux across the entire light curve. After this final fit to each light curve, we record the residual flux $F_{\text{res}} \equiv F_{\text{obs}}- F_{\text{calc}}$, and transform the time stamps into transit longitude $\Phi$ using Equation \[eqn:longitude\]. Searching for Transit Chord Correlations ---------------------------------------- We are now in the position to seek statistical evidence for the recurring pattern in the residuals that one would expect for a low-obliquity system. First we group together the residuals from a certain number $N_{\rm tra}$ of consecutive transits. The reason for grouping is two-fold: (1) each transit only probes the visible side of the photosphere, and we need at least a few transits to obtain complete longitude coverage; (2) grouping transits together enhances the SNR. The case $N_{\rm tra}$ = 1 corresponds to no grouping. To illustrate the method, Figure \[Kepler17\_residuals\] shows the data for Kepler-17, a young and active G star with a 1.5-day hot Jupiter. The system was found to have a low obliquity through an earlier application of the spot-crossing method [@Desert2011]. Shown are the residuals for five neighboring groups of transits, each of which is composed of five consecutive transits. All five of these groups show a very similar pattern as a function of longitude, as we would expect for a low-obliquity star. To quantify the significance of the correlations, we perform the following steps. First we bin the residual flux uniformly in transit longitude. We compute normalized residuals, $$r_{i, j} \equiv \frac{F_{\text{res}}(i,j)}{\sigma_{\text{res}}(j)},$$ where $F_{\text{res}}(i,j)$ is the median of the residual fluxes in the $i$th group of transits and $j$th bin in transit longitude, and $\sigma_{\text{res}}(i,j)$ is the standard deviation of the residuals of all the data in the $i$th group that contribute to the $j$th bin. Finally, we define the transit chord correlation (TCC) as the average of the dot products of the residuals between neighboring groups: $${\rm TCC} \equiv \frac{1}{N_{\text{gr}}-1} \sum_{i}^{N_{\text{gr}}-1} \frac{1}{N_{\Phi}} \sum_{j}^{N_{\Phi}}{ r_{i,j} \times r_{i+1, j} }$$ where $N_{\text{gr}}$ is the total number of groups observed, and $N_{\Phi}$ is the number of longitude bins. We selected $N_{\Phi}$ such that each bin contains at least five data points; typically $N_{\Phi} \approx 100$. In some cases, gaps in the data resulted in an empty bin, in which case we set $F_{\text{res}}(i,j)=0$. Note that the dot products of the residuals are only computed between neighboring groups. This is because each neighboring pair is only separated by a relatively short timescale (a few orbital periods). On longer timescales, we expect the correlation between different groups to dwindle because there might have been enough time for the active regions to undergo substantial evolution or migration. We will see later that Kepler-17 and many other systems do show evidence for evolution of the active regions. We compute the TCC on a 2-d grid of possible choices for the stellar rotation period $P_{\rm rot}$ and the number $N_{\rm tra}$ of neighboring transits that are included in each group. The stellar rotation period that gives rise to the strongest correlation is denoted as $P_{\text{tcc}}$. For a star with a low obliquity, the strongest TCC should be observed when we calculate the transit longitude using the independently measured stellar rotation period, i.e., for $P_{\rm tcc} = P_{\rm phot}$, within the uncertainties in the observations and the degree of differential rotation. Figure \[Kepler17\_grid\] shows the TCC grid for Kepler-17. In this case, $P_{\text{tcc}} = 11.8\pm0.7$ days agrees very well with $P_{\text{phot}} = 11.9\pm 1.1$ days, which is good evidence that this star has a low obliquity. We also need to establish the threshold value of TCC to be considered statistically significant. We do so with a Monte Carlo procedure. In each of $10^3$ trials, we scramble the transit epochs, i.e., we randomize the order of the individual transits without affecting the individual light curves. In this way, we remove any correlations between neighboring transits due to spot-crossing events, while preserving the structure of the residuals (the “red noise”) within each transit light curve. For each trial we recalculate the TCC using the values of $P_{\rm tcc}$ and $N_{\rm tra}$ that produced the strongest correlation in the real dataset. We find the resulting distribution of TCC to be nearly Gaussian with a well defined standard deviation, as one would expect from the central limit theorem. The TCC distribution for Kepler-17 is shown in Figure \[Scrambled\]. The statistical significance of the TCC of the real data is quantified by the number of standard deviations ($N_\sigma$) away from the median of the Monte Carlo distribution. The final question is, what threshold should be set on $N_{\sigma}$ for a statistically robust detection of correlation? We found that statistical fluctuations alone sometimes produce a TCC with $\text{N}_{\sigma} =$ 2–3 for a random value of $P_{tcc}$ (not necessarily close to $P_{phot}$). This is because when searching for the strongest TCC, we step through a grid of different stellar rotation periods. At each grid point, the transit longitudes and hence the TCC is recomputed with the new rotation period. Typically, we test a rotation period grid with about 200 different periods. It is often the case that a random rotation period among the 200 periods tested produce a 2 or 3-$\sigma$ outlier just by chance. We therefore set a threshold of low-obliquity detection at $\text{N}_{\sigma}>4$. Even more importantly, we guard against spurious detections by requiring that $P_{\rm tcc}$ must agree with $P_{\text{phot}}$. The problem of spin-orbit commensurabilities -------------------------------------------- When the ratio of the stellar rotation period and the planetary orbital period is close to the ratio of small integers, i.e., $P_{\rm rot}/P_{\rm orb} = n/m$, then the planet and the star return to the same sky-projected configuration every $n$ orbital periods or $m$ rotation periods. Such a spin-orbit commensurability may be the result of a physical process that drove a system into a spin-orbit resonance, or may simply occur by chance. Upon returning to the same sky-projected configuration, the planet will occult the same set of active regions. The TCC will therefore be strong regardless of the stellar obliquity. Fortunately, these cases for which the TCC is blind to obliquity can be easily recognized because $P_{\rm phot}$ and $P_{\rm orb}$ are both known in advance. Moreover, in such cases, the TCC will usually be strongest at $P_{\rm tcc} = mP_{\rm rot} = nP_{\rm orb}$, rather than $P_{\rm rot}$. In practice, our code raises an alarm whenever $P_{\rm phot}/P_{\rm orb}$ falls within 5% of a ratio of small numbers and $P_{\text{tcc}}$ also falls within 5% of $nP_{\rm orb}$. In these cases, we carry out a visual inspection of the light curve, which can easily distinguish the low-obliquity from the high-obliquity cases: in the low-obliquity case spot/faculae crossing events recur in all neighboring transits; whereas in the high-obliquity case spot/faculae crossing events only recur after $nP_{\rm orb}$. In summary, we declare a statistically significant detection of low stellar obliquity if the system satisfies: (1) $P_{\text{tcc}}$ agrees with $P_{\text{phot}}$ to within the uncertainties; (2) $N_{\sigma}>4$; and (3) the orbital and rotation periods are not in the ratio of small integers. For such systems we can place an upper bound on the allowed obliquity using a simple geometric argument. Between consecutive transits, the active regions move through an angle of $2\pi/P_{\text{rot}} \times P_{\text{orb}}$. The angular displacement of the active region between consecutive transits is therefore $$\Delta\theta = ~{\rm mod} (\frac{2\pi P_{\text{orb}}}{P_{\text{rot}}}, 2\pi).$$ In a well-aligned system, the active regions move along the transit chord, along the $x$ direction. However a non-zero obliquity introduces a vertical motion, in the $y$ direction. In order to observe a recurring pattern, the active regions must remain within the transit chord for at least two transits. For this to happen the vertical motion of the active regions must be smaller than the width of the transit chord, $$\label{eqn:upperlimit} \Delta\theta \times \text{sin}(\Psi) \lesssim \frac{R_p}{R_\star}.$$ For giant planets around Sun-like stars this leads to a typical upper limit of 10$^\circ$, though it is only approximate because we have neglected the effects of the stellar inclination toward or away from the observer. Also, in cases where the active regions are larger than the planet radius, we should replace $R_p/R_\star$ by the size of the active regions relative to the star. To allow a visual inspection of all the data we introduce the “transit tapestry”, shown in the left panel of Figure \[Kepler17\_tapestry\] for Kepler-17. This is a heat map in which the color scale represents the residual flux, the horizontal dimension is transit longitude, and the vertical dimension is the transit group number (essentially, the time). If a star has a low obliquity, then long-lived and stationary active regions will appear as vertical features in the image. Evolution of the active regions will cause the features to vary slowly in the vertical direction, and migration in longitude will cause them to vary in the horizontal direction. These effects produce a tapestry-like pattern of coherent features, which displays the properties of the active regions such as their sizes, relative intensities, lifetimes and migration patterns. If the obliquity is large or the active regions are weaker then the image will lose its artistic appeal and look like random static, as in the right panel of Figure \[Kepler17\_tapestry\], which was generated after scrambling the transit epochs. Retrograde Orbits ----------------- Although our emphasis in this description has been on the detection of low-obliquity orbits, it is worth noting that the same method can just as easily be used to identify planets on nearly perfectly retrograde orbits ($\Psi = 180^{\circ}$). This is because the transit chord of a retrograde-orbiting planet is also parallel to the lines of latitude on the stellar photosphere. In such a geometry, active regions on the star can be occulted multiple times in neighboring transits and produce strong correlations in the residual flux. To search for retrograde systems, we simply flip the sign of the second term in Equation \[eqn:longitude\] and proceed again with all the other steps in the analysis. In fact it is not strictly necessary to perform two completely different analyses, one for prograde and the other for retrograde systems. Even if the positive sign is always retained in Equation \[eqn:longitude\], a retrograde system will distinguish itself by producing a strong TCC for a value of $P_{\rm tcc}$ that is not equal to the stellar rotation period, but rather at the period given by $$\label{eqn:retro} \frac{1}{P'_{\rm tcc}} =\frac{1}{P_{\rm orb}} - \frac{1}{P_{\rm rot}}.$$ This is because in between two transits, the active regions on a retrograde star with rotation period $P'_{\rm tcc}$ will rotate through the same angle as a prograde star with rotation period $P_{\rm rot}$. In general $P'_{\rm tcc}$ differs significantly from $P_{\rm rot}$. The only exception is the case $P_{\rm rot}$ = 2$P_{\rm orb}$, which is already handled in the test for spin-orbit commensurabilities. Sample Selection ================ We tried to assemble a collection of all the transit datasets for which there seemed to be a reasonable chance of success for the TCC method. The requirement for high SNR and a large number of consecutive transits restricts us to data from the space missions [ *CoRoT*]{}, [*Kepler*]{}, and [*K2*]{}. We included all 29 confirmed [*CoRoT*]{} systems in our sample. We downloaded the light curves from the [*CoRoT*]{} N2 public archive[^1]. We used the monochromatic flux labelled “whiteflux” in the FITS header. We only retained data points with a Quality Flag of 0, i.e., those not affected by various known problems. Analyzing all of the thousands of [*Kepler*]{} Objects of Interest (KOIs) would have required too much computation time. Instead, we selected a sample of confirmed planets, planetary candidates, and eclipsing binaries with high SNR and for which there was some indication of spot-crossing anomalies. The SNR was calculated as the ratio between the transit or eclipse depth and the tabulated 6-hour CDPP (combined differential photometric precision). We identified the 360 KOIs with a single-transit SNR $>$ 20. Evidence for anomalies is based on the ratio of the standard deviation of the photometric residuals during the transit $\sigma_{\rm in}$, and the standard deviation of the out-of-transit data $\sigma_{\rm out}$. Stellar activity should cause this ratio to be significantly larger than unity. We retained the high SNR systems that also have $\sigma_{\rm in}/\sigma_{\rm out} >1.7$. A list of such systems had been compiled by @Sanchis-Ojeda_thesis. We also included any [ *Kepler*]{} systems for which the stellar obliquity had been previously measured using spot-crossing anomalies or any other method. We downloaded the light curves from the MAST website using the [ Python]{} package [kplr]{} [^2]. Whenever possible we used the short-cadence light curves, with one-minute time sampling. We used the light curves based on SAP (simple aperture photometry), and we only kept data points with a Quality Flag of 0. Our sample also includes nine stars with confirmed planets that were observed in the short-cadence mode during the [*K2*]{} mission [@Howell2014]. These systems are HATS-9, HATS-11, Qatar-2, WASP-47, WASP-55, WASP-75, WASP-85, WASP-107 and WASP-118. To produce the light curves, we downloaded the pixel files from the Mikulski Archive for Space Telescopes website (MAST). To remove the spurious intensity fluctuations caused by the uncontrolled rolling motion of the [*Kepler*]{} spacecraft, we used the photometry pipeline described by @Dai2017qatar2, which decorrelates the flux variations against the measured coordinates of the center of light on the detector. The resultant sample consists mainly of FGK dwarf stars, except for Kepler-13A (an A star) and Kepler-45 (an M dwarf). Because of the SNR requirement, the planets in our sample are almost all hot Jupiters. Three-quarters of the planets are larger than $8~R_{\oplus}$ and have orbital periods shorter than 10 days. We believe the resulting sample is fairly exhaustive, in the sense that it includes all of the data currently available for which the TCC method has a significant chance of revealing the obliquity. However, the sample is strongly biased toward close-in giant planets around active stars. It is not “complete” in any other sense, i.e., it is not amenable to any simple astrophysical description. Caution is therefore needed when trying to interpret the fraction of systems that are found to be aligned, retrograde, or indeterminate. Results {#Results} ======= Tables \[tab:planet\] and \[tab:eb\] summarize the results for the transiting planets and eclipsing binaries respectively. For each system we report the orbital period ($P_{\text{orb}}$); the scaled semi-major axis ($a/R_{\star}$); the mass and radius of the planet and the host star ($M_p$, $R_p$, $M_\star$, and $R_\star$); the effective temperature of the host star ($T_{\text{eff}}$); the rotation period measured from the quasiperiodic modulation in the light curve ($P_{\text{phot}}$); the rotation period that gives the strongest TCC ($P_{\text{tcc}}$); the statistical significance of the correlation based on the scrambling test ($N_{\sigma}$); the ratio between the rotation and orbital periods ($P_{\text{tcc}}/P_{\text{orb}}$); and any previous measurement of the sky-projected obliquity that has appeared in the literature $\lambda_{\text{lit}}$. We have arranged the table such that systems that show statistically significant TCC (those with a low stellar obliquity) appear in front. For those systems we report an upper limit on the true obliquity using Equation \[eqn:upperlimit\]. The other systems are arranged in descending order of $N_{\sigma}$. To check on the validity of the TCC method, we examined the results for all the systems for which stellar obliquities had been measured with other methods. For the low-obliquity systems (those that satisfy Equation \[eqn:upperlimit\]), we may detect a strong TCC with $P_{\text{tcc}}=P_{\text{phot}}$. However, a null result could also be obtained, if the host star is not sufficiently active, the planet’s transit chord misses the active latitudes, or the active regions undergo rapid evolution. Therefore we began by focusing on the low-obliquity systems where previous studies had revealed strong stellar activity through the clear identification of individual spot-crossing events: Kepler-17 [@Desert2011], CoRoT-2 [@Nutzman2011] and Qatar-2 [@Mocnik2016; @Dai2017qatar2]. These are hot Jupiters around active G or K stars, with orbital periods less than two days. For all of these systems, individual spot-crossing anomalies could be visually identified in the light curves. The recurrence of the spot-crossing events led previous authors to conclude that all three systems are aligned to within $\approx$$10^{\circ}$. Additionally, for the case of Qatar-2b, @Esposito2017 found $\lambda = 0 \pm 10^{\circ}$ based on the Rossiter-McLaughlin effect. Unsurprisingly, these three systems show the strongest TCC among all the systems in our sample, and $P_{\rm tcc}$ agrees well with $P_{\rm phot}$. Thus the TCC method easily confirms the low stellar obliquities for these systems. @Holczer2015 reported five systems with prograde obliquities (Kepler-17b, Kepler-71b, KOI-883.01, KIC 7767559b and Kepler-762b), based on the observed correlation between the apparent TTV and the local time derivative of the out-of-transit flux. This method was described in the Section 1. The TCC method shows that the orbits of all five systems are not only prograde ($\Psi < 90^{\circ}$) but also aligned to within about $\Psi \lesssim 10^{\circ}$. We now turn to the eight systems for which other methods unveiled a high stellar obliquity (see Table \[tab:planet\] for the measured obliquities and references). For CoRoT-3b, CoRoT-19b, Kepler-13Ab, Kepler-420b, HAT-P-7b, and WASP-107b, we did not detect a statistically significant TCC. For those cases in which the stellar rotation period could be measured from the light curve, $P_{\text{phot}}$ did not agree with $P_{\text{tcc}}$. This is as expected, for high-obliquity systems. However, our method did return a statistically significant TCC for the HAT-P-11b and Kepler-63b systems, which are known to be misaligned [@Sanchis-Ojeda2011Hat; @Sanchis-Ojeda2013]. We found $N_{\sigma} = 6.0$ for HAT-P-11b and 5.8 for Kepler-63b. These “false positives” are examples of the problem of spin-orbit commensurability described in the previous section. @Beky2014Hat reported a 6:1 spin-orbit period ratio for HAT-P-11b, and @Sanchis-Ojeda2013 reported a nearly 4:7 ratio for Kepler-63b. The strongest TCC for Kepler-63 was detected at $P_{\text{tcc}} = 37.7$ days (4$\times P_{\text{orb}}$) rather than $P_{\rm phot}$ = 5.401$\pm 0.014$ days. After completing the search for well-aligned systems, we used the same sample to search for perfectly retrograde systems. We flipped the sign of the second term in Equation \[eqn:longitude\] before repeating the TCC analysis. We did not find any cases of a statistically significant TCC. Hence, although we could have detected perfectly retrograde orbits just as easily as we detected prograde orbits, we did not find any evidence for perfectly retrograde orbits in our sample. In the following sections we discuss some of the individual systems in greater detail. Kepler-17 --------- Kepler-17b is a hot Jupiter orbiting a G star every 1.5 days. @Desert2011 identified and modeled a sequence of spot-crossing events in neighboring transits. The recurrence of the spot-crossing events led @Desert2011 to conclude that the Kepler-17b has an obliquity &lt; 15$^{\circ}$. Our analysis revealed a strong TCC ($N_{\sigma}$ = 59.6) at $P_{\text{tcc}} = 11.8 \pm 0.7$ days. The [*Kepler*]{} light curve shows rotational modulation in the out-of-transit light curve which offers an independent check on the stellar rotation period: $P_{\text{phot}} = 11.9 \pm 1.1$ days. These observations together suggest a low stellar obliquity of Kepler-17b; and we place an upper bound $\Psi \lesssim$ 10$^{\circ}$ with Equation \[eqn:upperlimit\]. Figure \[Kepler17\_tapestry\] shows the transit tapestry for Kepler-17. It displays clear clusterings of positive and negative residuals in transit longitude. We interpret the observed clustering as photometric signatures of active regions along the transit chord. These patterns can be used to infer the properties of the host star’s magnetic activity. At any particular time there are 1–4 active regions present on the transit chord. Each of these regions in the tapestry spans 20–30$^{\circ}$ in transit longitude. However, the nonzero size of the the planet (about 16$^{\circ}$ in longitude) acts to widen the apparent size of the active regions. After accounting for the size of the planet, the active regions span 5–20$^{\circ}$ in true longitude on the star. The active regions lasted for a 100–200 days before either disappearing or leaving the latitudes probed by the planet. The intensity of the active regions changed gradually over time; they did not burst into existence with maximum intensity. One point of interest is whether the active regions remain stationary in longitude or undergo longitudinal migration. However, there is a degeneracy between any constant rate of longitudinal migration, and the rotation period used to calculate transit longitudes. Specifically, if $P_{\rm tcc}$ is smaller than the actual stellar rotation period, all the photometric features in the transit tapestry would appear to undergo prograde longitudinal migration in the transit tapestry. The TCC method, by design, maximizes the correlation of residual flux. As a result, it always selects the stellar rotation period such that most photometric features remain fixed in stellar longitude. Therefore, the transit tapestry can only reveal relative migration rates between different active regions but not the absolute rate of migration. In the case of Kepler-17, there is indeed some indication of relative migration. The pattern enclosed by the green ellipse in Figure \[Kepler17\_tapestry\] suggests that an active region split apart into two smaller regions which separated longitudinally over time. This is reminiscent of the emerging magnetic flux tubes on the Sun which produce bipolar magnetic regions [@Spruit]. A flux tube initially gives rise to one footpoint on the host star when it just reaches the stellar surface. As the flux tube emerges further, it gives rise to two footprints that spread apart in longitude. From now on, we will define the contrast of an active region as the relative ratio between the brightness of the active regions to the average photosphere, in the observing bandpass. Specifically for Kepler-17, the photometric features of the active regions had amplitudes of about 0.002–0.004 in relative flux. By comparing this to the transit depth of about 0.02, the active regions on average are seen to be roughly 10-20% fainter than the average photosphere in the broad optical [*Kepler*]{} bandpass (approximately 420–900 nm). We note that the finite size of the transiting planet introduces a convolution effect that not only broadens the photometric signature of the active regions, but also reduces the contrast. The contrast calculated here and subsequently is the product of the relative brightness of individual magnetic features that constitute the active regions and the area filling factor within the shadow of the planet. Because of this degeneracy, we refrain from converting the contrast into an effective temperature (which would have allowed for more direct comparisons with sunspots). Figure \[Kepler17\_tapestry\] shows that the number, size, and relative intensity of active regions decreased towards the latter part of the [*Kepler*]{} mission. This decline might have been part of a magnetic activity cycle, or it might represent the latitudinal migration of the active regions away from the transit chord. In principle, the degree of differential rotation could be studied by comparing the rotation period on the transit chord $P_{\text{tcc}}$ with the photometrically-derived rotation period $P_{\rm phot}$, which is based on active regions across the entire star. However, this is difficult in practice because there is little information on how the active regions are distributed latitudinally outside the transit chord. Given the impact parameter of the transit $b=0.27$ [@Desert2011] and the size of the planet $R_p/R_\star \approx$0.13, the stellar latitude probed by the planet is $16\pm 8^{\circ}$ north or south of the stellar equator. At least, our analysis can show that this latitudinal range is magnetically active. CoRoT-2 ------- CoRoT-2b is a hot Jupiter orbiting a G star every 1.7 days [@Alonso2008]. The star rotates rapidly, with a period of about 4.5 days, and has strong magnetic activity which manifested as clear spot-crossing events in the [*CoRoT*]{} light curve. By modeling the recurrence of spot-crossing events, @Nutzman2011 found the sky-projected obliquity to be $\lambda = 4.7 \pm 12.3^{\circ}$. Our TCC method tells a consistent story. We find a strong TCC $N_{\sigma}$ = 22.6 at $P_{\text{tcc}} = 4.5 \pm 0.3$ days consistent with $P_{\text{phot}}$ of $4.530 \pm 0.068$ days. We conclude that CoRoT-2 has a low stellar obliquity Figure \[Corot2\_tapestry\] shows the transit tapestry. The clustering of residual flux indicates that there were two active regions present on the stellar latitude probed by the planet. They had similar sizes and intensities and were separated by about $180^{\circ}$ in longitude. These two regions span about 30–40$^{\circ}$ in the transit tapestry. After accounting for the blurring effect of the planet, which extends about 20$^{\circ}$ in longitude, we conclude the active regions are about 10-20$^{\circ}$ wide. Both active regions persisted throughout the $\approx$150 days of [*CoRoT*]{} observations, remaining nearly stationary in longitude. With Figure \[Corot2\_tapestry\], we can estimate that active regions produce flux variation of about 0.002–0.005 in relative flux. Taking the ratio with the transit depth (about 3%), we find the active regions to be roughly 7-17% dimmer than the rest of the photosphere in the [*CoRoT*]{} bandpass. Again, using the impact parameter of the transit $b\approx 0.22$ [@Gillon2010b] and the size of the planet $R_p/R_\star \approx 0.13$, the stellar latitude probed by the planet is likely $13\pm 10^{\circ}$ in either the northern or southern hemisphere. @Lanza2009 and @Huber2010corot also studied the magnetic activity of CoRoT-2. @Lanza2009 used a maximum entropy method to determine the longitudinal distribution of active regions using the rotational modulation in the out-of-transit light curve, while @Huber2010corot modeled the transit light curve by dividing the eclipsed and noneclipsed parts of photosphere into a number of longitudinal intervals with different intensities. Both of these works confirmed the presence of active regions on CoRoT-2. Qatar-2 ------- Qatar-2b is a hot Jupiter discovered by the Qatar Exoplanet Survey [@Bryan2012]. The planet orbits a K dwarf every 1.3 days. Recent [*K2*]{} observations in the short-cadence mode unveiled dozens of spot-crossing anomalies. By modeling the spot-crossing anomalies, both @Mocnik2016 and @Dai2017qatar2 showed that the system has a low stellar obliquity. @Esposito2017 independently confirmed the low stellar obliquity $\lambda = 0 \pm 10^{\circ}$ by observing the Rossiter-McLaughlin effect. As expected, we detect a strong TCC $N_{\sigma}$ = 18.0 in the [*K2*]{} light curve at $P_{\text{tcc}} = 18.2 \pm 0.4$ days which agrees with $P_{\text{phot}}$ = $18.5 \pm 1.5$ days. We place an upper bound $\Psi \lesssim 11^{\circ}$. Figure \[Qatar2\_tapestry\] shows the tapestry plot for Qatar-2. Although [*K2*]{} observations only spanned 80 days, the transit tapestry still reveals two active regions along the transit chord. One of them was located at 300$^{\circ}$ in longitude and was about 15–25$^{\circ}$ wide (after accounting for the broadening due to the nonzero size of the planet). Its intensity remained relatively constant throughout the 80 days of [*K2*]{} observations. The other active region was located near 20$^{\circ}$ in longitude. Its size and intensity underwent a significant increase during the [*K2*]{} observation. This might have been caused by the emergence of a magnetic flux tube, or latitudinal migration of an active region into the transit chord. These active regions maintained a constant separation in longitude. The active regions on average are roughly 3-7% dimmer than the photosphere. Given the impact parameter of the transit $b \approx 0.03$ [@Dai2017qatar2] and $R_p/R_\star= 0.15$, the transits probe the region close to the stellar equator ($2\pm9^{\circ}$). Kepler-71 --------- Kepler-71b is a 3.9-day hot Jupiter around a G star [@Howell2010]. With a V magnitude of 15.4, Kepler-71 is too faint for precise radial-velocity follow-up. @Holczer2015 detected a strong correlation between the detected TTV and the time derivative of the stellar flux, which suggested a prograde orbit and the presence of starspots. Our TCC analysis shows that the orbit of Kepler-71b is not only prograde but also well-aligned. We detect a strong TCC ($N_{\sigma}$ = 10.7; $P_{\text{tcc}} = 19.7 \pm 0.8$ days; $P_{\text{phot}}$ = $19.87 \pm 0.18$ days). The upper bound on the obliquity is $\Psi$ $\lesssim$ 6$^{\circ}$. We made a transit tapestry for Kepler-71 (Figure \[Kepler71\_tapestry\]). Two active regions are most discernible in the later part of the [*Kepler*]{} mission (BJD-2454833 &gt; 1200). These two active regions were separated by about 180$^{\circ}$. They each spanned about 5-15$^{\circ}$ in longitude after accounting for the convolution effect of the planet (16$^{\circ}$ in longitude). They lasted for about 150 days; and on average were roughly 10-20% dimmer than the photosphere. We estimate the latitude probed by the planet using the impact parameter of the transit [0.04, @Howell2010] and the size of the planet ($R_p/R_\star \approx$ 0.14). The result shows that the equatorial region of the photosphere ($2 \pm 8^{\circ}$) is magnetically active. KOI-883.01 ---------- KOI-883.01 is a planetary candidate discovered by the [*Kepler*]{} mission. The transit light curve is consistent with a hot Jupiter orbiting a K star every 2.7 days. As was the case with Kepler-71, @Holczer2015 detected a strong correlation between TTV and local flux variation and inferred a prograde orbit for the system. We show that the system is not only prograde but also well-aligned. The $N_{\sigma}$ is as high as 9.2 while $P_{\text{tcc}} = 9.1 \pm 0.2$ days and $P_{\text{phot}} = 9.11 \pm 0.11$ days agree well. We place an upper bound $\Psi$ $\lesssim$ 4$^{\circ}$. The transit tapestry is shown in Figure \[KOI-883.01\_tapestry\]. The photometric signatures of active regions are more diffuse than for the systems described earlier. In the first $\approx$100 days of [ *Kepler*]{} observations, an active region was located near 240$^{\circ}$ in longitude and was 20–30$^{\circ}$ wide (after accounting for the 18$^\circ$ angular size of the planet). After a gap in the data, another active region emerged at 120$^{\circ}$ in longitude with a similar size and intensity (10-20% dimmer than the average photosphere). Both active regions lasted for at least 100–200 days, and perhaps even longer if the active region near 240$^{\circ}$ persisted through the data gap. The impact parameter of the transit is about 0.01 and $R_p/R_\star$ is about 0.18 (ExoFoP [^3]). The transit chord overlaps with the equatorial region ($0 \pm 10^{\circ}$) which appears to be magnetically active. Kepler-45 --------- Kepler-45b is a 2.5-day hot Jupiter orbiting a M dwarf. The system was confirmed by a combination of radial velocity monitoring, adaptive optics imaging, and near-infrared spectroscopy [@Johnson2012]. It is one of the three hot Jupiters around M dwarfs that have been reported to date. The other two are HATS-6b [@Hartman2015] and NGTS-1b [@Bayliss2017]. No constraint on the stellar obliquity of Kepler-45b has been published yet. Application of the TCC method gave $N_{\sigma}$ = 6.3 at $P_{\text{tcc}} = 16.7 \pm 0.8$ days. The $P_{\text{tcc}}$ coincides with the measured $P_{\text{phot}}$ = $15.8 \pm 0.2$ days, to within the uncertainties. We conclude that Kepler-45b is very likely on a well-aligned orbit, with an upper bound on $\Psi$ $\lesssim$ 11$^{\circ}$. To our knowledge, this is only the second case of a stellar obliquity measured for an M dwarf. The first such report was for GJ 436b, a 2.6-day Neptune-mass planet [@Bourrier2017 $\lambda = 72^{+33}_{-24}~^{\circ}$]. Enlarging the number of such measurements would be interesting. Being fully convective, some M-dwarfs are structurally very different from solar-type stars for which most of the obliquity measurements have been obtained. The excitation and damping of stellar obliquities may operate in different ways for M dwarfs than for solar-type stars. Given that other, more traditional obliquity measurements tend to fail for M dwarfs, the TCC method may be particularly promising for constraining the stellar obliquities of M-dwarf planet hosts. We will return to this point in Section \[discussion\]. Kepler-45b has an impact parameter of about 0.6 [@Johnson2012] and a $R_p/R_\star$ of 0.18. Together this suggests that the transit chord probes the stellar latitude of $37 \pm 16^{\circ}$. The transit tapestry, shown in Figure \[Kepler-45\_tapestry\], does not show any visually compelling patterns. This is because we have organized this section in order of decreasing correlation strength. At this point, the correlations may not be visually obvious, even though our Monte Carlo procedure shows that they are statistically significant when summed over the entire dataset. Alternatively, it may be the case that active regions on M-dwarfs preferentially take the form of broad chromospheric plages and networks rather than dark photospheric spots on sun-like stars. The resultant photometric features in the transit light curves would be more spread out, more rapidly evolving and hence harder to discern visually, compared to the sun-like stars described earlier. The increased chromospheric activity of M dwarfs was observed in Ca II H and K lines [@Isaacson2010]. Moreover, the difference in magnetic behavior between M dwarfs and solar-type stars is also theoretically motivated: fully convective M dwarfs lack the tachocline, the strong shearing zone between the radiative and convective layer of the sun, which is thought to be important for the operation of sun-like dynamos and the formation of sunspots [@Charbonneau2014]. TrES-2 ------ TrES-2 is a G star hosting a 2.5-day hot Jupiter [@O'Donovan2006]. Rossiter-McLaughlin observations have shown that TrES-2b has a low stellar obliquity $\lambda = -9 \pm 12^{\circ}$ [@Winn2008], although the confidence in that measurement was lower than usual because of the star’s relatively low rotation velocity. We find a strong TCC $N_{\sigma}$ = 5.8 in the residual flux at $P_{\text{tcc}} = 29.9 \pm 1.4$ days. This period is in agreement with the independently measured photometric period, $P_{\text{phot}} = 28.35 \pm 0.34$ days. We conclude that TrES-2 likely has a low stellar obliquity with an upper bound $\Psi$ $\lesssim$ 10$^{\circ}$. We show the transit tapestry in Figure \[TrES-2\_tapestry\]. We can see correlations between neighboring groups, but not any distinct, long-lasting features of the type that appeared in some of the systems described earlier. According to @Montet2017, the photometric activity of solar-type stars with rotation periods greater than 25 days are more likely to be dominated by bright patches (faculae) than by dark spots. These bright patches and the active network, at least those on the Sun, are more extended and evolve more quickly than spots [@Foukal1998; @Pontieu; @Shapiro2016]. Given the slow stellar rotation period of TrES-2, we may be seeing the more extended and less persistent photometric features of the faculae or active networks. In particular, the blue patch spanning about $180^{\circ}$ near BJD-2454833 = 1000 may be associated with a bright, more extended active region. Kepler-762 ---------- Kepler-762b is a 3.8-day hot Jupiter around a G star. The planet went from being a “candidate” to being “validated” through the statistical analysis of @Morton2016. @Holczer2015 reported a prograde orbit for this system given the observed strong correlation between TTV and local flux variation. Our TCC analysis shows a good agreement between $P_{\text{tcc}} = 4.0 \pm 0.1$ days and $P_{\text{phot}}$ = $4.045 \pm 0.025$ days. The statistical strength of the correlation is $N_{\sigma}$ = 5.2. We thus argue that the orbit of Kepler-762 is not only prograde but also well-aligned ($\Psi$ $\lesssim$16$^{\circ}$). The transit tapestry, Figure \[Kepler-762\_tapestry\], is not impressive to the eye, even though the TCC is strong enough for a confident conclusion. Given the impact parameter of the transit of about 0.05 (ExoFop) and $R_p/R_\star$ of 0.10, the transit chord probes the stellar latitude of $3 \pm 7^{\circ}$. Kepler-423 ---------- Kepler-423b is a 2.7-day hot Jupiter orbiting a G star [@Gandolfi2015]. The maximum TCC has $N_{\sigma} = 4.7$ at $P_{\text{tcc}} = 23.0 \pm 0.9$ days. The stellar rotation period was also independently measured to be $P_{\text{phot}}$ = $22.047 \pm 0.121$ days based on the out-of-transit light curve. Therefore Kepler-762 likely has a low stellar obliquity ($\Psi$ $\lesssim$ 10$^{\circ}$). In the transit tapestry (Figure \[Kepler-423\_tapestry\]), there are hints of group-group correlations, however no large-scale, long-lasting active regions can be discerned by eye. As we said with regard to TrES-2, Kepler-423 may be also faculae-dominated given its slow stellar rotation period. WASP-85 ------- WASP-85Ab is a 2.7-day hot Jupiter around a G star in a visual binary system [@Brown2015]. The companion is a K star at an angular separation of about 1.5 arcseconds ($\sim$190 AU). @Mocnik2016wasp85 analyzed the spot-crossing anomalies identified in the [*K2*]{} short-cadence data for WASP-85A. They concluded that the recurrence of the spot-crossing events not only suggests a low stellar obliquity ($\lambda$ $\lesssim$ 14$^{\circ}$), but also constrains the stellar rotation period to be $15.1 \pm 0.6$ days. On the other hand, they measured the stellar rotation period to be $13.6 \pm 1.6$ days based on the rotational modulation seen in the out-of-transit light curve. We detect a TCC of $N_{\sigma} = 4.5$ at $P_{\text{tcc}} = 15.2 \pm 0.3$ days. This is consistent with the value of $15.1 \pm 0.6$ days reported by @Mocnik2016wasp85. We measure a $P_{\text{phot}}$ = $13.28 \pm 0.41$ days which was also consistent with the $13.6 \pm 1.6$ days reported by @Mocnik2016wasp85. Since the $N_{\sigma}$ of 4.5 is not as strong as the previous cases and the $P_{\text{tcc}}$ differ substantially from $P_{\text{phot}}$, we examined WASP-85 more closely. The left panel of Figure \[201862715\_resdiual\] shows the residual flux of [*K2*]{} short-cadence observation of WASP-85. After transforming time into transit longitude using $P_{\text{tcc}} = 15.2$ days, the spot-crossing anomalies clearly recur at fixed transit longitude from one transit to the next, except in cases where spot-crossing events happened near the ingress or egress. Geometrical foreshortening and limb darkening effect are most severe during the ingress/egress; and both effects tend to weaken the signal of spot-crossing anomalies. Therefore the spot-crossing anomalies are expected to be suppressed here. The recurrence of spot-crossing events are compelling enough that we agree with @Mocnik2016wasp85 that WASP-85Ab likely has a low stellar obliquity. The fact that $P_{\text{tcc}} = 15.2 \pm 0.3$ days disagrees with $P_{\text{phot}}$ = $13.28 \pm 0.41$ days may be a sign differential rotation. Alternatively, this may be the result of the short lifetime of starspots. The rotational modulation in the [*K2*]{} light curve underwent significant evolution during the $\approx$80 days of [ *K2*]{} observation suggesting a short spot lifetime. Furthermore, Figure \[201862715\_resdiual\] shows that many of crossed spots only persisted for days before disappearing. The spot-crossing anomalies are about 0.0004-0.0008 in relative flux while the transit depth is about 2%. If the active regions uniformly fills the shadow of the planet, their relative intensity is hence 2-4% dimmer than the averaged photosphere. The spot-crossing events last about $\sim 15$ min or $\sim 20^{\circ}$ in longitude. After accounting for the blurring effect of the planet ($R_p/R_\star$ of 0.13), the spots are smaller than $5^{\circ}$. For WASP-85, we are likely seeing smaller, shorter-lived, individual starspots rather than the more stable and extended active regions on Kepler-17. Curiously, on the Sun, the lifetime of a sunspot is proportional to its size [@Gnevyshev; @Waldmeier]. With this “GW rule” in mind, it is not surprising that the spots on WASP-85 have shorter lifetimes compared to the extended active regions on Kepler-17. We estimate the latitude probed by the transit chord as $3 \pm 7^{\circ}$ using the the impact parameter of the transit of about 0.05 [@Mocnik2016wasp85] and $R_p/R_\star$ of 0.13. HAT-P-11 -------- HAT-P-11b is a 4.9-day super-Neptune around a K dwarf. With a relatively large scaled orbital distance of $a/R_\star$ = 15.6, it is more akin to the “warm Jupiters” than hot Jupiters. Rossiter-McLaughlin observations revealed $\lambda = 103^{+23}_{-10}$ degrees. @Sanchis-Ojeda2011Hat confirmed that the orbit is nearly perpendicular to the stellar equator by analyzing the spot-crossing anomalies seen in the [*Kepler*]{} light curve. They showed that spot-crossing anomalies cluster near two specific phases of the transit. They attributed this phenomenon to the presence of two active latitudes on the photosphere. As the planet transits the host star on a perpendicular orbit, the two symmetric active latitudes on both hemisphere will be occulted sequentially. The photometric signatures of the active regions are two persistent brightening features in the residual flux that remain fixed relative to the mid-transit time. Recent work by @Morris mapped out the distribution of starspots on HAT-P-11 by explicitly modeling individual spot-crossing events. They drew attention to many similarities between sunspots and the starspots on HAT-P-11 such as size and latitudinal distribution. @Morris_activity showed that HAT-P-11 is chromospherically more active than other planet hosts with similar properties. @Beky2014Hat reported that the ratio of spin period to orbital period is very nearly 6 to 1. Because of this commensurability, every six transits the planet and star return to the same configuration. As a result, the transiting planet can revisit the same active regions despite the high stellar obliquity. This results in a repeating pattern in the residual flux. Consequently, our TCC method detects a strong correlation ($N_{\sigma}$ = 6.0) even though the system is known to have a high obliquity. Visual inspection of the [*Kepler*]{} light curve shows that spot-crossing events only recur at fixed transit longitude after 6 consecutive transits. The close spin-orbit commensurability distinguishes HAT-P-11 from the low-obliquity systems discussed earlier. Kepler-63 --------- Kepler-63b is a sub-Saturn orbiting a G dwarf every 9.4 days. @Sanchis-Ojeda2013 showed that the system has a high stellar obliquity ($\Psi$ = 104$^{+22}_{-14}~^{\circ}$) using both Rossiter-McLaughlin observations and spot-crossing events in the [ *Kepler*]{} light curve. Similar to HAT-P-11, Kepler-63 displays an apparent spin-orbit resonance (4:7 in this case). Similar to HAT-P-11, our TCC method finds a relatively strong signal ($N_{\sigma}$ = 5.8) even though the system has a high stellar obliquity. The strongest TCC is detected at $P_{\text{tcc}}$ = 37.7 $\pm$ 0.8 days (c.f. $P_{\text{phot}}$ = 5.041 $\pm$ 0.014 days) which coincide with the least common multiple between the stellar rotation period and the orbital period ($4 P_{\text{orb}}$ or $7 P_{\text{phot}}$) i.e. the period at which the system return to the same configuration. This is indicative of a spin-orbit commensurability rather than a low-obliquity orbit as we explained earlier. Kepler-25 --------- The two transiting planets b and c of Kepler-25, an F star, were confirmed by @Steffen2012 via transit timing variations. @Albrecht2013 performed Rossiter-McLaughlin measurement during the transits of planet c which is a sub-Saturn on a 12.7-day orbit. The Keck/HIRES data they obtained showed that Kepler-25c has a low stellar obliquity of $\lambda = -0.5 \pm 5.7^{\circ}$. We applied the TCC method to the short-cadence [ *Kepler*]{} light curve of Kepler-25. No statistically significant correlation was detected. The strongest TCC has $N_{\sigma}= 3.2$ at $P_{\text{tcc}} = 31 \pm 1$ days. We could not measure the stellar rotation period independently in the out-of-transit light curve. Here and elsewhere in similar cases, we use the measured rotational broadening $v \rm{sin} i_\star$ and the stellar radius to estimate the stellar rotation period. For Kepler-25, @Albrecht2013 reported $v \rm{sin} i_\star = 9.5\pm0.5$ km s$^{-1}$ and a stellar radius of $1.36\pm 0.13 R_\odot$. Assuming an edge-on geometry ($i_\star$ = 90$^{\circ}$), the stellar rotation period should be close to 7 days. The non-detection of correlation in the residual flux may be attributed to several possibilities. The SNR of any photometric signature of active regions is expected to be smaller since the transit depth is only about 0.1% (compared to the 1-4% for the strong TCC cases described earlier). The lack of rotational modulation in the out-of-transit light curve suggests that the host star ($T_{\text{eff}} \approx 6200\,K$) may not be magnetically active. Finally, the impact parameter of the planet is high ($b = 0.881\pm 0.004$); the transit chord might have missed any active latitudes that are closer to the stellar equator. CoRoT-11 -------- CoRoT-11b is a 3.0-day hot Jupiter orbiting an F star, first reported by [@Gandolfi2010]. Later, @Gandolfi2012 showed that the system has a low stellar obliquity ($\lambda = 0.1 \pm 2.6^{\circ}$) based on the Rossiter-McLaughlin effect. Our TCC analysis could not detect a statistically significant correlation ($N_{\sigma}= 3.0$; $P_{\text{tcc}} = 30 \pm 4$ days; $P_{\text{phot}}$ unconstrained). Assuming an edge-on geometry ($i_\star = 90^{\circ}$), the stellar rotation period should be close to 2 days [$v\rm{sin} i_\star = 40\pm 5$ km s$^{-1}$ and $R_\star = 1.37\pm 0.03 R_\odot$, from @Gandolfi2012]. Similar to Kepler-25, the CoRoT-11 star may be too massive and hot ($T_{\text{eff}} \approx 6400\,K$) to be magnetically active. In addition the impact parameter is high ($b = 0.8108\pm 0.0077$) and the transit chord might miss the active latitudes. CoRoT-19 -------- First reported by @Guenther2012, CoRoT-19b is a hot Jupiter orbiting a F star every 3.9 days. @Guenther2012 also measured the Rossiter-McLaughlin effect and found the sky-projected stellar obliquity to be $\lambda = -52^{+27}_{-22}~^{\circ}$. Our TCC analysis did not detect a statistically significant correlation ($N_{\sigma} = 3.0$; $P_{\text{tcc}} = 10 \pm 1$ days; $P_{\text{phot}}$ unconstrained). The $v\rm{sin} i_\star$ argument gives a stellar rotation period close to 14 days [$v\rm{sin} i_\star = 6\pm 1$ km s$^{-1}$ and $R_\star = 1.65\pm 0.04 R_\odot$, from @Guenther2012]. The non-detection of correlation is consistent with the high obliquity reported by @Guenther2012. However, it might also be because the host star is too massive and hot ($T_{\text{eff}} \approx 6100\,K$) to be magnetically active. WASP-47 ------- WASP-47b is 4.2-day hot Jupiter around a G star [@Hellier2012]. Observations by [*K2*]{} revealed two additional transiting planets in the system, thus making WASP-47b the first hot Jupiter known to have close-in planetary companions [@Becker2015]. @Sanchis-Ojeda2015 measured the Rossiter-McLaughlin effect induced by the hot Jupiter WASP-47b, finding $\lambda = 0 \pm 24 ^{\circ}$. We do not find a statistically significant correlation ($N_{\sigma} = 3.0$; $P_{\text{tcc}} = 6.3 \pm 0.4$ days; $P_{\text{phot}}$ unconstrained). The stellar rotation period may be close to 37 days based on the combination of $v\rm{sin} i_\star = 1.80^{+0.24}_{-0.16}$ km s$^{-1}$ and $R_\star = 1.16 \pm 0.26~R_\odot$ [@Hellier2012; @Sanchis-Ojeda2015]. WASP-47 may be magnetically quiet, which is consistent with the lack of rotational modulation in the [*K2*]{} light curve. Kepler-13 --------- Kepler-13Ab is 1.8-day hot Jupiter around an A star. @Barnes2011 reported a misaligned orbit based on the asymmetric transit profile induced by gravity darkening. Later Doppler tomography [@Johnson2014] and further analysis of gravity darkening [@Masuda2015] showed that the system has a high obliquity ($\lambda = 58.6 \pm 2.0 ^{\circ}$). Before applying our TCC method, we removed the best-fitting gravity darkening model [@Masuda2015], rather than the usual transit model, in order to obtain the residual flux time series. No statistically significant correlation was detected. The strongest TCC has $N_{\sigma} = 2.7$ at $P_{\text{tcc}} = 3.12 \pm 0.10$ days. However, @Masuda2015 reported a stellar rotation period close to one day. The lack of correlation is expected, given the high obliquity and the spectral type of the host star. @Szabo2014 reported the detection of a 3:5 spin-orbit resonance based on a pattern seen in the residual flux. We did not detect a correlation at the stellar rotation period implied by the 3:5 spin-orbit resonance using the best-fitting gravity-darkened model from [@Masuda2015]. CoRoT-3 ------- CoRoT-3b is a 4.3-day brown dwarf ($21.66 \pm 1.00~M_{\text{Jup}}$) transiting an F star [@Deleuil2008]. @Triaud2009 performed Rossiter-McLaughlin observations which gave a constraint on the obliquity ($\lambda = 37.6^{+10.0}_{-22.3}~^{\circ}$). We did not find a strong TCC; the maximum correlation has $N_{\sigma} = 2.7$ and $P_{\text{tcc}} = 24.3 \pm 0.5$ days. The photometric rotation period is unconstrained. The $v\rm{sin} i_\star$ argument gives a stellar rotation period close to 3.0 days [$v\rm{sin} i_\star = 17\pm1$ km s$^{-1}$ and $R_\star = 1.56 \pm 0.09 R_\odot$, from @Deleuil2008]. The non-detection of correlation is consistent with a mildly oblique orbit. Alternatively, it may be due to the magnetically inactive host star as indicated by the lack of rotational modulation in the [*CoRoT*]{} light curve. CoRoT-18 -------- CoRoT-18b is a 1.9-day hot Jupiter around a G star [@Hebrard2010]. Previous Rossiter-McLaughlin observations revealed a low stellar obliquity ($\lambda = -10 \pm 20 ^{\circ}$). We find the strongest TCC of $N_{\sigma} = 2.6$ at $P_{\text{tcc}} = 15.7 \pm 0.7$ days. On the other hand, the rotational modulation in the out-of-transit light curve suggests a rotation period of $P_{\text{phot}} = 5.4 \pm 0.4$ days. We cannot confirm the low stellar obliquity using our correlation method due to the lack of significant correlation and the disagreement between $P_{\text{tcc}}$ and $P_{\text{phot}}$. This may partly be attributed to the low SNR light curve of this 15th mag star. Kepler-448 ---------- Kepler-448b is a 17.9-day warm Jupiter around a rapidly rotating F star [$v\rm{sin} i_\star = 60.0\pm0.9$ km s$^{-1}$ @Bourrier2015]. The [*Kepler*]{} light curve shows a 7.5-hour periodic modulation, presumably the stellar rotation period. The sky-projected obliquity has been determined through Doppler tomography to be $\lambda = 12.6^{+3.0}_{-2.9}~^{\circ}$. We do not detect any statistically significant correlation ($N_{\sigma} = 2.7$; $P_{\text{tcc}} = 21.0 \pm 0.6$ days; $P_{\text{tcc}}$ unconstrained). If the stellar rotation period is indeed 7.5 hours, the fast rotation has a similar timescale as the transit duration. Any photometric anomalies caused by active regions would likely be smeared out by the fast rotation. Kepler-420 ---------- Kepler-420b is a 86.6-day warm Jupiter around a G dwarf. @Santerne2014 used radial velocity data to show that the orbit has a high eccentricity of $0.772 \pm 0.045$ and possibly a high stellar obliquity $\lambda = 75^{+32}_{-46}~^{\circ}$. There is no strong TCC ($N_{\sigma} = 2.5$; $P_{\text{tcc}} = 11.7 \pm 0.7$ days; $P_{\text{phot}}$ unconstrained). If the obliquity is low then the rotation period is close to 13.5 days, based on the reported values of $v\rm{sin} i_\star = 4.6 \pm 0.2$ km s$^{-1}$ and $R_\star = 1.13 \pm 0.14 R_\odot$ [@Santerne2014]. The non-detection of correlations in the residual flux is consistent with an oblique orbit, although the host star may simply lack strong surface magnetic activity. WASP-107 -------- WASP-107b is a warm Jupiter ($a/R_\star = 18.2$) around a K dwarf star [@Anderson2017]. @Dai2017 analyzed the short-cadence [ *K2*]{} light curve of WASP-107. They inferred that WASP-107 has a high stellar obliquity, with $\Psi = 40-140^{\circ}$, because the observed spot-crossing anomalies did not recur in neighboring transits. The lack of recurrence also implied that the stellar rotation period and the planet’s orbital period cannot be in an exact spin-orbit resonance as previously suggested by @Anderson2017. If the period ratio were exactly commensurate, spot-crossing anomalies would recur regardless of the stellar obliquity. The high obliquity of this system has now been confirmed by Rossiter-McLaughlin measurement which indicates a nearly perpendicular orbit (A. Triaud, private communication). We applied the TCC method to the [*K2*]{} light curve of WASP-107. The results show a weak signal, with $N_{\sigma} = 2.5$ at $P_{\text{tcc}} = 10.0 \pm 0.5 $ days. On the other hand, the rotational modulation in the out-of-transit light curve suggests a rotation period of $P_{\text{phot}} = 17.1 \pm 1.0$ days. The lack of significant correlation and the disagreement between $P_{\text{tcc}}$ and $P_{\text{phot}}$ are consistent with the oblique orbit. Kepler-8 -------- Kepler-8b is a 3.5-day hot Jupiter around an F star @Jenkins2010. @Albrecht2012 found $\lambda = 5 \pm 7^{\circ}$ based on observations of the Rossiter-McLaughlin effect. The strongest TCC analysis has $N_{\sigma} =2.2$ at $P_{\text{tcc}} = 27.1 \pm 0.3$ days. The rotational modulation in the out-of-transit light curve suggests a rotation period of $P_{\text{phot}} = 28.64 \pm 0.32$ days. The weak correlation in the residual flux may be attributed to the magnetically inactive host star ($T_{\text{eff}}$ = 6200K). Moreover, the impact parameter of the planet is high $b = 0.724\pm 0.020$. HAT-P-7 ------- HAT-P-7b is a hot Jupiter with an orbital period of 2.2 days and an F-type host star [@Pal2008]. @Albrecht2012 observed the Rossiter-McLaughlin effect and found a high obliquity, with $\lambda = 155 \pm 37~^{\circ}$. @Masuda2015 measured the true obliquity, $\Psi = 101 \pm 2~^{\circ}$, based on the observable manifestations of gravity darkening effects in the [*Kepler*]{} transit light curves. As we did with Kepler-13, we removed the best-fitting gravity-darkening model to isolate any anomalies in the residual flux time series. No statistically significant TCC was detected. The strongest signal has $N_{\sigma}= 2.0$ and $P_{\text{tcc}} = 0.8 \pm 0.2$ days, which conflicts with the rotation period of 1.5–2.1 days estimated by [@Masuda2015]. WASP-75 ------- WASP-75b is another hot Jupiter around an F star. The orbital period is 2.5 days. The system is distinguished by an unusually large transit impact parameter, $b =0.887\pm0.008$. The [*K2*]{} light curve reveals a stellar rotation period of $P_{\rm phot} = 13.7\pm1.1$ days. @Gomez reported the projected rotation velocity ($v~sin i_\star = 4.3\pm 0.8$ km s$^{-1}$) and stellar radius ($1.26 \pm 0.04~R_\odot$), leading to a weak constraint on the stellar inclination: $\sin i_\star > 0.36$ with 95% confidence. We do not see a strong correlation in our TCC analysis. The strongest TCC has $N_{\sigma} = 1.9$ at $P_{\text{tcc}} = 22 \pm 3$ days. KIC 6307537 ----------- KIC 6307537 is an eclipsing binary system discovered with [*Kepler*]{} data. The star listed in the Kepler Eclipsing Binary Catalogue is a K dwarf. It eclipses the companion star every 29.7 days, causing the system brightness to fade by 7% for about 23 hours. When the companion occults the catalogued star, the system fades by 17%. This indicates that the companion is likely an evolved star, with a larger radius and lower effective temperature than the K dwarf. @Prsa classified this system as an eclipsing binary of the Algol type (detached). The [*Kepler*]{} light curve also shows a clear rotational modulation of $78\pm 3$ days which we attribute to the slower rotation of the evolved star. From the separation between the primary and secondary eclipse, @vanEylenEB obtained a constraint of the eccentricity: $e\cos\omega = 0.0042 \pm 0.0008$. We apply the TCC method to the 7% eclipses of KIC 6307537, i.e., when the evolved star is being eclipsed. We detect a strong correlation ($N_{\sigma} = 11.4$) in the residual flux at $P_{\text{tcc}} = 80.4 \pm 1.6$ days. This is consistent with the $P_{\text{phot}} = 78 \pm 3$ days. We interpret the strong TCC as evidence for a low obliquity of the evolved star. We place an upper bound of $\Psi \lesssim 6^{\circ}$. The agreement between $P_{\text{tcc}}$ and $P_{\text{phot}}$ not only supports the low-obliquity interpretation but also suggests that the rotational modulation in the out-of-eclipse light curve originates from the evolved star. Figure \[6307537\_tapestry\] shows the eclipse tapestry. The tapestry suggests there are two active regions near longitudes of 40$^{\circ}$ and 120$^{\circ}$ each spanning about 5-10$^{\circ}$ in longitude. The active regions appear wider in Figure \[6307537\_tapestry\] because of the finite radius ratio. With $R_1/R_2 \approx 0.24$, the photometric features due to the active regions are broadened by about 30$^{\circ}$ in longitude. These photometric features in the residual flux had amplitudes of about 1-2% in relative flux. Comparing with the eclipse depth of about 7%, the active regions on average were roughly 15-30% dimmer than the photosphere. The active regions seem to have lasted the entire [*Kepler*]{} campaign indicating a lifetime of at least $10^3$ days. KIC 5193386 ----------- KIC 5193386 is another [*Kepler*]{} eclipsing binary system. One set of eclipses occurs every 21.4 days, during which the total light decreases by 8% for about 22 hours. The other set of eclipses is deeper (24%) and flat-bottomed. The secondary star (the star being eclipsed during the 8% fading events) is likely evolved, with a larger radius and lower effective temperature than its companion. We measure a rotational modulation of $26.0 \pm 0.8$ days in the [ *Kepler*]{} light curve. From the separation between the primary and secondary eclipse, @vanEylenEB constrained the eccentricity: $e\cos\omega = -0.0022 \pm 0.0015$. By analyzing the 8% eclipses we find a strong TCC, with $N_{\sigma}=9.5$ and $P_{\text{tcc}} = 25.8 \pm 0.3$ days, indicating a low obliquity ($\Psi \lesssim 14^{\circ}$). Figure \[5193386\_tapestry\] shows the tapestry. During the early stages of [*Kepler*]{} observations, one prominent active region was located at a transit longitude of 30$^{\circ}$. It gradually split into two distinct active regions that separated longitudinally from each other. Again, we interpret this phenomenon as the emergence of a magnetic flux tube and the subsequent separation of the two footprints of the tube on the stellar photosphere. The active regions are consistent of being about 10$^{\circ}$ wide and are 10–25% fainter than the photosphere. KIC 6603756 ----------- KIC 6603756 is an eclipsing binary system in the [*Kepler*]{} catalog of @Prsa, who classified this system as an eclipsing binary of the Algol type (detached). Only one set of eclipses was detected. These 4% eclipses last for about 6 hours and repeat every 5.2 days. The [*Kepler*]{} light curve shows a clear rotational modulation of 6.128 $\pm$ 0.054 days. Our TCC analysis gives a strong correlation, $N_{\sigma} = 9.2$, at $P_{\text{tcc}} = 6.0 \pm 0.1$ days. We interpret the strong TCC as a sign of a low obliquity for eclipsed star and place an upper bound of $\Psi\lesssim 12^{\circ}$. Figure \[6603756\_tapestry\] shows the eclipse tapestry. Active regions seemed to occur preferentially near a transit longitude of 230$^{\circ}$ throughout the 300 days of [ *Kepler*]{} observations. These photometric features in the residual flux had amplitudes of about 0.002–0.006 in relative flux which, when compared to the eclipse depth of 4%, implies that the implying that the active regions were on average 5–15% dimmer than the photosphere. KIC 5098444 ----------- KIC 5098444 is a [*Kepler*]{} eclipsing binary with 2% primary eclipses and 0.4% secondary eclipses (occultations), and an orbital period of 26.9 days. The [*Kepler*]{} light curve shows a clear rotational modulation of 23.49 $\pm$ 0.19 days. The primary eclipses, analyzed here, have a duration of about 11 hours. The TCC is strong, with $N_{\sigma} = 7.9$ and $P_{\text{tcc}} = 23.40 \pm 0.24$ days. We conclude the primary star has a low obliquity, with $\Psi\lesssim 8^{\circ}$. The agreement between $P_{\text{tcc}}$ and $P_{\text{phot}}$ also shows that the rotational modulation in the out-of-eclipse light curve originates from the primary. Figure \[5098444\_tapestry\] is the eclipse tapestry. An active region near transit longitude of 90$^{\circ}$ persisted throughout the 800 days of [*Kepler*]{} observations. The flux anomalies have amplitudes of 0.5–1%. Comparing with the eclipse depth of about 2%, the active regions on average were roughly 20-50% dimmer than the photosphere. The active regions span about 40$^{\circ}$ in longitude after accounting for the finite size of the secondary. KIC 7767559 ----------- KIC 7767559 was originally listed as a planetary candidate, KOI-895. This system was later classified as an eclipsing binary due to the detection of significant secondary eclipses. The primary eclipse occurs every 4.4 days with a depth of 1% and a duration of 3.9 hours. The [*Kepler*]{} light curve shows a clear rotational modulation of $P_{\text{phot}}$ = 5.02 $\pm$ 0.20 days. @Holczer2015 detected a strong correlation between TTV and local flux variation, indicating a prograde orbit. The 1% primary eclipses show a strong TCC with $N_{\sigma} = 7.7$ at $P_{\text{tcc}} = 5.1 \pm 0.1$ days. We place an upper bound on the obliquity, $\Psi\lesssim 7^{\circ}$. Figure \[7767559\_tapestry\] shows the Eclipse Tapestry. Although there are signs of group-group correlation, any large-scale, long-lasting active regions are not visually obvious. KIC 5376836, 3128793, 5282049, 5282049 -------------------------------------- KIC 5376836, 3128793, 5282049, 5282049 are all eclipsing binary systems discovered by [*Kepler*]{}. In all cases our analysis revealed strong TCC signals, with $N_{\sigma}$ = 5–7, and TCC periods that agree with the independently measured photometric periods. We put an upper bounds on the obliquity $\Psi \lesssim$ 3 to 20$^{\circ}$ (See Table \[tab:eb\]). Figure \[5376836\_tapestry\] shows the eclipse tapestries for each system, all of which lack high-contrast, long-lasting active regions. Discussion and Conclusion {#discussion} ========================= The TCC method -------------- As any other method to obtain information on stellar obliquities, the TCC method has strengths and weaknesses. One of the advantages of the TCC method is that it does not make any assumption about the size, shape and intensity distribution of the active regions. In contrast, the traditional spot-tracking methods [e.g. @Sanchis-Ojeda2011WASP; @Tregloan-Reed; @Beky2014] often assume circular and uniform starspots for ease of modeling. The TCC method looks for recurrence in the residual flux regardless of the shape of the recurring pattern. Therefore it can handle spots of arbitrary shape and intensity distribution and even bright active regions such as plages and faculae. We have created transit tapestries to allow the properties of the active regions on transit chord to be tracked and visualized in a model-independent manner. The TCC statistic combines all the data together, and can therefore be effective even in systems with lower SNR for which no features can be discerned in the transit tapestry. In contrast, the traditional spot-tracking method is based on the visual identification of individual spot-crossing events. This is more subjective and is only possible in the systems for which the data have the highest SNR. The TCC method is largely automated and does not require additional follow-up observations, facilitating the application to a large sample of systems. One of the limitations of the TCC method is that it requires light curves with a high temporal sampling rate. This is because the durations of spot-crossing events are similar to the brief durations of the ingress and egress phases of the transit. Continuous monitoring of the system is also very important for the TCC method. Continuous monitoring enables an independent check on the stellar rotation period $P_{\text{phot}}$ from the rotational modulation in the out-of-transit light curve. Moreover, active regions have a limited lifetime, and may drift in longitude and latitude. The recurring pattern in residual flux caused by active regions may quickly change or disappear after a few transits. This is why it is important to observe many pairs of neighboring transits, separated by a relatively short amount of time. In the TCC method, the detection of a strong correlation requires a confluence of factors. The host star must be magnetically active, with stable active regions. Unless the system is close to a spin-orbit commensurability, the stellar obliquity $\Psi$ must be nearly 0$^\circ$ or 180$^\circ$, such that the planet repeatedly transits the same active regions. In addition, the impact parameter of a transiting planet must be such that the transit chord overlaps with the active latitudes on the host star. On the other hand, when no strong TCC is observed, the system is not guaranteed to have a high obliquity, as the failure of any of the preceding conditions could explain the lack of correlation. Therefore the TCC method is best at picking out stars with obliquities that are very low or are close to exactly retrograde. In contrast, the Rossiter-McLaughlin effect can lead to tight constraints for just about any angle, although it is only sensitive to the sky-projected obliquity $\lambda$. Detected low-obliquity systems ------------------------------ We applied the TCC method to selected [*CoRoT*]{}, [*Kepler*]{} and [*K2*]{} transiting planets and planetary candidates. We found 10 cases in which the star has a low obliquity (see Table \[tab:planet\]). Among these, five were not previously known to have a low obliquity: Kepler-71b, KOI-883.01, Kepler-45b, Kepler-762b and Kepler-423b. For the other five systems, we confirmed the low stellar obliquities reported previously. Notably, all of these low-obliquity detections are for stars with giant planets ($R_p>0.9~R_{\text{Jup}}$), with scaled semi-major axis $a/R_{\star} <12$. This is not too surprising; the requirements for a relatively high SNR and many consecutive transits are most easily met for close-in giant planets. All of these low-obliquity stars are G and K dwarfs, with the sole exception of Kepler-45, which is an M dwarf. The low obliquities of the G and K dwarfs are consistent with the general trend that hot Jupiters around relatively low-mass stars below the Kraft break tend to have low obliquities [@Winn2010a]. The presence of a convective outer layer in these low-mass stars may be also crucial for the generation of stellar magnetic activity. In contrast, several previously reported low-obliquity systems (Kepler-25c, CoRoT-11b and Kepler-8b) with more massive host stars did not show strong TCC. This may be ascribed to the lack of magnetic activity; these stars have effective temperatures exceeding the Kraft break, and do not have thick outer convective zones. Kepler-45 is only the second M dwarf for which the stellar obliquity has been measured, the other one being GJ436 [@Bourrier2017]. As noted in the introduction, the distribution of stellar obliquities has been observed to depend on the properties of the host star. @Winn2010a found that the stars above and below the @Kraft break have different obliquity distributions. Stars cooler than about 6250K have thick convective envelopes, while hotter stars have radiative envelopes. The differing obliquity distributions may be related to the differing magnetic fields, tidal dissipation rates, or rotational histories of the stars on either side of this boundary. Performing obliquity measurements on a wider range of stellar types may help to clarify the situation. This is especially true for M dwarfs which may be completely convective and for which obliquity measurements have been very limited. Rossiter-McLaughlin observations have not been very successful for M dwarfs, because of their higher stellar variability, slower stellar rotation, and faint optical magnitudes, all of which hinder the acquisition and interpretation of high-resolution spectra. The TCC method may be particularly useful for M dwarfs because high levels of activity and slow rotation are actually favorable for the technique, and because high-resolution spectroscopy is not needed. As we mentioned previously, the TCC method is capable of identifying both well-aligned and perfectly retrograde systems. However, among the 64 planetary systems and 24 eclipsing binaries in our sample, none were found with a strong TCC corresponding to a retrograde orbit. In particular, all 10 of the planetary systems that had a strong TCC were found to have very well-aligned orbits, even though retrograde orbits would have been equally easy to detect. All 10 of the systems feature cool stars, below the Kraft break. Thus the results for these 10 systems is further evidence that cool stars with hot Jupiters mainly have prograde orbits rather than retrograde orbits. Constraints on stellar magnetic activity ---------------------------------------- The properties of sunspots have been tracked since the 18th century. It is well known that sunspots preferentially emerge at latitudes of about $30^{\circ}$ north or south of the equator at the start of a magnetic cycle [@Hathaway]. These active latitudes gradually drift towards the equator as the magnetic cycle continues, giving rise to the “butterfly diagram” when spot latitudes are plotted against time. It has also been noted that sunspots tend to cluster at certain active longitudes [see e.g. @Bumba1965]. Other solar magnetic phenomena such as flares, coronal mass ejections, and X-ray emission are also associated with active longitudes [@Zhang2008; @Gyenge2017; @Zhang2007]. Typically there are two active longitudes on the Sun separated by 180$^{\circ}$, each of which has a width of order 20$^{\circ}$ [@BU2003]. At solar maximum there can be as many as four active longitudes, and at solar minimum active longitudes may disappear [@deToma2000]. The lifetime of active longitudes has been found to be seven rotation cycles [@deToma2000] and perhaps even as long as a century [@BU2003]. While the exact physical mechanism responsible for persistently active longitudes is still unknown, they do not seem to be a unique property of our Sun. @Lanza2009 and @GA2011 found evidence for active longitudes on other stars too. The transit tapestries of several low-obliquity systems with high SNR and strong magnetic activity — Kepler-17b, CoRoT-2b, Qatar-2b, Kepler-71b and KOI-883.01 — allow us to track the size, contrast, lifetime and latitudinal distribution of the active regions along the transit chords of these planet-hosting stars. These are the most magnetically active among the systems we have studied. We summarize the properties of the active regions of these systems in Table \[active\_regions\]. We note the following trends. In all cases the Rossby number — the ratio between the rotation period and the convective overturn time — is smaller than that of the Sun. The Rossby number is routinely used for expressing thresholds and scaling relations in the study of stellar magnetic activity and magnetic braking [see, e.g., @vanSaders]. The strong magnetic activity that is characteristic of our sample is linked to their fast rotation. All five stars are below the Kraft break. The typical contrast of the active regions is on the order of 10% in the wide optical bandpasses of the [*Kepler*]{} and [*CoRoT*]{} data. The typical lifetime of the active regions is on the order of 100 days, although we note that the inferred lifetime of the active longitudes is degenerate with the rate of latitudinal migration. If the active regions migrate away from the transit chord, the active regions will no longer cause photometric signatures in the transit light curve. For all of the systems in Table \[active\_regions\], there are typically two active longitudes that are present on the same stellar latitude simultaneously. At a given time there may be no active longitudes or as many as four, similar to the Sun. Interestingly, the size, contrast, and the lifetime of the different active regions on the same stars are very similar to each other. The latitude probed by the planet is $\lesssim$20$^{\circ}$ in all cases, suggesting that active regions tend to occur at these low latitudes, which is also similar to the situation on the Sun. If active regions had no preferred latitudes, the systems in Table \[active\_regions\] might have shown a broader distribution of transit impact parameters. @Montet2017 noted a pattern in the origin of photometric variability of Sun-like stars observed by [*Kepler*]{} ($T_{\rm eff}$ within 150K of the solar value and $\log g >4.2$). The brightness variations of stars with rotation periods shorter then 25 days are consistent with those produced by dark spots, while slower rotators are more consistent with bright spots (plages and faculae). The most magnetically active systems in our sample (Table \[active\_regions\]) are all consistent with being spot-dominated. This is inferred from the prevalence of positive residuals seen in the transit tapestries. For the slower rotators in our sample, TrES-2 and Kepler-423, we did not observe strong features in the transit tapestries, perhaps because the active regions are faculae spanning a larger angular size. Eclipsing binaries ------------------ We now turn to the eclipsing binaries in our sample. We found eight [*Kepler*]{} eclipsing binaries that show strong correlation and good agreement between $P_{\text{phot}}$ and $P_{\text{tcc}}$ (see Table \[tab:eb\]). It is more difficult to detect and interpret the TCC signals for these systems. Due to the lack of follow-up observations, little is known about the properties of the stars in these binaries. The stellar parameters in Table \[tab:eb\] were estimated from fitting stellar models and evolutionary tracks to archival broadband photometry [@Prsa]. Moreover, only the properties of the primary stars were reported. Second, because of the larger radius ratio in these binaries, the TCC method provides poor angular resolution for probing the active regions. Furthermore, some of the eclipsing binaries were only observed in the long-cadence mode. The 30-minute time averaging smears out the photometric patterns of active regions. Finally the orbital periods of some binaries are longer than 20 days. This means we have a smaller number of eclipses to analyze, and leaves the active regions more time in between eclipses for their properties to change, weakening the TCC signal. Nonetheless, we do see good evidence of active regions in the eclipse tapestries of KIC 6307537, KIC 5193386 and KIC 5098444. Given the shallow eclipse depths and long eclipse durations, one of the stars in each of these binaries is likely evolved. This is further supported by the deeper and flat-bottomed eclipse seen in KIC 6307537 and KIC 5193386 which indicates the eclipsed star has a larger radius and lower effective temperature than the eclipsing star. The active regions extend by about $10^{\circ}$ in longitude and are on the order of 10% fainter than the rest of the photosphere. We did not detect significant longitudinal migration except for KIC 5193386, where one active region gradually split into two regions that drifted apart in longitude. These regions may be the two footprints of an emerging magnetic flux tube. The active regions on these evolved stars persisted for longer than those on the planet-hosting dwarf stars, up to at least a thousand days. The low obliquities of these binaries could be a consequence of tidal evolution, which tends to bring stars into a double-synchronous aligned configuration. However we note that the all eight of the low-obliquity binary systems have not achieved a state of synchronized rotation: $P_{\rm{orb}} \ne P_{\rm{rot}}$ (see Table \[tab:eb\]). Thus the low-obliquity state may reflect the initial conditions of formation, rather than the outcome of tidal evolution. Future prospects ---------------- Crucial to the success of the TCC methods are light curves with high SNR, high cadence, and continuous time coverage for many orbital and rotation periods. New opportunities for applying TCC will be provided by the existing and upcoming space-based missions [*TESS*]{} [@Ricker2014], [*CHEOPS*]{} [@Cheops], and [*PLATO*]{}. The [*TESS*]{} mission, scheduled for launch in 2018, will provide a fresh sample of transiting planets which will be suitable targets for the TCC analysis. The plan for the mission is to obtain data with 2-minute cadence for several hundred thousand preselected target stars. This will be ideal for resolving the photometric signatures of active regions. Most [*TESS*]{} stars will only be observed for one month during the two-year primary mission, but a small fraction of the stars (near the ecliptic poles) will be observed for as long as a year. Our best hopes are for [*TESS*]{} targets that are observed for at least a few months, similar in duration to the 80-day [*K2*]{} campaigns that have already yielded some strong detections with the TCC method. Using the [*TESS*]{} mission simulations of @Sullivan and @Bouma, we have made a rough estimate that there will be a few dozen [*TESS*]{} planets with single-transit SNR $>50$, amenable to TCC analysis. We are grateful to Ronald Gilliland for helping to arrange short-cadence [*Kepler*]{} observations for many of the targets analyzed here. Work by J.N.W. was supported through the [ *Kepler*]{} Participating Scientist program and by NASA award NNN12AA01C. [^4] Status $P_{\text{orb}}$ (days) $a/R_{\star}$ $M_p$ ($M_{\text{Jup}}$) $R_p$ ($R_{\text{Jup}}$) $M_\star$ ($M_{\odot}$) $R_\star$ ($R_{\odot}$) $T_{\text{eff}}$ (K) $P_{\text{phot}}$ (days) $P_{\text{tcc}}$ (days) $N_{\sigma}$ $P_{\text{tcc}}$/$P_{\text{orb}}$ $\Psi_{\text{Upper}}$ ($^{\circ}$) $\lambda_{\text{lit}}$($^{\circ}$) Ref. --------------------------------- ------------- ------------------------- --------------------------- -------------------------- --------------------------- ------------------------- ------------------------- ---------------------- -------------------------- ------------------------- -------------- ----------------------------------- ------------------------------------ ------------------------------------ ----------------------------- -- Kepler-17b Confirmed 1.4857108(2) 5.48(2) 2.450(14) 1.312(18) 1.16(6) 1.05(3) 5781(85) 11.9$\pm$1.1 11.8(7) 59.6 7.94 &lt;10 &lt; 15 @Desert2011 CoRoT-2b Confirmed 1.7429964(17) 6.70(3) 3.31(16) 1.465(29) 0.97(6) 0.902(18) 5575(66) 4.530(68) 4.5(3) 22.6 2.58 &lt;4 4.7$\pm$12.3 @Alonso2008; @Nutzman2011 Qatar-2b Confirmed 1.3371182 (37) 6.52(10) 2.487(86) 1.144(35) 0.740(37) 0.713(18) 4645(50) 18.5 $\pm$1.9 18.2(4) 18.0 13.61 &lt;11 0$\pm$10 @Esposito2017 Kepler-71b Confirmed 3.905081(72) 11.92(15) 1.11(2) 0.95(5) 0.86(2) 5591(105) 19.87(18) 19.7(8) 10.7 5.04 &lt;6 Prograde @Howell2010; @Holczer2015 KOI-883.01 Candidate 2.688899317(109) 10.40 1.05(13) 0.702(63) 0.643(81) 4809(151) 9.11(11) 9.1(2) 9.2 3.38 &lt;4 Prograde ExoFOP; @Holczer2015 Kepler-45b Confirmed 2.455239(5) 10.6 $\pm$ 1.0 0.51(9) 0.96(11) 0.59(6) 0.55(11) 3820(90) 15.8(2) 16.7(8) 6.3 6.80 &lt;11 @Johnson2012 TrES-2b Confirmed 2.4706133738(187) 7.98(17) 1.253(52) 1.189(25) 0.98(62) 1.000(36) 5850(50) 28.35(34) 29.9$\pm$1.4 5.8 12.10 &lt;10 9$\pm$12 @Winn2008 Kepler-762b Confirmed 3.7705521(94) 8.16 1.10(56) 1.06(7) 1.08(23) 5944(124) 4.045(25) 4.0(1) 5.2 1.06 &lt;16 Prograde ExoFOP; @Holczer2015 Kepler-423b Confirmed 2.68432850(7) 8.106$^{+0.117}_{-0.259}$ 0.595(81) 1.192(52) 0.85(4) 0.95(4) 5560(80) 22.047(121) 23.0(9) 4.7 8.57 &lt;10 @Gandolfi2015 WASP-85b Confirmed 2.6556777(44) 8.79(23) 1.265(65) 1.24(3) 1.04(7) 0.96(13) 5685(65) 13.28(41) 15.2(3) 4.5 5.69 &lt;7 &lt; 14 @Mocnik2016wasp85 HAT-P-11b Confirmed 4.887802443(30) 15.58$^{+0.17}_{-0.82}$ 0.08243(90) 0.42(13) 0.81(3) 0.75(2) 4780(50) 29.472(134) 29.34(30) 6.0 6.00 $103^{+26}_{-10}$ @Winn2010hat Kepler-63b Confirmed 9.4341505(10) 19.12(8) &lt;120 0.545(20) 0.984(40) 0.901(27) 5576(50) 5.401(14) 37.7(8) 5.8 4.00 $-110^{+22}_{-14}$ @Sanchis-Ojeda2013 CoRoT-16b Confirmed 5.35227(20) 11.20$^{+1.21}_{-1.09}$ 0.535(85) 1.17(14) 1.098(78) 1.19(14) 5650(100) 4.0(4) 3.9 0.75 @Ollivier2012 WASP-55b Confirmed 4.465633(4) 10.81(14) 0.57(4) 1.30(5) 1.013(37) 1.011(36) 5900(100) 27(2) 3.5 6.05 @Hellier2012 CoRoT-17b Confirmed 3.7681(3) 6.23(24) 2.43(30) 1.02(7) 1.04(1) 1.59(7) 5740(80) 17(2) 3.5 4.51 @Csizmadia2011 CoRoT-22b Confirmed 9.75598(11) 17.30$^{+1.3}_{-0.57}$ 12.2$^{+14}_{-8.8}$ 0.435$_{-0.035}^{+0.015}$ 1.099(49) 1.136(90) 5939(120) 25(2) 3.2 2.56 @Moutou2014 CoRoT-24b Confirmed 5.1134(6) 10.9 $\pm$ 2.8 &lt;0.018 0.33(4) 0.91(9) 0.86(9) 4950(150) 14(1) 3.2 2.74 @Alonso2014 Kepler-25c[^5] Confirmed 12.7203678(35) 18.52(24) 0.045(8) 0.4598(54) 1.22(6) 1.36(13) 6190(80) 23.21(27) 29.2(7) 3.2 2.30 0.5$\pm$5.7 @Albrecht2013 CoRoT-13b Confirmed 4.03519(3) 10.81(32) 1.308(66) 0.885(14) 1.09(2) 1.01(3) 5945(90) 20.6(6) 3.1 5.11 @Cabrera2010 CoRoT-11b$^{\rm \ref{note1}}$ Confirmed 2.994330(11) 6.890(80) 2.33(34) 1.430(33) 1.27(5) 1.36(13) 6440(120) 30(4) 3.0 10.02 0.1$\pm$2.6 @Gandolfi2012 CoRoT-19b Confirmed 3.89713(2) 6.7(1) 1.11(6) 1.29(5) 1.21(6) 1.65(4) 6090(70) 10(1) 3.0 2.57 -52$^{+27}_{-22}$ @Guenther2012 WASP-47b$^{\rm \ref{note1}}$ Confirmed 4.1591287(49) 9.715(50) 1.123(510) 1.17(79) 1.11(49) 1.16(26) 5576(67) 6.3(4) 3.0 1.51 0$\pm$24 @Sanchis-Ojeda2015 CoRoT-14b Confirmed 1.51214(13) 4.78(28) 7.6(6) 1.09(7) 1.13(9) 1.21(8) 6035(100) 5.4(8) 2.9 3.57 @Tingley2011 CoRoT-8b Confirmed 6.21229(3) 17.6(4) 0.22(3) 0.569(20) 0.88(4) 0.77(2) 5080(80) 25.3(0.3) 2.8 4.07 @Borde2010 CoRoT-25b Confirmed 4.86069(6) 10.2$^{+1.1}_{-0.5}$ 0.27(4) 1.08 $_{-0.1}^{+0.3}$ 1.09$_{-0.05}^{+0.11}$ 1.19 $_{-0.03}^{+0.14}$ 6040(90) 31(1) 2.8 6.38 @Almenara2013 KOI-212.01 Candidate 5.6959023(15) 12.2 0.58(23) 1.013(135) 0.94(38) 6106(156) 16.295(336) 33(1) 2.8 5.79 ExoFOP CoRoT-6b Confirmed 8.886593(4) 17.94(33) 2.96(34) 1.166(35) 1.055(55) 1.025(26) 6006(73) 6.35(13) 5.4(2) 2.8 0.61 @Fridlund2010 Kepler-13Ab Confirmed 1.7633587(20) 3.16(8) 6.6$\pm$1.5 1.335$_{-0.264}^{+0.484}$ 1.49$_{-0.18}^{+0.22}$ 1.80$_{-0.36}^{+0.65}$ 8500(400) 3.12(10) 2.7 1.77 58.6$\pm$2.0 @Masuda2015 CoRoT-3b Brown Dwarf 4.256800(5) 7.8(4) 21.7 $\pm$ 1.0 1.01(7) 1.41(8) 1.44(8) 6740(140) 24.3(5) 2.7 5.71 -37.6$^{+22.3}_{-10.0}$ @Deleuil2008; @Triaud2009 HATS-9b Confirmed 1.9153073(52) 4.36$^{+0.10}_{-0.25}$ 0.837(29) 1.40(6) 1.28(13) 1.32(7) 5366(70) 13.6$\pm$1.1 2.7 7.10 @Brahm2015 Kepler-471b Confirmed 5.01423457(59) 8.3 0.94(69) 1.314(257) 1.271(927) 6475(169) 26.51(57) 20.5(8) 2.7 4.09 ExoFOP Kepler-494b Confirmed 8.0251182(23) 11.5 0.637(118) 1.10(7) 1.26(23) 5513(184) 8.985(186) 30.1(4) 2.6 3.75 ExoFOP CoRoT-12b Confirmed 2.828042(13) 7.74(18) 0.917(70) 1.44(13) 1.078(72) 1.116(92) 5675(80) 9.8(8) 2.6 3.47 @Gillon2010 CoRoT-18b$^{\rm \ref{note1}}$ Confirmed 1.9000693(28) 6.35(40) 3.47(38) 1.31(18) 0.95(15) 1.00(13) 5440(100) 5.4(4) 15.7(7) 2.6 8.26 -10$\pm$20 @Hebrard2010 CoRoT-28b Confirmed 5.20851(38) 7.29(16) 0.484(87) 0.955(66) 1.01(14) 1.78(11) 5150(100) 41(3) 2.6 7.87 @Cabrera2015 Kepler-448b$^{\rm \ref{note1}}$ Confirmed 17.8552333(9) 18.8(4) &lt; 10 1.43(13) 1.452(93) 1.63(15) 6820(120) 1.245(124) 21.0(6) 2.6 1.18 12.6$^{+3.0}_{-2.9}$ @Bourrier2015 CoRoT-21b Confirmed 2.72474(14) 4.60(26) 2.26(31) 1.30(14) 1.29(9) 1.95(21) 6200(100) 16.7(5) 2.5 6.13 @Patzold2012 Kepler-420b Confirmed 86.647661(34) 155.4$\pm$2.4 1.45(35) 0.94(12) 0.99(5) 1.13(14) 5520(80) 11.7(7) 2.5 0.14 75$^{+32}_{-46}$ @Santerne2014 Kepler-74b Confirmed 7.340718(1) 11.8$^{+1.4}_{-0.8}$ 0.68(9) 1.32(14) 1.4$_{-0.11}^{+0.14}$ 1.51(14) 6050(110) 26.7(8) 36$\pm$1.2 2.4 4.90 @Hebrard2013 Kepler-96b Confirmed 16.2385 21.2 0.027(11) 0.238(20) 1.00(6) 1.02(9) 5690(73) 14.922(83) 42.7$\pm$1.7 2.3 2.63 @Marcy2014 Kepler-485b Confirmed 3.2432598(18) 8.2 1.286$^{+0.302}_{-0.145}$ 1.07(8) 1.09(26) 5958(153) 30.72(45) 22.8$\pm$1.1 2.3 7.03 ExoFOP CoRoT-7b Confirmed 0.85353(2) 4.27(20) 0.0149(30) 0.149(8) 0.93(3) 0.87(4) 5275(75) 22.4$\pm$3.6 2.5(2) 2.3 2.93 @Leger2009 CoRoT-29b Confirmed 2.850570(6) 9.22(19) 0.85(20) 0.90(16) 0.98(14) 0.90(12) 5260(100) 16.1$\pm$1.5 2.3 5.65 @Cabrera2015 WASP-107b Confirmed 5.7214742(43) 18.164(37) 0.12(1) 0.948(30) 0.69(5) 0.66(2) 4430(120) 17.1$\pm$1.0 10.0(5) 2.3 1.75 40-140 @Anderson2017 [@Dai2017] Kepler-7b Confirmed 4.885525(40) 6.623(49) 0.443(42) 1.614(15) 1.36(3) 2.02(2) 5933(44) 15.02(21) 54.1(9) 2.2 11.07 @Demory2011 Kepler-8b$^{\rm \ref{note1}}$ Confirmed 3.52254(5) 6.97(24) 0.60(19) 1.419(58) 1.213(63) 1.486(62) 6213(150) 28.64(32) 27.1(3) 2.2 7.69 5$\pm$7 @Jenkins2010; @Albrecht2012 CoRoT-23b Confirmed 3.6313(1) 6.85(60) 2.8(3) 1.08(13) 1.14(8) 1.61(18) 5900(100) 7.3$\pm$1.5 2.2 2.01 @Rouan2012 Kepler-522b Confirmed 38.58422(46) 28.4 0.616(134) 1.54(14) 1.98(43) 6267(109) 4.895(34) 27.7(5) 2.2 0.72 ExoFOP CoRoT-26b Confirmed 4.20474(5) 6.28$^{+0.01}_{-0.52}$ 0.52(5) 1.26(13) 1.09(6) 1.79(18) 5590(100) 15.1(9) 2.2 3.59 @Almenara2013 CoRoT-27b Confirmed 3.57532(6) 9.48(95) 10.39(55) 1.007(44) 1.05(11) 1.08(18) 5900(120) 35.5(5) 2.2 9.93 @Parviainen2014 HATS-11b Confirmed 3.6191613(99) 6.88(27) 0.85(12) 1.510(78) 1.00(6) 1.44(6) 6060(150) 14(3) 2.1 3.87 @Rabus2016 Kepler-539b Confirmed 125.63243(71) 103.3$\pm$1.1 0.97(29) 0.747(16) 1.048(45) 0.952(20) 5820(80) 11.769(16) 16.0(2) 2.0 0.13 @Mancini2016 CoRoT-4b Confirmed 9.20205(37) 16.54(18) 0.72(8) 1.19(5) 1.10(2) 1.15(3) 6190(60) 8.7$\pm$1.1 31(2) 2.0 3.37 @Moutou2008 HAT-P-7b Confirmed 2.204737(17) 4.1502(39) 1.741(28) 1.431(11) 1.51(5) 2.00(2) 6259(32) 28.10(36) 0.8(2) 2.0 0.36 155$\pm$37 @Albrecht2012; @Masuda2015 WASP-75b Confirmed 2.484193(3) 6.40(14) 1.07(5) 1.270(48) 1.14(7) 1.26(4) 6100(100) 13.7$\pm$1.1 22$\pm$3 1.9 8.87 @Gomez2013 Kepler-548b Confirmed 4.45419434(38) 14.7 1.070(195 ) 0.93(9) 0.90(16) 5359(178) 31.6(3.1) 115(2) 1.9 25.82 ExoFOP CoRoT-10b Confirmed 13.2406(2) 31.33 $\pm$ 2.15 2.75(16) 0.97(5) 0.89(5) 0.79(5) 5075(75) 30(2) 1.9 2.27 @Bonomo2010 CoRoT-1b Confirmed 1.5089557(64) 4.92(8) 1.03(12) 1.49(8) 0.95(15) 1.11(5) 5950(150) 14(3) 1.8 9.28 @Barge2008 CoRoT-15b Brown Dwarf 3.06036(3) 6.68$^{+0.49}_{-1.04}$ 63.3 $\pm$ 4.1 1.12(30) 1.32(12) 1.46(31) 6350(200) 19.7(6) 1.7 6.44 @Bouchy2011 KOI-425.01 Candidate 5.42834472(49) 22.1 1.23(53) 1.075(187) 0.965(419) 5936(166) 5.301(177) 11.2(8) 1.7 2.06 ExoFOP WASP-118b Confirmed 4.0460435(44) 6.92(11) 0.514(20) 1.440(36) 1.320(35) 1.696(29) 6410(125) 7.8(6) 1.6 1.93 @Hay2016 CoRoT-5b Confirmed 4.0378962(19) 8.970(47) 0.467(47) 1.330(47) 1.00(2) 1.186(40) 6100(65) 2.5(2) 1.5 0.62 @Rauer2009 CoRoT-20b Confirmed 9.24285(30) 18.95(73) 4.24(23) 0.84(4) 1.14(8) 1.02(5) 5880(90) 12(4) 1.1 1.30 @Deleuil2012 CoRoT-9b Confirmed 95.2738(14 ) 93(3) 0.84(7) 0.94(4) 0.99(4) 0.94(4) 5625(80) 26(1) 0.7 0.27 @Deeg2010 \[tab:planet\] ------------ ------------------------- ------------------------- --------------------- --------------- --------------- ------------------------- ------------------------- ---------------------- -------------------------- ------------------------- -------------- ----------------------------------- ------------------------------------ -------- -- -- KIC Status $P_{\text{orb}}$ (days) ecos($\omega$) [^6] $a/R_{\star}$ $R_1/R_2$ $M_\star$ ($M_{\odot}$) $R_\star$ ($R_{\odot}$) $T_{\text{eff}}$ (K) $P_{\text{phot}}$ (days) $P_{\text{tcc}}$ (days) $N_{\sigma}$ $P_{\text{tcc}}$/$P_{\text{orb}}$ $\Psi_{\text{Upper}}$ ($^{\circ}$) Ref. 6307537 EB (Algol Detached) 29.74455(33) 0.0042 (8) 11.54 0.23861(45) 0.839(73) 0.788(94) 4892(192) 78(3) 80.4 $\pm$1.6 11.4 2.70 &lt;6 ExoFOP 5193386 EB 21.378308(138) -0.0022(15) 9.31 0.261272(288) 0.883(115) 3.484(909) 4905(66) 26.0(8) 25.8(3) 9.5 1.21 &lt;14 ExoFOP 6603756 EB (Algol Detached) 5.204283146(797) 7.09 0.177941(42) 0.798(134) 0.86(34) 5332(188) 6.128(54) 6.0(1) 9.2 1.15 &lt;12 ExoFOP 5098444 EB 26.94923943(977) 19.24 0.131908(196) 0.526(30) 0.512(26) 4723(135) 23.49(19) 23.40(24) 7.9 0.87 &lt;8 ExoFOP 7767559 EB 4.409409352(787) 9.15 0.106974(96) 0.811(144) 0.975(405) 5600(186) 5.02(20) 5.1(1) 7.7 1.16 &lt;7 ExoFOP 5376836 EB 3.479425021(583) 7.0 0.136233(295) 0.929(108) 0.878(248) 5907(158) 3.63(21) 3.69(16) 7.0 1.06 &lt;22 ExoFOP 3128793 EB 24.68637(57) 12.24 0.0909(5) 0.950(201) 5.132(1.336) 4648(69) 67.1 $\pm$2.2 67.8 $\pm$2.0 5.4 2.75 &lt;3 ExoFOP 5282049 EB 5.91037138(84) 12.48 0.12539(85) 1.029(134) 0.958(291) 6200(200) 17.5 $\pm$1.0 16.9(5) 5.0 2.86 &lt;4 ExoFOP 4737267 EB 9.52407775(158) 8.71 0.416822(283) 1.112(189) 3.371(0.641) 5180(62) 9.52(7) 9.25(10) 12.6 0.97 ExoFOP 5270698 EB (Algol Detached) 3.964333792(178) 7.01 0.15095(64) 1.258(106) 2.318(380) 5758(77) 4.004(33) 15.38(25) 10.8 3.88 ExoFOP 8081482 EB (Algol Detached) 2.819454482(498) 5.42 0.254718(261) 1.074(305) 1.074(370) 5767(161) 2.762(86) 2.8(2) 9.0 0.99 ExoFOP 6548447 EB 10.76838965(176) -0.0873(48) 6.55 0.297635(114) 0.921(101) 1.643 (851) 5226(173) 5.380(39) 9.6(1) 7.5 0.89 ExoFOP 7940533 EB 3.905532661(473) 0.0070(36) 5.04 0.242875(1) 0.812(100) 0.798(152) 5495(169) 3.842(96) 7.5(2) 5.9 1.92 ExoFOP 3955867 EB 33.6575(83) 6.54 0.110724(28) 0.871(151) 6.447(1.781) 4774(71) 33.26(39) 33.4 $\pm$1.1 5.2 0.99 ExoFOP 8180020 EB 5.803127542(553) 8.93 0.231357(76) 0.876(092) 0.795(188) 5533(182) 6.54(10) 16.2(2) 5.1 2.79 ExoFOP 8230809 EB (Algol Detached) 4.078353619(475) 0.0108(113) 5.7 0.182022(95) 0.83(105) 0.867(221) 5872(177) 3.989(81) 7.8(4) 5.1 1.91 ExoFOP 5370302 EB 3.904344085(366) 0.0114(118) 7.812 0.127552(175) 0.989(116) 0.898(260) 5810(157) 3.63(21) 7.8(6) 4.3 2.00 ExoFOP 6613006 EB (Algol Detached) 7.388833197(504) 10.04 0.117062(64) 0.926(99) 0.838(241) 5845(139) 34.2(2) 4.3 4.63 ExoFOP 5306862 EB 2.025580507(270) -0.0052(27) 4.44 0.197412(202) 1.431(201) 1.990(578) 6348(170) ELV[^7] 1.0(5) 4.0 0.49 ExoFOP 6205460 EB 3.7228051(92) 0.0155(0.0018) 1.96 0.324154(195) 1.166(177) 2.59(1.10) 5425(296) 2.47(1) 0.6(2) 3.3 0.16 ExoFOP 5700330 EB 53.2204218(52) 22.96 0.25037(46) 1.240(96) 2.057(292) 5653(84) 26.18(27) 17.5 $\pm$1.2 3.0 0.33 ExoFOP 6694186 EB 5.554226627(175) 0.0142(12) 10.98 0.349048(83) 0.756(119) 0.843(181) 5433(145) 5.472(83) 7.24(20) 2.9 1.30 ExoFOP 5014753 EB 3.17062650(50) 0.0057(111) 8.96 0.12876(30) 1.042(159) 1.034(330) 6068(190) 3.66(11) 21(4) 2.6 6.62 ExoFOP 2438502 EB (Variable of RS CVn) 8.36061289(25) 7.45 0.16632(162) 1.343(144) 2.445(420) 5463(87) 8.31(28) 8.4(5) 1.9 1.00 ExoFOP \[tab:eb\] ------------ ------------------------- ------------------------- --------------------- --------------- --------------- ------------------------- ------------------------- ---------------------- -------------------------- ------------------------- -------------- ----------------------------------- ------------------------------------ -------- -- -- System $T_{\rm eff}$ (K) $P_{\rm rot}$ (days) Ro[^8] Longitudinal Size[^9]($^{\circ}$) Contrast[^10] Lifetime (days) $N_{\rm AL}$[^11] Active Latitude ($^{\circ}$) ----------- ------------------- ---------------------- -------- ----------------------------------- ------------------------ ------------------- ------------------- ------------------------------ Sun[^12] 5777 25.05 2.3 $\lesssim$20 $\approx$3-33% (spots) $\approx$200[^13] 2 (0-4) 30-10 Kepler-17 5800 11.9 0.81 5-20 10-20% 100-200 1-4 16$ \pm $8 CoRoT-2 5600 4.5 0.22 10-20 7-17% &gt;150 2 13$\pm$10 Qatar-2 4600 18.5 0.76 15-25 3-7% &gt;80 2 2$\pm$9 Kepler-71 5600 19.9 1.04 5-15 10-20% 100-200 2 2$\pm$8 KOI-883 4800 9.11 0.40 10-40 10-20% $\approx$100 0-2 0$\pm$10 \[active\_regions\] [^1]: <http://idoc-CoRoT.ias.u-psud.fr> [^2]: <http://dfm.io/kplr> [^3]: <https://exofop.ipac.caltech.edu> [^4]: The columns are status of the planet, the orbital period $P_{\text{orb}}$, the scaled semi-major axis $a/R_{\star}$, the planetary mass $M_p$, the planetary radius $R_p$, the stellar mass $M_\star$, the stellar radius $R_\star$, the stellar effective temperature $T_{\text{eff}}$, the stellar rotation period measured from rotational modulation in the light curve $P_{\text{phot}}$, stellar rotation period measured from correlation in the residual flux $P_{\text{tcc}}$, the significance of correlation compared to the results of scrambling test $N_{\sigma}$, ratio between the stellar rotation period and orbital period $P_{\text{tcc}}$/$P_{\text{orb}}$, the upper limit on the true obliquity $\Psi_{\text{Upper}}$, the obliquity constraint from literature $\lambda_{\text{lit}}$ and the references. The systems are sorted by the significance of correlation in the residual flux. Upper limit on the true obliquity is only calculated when a low stellar obliquity is detected. [^5]: \[note1\]Systems whose stellar obliquity was previously reported to be low yet did not show strong TCC. These systems are likely magnetically inactive: the host stars are above the Kraft break; the light curve lacks rotational modulation. Alternatively the high impact parameter of the planets indicate that the transit chord might have missed the active latitude. [^6]: Eccentricity constraints from @vanEylenEB [^7]: unable to measure $P_{\rm phot}$ because ellipsoidal light variation dominates the flux variation [^8]: Rossby Number, using the prescription for convective overturn time of @Noyes [^9]: Longitudinal width of the active regions. [^10]: Average contrast between the active regions and the surrounding photosphere, in the [*Kepler*]{} band. [^11]: Number of the active longitudes [^12]: https://nssdc.gsfc.nasa.gov/planetary/factsheet/sunfact.html; @BU2003, @deToma2000, @Kilcik [^13]: Lifetime of active longitudes instead of sunspots [@deToma2000].

--- abstract: 'Black-box risk scoring models permeate our lives, yet are typically proprietary or opaque. We propose Distill-and-Compare, a model distillation and comparison approach to audit such models. To gain insight into black-box models, we treat them as teachers, training transparent student models to mimic the risk scores assigned by black-box models. We compare the student model trained with distillation to a second un-distilled transparent model trained on ground-truth outcomes, and use differences between the two models to gain insight into the black-box model. Our approach can be applied in a realistic setting, without probing the black-box model API. We demonstrate the approach on four public data sets: COMPAS, Stop-and-Frisk, Chicago Police, and Lending Club. We also propose a statistical test to determine if a data set is missing key features used to train the black-box model. Our test finds that the ProPublica data is likely missing key feature(s) used in COMPAS.' author: - Sarah Tan - Rich Caruana - Giles Hooker - Yin Lou bibliography: - 'main.bib' title: | Distill-and-Compare: Auditing Black-Box Models\ Using Transparent Model Distillation --- &lt;ccs2012&gt; &lt;concept&gt; &lt;concept\_id&gt;10010147.10010341.10010342.10010344&lt;/concept\_id&gt; &lt;concept\_desc&gt;Computing methodologies Model verification and validation&lt;/concept\_desc&gt; &lt;concept\_significance&gt;500&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10002950.10003648.10003662.10003666&lt;/concept\_id&gt; &lt;concept\_desc&gt;Mathematics of computing Hypothesis testing and confidence interval computation&lt;/concept\_desc&gt; &lt;concept\_significance&gt;300&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;concept&gt; &lt;concept\_id&gt;10010405.10010455&lt;/concept\_id&gt; &lt;concept\_desc&gt;Applied computing Law, social and behavioral sciences&lt;/concept\_desc&gt; &lt;concept\_significance&gt;100&lt;/concept\_significance&gt; &lt;/concept&gt; &lt;/ccs2012&gt; Introduction {#sec:introduction} ============ Risk scoring models have a long history of usage in criminal justice, finance, hiring, and other critical domains [@corbett2017algorithmic; @louzada2016classification]. They are designed to predict a future outcome, for example defaulting on a loan. Worryingly, risk scoring models are increasingly used for high-stakes decisions, yet are typically proprietary or opaque. Several approaches have been proposed [@henelius2014peek; @Feldman2015Certifying; @Adler2016AuditingBM; @Adebayo2016Iterative; @datta2016algorithmic; @Wachter2018explanations] to audit black-box risk scoring models: remove, permute, or obscure a protected feature, then see how the the model’s predictions change after retraining the model or probing the model API with the transformed data. However, with many risk scoring models being proprietary, commercial model creators often do not provide unrestricted access to model APIs, much less release the model form or training data. In this paper, we study a more realistic setting where we only have a data set labeled with the risk score as produced by the risk scoring model, the information on the ground-truth outcome, and some or all features; we are not able to probe the model API with new data. We call this data set the *audit data*. We add two potential complications: the audit data may not be the original training data, and the audit data may not have all features used to train the risk scoring model. For example, ProPublica obtained data for their COMPAS study [@propublica_story] not from the company that created COMPAS, but through a public records request to Broward County, a US jurisdiction that used COMPAS in their criminal justice system [@propublica_method], and ProPublica may not have had access to all the features Broward County used to get COMPAS scores. We propose a two-step approach to audit black-box risk scoring models, using audit data with both black-box risk scores and ground-truth outcomes. First, we use a Distill-and-Compare approach (Section \[sec:distillation\]) to understand how features (in the audit data) affect the risk scores. Then, we use a statistical test (Section \[sec:correlation\]) to determine if the black-box model used additional features we do not have access to (i.e. features not in the audit data). The contributions of this paper are: 1) We propose an approach using model distillation and comparison to audit black-box risk scoring models under realistic conditions. 2) We show the importance of calibrating risk scores to remove scale distortions that may have been introduced during the training of the black-box model. 3) We apply the approach to audit four risk scoring models. 4) We propose a statistical test to determine if the audit data is missing key features used to train the black-box model. 5) An ancillary contribution of this paper is a new confidence interval estimate for iGAM [@lou2012intelligible; @lou2013accurate; @caruana2015intelligible], a type of transparent model, to compare two models of this class. Audit Approach {#sec:approach} ============== Our goal is to gain insight into the input-output relationships of a black-box risk scoring model. We draw from model distillation and model comparison techniques to develop our approach. Distill and Compare {#sec:distillation} ------------------- Model distillation was first introduced to transfer knowledge from a large, complex model (teacher) to a faster, simpler model (student) [@bucilua2006compression; @hinton2014distilling; @ba2014deep]. This was done by running unlabeled samples (either new unlabeled data or training data with labels discarded) through the teacher model to obtain the teacher’s outputs, then training the student model to mimic the teacher’s outputs. We draw parallels to our setting, taking the risk scoring model to be the teacher and the audit data to be unlabeled samples ran through the teacher (risk scoring model) to obtain the teacher’s output (risk scores). We train the mimic model to minimize mean squared error between the teacher and student, i.e. $$L(S, \hat{S}) = \frac{1}{T} \sum_{t=1}^T \left(S(x^t) - \hat{S}(x^t)\right)^2$$ where $x^t$ is the t-$th$ sample in the audit data, $S(x^t)$ is the output of the teacher model (risk scores) for sample $x^t$, $\hat{S}(x^t)$ is the output of the mimic model for sample $x^t$, and $T$ is the number of samples. Throughout this paper, we will call the teacher model the *black-box model* and the student model the *mimic model*. Next, we leverage the ground-truth outcome information. We train *our own risk scoring model* on the audit data to predict the ground-truth outcome, i.e. $$\begin{aligned} L(O, \hat{O}) &= \frac{1}{T}\sum_{t=1}^T \Bigl\{O(x^t)\log\left(P(\hat{O}(x^t)=1)\right) + \nonumber \\ &\quad(1-O(x^t))\log\left(P(\hat{O}(x^t)=0)\right)\Bigr\}\end{aligned}$$ where $O(x^t) \in \{0,1\}$ is the ground-truth outcome for sample $x^t$ and $\hat{O}(x^t) \in \{0,1\}$ is the output of the model for sample $x^t$. Throughout this paper, we call this model the *outcome model*. Note that the outcome model is not a mimic model. ![Auditing a loan risk scoring model by training transparent models on data labeled with the risk scores and with ground-truth outcomes for loan defaults.[]{data-label="fig:distillation"}](approach.pdf){width="0.9\linewidth"} It is critical that both the mimic model and outcome model are trained using the same model class that allows for interpretation and comparison. Not all model classes have the property that two models of that class can be compared. E.g. it is not obvious how to compare two decision trees, random forests or neural nets. We want a model class that is as rich and complex as possible so that the mimic model can be faithful to the black-box model and the outcome model can accurately predict ground-truth outcomes. But this model class should still be transparent [@doshivelez2017towards] so that we can understand how the input features affect the models’ outputs—the goal of our audit. We focus on a particular transparent model class, iGAM (Section \[sec:modelclass\]) in this paper; other choices are possible. The risk score and the ground-truth outcome are closely related—the ground-truth outcome is what the black-box model was meant to predict. If the black-box model is accurate *and* generalizes to the audit data, it would predict the ground-truth outcomes in the audit data correctly; the converse is true if the black-box model is not accurate *or* does not generalize to the audit data. Because both the mimic and outcome models are trained with the same model class on the same audit data using the same features, the more faithful the mimic model, and the more accurate the outcome model, the more likely it is that differences observed between the mimic and outcome models result from differences between the ground-truth outcomes and risk scores assigned by the black-box risk scoring model. Moreover, similarities between the mimic and outcome models (e.g. on COMPAS in Section \[sec:COMPAS\], the Number of Priors feature is modeled very similarly by the two models) increases confidence that the mimic model is a faithful representation of the black-box model, and that any differences observed on other features are meaningful. **Related work.** Adler et al. [@Adler2016AuditingBM] trained a model to predict outcomes and then a second model to predict the first’s predictions. This is a different distillation setup from ours, as we use both risk scores and outcomes. Adebayo and Kagal [@Adebayo2016Iterative] also learned their own risk scoring models when the black-box model cannot be queried. Some papers also compare two models, but not risk scores and outcomes at the same time. Wang et al. [@wang2018direction] trained a model to predict outcomes and another to predict membership in a protected subgroup. Chouldechova and G’Sell [@Chouldechova2017Fairer] trained two different outcome models then identified subgroups where the two models differed. Other papers work on a single outcome or risk score model [@Zhang2017Predictivebias; @tramer2017fairtest]. Testing for Missing Features {#sec:correlation} ---------------------------- If the audit data is missing features used by the black-box model, the audit data alone may be insufficient to audit the black-box model. We propose a statistical test to test if there are important missing features based on the following observation: > *If the black-box model used features that are missing from the audit data but are useful for predicting the ground-truth outcome, the error between the mimic model (learned on the audit data) and the risk score, $||\hat{S}-S||_E$, should be positively correlated with the error between the outcome model (learned on the audit data) and ground-truth outcome, $||\hat{O}-O||_E$.* [90]{} $\quad$ [Feature Contribution, $h_i(x_i)$]{} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} [90]{} $\quad$ [Feature Contribution, $h_i(x_i)$]{} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} [90]{} $\quad$ [Feature Contribution, $h_i(x_i)$]{} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} [90]{} $\quad$ [Feature Contribution, $h_i(x_i)$]{} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} {width="23.50000%" height="0.8in"} Since the test involves predictions from the mimic and outcome models, both models need to be trained (Sections \[sec:distillation\]) prior to performing this test; this test merely checks if the results of our audit are significantly affected by missing features. We perform this test on all the risk scoring models we audit in this paper in Section \[sec:missing\]. Independent of our audit approach, this test can standalone as a test for whether a data set is missing key features that were used to train a black-box model. Transparency of the mimic and outcome models is not a requirement for this test. Comparing Mimic and Outcome Models {#sec:compare} ---------------------------------- In this section, we provide technical details on how to train the mimic and outcome models so that they are comparable, and detect statistically significant differences. ### Choice of model class {#sec:modelclass} As noted in Section \[sec:distillation\], we train the mimic model and outcome model using the same transparent model class—in this paper, iGAM [@lou2012intelligible; @lou2013accurate; @caruana2015intelligible]. We point the reader to [@lou2012intelligible; @lou2013accurate; @caruana2015intelligible] to learn more about iGAMs and to [@tan2018transparent] for a distillation example where it was used as a student. Briefly, iGAM has the form $$\label{eq:ga2m} g(y) = h_0 + \sum_i h_i(x_i) + \sum_{i \neq j} h_{ij}(x_i,x_j)$$ where the contribution of any one feature $x_i$ or pair of features $x_i$ and $x_j$ to the prediction can be visualized in graphs such as Figure \[fig:police\_feats\] that plot $x_i$ on the x-axis and $h_i(x_i)$ on the y-axis. For classification, $g$ is the logistic function. For regression, $g$ is the identity function. For classical GAMS [@hastie1990generalized], feature contributions $h(\cdot)$ are fitted using splines; for iGAM, they are fitted using ensembles of short trees. Crucially, since iGAM is an additive model, two iGAM models can be compared by simply taking a difference of their feature contributions $h(\cdot)$, which we exploit in Section \[sec:differences\] to detect differences between the mimic and outcome models. ### Calibrating model inputs {#sec:calibration} Calibration is the process of matching predicted and empirical probabilities [@degroot1983reliability; @niculescu05predicting]. If a risk score is well-calibrated, the relationship between the risk score and empirical probabilities will be linear (e.g. COMPAS and Stop-and-Frisk in the top row of Figure \[fig:reliability\] in the Appendix). While developing our approach, we discovered that not all risk scores exhibit the desired linear relationship with outcomes in the audit data. For example, the Chicago Police risk score (third column of Figure \[fig:reliability\] in the Appendix) is rather flat for risk scores less than 350, then exhibits a sharp kink upwards. One possible explanation for any nonlinear relationship is that the risk score was well-calibrated on its original training data, but the audit data has a different distribution (data shift) [@riley2016external]. Another explanation is post-processing by risk scoring model creators to reduce sensitivity in less important parts of the risk score scale and enhance separation in more important parts of the scale [@lingo2008discriminatory]. We make the reasonable assumption that risk scores should be monotonic and well-calibrated [@lingo2008discriminatory] and use this assumption to undo distortions or audit data shift. Specifically, we learn a nonlinear transformation of the risk score (the blue line in Figure \[fig:reliability\] in the Appendix), similar to isotonic regression [@niculescu05predicting], to make the risk scores and outcomes linearly related on a scale of choice. The mimic model is then trained on these transformed risk scores. This calibration step is necessary to compare the mimic and outcome models, as it makes the targets for the two models (risk scores and outcomes) linearly related on a scale that their outputs will later be compared on. We select this scale to be logit probabilities (since the outcome model’s outputs are already on this scale), and perform this calibration step for Chicago Police and Lending Club but not COMPAS and Stop-and-Frisk, since the latter two already exhibit the desired linear relationships. ### Detecting differences {#sec:differences} To not mistake random noise for real differences between the mimic and outcome models, we control potential sources of noise during the training process. To avoid data sample-specific effects, we train the mimic and outcome models on the same data sample. We then calculate the difference in feature $x_i$’s contribution to the models, $sh_i(x_i) - oh_i(x_i)$. If this number is positive, the mimic model assigns more risk than the outcome model for feature $x_i$; the converse is true if this number is negative. We construct a confidence interval for this difference to tell if it is statistically significant. One ancillary contribution of this paper is a new method to estimate confidence intervals for the iGAM model class, by employing a [*bootstrap-of-little-bags*]{} approach [@sexton2009standard] to estimate the variance of $h_i(x_i)$ and $sh_i(x_i)-oh_i(x_i)$. See Appendix \[sec:variance\] for details. The resulting confidence intervals are the dotted lines in Figures \[fig:police\_feats\], \[fig:police\_not\_feats\], and \[fig:recid\_feats\]. Results {#sec:results} ======= Validating the Audit Approach {#sec:validating} ----------------------------- [ccccc]{} [90]{} $\quad$ {width="23.50000%" height="1.6in"} & {width="23.50000%" height="1.6in"} & {width="23.50000%" height="1.6in"} & {width="23.50000%" height="1.6in"} & In this section, we demonstrate the approach on risk scoring models where we have some information on how they were trained, and check that our approach can recover this information. ### Stop-and-Frisk. {#sec:stopfrisk} Using the New York Police Department’s Stop-and-Frisk data[^1], Goel et al. [@goel2016precinct] proposed a simple risk scoring model for weapon possession: $S = 3 \times \mathbbm{1}_{PS} + 1 \times \mathbbm{1}_{AS} + 1 \times \mathbbm{1}_{Bulge}$, where $S$ is the risk score, $PS$ denotes primary stop circumstance is presence of suspicious object, $AS$ denotes secondary stop circumstance is sight or sound of criminal activity, and $Bulge$ denotes bulge in clothing [@goel2016precinct]. Since we know the model’s functional form, we can verify that the mimic model correctly recovers its coefficients. We apply the risk scoring model to label 2012 data (T=126,457, p=40) after following Goel et al.’s data pre-processing steps [@goel2016precinct]. **Result.** The mimic model recovers the coefficients of (3, 1, 1) for the three features used in the risk scoring model ($PS$, $AS$, $Bulge$) and 0 for the remaining features. ### Chicago Police “Strategic Subject”. The Chicago Police Department released arrest data[^2] from 2012 to 2016 that was used to create a risk score for the probability of an individual being involved in a shooting incident as a victim or offender. This data set contains 16 features, but only 8 are used by the black-box model, which gives us an opportunity to test if our approach can accurately detect which features are and are not used by a black-box model. We trained a mimic model, intentionally including all 16 features. Figure \[fig:police\_feats\] shows the feature contributions of the mimic model (in red) and outcome model (in green) for the 8 features the Chicago Police says were used by the black-box model; Figure \[fig:police\_not\_feats\] shows the 8 features the Chicago Police says were *not used* in their model. **Result.** There is a striking difference between Figure \[fig:police\_feats\] and Figure \[fig:police\_not\_feats\]: the mimic model (in red) assigns importance to the features in Figure \[fig:police\_feats\], but does not assign any importance to the features in Figure \[fig:police\_not\_feats\]. This agrees with Chicago Police’s statement about which features were and were not used in the black-box model. We also note that the outcome model (in green) does assign importance to the unused features (Figure \[fig:police\_not\_feats\]), suggesting that there is signal available in the 8 unused features that the Chicago Police risk scoring model could have used, but chose not to use. Race and sex are 2 of these 8 features, which the Chicago Police especially emphasize are not used. These experiments confirm that the mimic models can provide insights into the black-box models, and demonstrate the advantages of using the outcome information available in the audit data. Auditing COMPAS {#sec:COMPAS} --------------- COMPAS, a proprietary score developed to predict recidivism risk, has been the subject of scrutiny for racial bias [@propublica_story; @kleinberg2016inherent; @chouldechova2017fair; @corbett2017algorithmic; @blomberg2010validation; @dieterich2016COMPAS]. We do not know what model class, input features or data were used to train COMPAS. As described in Section \[sec:introduction\], the COMPAS audit data[^3] was collected by ProPublica; it is likely different from the original COMPAS training data. Figure \[fig:recid\_feats\] compares the COMPAS mimic model (in red) and outcome model (in green) for four features: Age, Race, Number of Priors, and Gender. The dotted lines are 95% pointwise confidence intervals. We observe the following: **COMPAS agrees with ground-truth outcomes regarding the number of priors.** In the 3rd plot in Figure \[fig:recid\_feats\], the mimic model and outcome model agree on the impact of Number of Priors on risk; their confidence intervals overlap through most of its range. **COMPAS disagrees with ground-truth outcomes for some age and race groups.** The 1st and 2nd plots in Figure \[fig:recid\_feats\] show the effect of Age and Race on the mimic and outcome models. The mimic model (red) and the outcome model (green) are very similar between ages 20 to 70, and their confidence intervals overlap. For ages greater than 70, there is evidence that the models disagree as the confidence intervals do not overlap. The mimic and outcome models are also different for ages 18 and 19: the mimic model predicts low risk for young individuals, but we see no evidence to support this in the outcome model, with risk appearing to be highest for young individuals. The mimic model predicts that African Americans are even higher risk, and Caucasians lower risk, than the ground-truth outcomes suggest is warranted. Note that the ground-truth outcomes might themselves be biased due to historical discrimination against African Americans. **Gender has opposite effects on COMPAS compared to ground-truth outcome.** In the 4th plot in Figure \[fig:recid\_feats\], we see a discrepancy between the mimic model and outcome model on Gender. The mimic model predicts higher risk than warranted by ground-truth outcomes for females, and conversely for males. **Explaining differences between mimic and outcome models.** Each observed difference between the two models could have several different explanations. We propose to leverage the observed differences to gain insight into the black-box model, by asking the question, “what could be happening in the black-box model, that could explain the differences we are seeing”? We provide several examples below. [ccccc]{} [90]{} $\quad$ & &\ 1. Some feature regions may be underrepresented in the black-box model’s training data and/or the audit data. For example, in the COMPAS audit data (collected by ProPublica), only 3% of samples are between 18 and 20 years old, only 0.5% are older than 70 years old, and only 19% are female, which makes learning accurate models in these regions harder. 2. The black-box model may have a more simple or complex form than the outcome model. For example, we saw that the mimic model predicts low risk for young individuals, but there is no evidence to support this in the outcome model. We trained an iGAM model with interactions between pairs of features, and observed strong interactions between very young age and other variables such as Gender, Charge Degree, and Length of Stay. If COMPAS does not model interactions well, this may explain why COMPAS needs to predict low risk for very young individuals (because it cannot otherwise predict a reduced risk via interactions of age with other variables). 3. The black-box model may be deliberately simple for some feature regions. For example, for ages greater than 70, the outcome model has much wider confidence intervals than the mimic model. The ground-truth outcomes are potentially high-variance in this region, yet the black-box model’s scoring function may have been kept deliberately simple for extreme feature values like this. 4. The black-box model may be using additional features missing from the audit data, that interact with the non-missing features. We provide a statistical test to determine the likelihood of this (Section \[sec:missing\]). and so on. We cannot tell (without further testing) the definitive reason that explains a particular difference between the mimic and outcome models. However, we suggest that this form of reasoning coupled with the use of transparent models surfaces differences that we did not *a priori* know, but can then proceed to check and reason about, to gain insight into the black-box model. Auditing Lending Club --------------------- Lending Club, an online peer-to-peer lending company, rates loans it finances and releases data[^4] on these loans. We use a subset of five years (2007-2011) of loans that have matured, so that we have ground-truth on whether the loan defaulted. We do not know what model class or input features Lending Club used to train their risk scoring model. We believe the data sample we have is similar to the data they would have used to train their models. According to Lending Club, their models are refreshed periodically. We use this Lending Club example to discuss an insight we gained into the black-box model from examining interactions revealed by our transparent models. Figure \[fig:loan\_feats\] shows the interaction of loan issue year and home ownership in the Lending Club risk scoring mimic model (in red) and ground-truth outcome model (in green). Having a home mortgage in 2007-2008 increases the loan default risk more than having a home mortgage in 2009 and beyond. Recall that 2007-2008 is around the time of the subprime housing crisis. Note the difference in ranges between the two plots—the range goes up to 0.2 for the outcome model (in green) whereas the range is much lower for the mimic model (in red). This could indicate that the Lending Club risk scoring model is updated conservatively, with some lag time, instead of being rapidly updated as economic conditions and behavior change. Testing for Features Missing from\ Audit Data {#sec:missing} ---------------------------------- As black-box models may use additional features we do not have access to, we developed a test (Section \[sec:correlation\]) to assess the impact missing features could have on our analysis. Table \[table: missingfeatures\] provides 95% confidence intervals for three linear and nonlinear measures of correlation used in the test. If zero is in the confidence interval, the error of the mimic model (trained on the audit data) is not correlated to the error of the outcome model (also trained on the audit data). Then, it is unlikely that the audit data is missing key feature(s) that are a) predictive of outcomes (and hence will negatively affect the error of the outcome model if missing); and b) used in the black-box model (and hence will negatively affect the error of the mimic model if missing). **Risk Score** **Pearson** $\rho$ **Spearman** $\rho$ **Kendall** $\tau$ ---------------- -------------------- --------------------- -------------------- COMPAS \[0.10, 0.13\] \[0.10, 0.14\] \[0.08, 0.10\] Lending Club \[0.00, 0.03\] \[-0.01, 0.01\] \[-0.01, 0.01\] Stop-and-Frisk \[0.00, 0.01\] \[-0.03, 0.01\] \[-0.02, 0.01\] Chicago Police \[0.00, 0.01\] \[0.01, 0.03\] \[0.01, 0.02\] : Statistical test for likelihood of audit data missing key features used by black-box model.[]{data-label="table: missingfeatures"} In Lending Club and Stop-and-Frisk we cannot distinguish these correlations from zero, suggesting that no key features are missing from the audit data. For Chicago Police, the confidence intervals contain 0 or are very close to 0 (lower limit 0.01), hence there is little evidence of the audit data missing key features. For COMPAS there is evidence of positive correlation, indicating that the ProPublica data may be missing key features used in the COMPAS model. This is supported by our findings in Section \[sec:fidelity\] that no mimic models trained on the ProPublica data, however powerful (e.g. random forests), could mimic COMPAS well. **Risk Score** **Metric** **Linear model** **iGAM** **iGAM w/ interactions** **Random Forest** -- ---------------- -------------- ------------------------------------- ------------------------------------ ------------------------------------ --------------------------- COMPAS RMSE (1-10) $2.11 \pm 0.057$ $2.01 \pm 0.045$ $\mathbf{2.00 \pm 0.047}$ $2.02 \pm 0.053$ Lending Club RMSE (2-36) $3.27 \pm 0.037$ $2.60 \pm 0.049$ $2.52 \pm 0.051$ $\mathbf{2.48 \pm 0.033}$ Chicago Police RMSE (0-500) $17.4 \pm 0.102$ $17.2 \pm 0.125$ $16.5 \pm 0.130$ $\mathbf{14.0 \pm 0.280}$ Stop-and-Frisk RMSE (0-5) $\mathbf{0.00\pm 2\times10^{-15}}$ $\mathbf{0.00 \pm 1\times10^{-5}}$ $\mathbf{0.00 \pm 2\times10^{-5}}$ $0.01 \pm 2\times10^{-3}$ COMPAS AUC $0.73 \pm 0.029$ $0.74 \pm 0.027$ $\mathbf{0.75 \pm 0.029}$ $0.73 \pm 0.026$ Lending Club AUC $0.69 \pm 0.006$ $\mathbf{0.69 \pm 0.016}$ $\mathbf{0.69 \pm 0.014}$ $0.68 \pm 0.020$ Chicago Police AUC $\mathbf{0.95 \pm 0.007}$ $\mathbf{0.95 \pm 0.007}$ $\mathbf{0.95 \pm 0.007}$ $0.929 \pm 0.009$ Stop-and-Frisk AUC $0.84 \pm 0.020$ $0.85 \pm 0.020$ $0.85 \pm 0.020$ $\mathbf{0.87 \pm 0.024}$ Fidelity and Accuracy {#sec:fidelity} --------------------- To quantitatively evaluate our audit approach, we report fidelity (how well the mimic model predicts the black-box model’s risk scores, measured in RMSE) and accuracy (how well the outcome model predicts the ground-truth outcomes, measured in AUC) for all the risk scoring models we audit in Table \[table: studenterror\]. For comparison, we also train linear models (a simpler model class than iGAM) and random forests (more complex, but less interpretable). For COMPAS, all model classes (linear model, iGAM, random forest) have roughly equal fidelity and accuracy. Interestingly, none obtained RMSE lower than 2 on a 1-10 scale. Comparing outcome model AUCs across different model classes, iGAM’s results are generally comparable to (or slightly better than) more complex random forests (Table \[table: studenterror\]). For the risk score mimic models, random forests are competitive for Lending Club and Chicago Police. Linear mimic models are not far behind iGAMs for several risk scoring models (COMPAS, Chicago Police, Stop-and-Frisk), suggesting that the black-box model’s functional form might be very simple. We know this to be true for Stop-and-Frisk from Section \[sec:stopfrisk\] where the model was a simple linear model. **COMPAS deep dive.** One possible reason why COMPAS is challenging to mimic may be that the ProPublica data set is missing key features. This agrees with the results of the statistical test in Section \[sec:missing\]. Another possible reason is the small sample size (less than 7,000 samples). One advantage of using a model distillation approach to inspect black-box models is that the approach may be able to benefit from additional unlabeled data if the black-box teacher model can be queried to label the additional data [@bucilua2006compression]. We found an additional 3,000 individuals in the ProPublica data with COMPAS risk scores *but no ground-truth outcomes*. Adding them to the training (not testing) data for the mimic model and retraining the mimic model, we find marginal improvement in the mimic model’s fidelity (from RMSE 2.0 to 1.98). Doing the opposite—removing individuals from the training data in 1,000 increments—decreased the mimic model’s fidelity only marginally (to RMSE 2.1, training on only 1,000 individuals). These analyses suggest that for COMPAS, missing key features is a more pressing issue than insufficient data. Discussion ========== Sometimes we are interested in detecting bias on features intentionally excluded from the black-box model. For example, a credit risk scoring model is probably not allowed to use race as an input. Unfortunately, not using race does not prevent the model from learning to be biased. Racial bias in a data set is likely to be in the outcomes — the labels used for learning; not using race as an input *feature* does not remove the bias from the *labels*. If race were uncorrelated with all other features (and combinations of features) provided to the model, then removing race would prevent the model from learning to be racially biased because it would not have any input features on which to model this bias. Unfortunately, in any large, real-world data set, there is massive correlation among the high-dimensional features, and a model trained to predict credit risk will learn to be biased from correlation of the *excluded* race feature with other features that likely remain in the model (e.g., income or education). Hence, removing a protected feature like race or gender does not prevent a model from learning to be biased. Instead, removing protected features make it harder to detect how the model is biased, or correct the bias, because the bias is now spread in a complex way among all the correlated features throughout the model instead of being localized to the protected features. The main benefit of excluding protected features like race or gender from the inputs of a machine learning model is that it allows the group deploying the model to claim (incorrectly) that their model is not biased because it did not use these features. When training a transparent student model to mimic a black-box model, we intentionally include all features—even protected features like race and gender—specifically because we are interested in seeing what the models [*could*]{} learn from them. If, when the mimic model mimics the black-box model, it does not see any signal on the race or gender features and learns to model them as flat (zero) functions, this indicates whether the teacher model (the black-box model) did or did not use these features, but also if the teacher model exhibits race or gender bias even if the model did not use race or gender as inputs. Conclusion {#sec:conclusion} ========== Our Distill-and-Compare approach to auditing black-box models was motivated by a realistic setting where access to the black-box model API is not available; only a data set labeled with the risk score as produced by the risk scoring model and the ground-truth outcome is available. The efficacy of the Distill-and-Compare audit approach increases when a model class that can be highly faithful to the black-box model and highly accurate at predicting the ground-truth outcomes is used, and when the audit data is not missing key features used in the black-box model. A key advantage of using transparent models to audit black-box models is that we do not need to know in advance what to look for. Many current auditing approaches focus on one or two protected features selected in advance, and thus are less likely to detect biases that are not *a priori* known. The Distill-and-Compare audit approach, using transparent models, can hence be most useful for complicated, real-world data with unknown sources of biases. A new confidence interval estimate for iGAM {#sec:variance} =========================================== It is not trivial to estimate confidence intervals for nonparametric learners such as trees [@mentch2016quantifying]; iGAM’s base learners are short trees. We employ a [*bootstrap-of-little-bags*]{} approach originally developed for bagged models in [@sexton2009standard] to estimate the variance of feature $x_i$’s contribution to the model, $h_i(x_i)$, and difference in feature $x_i$’s contribution to the mimic and outcome models, $sh_i(x_i) - oh_i(x_i)$. Bootstrap-of-little-bags exploits two-level structured\ cross-validation (e.g. 15% of data points are selected for the test set, with the remaining 85% split into training (70%) and validation (15%) sets). Repeating this inner splitting $L$ times and outer splitting $K$ times gives a total of $KL$ bags on which we train the mimic model and outcome model. Let $h^{lk}_i (x_i)$ be $x_i$’s feature contribution estimated by the model in the $l$th inner and $k$th outer fold. Its variance can then be estimated as $$\widehat{\mbox{Var}}(h_i(x_i)) = \frac{1}{K} \sum_{k=1}^K \left( \frac{1}{L} \sum_{l=1}^L h^{kl}_i(x_i) - \frac{1}{KL} \sum_{l=1}^l \sum_{k=1}^K h^{kl}_i(x_i) \right)^2$$ and its mean $\overline{h_i(x_i)}$ is the average of the $KL$ models. We can now construct pointwise confidence intervals for feature contributions to iGAM models. The confidence interval for feature $x_i$’s contribution to the model, $h_i(x_i)$, is $\overline{h_i(x_i)} \pm 1.96 \sqrt{ \widehat{\mbox{Var}}(h_i(x_i))}$ and the confidence interval for the difference in feature $x_i$’s contribution to the mimic and outcome models, $sh_i(x_i)-oh_i(x_i)$, is $\overline{sh_i(x_i) - oh_i(x_i)} \pm 1.96 \sqrt{\widehat{\mbox{Var}}(sh_i(x_i)) + \widehat{\mbox{Var}}(oh_i(x_i))- 2\widehat{\mbox{Cov}}(sh_i(x_i),oh_i(x_i))}$ with $\widehat{\mbox{Cov}}(sh_i(x_i),oh_i(x_i))$ also estimated using bootstrap-of-little-bags. This variance estimate is conservative (meaning it overestimates true variability), however, given that we are trying to detect differences, overestimating means we are less likely to mistake random noise for real differences. For large $K$ and $L$, consistency of this estimate was established in [@wager2017grf]. Note that are pointwise, not uniform, confidence intervals. That is, they capture the variability of the effect of age at age=50, not the entire effect of age. Uniform confidence intervals can be constructed by adjusting the critical value of the confidence interval. Calibration Plots ================= [ccccc]{} [90]{} {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"}\ [90]{} {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"}\ [90]{} {width="23.00000%" height="0.85in"} & {width="23.00000%" height="0.85in"} & {width="23.00000%" height="0.85in"} & {width="23.00000%" height="0.85in"} [ccc]{} [90]{} {width="23.00000%" height="1.4in"} & {width="23.00000%" height="1.4in"} [^1]: http://www1.nyc.gov/site/nypd/stats/reports-analysis/stopfrisk.page [^2]: <https://data.cityofchicago.org/Public-Safety/Strategic-Subject-List/4aki-r3np> [^3]: <https://github.com/propublica/COMPAS-analysis> [^4]: <https://www.lendingclub.com/info/download-data.action>

--- address: | Institut für Theoretische Physik, Universität Heidelberg\ Philosophenweg 16, 69120 Heidelberg, Germany\ E-mail: c.ewerz@thphys.uni-heidelberg.de author: - 'C. EWERZ' title: 'Gribov’s Light Quark Confinement Scenario' --- Gribov’s Confinement Scenario ============================= An interesting physical mechanism for explaining the phenomenon of confinement was proposed by V.N.  Gribov in Refs. . This mechanism is based on the idea the colour charges in QCD are supercritical. The analogous phenomenon is well known in QED where a pointlike charge $Z>137$ is unstable and captures an electron from the vacuum while a positron is emitted. In QCD a similar phenomenon can occur due to the existence of very light quarks. If already a single quark has a supercritical colour charge this leads to a dramatic change in the vacuum structure of light quarks (see Refs. ). The properties of the new vacuum structure are such that a single quark is unstable and can be observed as a resonance but not as an asymptotic state. Interestingly, confinement arises in this picture due to the Pauli principle rather than due to very strong forces. The Gribov-Dyson-Schwinger Equation for Light Quarks ==================================================== A new approximation scheme for the treatment of the exact Dyson–Schwinger equation for quarks was suggested in Ref.  . This scheme is motivated by the physical picture of supercritical charges, but is also interesting in its own right. One uses Feynman gauge in which the Green function of the gluon has the form $D_{\mu\nu}(k) = - \alpha_s(k)g_{\mu\nu}/k^2$, and one assumes the effective strong coupling constant $\alpha_s(k)$ to be a slowly varying function that remains finite at low momenta — an assumption that is in agreement with recent results obtained in the dispersive approach to power corrections[@Dokshitzer:1995qm]. The choice of the Feynman gauge allows one to collect in a very elegant way those contributions to the Dyson–Schwinger equation which are most singular in the infrared. Clearly, it is this region in which the dynamics of chiral symmetry breaking and confinement mainly resides. The corresponding approximation scheme then leads to the Gribov–Dyson–Schwinger equation for the Green function $G$ of a light quark, G\^[-1]{}(q) = C\_F (\^G\^[-1]{}(q)) G(q) (\_G\^[-1]{}(q)) , \[gds\] where $C_F=4/3$. A very interesting observation is that the running coupling $\alpha_s(q)$ in this equation ensures that at large space–like momenta the solutions exactly reproduce the correct mass and wave function renormalization as it is known in perturbation theory. In principle one can also include subleading terms in the equation, i.e. terms which are less singular in the infrared. Chiral Symmetry Breaking and Critical Coupling ============================================== A detailed study of the Gribov–Dyson–Schwinger equation was performed in Ref. . It turns out that the resulting Green function shows a characteristically different behaviour if the effective strong coupling constant $\alpha_s(q)$ exceeds a critical value in some interval in the small momentum region. The critical coupling $\alpha_c$ can be determined analytically by considering asymptotic expansions of eq. (\[gds\]) and is \[critcoupl\] \_c = ( 1- ) 0.43 . The change in the behaviour of the Green function corresponds to chiral symmetry breaking. It is best seen in the behaviour of the dynamical mass function $M(q^2)$ of the quark at space–like momenta, $q^2<0$. Let us define a ‘perturbative’ mass $m_P=M(\lambda^2)$ at some large momentum scale $\lambda$ where perturbation theory holds. Further we define a ‘renormalized’ mass $m_R=M(0)$ as the low–momentum limit of the dynamical mass function. If the coupling $\alpha_s$ remains subcritical in the infrared the relation between $m_P$ and $m_R$ is monotonic, and $m_R$ vanishes for vanishing $m_P$. If the coupling becomes supercritical in the infrared, $\alpha> \alpha_c$, that behaviour changes dramatically. The relation between $m_P$ and $m_R$ for this case is shown in fig. \[fig:m2m\]. One sees that the renormalized mass remains finite even for vanishing perturbative mass, clearly signalling chiral symmetry breaking. Outlook ======= One expects that the confinement of quarks is reflected in the analytic structure of their Green function. The analytic structure resulting from the Gribov–Dyson–Schwinger equation (\[gds\]) has been shown[@Ewerz:2000qb] to exhibit poles and cuts at positive $q^2$, and hence does not correspond to confined quarks. This result had been anticipated[@Gribov:1998kb] because this equation does not take into account the back-reaction of the pions which are generated as Goldstone bosons in the process of chiral symmetry breaking. This back-reaction is expected to be crucial for the emergence of confinement, and there are indications that its inclusion in the equation in fact induces an analytic behaviour of the Green function corresponding to confined quarks[@Gribov:1999ui]. [0]{} V. N. Gribov, Eur. Phys. J. C [**10**]{} (1999) 71 \[arXiv:hep-ph/9807224\]. V. N. Gribov, Eur. Phys. J. C [**10**]{} (1999) 91 \[arXiv:hep-ph/9902279\]. C. Ewerz, Eur. Phys. J. C [**13**]{} (2000) 503 \[arXiv:hep-ph/0001038\]. Y. L. Dokshitzer, G. Marchesini and B. R. Webber, Nucl. Phys. B [**469**]{} (1996) 93 \[arXiv:hep-ph/9512336\].

Manganese oxides with the general composition R$_{1-x}$A$_x$MnO${_3}$ (where R and A are rare- and alkaline-earth ions, respectively) have attracted considerable attention because of their unusual magnetic and electronic properties. In some of these materials, metal-insulator transitions can be observed where both conductivity and magnetization change markedly. The $x = 0$ and $x = 1$ end members of the R$_{1-x}$A$_x$MnO${_3}$ family are insulating and antiferromagnetic (AF) with the Mn-ion in the Mn$^{3+}$ and Mn$^{4+}$ state, respectively. For intermediate $x$, the average Mn-valence is non-integer and the material is generally metallic or semiconducting. Most of the perovskite manganites show a ferromagnetic (FM) groundstate when the holes are optimally doped (usually $0.2<x<0.5$) and anisotropic antiferromagnetic (AFM) phases for $x>0.5$. The half-doped manganites, with $x=\frac{1}{2}$, are very particular. Magnetically these systems form FM zig-zag chains that are coupled AFM (see Fig.\[fig:ce\_phase\]) at low temperatures, the so-called magnetic CE-phase [@Wohlan55]. The groundstate is, moreover, an orbital-ordered and charge-ordered insulator. This behavior is generic and is experimentally observed in Nd$_{1/2}$Sr$_{1/2}$MnO$_3$ [@Kuwahara95; @Kawano97], Pr$_{1/2}$Sr$_{1/2}$MnO$_3$  [@Tomioka95], Pr$_{1/2}$Ca$_{1/2}$MnO$_3$ [@Tomioka96], La$_{1/2}$Ca$_{1/2}$MnO$_3$  [@Urushibara95; @Mori98], Nd$_{1/2}$Ca$_{1/2}$MnO$_3$ [@Kimura_prep] and in the half-doped layered manganite La$_{1/2}$Sr$_{3/2}$MnO$_4$ [@Moritomo95]. The insulating charge-ordered state can be transformed into a metallic FM state by application of a external magnetic field, a transition that is accompanied by a change in resistivity of several orders of magnitude [@Kuwahara95; @Tokunaga98]. The occurrence of charge order, orbital order and large magneto-resistance in the half-doped manganites is experimentally well established. Theoretically, however, the nature of the charge ordering at $x=\frac{1}{2}$ and the origin of the unconventional zig-zag magnetic structure remain unclear. We address these points here and show that $(i)$ the insulating CE-phase results from a particular ordering of orbitals, $(ii)$ this state is stable only in a narrow concentration range around the commensurate value and that $(iii)$ the zig-zag chains are intrinsically charge ordered due to on-site Hubbard correlations. The competition between kinetic energy (double-exchange) and superexchange between the manganese core-spins, combined with the orbital degeneracy leads to the formation of the magnetic CE-phase, for the same reasons as it leads to the anisotropic magnetic phases in the highly doped manganites with $x>\frac{1}{2}$ [@Brink99]. The CE-phase is found to be stable close to $x=\frac{1}{2}$; for $x>0.57$ the C-phase is stable and for $x<\frac{1}{2}$ we find phase separation between the CE-phase and FM-phase. As in the double-exchange framework electrons can only hop between sites with FM aligned core-spins, in the CE-phase only hopping processes within the zig-zag chains are possible, rendering the system one-dimensional for low-energy charge fluctuations. The essential observation is that an electron that passes a corner-site of the zig-zag chain, acquires a phase that depends on the orbital through which it passes. This leads to an effective dimerization that splits the bands and opens a gap at the Fermi-surface. The gap is very robust as it is a consequence of the staggered phase-factor that is itself fully determined by the topology of the system. At the same time not all orbitals are fully occupied, leading to an orbital-polarized groundstate. Our most surprising observation is that the experimentally observed charge order is directly obtained from the degenerate double-exchange model when the Coulomb interaction (the Hubbard $U$) between electrons in different orbitals, but [*on the same site*]{} is included. This can be understood from the fact that in the band-picture on the corner-sites both orthogonal orbitals are partially occupied, but on the bridge-site only one orbital is partially filled. The on-site Coulomb interaction acts therefore differently on the corner- and bridge-sites: charge is pushed away from the effectively correlated corner-sites to the effectively uncorrelated bridge-sites. The theoretical explanation for the ferromagnetic metallic state in doped manganites was already developed in 50’s and 60’s [@Zener51; @Anderson55; @DeGennes60]. The number of $d$ electrons per manganese site is $4-x$. Three electrons occupy the low lying $t_{2g}$ orbitals, having parallel spins. The other electrons occupy the $e_g$ orbitals with their spin parallel to the $t_{2g}$ spin because of the large ferromagnetic Hund’s rule exchange $J_H$. In the double-exchange framework electrons can only hop from one Mn-site to a neighboring one, when the $t_{2g}$ spins of these sites are aligned ferromagnetically because otherwise, in the classical treatment of the $t_{2g}$ core-spin the electron would have to overcome an energy barrier proportional to $J_H$. On the other hand, the conventional superexchange favors AF alignment the spins. This leads to an intricate competition between the kinetic energy and the superexchange [@DeGennes60] that is amplified by the two-fold degeneracy of the $e_g$ levels in a cubic crystal. As a basis for the $e_g$ wavefunctions we take the $x^2-y^2$ ($| \chi \rangle$) and $3z^2-r^2$ ($| \zeta \rangle$) orbitals. The hopping of the electrons between neighboring Mn-sites depends strongly on the kind of the orbitals involved and on the direction of the bond. The Hamiltonian for the kinetic energy is $$H_t= t \Sigma_{\alpha,\beta, \langle ij \rangle_{\Gamma}} \ \ M^{\Gamma}_{\alpha,\beta} \ c^+_{i,\alpha} c_{j,\beta},$$ where $c^+_{i,\alpha}$ creates an electron on site $i$ in orbital $\alpha$ and $\langle ij \rangle_{\Gamma}$ denotes a nearest neighbor pair along the $\Gamma$-direction ($\Gamma=x,y,z$) and $t$ is the hopping integral. The hopping is constrained to nearest neighbor Mn-ions that have FM aligned $t_{2g}$ core-spins (canted phases are not important in the doping range we consider here.) The matrices $M^{x}_{\alpha,\beta}$ and $M^{y}_{\alpha,\beta}$ can be found by applying the cubic symmetry operations on $M^{z}_{\alpha,\beta}$, where $M^{z}_{| \zeta \rangle ,| \zeta \rangle }=1$ and all other matrix elements are zero. This means physically that along the $z-$direction, for instance, electrons can only hop between $| \zeta \rangle $ orbitals and that all other hoppings (involving at least one $| \chi \rangle $ orbital) are zero by symmetry, whereas in the $x-$ and $y-$direction all hoppings are allowed (and related by symmetry). The strong spatial anisotropy of the hopping, combined with the competing kinetic and superexchange energy $J$ can, depending on the $e_g$ bandwidth and $J$, give rise to low dimensional spin-structures that, for instance, optimize kinetic energy by forming FM chains and optimize magnetic energy by a AFM coupling of these chains (C-phase), or AFM coupled FM planes (A-phase) [@Brink99]. =85.7mm In the homogenous FM state at $x=\frac{1}{2}$ the magnetic energy per site is $E_{mag}^{FM}=3J$ and the kinetic energy per site $E_{kin}^{FM}=-t$, if the $t_{2g}$ spin is treated as a classical spin. Here the magnetic energy is found by determining the number of AFM bonds, and the kinetic energy from the filling of the bands, as is described in Ref. [@Brink99]. For AFM coupled FM chains, the C-phase, $E_{kin}^{C}+E_{mag}^{C}=-0.6366t+J$, so that the C-phase is stable with respect to the FM-phase for $J>0.1816t$. A third possibility is for the spins to form FM zig-zag chains that are coupled AFM, the CE-phase [@Wohlan55], see Fig.\[fig:ce\_phase\]. The CE-phase is always higher in energy than the C-phase, except close to $x=\frac{1}{2}$. For the half-doped system $E_{kin}^{CE}+E_{mag}^{CE}=-0.695t+J$. It is easily shown that the A-phase is higher in energy. Thus for the half-doped manganites the experimentally observed CE-phase is indeed lower in energy than the C-phase, and lower than the FM-phase when $J>0.1524t$. Let us address our main findings that the CE-phase is charge-ordered, orbital-ordered and insulating, and come back to the stability of the CE-phase at the end of the paper. As shown in Fig.\[fig:ce\_phase\] the CE-phase contains two geometrically inequivalent sites, so called bridge- and corner-sites. Note that our specific choice of basis-orbitals as shown in Fig.\[fig:ce\_phase\] is motivated by the convenience of this basis for the calculations. The expectation value of actual observables is, of course, independent from the choice of basis Wannier orbitals. The topology of the electron hopping integrals is shown in Fig.\[fig:hop\]. The crucial observation is that an electron that hops from one bridge-site to another bridge-site via a $| \chi \rangle $ corner-orbital obtains a phase-factor $-1$, while if the hopping takes place via a $| \zeta \rangle $ corner-orbital, the phase-factor is $+1$. The bands are obtained by the solution of a 3x3 matrix. There are two bands with energy $\epsilon_{\pm}= \pm t\sqrt{2 + \cos k}$, where $k$ is the wave vector, and two nondispersive bands at zero energy, see Fig.\[fig:hop\]. [@Solovyev_prep]. At $x=\frac{1}{2}$ the $\epsilon_{-}$ band is fully occupied, and all other bands are empty. The system is insulating as the occupied and empty bands are split by a gap $\Delta=t$. In the C- and CE-phase the magnetic energies are equal, but the opening of the gap at the Fermi-energy in the CE-phase lowers its energy if the bands are half filled. This mechanism is equivalent to the situation in the lattice-Peierls problem, where the opening of a gap stabilizes a groundstate with a lattice deformation. In the insulating state the average occupancy of the bridge-site $3x^2-r^2$ ($3y^2-r^2$) orbital is $\frac{1}{2}$ while the orthogonal bridge-orbital is empty. The average occupancy of the corner-site is also $\frac{1}{2}$, with the ratio between $| \chi \rangle $ and $| \zeta \rangle $ occupancy of $\sqrt{3}:1$. The system is therefore orbital-ordered, but charge is homogeneously distributed between corner and bridge-sites. This changes drastically when also the Coulomb interaction $U$ between electrons on the same site is taken into account. The Hamiltonian is $$H_U= U \Sigma_{\alpha < \beta,i} \ n_{i,\alpha} n_{i,\beta},$$ where $n_{i,\alpha}= c^+_{i,\alpha}c_{i,\alpha}$. For the $e_g$ electrons in the manganites $U \approx 10t$, so that the system is strongly correlated. The correlations, however, have a very different effect on the corner and bridge-sites. On the bridge-site, one orbital is always empty, so that the Coulomb repulsion is ineffective, whereas on the corner-sites both orbitals are partially occupied. The consequence is that charge is pushed away from the correlated corner-sites to the effectively uncorrelated bridge-sites, causing charge order. We treat the the full Hamiltonian $H=H_t+H_U$ for $x=\frac{1}{2}$ with three different methods: exact diagonalization (ED), in mean-field (MF) and with the Gutzwiller-projection (GP) [@Fulde91]. In the ED calculation we consider a ring of 12 and 14 sites (or 18 and 21 orbitals, respectively) with the topology as in Fig.\[fig:hop\] . In Fig.\[fig:co\] the charge-disproportionation $\delta$ as a function of $U/t$ is shown, where $\delta$ is defined as $\langle n_B-n_A \rangle$, where $\langle n_B \rangle$ ($\langle n_A \rangle$) is the expectation value for finding an electron on a bridge- (corner-) site. The results for the 12 and 14-site cluster differ by less than 2% and finite size effects are therefore very small. For $U/t=0$ also $\delta=0$, as explained above, and for finite Coulomb interaction $\delta$ monotonously increases with $U$. For $U/t=10$, $\delta=0.171$ and $\delta$ reaches its maximum value for $U \rightarrow \infty$. In a ED calculation for a 16-site cluster with $U \rightarrow \infty$, we find that $\delta^{ED}_{\infty}=0.185$. For small values of $U/t$ a MF treatment is expected to be very reliable because the charge fluctuations are gapped. We decouple the quartic term in Eq.2 and determine the charge densities in the orbitals self-consistently. The results for small $U/t$ are shown with a dashed line in Fig.\[fig:co\] and they agree very well with the ED results for $U<\Delta$. In the GP, valid for $U \rightarrow \infty$, we introduce constrained electrons on the correlated corner-site. For the $| \chi \rangle$ orbital we introduce $\bar{x}_i=(1-n_i^z) x_i$ and for the $| \zeta \rangle$ orbital $\bar{z}_i=(1-n_i^x) z_i$, where $x_i$ ($z_i$) denotes an electron operator on corner-site $i$, with corresponding density $n_i^x$ ($n_i^z$) and $\bar{x}_i$ ($\bar{z}_i$) a constrained electron operator. The Hamiltonian is $$\begin{aligned} H_{GP} &=& \Sigma_{i \in B} \ c^{\dag}_i \left[ t_1 (\bar{z}_{i+1} +\bar{z}_{i-1}) \right. \nonumber \\ &+& t_2 (\bar{x}_{i+1} - \bar{x}_{i-1}) \left. \right] +h.c.,\end{aligned}$$ which physically means that an electron can only hop to a $| \chi \rangle$-orbital on a corner-site when there is no electron in the $| \zeta \rangle$-orbital on that site and vice versa, so that a corner-site can never be occupied by more than one electron, as is required in the $U$-infinity limit. We decouple the quartic terms in $H_{GP}$ as $c^{\dagger}_{i+1}z_in^x_i \rightarrow c^{\dagger}_{i+1} z_i \langle n^x_i \rangle + \langle c^{\dagger}_{i+1} z_i \rangle n^x_i$. After a Fourier transform we find $$\begin{aligned} H_{GP}^{mf} &=& \Sigma_{k} \ \epsilon_x x^{\dag}_k x_k +\epsilon_z z^{\dag}_k z_k \nonumber \\ &+& \tilde{t}_{1k} (c^{\dag}_k z_k + z^{\dag}_k c_k)+ \tilde{t}_{2k} (c^{\dag}_k x_k + x^{\dag}_k c_k) ,\end{aligned}$$ with $\epsilon_x = -4 t_1 \Sigma_k \cos{k} \ \langle c^{\dag}_k z_k \rangle$, $\epsilon_z = 4 i t_2 \Sigma_k \sin{k} \ \langle c^{\dag}_k x_k \rangle$, $\tilde{t}_{1k} = 2 t_1 \langle 1-n^x \rangle \cos{k}$ and $\tilde{t}_{2k} = -2 i t_2 \langle 1-n^z \rangle \sin{k}$. We obtain the solution of this system of equations by iteration, and find the charge-disproportionation $\delta^{GP}_{\infty}=0.191$, which is represented by the dashed line in Fig.\[fig:co\] for large $U/t$, and agrees well with the ED results. The on-site Coulomb interaction causes a charge-surplus on the bridge-sites in the x-y plane. In the CE-structure the zig-zag chains are stacked AFM along the z-direction, which implies that above each bridge-site there is another bridge-site in the next plane. So we find that the charges actually accumulate on sheets formed by the bridge-sites along the z-direction. This is in remarkable agreement with experiment and at the same time excludes the possibility that the charge-order is driven by longer range Coulomb interactions because the Madelung-sum is always minimized for a rocksalt-type charge-order. Similar physics may also apply to the situation at $x>0.5$, e.g. for the stripe phases observed in Ref. [@Mori98]. Fig.\[fig:co\] shows that the charge disproportionation is strongly doping dependent. For $x>\frac{1}{2}$ the holes that are doped into the lower $\epsilon_-$-band efficiently suppress charge-order. In this doping range, however, the CE-phase becomes unstable with respect to the C-phase. In Fig.\[fig:hop\] is shown that the kinetic energy of the C-phase is lower for $x>0.57$. For $x<\frac{1}{2}$ the energy per site of the CE-phase, $E_{CE}$, is constant because the extra electrons are doped in the non-dispersive bands at zero energy, which causes a kink of $E_{CE}$ at $x=\frac{1}{2}$. For lower hole-doping (higher electron concentration) the homogeneous FM-phase is lower in energy, as is expected. The stability of the phases can be tested by performing a Maxwell-construction. Due to the kink in $E_{CE}$ at $x=\frac{1}{2}$, one finds that for $x<\frac{1}{2}$ the FM-phase and CE-phase coexist. The doping region where this phase separation occurs depends strongly on the actual value of the electron bandwidth and magnetic exchange. We conclude that there is a strong asymmetry for doping lower and higher than $\frac{1}{2}$, as phase separation between the CE- and FM-phases is only present on the lower doping side. Inclusion of the correlations beyond the mean-field approximation and longer range Coulomb interactions, which we did not consider in the calculation of total energies, may modify these results. On the basis of the ED results we believe, however, that $U$ has only a small effect on the total energies. A more relevant contribution to the total energy comes from lattice deformations, as it is obvious that in the Jahn-Teller distorted C- and, to a lesser extend, CE-phases [@JT_CE] there is a large gain in lattice energy with respect to the undistorted FM-phase. It can be shown that the lattice contributions to the groundstate energy can, to a large extend, be described by using effective values of $J$ [@Brink_tbp], so that our general conclusions about phase separation for $x<\frac{1}{2}$ and about the asymmetry of the phase diagram for $x<\frac{1}{2}$ and $x>\frac{1}{2}$ remain valid. Recently this phase separation into a FM metallic and a charge ordered AFM insulating phase was observed experimentally [@Babushkina99]. In summary, we have given an explanation for the charge order, orbital order and insulating state of the half-doped manganese oxides. The CE-phase with one-dimensional ferromagnetic zig-zag chains is stable for $x=\frac{1}{2}$ and due to staggered phase factors in the hopping, the chains are insulating and orbital-ordered. The striking feature of our model is that the strong Coulomb interaction between electrons on the same Mn-site leads to the experimentally observed charge ordering. In a magnetic field we expect the chains to be unstable with respect to the ferromagnetic metallic state. This might offer a likely explanation for the large magneto-resistance at the metal-insulator transition for the half-doped manganites. We thank I. Solovyev and P. Horsch for useful discussions. J.v.d.B. acknowledges with appreciation the support by the Alexander von Humboldt-Stiftung, Germany. This work was financially supported by the Nederlandse Stichting voor Fundamenteel Onderzoek der Materie (FOM) and by the European network OXSEN. [99]{} Permanent address: Kazan Physicotechnical Institute, 420029 Kazan, Russia. E.O. Wohlan and W.C. Koehler, Phys. Rev. [**100**]{}, 545 (1955). H. Kuwahara [*et al.*]{}, Science [**270**]{}, 961 (1995). H. Kawano [*et al.*]{}, , 4253 (1997). Y. Tomioka [*et al.*]{}, , 5108 (1995). Z. Jirac [*et al.*]{}, J. Magn. Magn. Mater. [**53**]{}, 153 (1985), Y. Tomioka [*et al.*]{}, , R1689 (1996). Y. Okimoto [*et al.*]{}, , 109 (1995), P. Schiffer [*et al., ibid.*]{} [**75**]{}, 3336 (1995), P. G. Radaelli [*et al., ibid.*]{} [**75**]{}, 4488 (1995), Y. Okimoto [*et al.*]{}, , 4206 (1997). S. Mori, C.H. Chen and S.-W. Cheong, Nature [**392**]{}, 473 (1998). T. Kimura [*et al.*]{}, unpublished. Y. Moritomo [*et al.*]{}, , 3297 (1995), B.J. Sternlieb [*et al.*]{}, , 2169 (1996), Y. Moritomo [*et al.*]{}, Nature [**380**]{}, 141 (1996). M. Tokunaga [*et al.*]{}, , 5259 (1998). J. van den Brink and D.I. Khomskii, , 1016 (1999). C. Zener, Phys. Rev. [**82**]{}, 403 (1951). P.W. Anderson and H. Hasegawa, Phys. Rev. [**100**]{}, 675 (1955). P.G. de Gennes, Phys. Rev. [**118**]{}, 141 (1960). This result was first obtained by I. Solovyev and K. Terakura, private communication and unpublished and also found by T. Hotta [*et al.*]{}, unpublished. See, for instance, P. Fulde, Electron Correlations in Molecules and Solids, Springer-Verlag (Berlin), 1991. J. van den Brink and G. Khaliullin, unpublished. Note that the orbital polarization on the corner-sites in the CE-phase is not complete. N.A. Babushkina [*et al.*]{}, , 6994 (1999), M. Urhara [*et al.*]{}, Nature [**399**]{}, 560 (1999).

--- author: - | [Rabin Banerjee]{}$^{a,b}$[^1],   [Sunandan Gangopadhyay]{}$^{c,d}$[^2],   [Shailesh Kulkarni]{}$^{a,e}$[^3]\ $^a$*S.N. Bose National Centre for Basic Sciences,*\ *JD Block, Sector III, Salt Lake, Kolkata-700098, India*\ $^b$ *Department of Physics, University of Helsinki*\ *and Helsinki Institute of Physics,P.O.Box 64, FIN-00014 Helsinki, Finland*\ $^c$ *Department of Physics and Astrophysics, West Bengal State University, Barasat, India*\ $^d$ *Visiting Associate in S.N. Bose National Centre for Basic Sciences, Kolkata*\ $^e$ *Harish-Chandra Research Institute, Chhatnag Road, Jhusi, Allahabad 211019, India* title: Aspects of Diffeomorphism and Conformal invariance in classical Liouville theory --- > The interplay between the diffeomorphism and conformal symmetries (a feature common in quantum field theories) is shown to be exhibited for the case of black holes in two dimensional classical Liouville theory. We show that although the theory is conformally invariant in the near horizon limit, there is a breaking of the diffeomorphism symmetry at the classical level. On the other hand, in the region away from the horizon, the conformal symmetry of the theory gets broken with the diffeomorphism symmetry remaining intact. It is a well known fact from general quantum field theoretic considerations [@polyakov] that in two dimensions there is an interplay between the diffeomorphism and the trace anomaly of the energy momentum tensor in the sense that it is possible to remove one of them but not both. This is because for $D=2$ case, there is no regularization which simultaneously preserves conformal as well as diffeomorphism symmetries [@polyakov; @green; @bertlman; @witten]. Interestingly, we find that such a situation arises in a classical Liouville theory [@liou] used to compute the black hole entropy owing to the presence of a nontrivial central charge in the theory [@solodukhin; @cvitan; @kang; @pinamonti; @pallua; @chen]. Further, Hawking effect was also studied by using boundary Liouville model [@solodukhin2]. In classical Liouville theory (which is relevant for black hole physics) the energy momentum tensor is covariantly conserved but has a nontrivial trace anomaly. However, it turns out that if we consider terms upto leading order in the expansion of the metric, near the horizon, the trace anomaly vanishes and hence the theory becomes conformally invariant (which explains its utility for studies on black hole physics), but the energy-momentum tensor fails to remain covariantly conserved. Hence, the interplay between the diffeomorphism and the trace anomaly of the energy momentum tensor gets exhibited at the classical level. To begin with, we consider a four-dimensional spherically symmetric metric of the form $$\begin{aligned} ds^2 = \gamma_{ab}(x^0, x^1) dx^{a} dx^{b} + r^{2}(x^0, x^1) d\Omega^{2} \label{0}\end{aligned}$$ where $\gamma_{ab}(x^0, x^1)$ is the metric on an effective $2$-d spacetime $M^2$ with coordinates $(x^{0}, x^{1})$. We now start from the four-dimensional Einstein-Hilbert action defined by $$S_{EH} = -\frac{1}{16\pi G} \int_{M^4} \ d^{4}x \sqrt{-g} R_{(4)} \label{0.1}.$$ Considering the above action on the class of spherically symmetric metrics (\[0\]), we obtain an effective two-dimensional theory described by the action $$S = - \int_{M^2} d^{2}x \sqrt{-\gamma} \left( \frac{1}{2}(\nabla\Phi)^2 + \frac{1}{4}\Phi^2 R_{(2)} + \frac{1}{2G}\right) \label{0.2}$$ where $\Phi = rG^{-1/2}$ and $R_{(2)}$ is the two-dimensional scalar curvature [^4]. This action represents dilaton gravity in two dimensions with $r$ playing the role of dilaton field. The above action can now be transformed to a form similar to that of the Liouville theory [@russo] $$S_{L} = - \int_{M^2} d^{2}x \sqrt{-\bar\gamma} \left(\frac{1}{2} (\bar\nabla \phi)^2 + \frac{1}{4}q\Phi_{h}\phi \bar R + U(\phi)\right) \label{0.4}$$ with the aid of the following transformations $$\gamma_{ab} = \left(\frac{\phi_{h}}{\phi} \right)^{\frac{1}{2}}e^{\frac{2\phi}{q\phi_{h}}} \bar\gamma_{ab}\quad, \ \phi = \frac{\Phi^{2}}{q\Phi_{h}} \label{0.3}$$ where $\Phi_{h}=r_{h}G^{-1/2}$ is the classical value of the field $\Phi$ at the horizon and $U(\phi)=\frac{1}{2G} \left(\frac{\phi_{h}}{\phi}\right)^{1/2}e^{\frac{2\phi}{q\phi_{h}}}$. In (\[0.4\]), $\bar\nabla$ is the covariant derivative compatible with the metric $\bar\gamma_{ab}$ while $\bar R$ is the corresponding curvature scalar. Varying the above action with respect to the scalar field $\phi$ yields the equation of motion for $\phi$ $$\bar \nabla^{a}\bar \nabla_{a} \phi = \frac{q\Phi_{h}}{4} \bar R + \frac{dU}{d\phi}~. \label{new1}$$ Similarly, varying the action with respect to the metric $\bar\gamma_{ab}$, we obtain the constraints $$T_{ab}\equiv {1\over 2} \partial_a\phi\partial_b \phi-{1\over 4}\bar{\gamma}_{ab}(\bar\nabla\phi )^2 +{1\over 4}q\Phi_h (\bar{\gamma}_{ab}\bar \nabla^{c}\bar \nabla_{c} \phi -\bar\nabla_a\bar\nabla_b\phi)-{1\over 2}\bar{\gamma}_{ab}U(\phi)=0~. \label{constra}$$ The theory of the scalar field $\phi$ described by the action (\[0.4\]) is not conformal in general. This can be easily seen by contracting (\[constra\]) with metric $\bar\gamma^{ab}$ to obtain $${T^{a}}_{ a} = \frac{q\Phi_{h}}{4} \bar \nabla^{a}\bar \nabla_{a} \phi - U(\phi) \label{new2}.$$ By substituting (\[new1\]) in the above equation we obtain on shell expression for the trace of the energy-momentum tensor $${T^{a}}_{ a} = \left(\frac{q\Phi_{h}}{4}\right)^2 \bar R + \frac{q\Phi_{h}}{4}\frac{dU}{d\phi} - U(\phi)~. \label{new2'}$$ Thus, we have a nonvanishing trace of the energy-momentum tensor, leading to the breaking of the conformal symmetry in the classical Liouville theory. We would like to point out that such a violation in the conformal invariance generally occurs when we quantize the theory on the curved background. We now take the divergence of (\[constra\]) and obtain $$\bar\nabla_{a}{T^{a}}_{b} = \frac{1}{2}\left[\bar \nabla^{c}\bar \nabla_{c}\phi \bar\gamma_{be} -\frac{q\Phi_{h}}{4}\bar R \bar\gamma_{be}\right] \bar\nabla^{e}\phi - \frac{1}{2}\partial_{b}U(\phi). \label{new3}$$ Substituting the equations of motion for $\phi$ (\[new1\]) in (\[new3\]), leads to the conservation of energy momentum tensor $$\bar\nabla_{a}{T^{a}}_{b} = 0. \label{new4}$$ Hence, we find that although there is a breaking of the classical conformal invariance as the energy-momentum tensor has a nonvanishing trace (\[new2’\]), the diffeomorphism symmetry remains intact. This is a feature which has been observed earlier only in a quantum theory. Now we consider the near horizon behavior of the theory defined by (\[0.4\]). For simplicity we consider the metric $$ds^2 = \bar{\gamma}_{ab}dx^{a}dx^{b}= -g(x)dt^2 + \frac{1}{g(x)} dx^2 \label{new5}$$ where $g(x =x_{h}) =0$ defines the location of the event horizon $x_{h}$ in Schwarzschild coordinates $(t,x)$. Since the metric coefficient $g(x)$ is a well behaved function, we can expand it about $x=x_{h}$ $$\begin{aligned} g(x)=g'(x_h)(x-x_{h})+ \frac{1}{2!}(x-x_{h})^2 g''(x_h)+\frac{1}{3!}(x-x_{h})^3 g'''(x_h) + \cdots~. \label{expan}\end{aligned}$$ In the vicinity of the horizon, we keep terms proportional to $(x-x_{h})$ only. With this approximation, the above function (\[expan\]) becomes $$g(x) \approx g'(x_{h})(x-x_{h})=\frac{2}{\beta_{H}}(x-x_{h}) \label{new6}$$ where $\beta_{H}$ is related to the Hawking temperature [@solodukhin]. Now we would like to see how the metric function vanishes in the region near to the horizon. For that it is convenient to introduce coordinate $z$ defined as $$z= \int^x{dx\over g(x)}~. \label{coord}$$ Substituting (\[new6\]) in (\[coord\]) we get $$g(x) \equiv g(z) = \frac{2}{\beta_{H}}e^{\frac{2z}{\beta_{H}}}. \label{new7}$$ Note that in the $(t, z)$ coordinates, large negative $z$ corresponds to near horizon limit. We now consider the equation of motion for scalar field (\[new1\]) in the vicinity of the horizon. Expressing (\[new1\]) in $(t,z)$ coordinates we get $$-\partial_{t}^{2}\phi + \partial_{z}^{2}\phi = g(z)[\frac{q\Phi_{h}}{4}\bar R + U'(\phi)].\label{new8}$$ In the region near to the horizon, in view of (\[new7\]), the right hand side of above equation vanishes exponentially. Hence, in the vicinity of the horizon we have $$-\partial_{t}^{2}\phi + \partial_{z}^{2}\phi = 0 .\label{new9}$$ Therefore, in the near horizon region Liouville theory gets effectively described by free massless scalar field. Noting that the trace of the energy-momentum tensor (\[new2\]) can be written in the form (dropping the interaction term $U(\phi)$ as it is not important for our present purpose) $${T^{a}}_{a}=\frac{1}{g(x)}(-T_{00}+T_{zz}) =-\frac{q\Phi_{h}}{4}\left[\frac{1}{g(x)}(\partial^{2}_{t} \phi-\partial^{2}_{z} \phi)\right] \label{new10az}$$ we find that in the near horizon limit, using the near horizon equation of motion (\[new9\]), the trace of the energy-momentum tensor as defined in [@solodukhin] becomes $$-T_{00} + T_{zz} = 0. \label{new10}$$ This indicates the fact that near horizon theory is conformally invariant. However, one can rewrite the right hand side of (\[new10az\]) in terms of $\bar{R}$ using (\[new8\]). This simply reproduces the exact result for the trace of the energy-momentum tensor (\[new2’\]). The near horizon result is trivially obtained to be $${T^{a}}_{a}=-\left(\frac{q\Phi_{h}}{4}\right)^{2}g''(x_h). \label{new10z}$$ This shows that the above way of computing the trace near the horizon leads to a nonvanishing result which is incompatible with (\[new10\]). To make the trace ${T^{a}}_{a}$ compatible with (\[new10\]), we shall use the equation of motion (\[new9\]) describing free massless scalar field as the theory near the horizon has conformal invariance. This leads to ${T^{a}}_{a}=0$ since the term in the braces in (\[new10az\]) vanishes as a consequence of (\[new9\]). Hence, in the near horizon limit, we take the equation of motion to be (\[new9\]) with $g(x)$ vanishingly small but not zero. We shall consistently apply this approximation in the subsequent near horizon analysis. Now we show that, though we are able to keep the conformal invariance intact, the near horizon theory does not preserve the diffeomorphism invariance. For that, we write the energy-momentum tensor (\[constra\]) in the $(t,z)$ coordinates as $$\begin{aligned} T_{00}(z) &=& \frac{1}{4}\left(\dot\phi^2 + (\partial_{z}\phi)^2 \right) - \frac{q\Phi_{h}}{4}\left( \partial^2_{z} \phi-\frac{g'(x)}{2}\partial_{z}\phi\right) \label{new15}\\ T_{0z}(z) &=& \frac{1}{2}\dot\phi~\partial_{z}\phi - \frac{q\Phi_{h}}{4} \left(\partial_{z}\dot\phi - \frac{g'(x)}{2}\dot\phi\right)\label{new16}\\ T_{zz}(z) &=& \frac{1}{4}\left(\dot\phi^2 + (\partial_{z}\phi)^2 \right) + \frac{q\Phi_{h}}{4}\left(-\ddot\phi + \frac{g'(x)}{2}\partial_{z}\phi\right) \label{new17} \end{aligned}$$ where the ‘dot’ represents derivative with respect to time and ‘prime’ represents derivative with respect to $x$. Note that once again we do not consider the interaction term $U(\phi)$ since it is not important for our purpose. Now, let us consider the near horizon approximation of the above equations. First, we note that $T_{00}$, $T_{0z}$ and $T_{zz}$ contain a term proportional to a derivative of the metric function. Therefore, it will be inappropriate to substitute the form of the metric function to first order in $(x-x_h)$ (see (\[new6\])) in (\[new15\]-\[new17\]), rather, we have to take into account the next order term in the metric expansion (\[expan\]). In other words, we put $$g'(x) = \frac{2}{\beta_{H}} + g''(x_{h})(x-x_{h}) \label{new21}$$ in the expressions for $T_{00}$, $T_{0z}$ and $T_{zz}$. Then we have $$\begin{aligned} T_{00}(z) &=& \frac{1}{4}\left(\dot\phi^2 + (\partial_{z}\phi)^2 \right) - \frac{q\Phi_{h}}{4}\left( \partial^2_{z} \phi-\frac{1}{2}[\frac{2}{\beta_{H}}+g''(x_{h})(x-x_{h})] \partial_{z}\phi\right) \label{new29}\\ T_{0z}(z) &=& \frac{1}{2}\dot\phi~\partial_{z}\phi - \frac{q\Phi_{h}}{4} \left(\partial_{z}\dot\phi - \frac{1}{2}[\frac{2}{\beta_{H}}+g''(x_{h})(x-x_{h})] \dot\phi\right)\label{new30}\\ T_{zz}(z) &=& \frac{1}{4}\left(\dot\phi^2 + (\partial_{z}\phi)^2 \right) + \frac{q\Phi_{h}}{4}\left(-\ddot\phi + \frac{1}{2}[\frac{2}{\beta_{H}}+g''(x_{h})(x-x_{h})] \partial_{z}\phi\right) \label{new31}~. \end{aligned}$$ Further justification for this approximation will become clear in the subsequent discussion. We move on to compute the covariant divergence of the energy-momentum tensor (\[new29\], \[new30\], \[new31\]). This can be written as : $$\begin{aligned} \bar\nabla_{a}{T^{a}}_{b} &=& \Lambda^{ac}\bar\nabla_{a}T_{cb} \label{new32a+}\end{aligned}$$ where, $\Lambda_{ab}$ is the metric in $(t, z)$ coordinates[^5]. For $b=t$, we obtain : $$\begin{aligned} \bar\nabla_{a}{T^{a}}_{t} &=& -\frac{1}{g(x)} \left[\frac{1}{2}\dot{\phi}(\ddot{\phi}-\partial_{z}^{2}\phi) -\frac{q\Phi_{h}}{8}g(x)g''(x_h)\dot{\phi}\right]~. \label{new32b+}\end{aligned}$$ It is worth mentioning now that rewriting the first term in (\[new32b+\]) in terms of $\bar{R}$ using the exact equation (\[new8\]) yields a structure that precisely cancells the second term in (\[new32b+\]) thereby leading to a vanishing covariant divergence of the energy-momentum tensor. This way, however, the effects of the near horizon approximation are bypassed and the exact result is expectedly reproduced. This is just the analogue of computing the trace (\[new10z\]) by using the exact equation of motion (\[new8\]). In order to systematically implement the near horizon approximation that would be consistent with getting a vanishing trace (\[new10\]) [@solodukhin], we adopt the previous interpretation, that is take the equation of motion as (\[new9\]) with $g(x)$ vanishingly small but not zero. This leads to $$\begin{aligned} \bar\nabla_{a}{T^{a}}_{t} &=& \frac{q\Phi_{h}}{8}g''(x_h)\dot{\phi}~. \label{new32c+}\end{aligned}$$ For $b=z$, we obtain (after using the near horizon equation of motion (\[new9\])): $$\begin{aligned} \bar\nabla_{a}{T^{a}}_{z} &=& \frac{q\Phi_{h}}{8}g''(x_h)\partial_{z}\phi~. \label{new32d+}\end{aligned}$$ The above equations can be compactly written as : $$\begin{aligned} \bar\nabla_{a}T^{a}_{\quad b} &=& \frac{q\Phi_{h}}{8}g''(x_{h})\partial_{b}\phi~. \label{new32}\end{aligned}$$ It is important to observe that the near horizon equation of motion for $\phi$ (\[new9\]) (which is different from the equation of motion satisfied by $\phi$ away from the horizon (\[new1\])) plays an important role in the derivation of (\[new32\]). We therefore conclude that classical Liouville theory, when considered in the region near to the horizon, respects conformal symmetry but it does not preserve the diffeomorphism invariance. This is a new result in our paper. We find that though we are able to keep the conformal invariance intact, the near horizon theory does not preserve the diffeomorphism invariance thereby clearly pointing out the interplay between the diffeomorphism and conformal invariance exhibited in the case of black holes in two dimensional classical Liouville theory. To put this result in a proper perspective, we recall the findings of [@solodukhin] where it was shown that the conformal symmetry near the horizon of the black hole leads to a Virasoro algebra among the Fourier transform of specific combinations of the components of the energy-momentum tensor. Here, we show that the Virasoro algebra which is a reflection of conformal symmetry near the horizon of the black hole can also be understood as the breaking of diffeomorphism symmetry near the black hole horizon. This observation is similar to that of quantum anomalous theories. Indeed, when we quantize the scalar field theory on general curved background, the trace of $\langle T_{ab}\rangle$ turns out to be nonzero. In particular, for the nonchiral theory in $1+1$ dimensions, $\langle {T^{a}}_{a}\rangle$ is proportional to the curvature scalar. However, the regularization adopted to compute the trace anomaly preserves the diffeomorphism invariance. One may adopt different type of regularization which spoils the diffeomorphism symmetry but keeps conformal symmetry intact. It turns out that in $D=2$ dimensions there is no regularization prescription which preserves the conformal as well as diffeomorphism invariance simultaneously [@polyakov; @green]. As a side remark we mention that it might be possible to write an improved stress tensor from (\[new32\]) that is conserved, $$\hat T^{a}_{\quad b} = T^{a}_{\quad b} - \frac{q\Phi_{h}}{8}g''(x_{h})\delta^{a}_{b}\phi~. \label{improvedEM}$$ However, as is easily observed, this tensor is no longer traceless since, $$\hat T^{a}_{\quad a} = \frac{q\Phi_{h}}{4}\bar R \phi~. \label{improveEMtrace}$$ Consequently, simultaneous imposition of both Ward identities is not feasible. This shift of anomaly by using counterterms is more akin to what is done in quantum field theory. A similar phenomenon for the Liouville theory was also observed, though in a different context [@jackiw]. The fact that in the region near the horizon, the energy-momentum tensor is not covariantly conserved (\[new32\]) can also be inferred by computing the classical Poisson algebra among the various light cone components of the energy-momentum tensor. It turns out that the Poisson algebra do not close which in general, is related to the breaking of either diffeomorphism or conformal invariance. This in fact justifies our approximation of keeping terms up to first order in $(x-x_h)$ in $g'(x)$ (\[new21\]) in the evaluation of the covariant divergence of the energy-momentum tensor. It is important to note that if we compute $g'(x)$ from the near horizon expansion of $g(x)$ (\[new6\]), it would finally lead to the conservation of the energy-momentum tensor. This would be in direct clash with the non-closure of the Poisson algebra. We begin our analysis of the classical Poisson algebra by writing (\[new29\]-\[new31\]) in the light cone coordinates $$\begin{aligned} x^{+}&=&\frac{1}{\sqrt{2}}(t+z) \nonumber\\ x^{-}&=& \frac{1}{\sqrt{2}}(t-z)~. \end{aligned}$$ Then we have $$\begin{aligned} T_{++} &=& T_{00} + T_{0z} \label{1.15}\nonumber\\ &=& \frac{1}{4} \left( \dot\phi + \partial_{z}\phi\right)^2 - \frac{q\Phi_{h}}{4}\left( \partial^{2}_{z}\phi + \partial_{z}\dot\phi -\frac{1}{2}[\frac{2}{\beta_{H}}+\frac{g''(x_{h})}{2}(x-x_{h})](\partial_{z}\phi + \dot\phi)\right) \label{new18}\\ T_{--} &=& T_{00} - T_{0z}\label{1.25}\nonumber\\ &=& \frac{1}{4} \left( \dot\phi - \partial_{z}\phi\right)^2 - \frac{q\Phi_{h}}{4} \left( \partial^{2}_{z}\phi - \partial_{z}\dot\phi - \frac{1}{2}[\frac{2}{\beta_{H}}+\frac{g''(x_{h})}{2}(x-x_{h})](\partial_{z}\phi - \dot\phi)\right) ~. \label{new20}\end{aligned}$$ Next, we give the basic Poisson brackets among the canonical variables [@solodukhin] $$\begin{aligned} \{\phi(z),\dot\phi(z')\} &=& \delta(z-z') \label{new22}\\ \{\phi(z),\phi(z')\} &=& 0\label{new23}\\ \{\dot\phi(z),\dot\phi(z')\} &=& 0 \label{new24}~.\end{aligned}$$ The Poisson algebra between $T_{++}(z)$ and $T_{++}(z')$ therefore reads $$\begin{aligned} \{T_{++}(z), T_{++}(z')\} &=& \left(T_{++}(z) + T_{++}(z')\right)\partial_{z}\delta(z-z') \nonumber\\ && + \frac{q^2\Phi_h^{2}}{8}\left[ -\partial_{z}^3 \delta(z-z') + \frac{g'(x)g'(x')}{4} \partial_{z}\delta(z-z') + \frac{1}{2}[g'(x)-g'(x')] \partial_{z}^2 \delta(z-z')\right] \label{new25}\end{aligned}$$ with $g'(x)$ given by (\[new21\]). Note that here, we have only given the Poisson bracket among $T_{++}$ component of EM tensor because the algebra among $T_{--}(z)$ and $T_{--}(z')$ is identical in structure, while the Poisson bracket among $T_{++}$ and $T_{--}$ is zero. Now neglecting terms proportional to $(x-x_{h})$ in the above algebra leads to $$\begin{aligned} \{T_{++}(z), T_{++}(z')\} &=& \left(T_{++}(z) + T_{++}(z')\right)\partial_{z}\delta(z-z') \nonumber\\ &&+\frac{q^2\Phi_h^{2}}{8}\left[ -\partial_{z}^3 \delta(z-z') + \frac{1}{\beta_{h}^2}\partial_{z}\delta(z-z')\right]~.\label{new26}\end{aligned}$$ This algebra would have been the same if we had substituted $g'(x) = \frac{2}{\beta_{H}}$ in (\[new15\]-\[new17\]). However, in that case we would not have found any violation in the conservation of the energy-momentum tensor (\[new32\]) which would be inconsistent with the fact that the Poisson algebra (\[new26\]) does not close. For the sake of completeness, we now move on to investigate the Virasoro algebra. The Virasoro generator is defined as [@solodukhin] $$L_{n} = \frac{L}{2\pi} \int_ {L/2}^{L/2} dz~ T_{++}(z) e^{2i\pi nz/L}~. \label{new27}$$ From the algebra given in (\[new26\]), we compute the algebra between the Virasoro generators which yields : $$\begin{aligned} \{L_n, L_m\}&=&(n-m)L_{n+m} + \frac{c}{12}n\left(n^2+\left(\frac{L}{2\pi\beta_h}\right)^2\right) \delta_{n+m,0}~. \label{vira_alg}\end{aligned}$$ This is the expression for the classical Virasoro algebra with the central charge $c=3\pi q^2\Phi^{2}_{h}$ [@solodukhin].\ Discussions :\ In this article we have studied thoroughly the interplay between the diffeomorphism and conformal symmetries for black holes in $D=2$ classical Liouville theory. The energy-momentum tensor derived from $D=2$ classical Liouville action, in general, is covariantly conserved. However, the trace of the energy-momentum tensor becomes nonzero leading to violation of the conformal invariance. In the region near to the horizon, Liouville theory shows interesting behavior. In the vicinity of the horizon we expand the metric function about the horizon and keep only the leading order terms. In this approximation, the equations of motion for Liouville field takes the form of $D=2$ free, massless Klein-Gordon equation and eventually makes the energy-momentum tensor traceless. Hence, in the vicinity of the horizon Liouville theory is conformally invariant. However, we observe that, in the near horizon limit the energy-momentum tensor is not covariantly conserved. In fact, we showed that the covariant divergence of energy-momentum tensor is proportional to the curvature scalar. This fact is quite well known in the context of quantization of fields on the general curved background. Classically, the energy-momentum tensor, of a field theory under consideration, is traceless and also covariantly conserved. However, during the process of quantization, either there is a trace or diffeomorphism anomaly, depending upon the choice of regularization. It is impossible to preserve both, the conformal as well as the general coordinate (diffeomorphism) invariance [@polyakov; @green]. In this paper, for the classical Liouville theory, we have a similar kind of behavior of the energy-momentum tensor. Away from the horizon, the classical energy-momentum tensor is covariantly conserved signalling the presence of diffeomorphism invariance but the conformal symmetry is lost due to a nonvanishing trace. Near the horizon, on the contrary, the diffeomorphism symmetry is broken but the conformal symmetry is intact. This interplay is a new finding in the context of classical Liouville theory in the near horizon approximation. Another feature in our paper is that the Virasoro algebra which is a reflection of the conformal invariance near the horizon of the black hole can also be understood as the breaking of the diffeomorphism invariance near the black hole horizon. Here we would like to mention that similar results, though in a different context, were discussed in [@jackiw]. To be more specific counterterms were added in the parent $D=2$ Liouville action in order to restore Weyl invariance which however, breaks diffeomorphism symmetry. The point is that counterterms, taken in [@jackiw], are a manifestation of regularization ambiguities. However, the issue of regularization is meaningful only in the quantum field theory. In our example no counterterms are necessary. The interplay of the symmetries occurs very naturally with the near horizon results displaying one feature while the away from the horizon displaying another feature. The whole analysis, either at the technical or conceptual level, is completely classical.\ [**Acknowledgements**]{}\ One of the authors (RB) thanks Masud Chaichian and Anca Tureanu for their kind hospitality at the Physics Department, University of Helsinki, where part of this work was done. The authors would also like to thank the referee for useful comments. [99]{} A. Polyakov, Gauge Fields and Strings, (Harwood, Chur, Switzerland, 1987). M. Green, J. Schwartz, E. Witten, “Superstring Theory, Vol.1" ,(Cambridge University Press, 1987). R. Bertlmann, “Anomalies In Quantum Field Theory", (Oxford Sciences, Oxford, 2000). L. Alvarez-Gaume and E. Witten, Nucl. Phys. B 234, (1984) 269. O. Coussaert, M. Henneaux, P. van Driel, Class. Quant. Grav. 12, (1995) 2961. S. Solodukhin, Phys. Lett. B 454, (1999) 213, \[hep-th/9812056\]. M. Cvitan, S. Pallua, P. Prester, Phys. Lett. B 546, (2002) 119, \[hep-th/0207265\]. G. Kang, J. Koga, M. Park, Phys.Rev.D 70, (2004) 024005, \[hep-th/0402113\]. N. Pinamonti, L. Vanzo, Phys.Rev.D 69, (2004) 084012, \[hep-th/0312065\]. M. Cvitan, S. Pallua, Phys.Rev.D 71,(2005) 104032, \[hep-th/0412180\]. Y. Chen, Class.Quant.Grav.21, (2004) 1153, \[hep-th/0310234\]. S. Solodukhin, Phys.Rev. Lett. 92, (2004) 061302, \[hep-th/0310012\]. J. G. Russo, Phys. Lett. B 359, (1995), 69. R.Jackiw, Theor.Math.Phys.148, (2006) 941, \[hep-th/0511065\]. [^1]: e-mail: rabin@bose.res.in [^2]: e-mail: sunandan.gangopadhyay@gmail.com, sunandan@bose.res.in [^3]: e-mail: shailesh@bose.res.in [^4]: From now onwards we shall suppress the suffix on $R$ but it is understood as $2D$ curvature scalar. [^5]: Note that in the $(t, z)$ coordinates, the metric (\[new5\]) reads $ds^2=-g(x)dt^2 + g(x)dz^2$. Hence, $\Lambda_{tt}=-g(x)$ and $\Lambda_{zz}=g(x)$.

--- abstract: 'A system of two exchange-coupled Kondo impurities in a magnetic field gives rise to a rich phase space hosting a multitude of correlated phenomena. Magnetic atoms on surfaces probed through scanning tunnelling microscopy provide an excellent platform to investigate coupled impurities, but typical high Kondo temperatures prevent field-dependent studies from being performed, rendering large parts of the phase space inaccessible. We present an integral study of pairs of Co atoms on insulating Cu$_2$N/Cu(100), which each have a Kondo temperature of only 2.6 K. In order to cover the different regions of the phase space, the pairs are designed to have interaction strengths similar to the Kondo temperature. By applying a sufficiently strong magnetic field, we are able to access a new phase in which the two coupled impurities are simultaneously screened. Comparison of differential conductance spectra taken on the atoms to simulated curves, calculated using a third order transport model, allows us to independently determine the degree of Kondo screening in each phase.' author: - 'A. Spinelli' - 'M. Gerrits' - 'R. Toskovic' - 'B. Bryant' - 'M. Ternes' - 'A. F. Otte' bibliography: - 'library.bib' title: 'Exploring the phase diagram of the two-impurity Kondo problem' --- [^1] The coupling of individual magnetic atoms to the itinerant host electrons of a metal substrate can lead to the formation of a correlated Kondo state in which the magnetic moment is effectively reduced [@hewson1997kondo]. This has been shown for 3$d$-atoms on bare metal surfaces [@Madhavan1998; @Li1998] as well as on thin decoupling layers [@Otte2008; @vonBergmann2015] and leads to a strong spectroscopic feature at the Fermi energy. Pairs of magnetic atoms are considerably more complex because in addition to the Kondo coupling they can also couple to each other through exchange interactions mediated by the substrate electrons. Depending on their spatial distance, this oscillatory Rudermann-Kittel-Kasuya-Yoshida (RKKY) interaction results in a ferromagnetic (FM) or antiferromagnetic (AFM) coupling [@Ruderman1954; @Kasuya1956; @Yosida1957]. As shown in Fig. \[Figure1\], the competition between these two effects in combination with an external magnetic field embodies rich physics ranging from a correlated singlet or triplet state to complex Kondo states and has been of considerable theoretical interest for decades [@Jayaprakash1981; @Jones1987; @Jones1988; @Jones1989; @Silva1996; @Simon2005; @Zitko2010; @Jabben2012]. The lower part of this phase space has previously been investigated experimentally through studies at zero magnetic field on coupled quantum dots [@Craig2004], molecules [@Wegner2009; @Tsukahara2011], and atoms on top of [@Jamneala2001; @Wahl2007; @Neel2011] as well as buried below [@Pruser2014] a metal surface. In particular, controlled mechanical separation of two Kondo atoms enabled continuous tuning between the single impurity Kondo screening and two-impurity singlet phases [@Bork2011]. However, field-dependent measurements have so far been hampered by high Kondo temperatures for atoms in direct contact with a metal, resulting in wide resonances that are impossible to split or recombine with experimentally available magnetic fields. Kondo temperatures can be reduced by decoupling the atoms from their host either by means of molecular ligands [@Tsukahara2011; @Dubout2015] or a thin insulating layer [@Otte2008]. ![**Phase diagram of the two-impurity Kondo problem.** Schematic phase diagram of two coupled Kondo-screened spins with varying interaction strength $J$ and external field $B$ transverse to the main anisotropy axis of the individual spins. When $|J|$ is small compared to the characteristic Kondo energy $\epsilon_{\rm K}$, at $B=0$ the two spins are independently screened by the substrate electrons, while for $|J|\gg\epsilon_{\rm K}$ a non-magnetic singlet or high-spin triplet state forms. For $J>0$ a sufficient $B$-field can lead to the formation of a new, combined correlated state in which both spins are screened.[]{data-label="Figure1"}](figure1.png) Here we use a low-temperature scanning tunnelling microscope (STM) in ultra-high vacuum to assemble pairs of Kondo-screened Co atoms on a thin insulating Cu$_2$N/Cu(100) substrate. By adjusting the relative position and orientation of the atoms on the underlying crystal lattice we are able to tune their exchange interaction strength, and by applying magnetic fields up to 8 T we can access the complete phase space of Fig. \[Figure1\]. Furthermore, by simulating the spectra using a transport model accounting only for weak Kondo coupling, we independently estimate the Kondo screening. We find that when in AFM coupled dimers the field exactly cancels the exchange interaction, Kondo resonances are re-established, signifying a phase transition through the two-impurity Kondo-screened phase. Results ======= A single Co atom on Cu$_2$N can be described by an effective spin $S=3/2$, and the crystal field induced by the surface is such that the two lowest energy states have magnetization $m=\pm1/2$, degenerate in absence of external magnetic field [@Otte2008]. Its differential conductance (${\rm d}I/{\rm d}V$) spectrum shows a sharp peak at zero excitation voltage, corresponding to a Kondo resonance with a Kondo temperature $T_{\rm K}=2.6 \pm 0.2$ K [@Otte2008], equivalent to a characteristic Kondo energy $\varepsilon_{\rm K}=k_{\rm B}T_{\rm K}=0.22\pm0.02$ meV, which is much lower than for single atoms directly on [@Madhavan1998; @Li1998] or inside a metal [@Pruser2014]. This peak splits when a magnetic field is applied due to the Zeeman effect removing the degeneracy of the states with $m=\pm1/2$ [@vonBergmann2015]. If a second magnetic atom is placed adjacent to the first, the Kondo peak splits similarly to when an effective magnetic field is applied; the intensity of this field scales with the strength of the exchange interaction between the two atoms [@Otte2009]. To model the differential conductance spectra we use a single spin anisotropy Hamiltonian in presence of an external magnetic field with energy $\mu_{\rm B}{\bf B}$ ($\mu_{\rm B}$ being the Bohr magneton), and we assume isotropic Heisenberg exchange $J$ [@Fernandez-Rossier2009] between the atoms in the dimer: $$\label{H} \mathcal{H}=J{\bf S}^{\rm A}\cdot{\bf S}^{\rm B}-\sum_{\substack{i,\mu}}\left(\lambda^2\Lambda_{\mu\mu}S_{\mu}^{i}S_{\mu}^{i}+2(1-\lambda\Lambda_{\mu\mu})\mu_{\rm B} B_{\mu}S_{\mu}^{i}\right).$$ Here the parameters $\Lambda_{\mu\mu}$ ($\mu=x,y,z$) come from second order expansion of the spin-orbit coupling $\lambda {\bf L}^i\cdot {\bf S}^i$ [@Dai2008] (with spin-orbit constant $\lambda=-21$ meV) and represent the extent to which the angular momentum $\mathbf{L}$ (which for a free Co atom has magnitude $L=3$) is unquenched along the $\mu$-direction, while the indices $i={\rm A},{\rm B}$ refer to the different atoms in the dimer. Individual Co atoms experience a strong hard-axis anisotropy $\Lambda_{xx}=0$, $\Lambda_{yy}=\Lambda_{zz}=6.4~{\rm eV}^{-1}$ along the vacancy direction ($x$) of the Cu$_2$N lattice (shown in Fig. \[Figure2\]e), as a result of which the $m=\pm3/2$ doublet is split off by $2\lambda^2 \Lambda_{zz}=5.6$ meV from the lower energy $m=\pm1/2$ doublet [@Otte2008]. When written in terms of the conventional phenomenological anisotropy parameters $D$ and $E$ [@Hirjibehedin2007; @Brune2009; @Gatteschi2006], this corresponds to $D=+2.8$ meV, $E=0$. In most dimers studied here, the intra-dimer exchange coupling $J$, which can be either antiferromagnetic (AFM, $J>0$) or ferromagnetic (FM, $J<0$), is much weaker than the anisotropy energy ($|J|\ll\lambda^2\Lambda_{zz}$). As a result, each dimer is characterised by a low energy quartet that consists of combinations of $m=\pm 1/2$ states of both atoms. Exchange interaction -------------------- For either sign of $J$, the zero-field energy difference between the lowest and the highest state of the quartet is $4J$ (Figs. \[Figure2\]a–b). In the limit of $J$ being much smaller than the magnetic anisotropy, the energy difference between the ground state and the first excited state is $\sim(5/2)J$ or $\sim(3/2)J$ for AFM or FM coupling respectively. Figs. \[Figure2\]c–d show two ${\rm d}I/{\rm d}V$ spectra taken on Co atoms in two different dimers, where each step-like increase in the differential conductance corresponds to a spin excitation from the ground state to an excited state, and in which the two energy scales $4J$ and $2\lambda^2\Lambda_{zz}$ are clearly distinguishable. The evolution of the energetically lowest excitation as a function of magnetic field (in our experiment applied along the transverse direction $y$) reveals the sign of $J$. For AFM coupling, the step-energy decreases until the critical field $B_{\rm c} =(13/8)\times J/(g\mu_{\rm B})$ is reached, at which the ground state and the first excited state become degenerate (Fig. \[Figure2\]a). At $B>B_{\rm c}$ these states change their order and their energy difference increases with field. For FM coupling, the step-energy will only increase with increasing field and a crossing between the ground state and the first excited state does nor occur (Fig. \[Figure2\]b). Following these considerations, we can extract the coupling value between any two atoms placed on the Cu$_2$N surface through differential conductance spectroscopy. In Fig. \[Figure2\]e we show the obtained coupling map with respect to a reference Co atom drawn in the center. Depending on the relative distance and orientation between the atoms in the pair, that can be placed only in specific positions allowed by the substrate, the coupling can be tuned in strength and sign. The obtained results are in good agreement with the coupling values that were previously extracted for Fe dimers [@Bryant2013]. ![**Sign and strength of the exchange interaction between atoms.** (**a–b**) Energy vs transverse magnetic field of the four lowest energy states for two example dimers, corresponding to the instances presented in panels **c–d**, with AFM ($J>0$) and FM ($J<0$) coupling respectively. In the AFM case, a state crossing is expected at a field $B_{\rm c}$. (**c–d**) Example of zero-field differential conductance spectra taken on a Co atom in two different dimers. Red and green arrows show respectively the exchange and anisotropy energies. In the FM spectrum, an additional step can be resolved at $3/2J$. (**e**) Colour map of the experimentally extracted coupling strength between Co atoms on Cu$_2$N. Each circle corresponds to the position of the second Co atom atop a Cu atom of the lattice, with respect to a reference Co atom (blue sphere). Light gray represent Cu positions too close to allow a Co dimer to be built while the small grey circles are the N atoms. The axes identify the coordinate system for the reference atom, with the nitrogen direction defined as $z$-axis. (**f**) Coupling strength vs. inter-dimer distance. For comparative purposes, an isotropic three-dimensional RKKY curve with Fermi wavelength of bulk Cu and horizontal offset of $0.15$ nm, corresponding to a phase shift of 1.3$\pi$, is shown (dashed line). The point represented with an open circle is the $\left\{2,0\right\}$ dimer, whose coupling cannot be explained in terms of RKKY interaction only.[]{data-label="Figure2"}](figure2.png) In previous experiments [@Khajetoorians2012] and theoretical calculations [@Simon2005; @Simon2011] on magnetic atoms directly on a metal surface, the dominant exchange mechanism was found to be RKKY interaction [@Ruderman1954; @Kasuya1956; @Yosida1957]. It is not known to what extent this is still true in the current case, where the adatoms are embedded in a covalently bonded network [@Hirjibehedin2007] which decouples them from the substrate conduction electrons mediating this interaction. In Fig. \[Figure2\]f we plot the extracted coupling value of each dimer as a function of the absolute separation distance. We classify each dimer by the number of unit cells separating the Co atoms in the two symmetry directions of the underlying Cu$_2$N lattice, first stating the distance along the nitrogen direction ($z$) and then along the vacancy direction ($x$): $\{\rm N,\rm v\}$. For all dimers studied except the $\left\{2,0\right\}$ type, depicted with a different symbol, the coupling dependence on the distance seems to be qualitatively comparable to isotropic bulk RKKY [@kittel1963quantum], shown in the same figure for comparative purposes only. A full theoretical treatment of the RKKY coupling on this composite surface, in which also anisotropic RKKY interaction [@Zhou2010] could play a role, is beyond the scope of this paper. In the case of the $\left\{2,0\right\}$ dimer, the exchange coupling is an order of magnitude larger than for the equally spaced $\left\{0,2\right\}$ dimer. According to isotropic RKKY theory the coupling strength and sign depend only on absolute distance, and should therefore be identical for these two dimers. In this particular case we believe that, in addition to RKKY interaction, superexchange coupling plays an important role [@Koch2012]. Being mediated by the Co–N and Cu–N bonds separating the two Co atoms, superexchange coupling should be strongly direction-dependent: every $90^{\circ}$ corner in the coupling path significantly reduces its magnitude [@Goodenough1958; @Kanamori1959]. Of all dimers studied, the $\left\{2,0\right\}$ geometry is the only case where the two Co atoms are connected by a series of bonds without any corner. To further investigate the role of superexchange interaction, dimers would have to be built in which Co atoms are separated by a single N atom only, which so far we have not been able to do controllably. Spectroscopy measurements ------------------------- We measured ${\rm d}I/{\rm d}V$ spectra on every atom of each dimer at different values of external magnetic field applied perpendicular to the sample plane, transverse to the magnetic hard axis (Figs. \[Figure3\]b–f). For comparison, spectra taken on a single Co are shown in Fig. \[Figure3\]a. Additionally, we simulate the experimental data by diagonalizing the Hamiltonian (\[H\]) and using a perturbative scattering model in which the tunnelling electrons interact with the localized spin $\mathbf{S}$ via an exchange interaction $\mathbf{S}\cdot$${\sigma}$, with $\mathbf{\sigma}$$=(\sigma_x,\sigma_y,\sigma_z)$ the standard Pauli matrices for spin-1/2 electrons [@LothNature2010]. In the differential conductance curves of Figs. \[Figure3\]b–e, the magnetic anisotropy is best described with $\Lambda_{xx}=0$, and $\Lambda_{yy}$ and $\Lambda_{zz}$ varying between 6.1 eV$^{-1}$ and 7.0 eV$^{-1}$, very close to the single Co values. However, this is not the case for the $\left\{2,0\right\}$ dimer, presented in Fig. \[Figure3\]f, for which $\Lambda_{zz}=2\Lambda_{yy}$. We believe that in this structure the strain induced by the presence of the second atom is playing a critical role, enhancing the magnetic anisotropy as shown in Ref. [@Bryant2013] (see Supplementary Note 1 and Supplementary Fig. S1). We account for scattering up to third order in the matrix elements by additionally considering an AFM spin-spin Kondo-exchange coupling $J_{\rm K} \rho_{\rm s}$ between the substrate electrons and the localized spin system [@Zhang2013; @Ternes2015], with $\rho_{\rm s}$ the density of states in the substrate around the Fermi energy. When the exchange interaction $J$ between the Co spins is large compared to the Kondo energy $\varepsilon_{\rm K}$, i.e. $|J/\varepsilon_{\rm K}|\gg1$, we find very good agreement between model and experimental data when using a constant exchange coupling of $J_{\rm K} \rho_{\rm s}=0.15$. This value is in accordance with previous results [@Delgado2014; @Spinelli2014; @Oberg2014]. Additional parameter values are listed in Supplementary Table S1. ![**Overview of dimer configurations: experimental data and simulated spectra.** (**a–f**) Measured ${\rm d}I/{\rm d}V$ spectra (blue dots) and simulated curves (green lines) on a single Co atom and on all five types of dimers showing a measurable coupling, presented with increasing value of $|J/\varepsilon_{\rm K}|$, with $\varepsilon_{\rm K}=0.22\pm 0.02$ meV, for zero magnetic field and fields applied in the $y$ direction, perpendicular to the sample surface. Field spectra have been offset for clarity; all measurements were performed at 330 mK. Above each panel, a corresponding atomic lattice diagram and topographic STM image of the measured dimer are shown. Large and small grey circles represent respectively Cu and N atoms; the blue spheres the Co atoms. The arrows in panels b and d indicate which atom of the dimer is presented in the figure; those are the only two cases in which the two atoms are not equivalent. (**g**) Mean square error between the data and the simulated conductance curve as a function of $J/\varepsilon_{\rm K}$ for zero field (blue), 4 T (green) and 8 T (red).[]{data-label="Figure3"}](Figure3.png){width="\textwidth"} This model cannot account for the additional spectral weight which arises when the localized spin system enters a state in which it is strongly screened by the substrate electrons. For example, the zero-field Kondo peak of a single Co atom (Fig. \[Figure3\]a) is strongly underestimated in the simulations. The observed discrepancy arises most likely from the fact that the model accounts only for correlations induced by scattering events up to third order during the tunnelling process, meaning that processes involving interaction with more than a single substrate electron are not taken into account [@Ternes2015]. In a full theoretical treatment of the Kondo interaction, higher order scattering between the localized spin and the substrate electrons produces correlations which strongly influences the spectral weight around the Fermi energy, leading to an increased tunnelling probability and therefore a higher conductance at zero bias. The difference between measured and simulated spectra provides an easily accessible quantitative measure for the degree of Kondo screening in the system. In Fig. \[Figure3\]g we have plotted the mean square error (MSE), defined as ${\rm MSE}=\frac{1}{n}\sum_{n} {\left(G_{\rm exp}-G_{\rm sim}\right)}^2$ (with $n$ the number of data-points in the voltage range and $G_{\rm exp}$ and $G_{\rm sim}$ respectively the normalized measured and simulated conductance), as a function of $J/\varepsilon_{\rm K}$. The MSE, which was determined for spectra at zero field, 4 T and 8 T in the voltage range from $-3$ mV to $3$ mV, scales with the areas coloured in yellow in Figs. \[Figure3\]a–f (see inset Fig. \[Figure3\]g). We find that for larger field values, the region of largest MSE shifts in the direction of positive $J/\varepsilon_{\rm K}$, as expected based on the phase diagram in Fig. \[Figure1\]. ![**Two-impurity Kondo screening phase.** (**a–b**) Differential conductance spectra for the $\left\{0,2\right\}$ and the $\left\{3/2,1/2\right\}$ dimer respectively, around the critical field of each pair. The measurement settings are the same as in Fig. \[Figure3\]. (**c**) Phase diagram of Fig. \[Figure1\] showing the locations of measurements presented in Fig. \[Figure3\] (blue markers) and panels a and b of this figure (white markers). The data points marked with a square indicate the presence of a Kondo resonance. The solid line marks the position of the ground state degeneracy.[]{data-label="Figure4"}](Figure4.png) Intriguingly, in the case of AFM coupling, the combined ground state at zero field is the non-magnetic singlet which we can deliberately tune to different phases by applying an external magnetic field. The field effectively counteracts the exchange interaction leading to a crossing of the singlet state with the lowest triplet state at the critical field $B_{\rm c}$. Around this degeneracy point, the system enters a new phase in which the two impurities are together Kondo-screened by the substrate electrons (see Fig. \[Figure1\]). Among the dimers studied, there are three with AFM coupling. For the $\left\{2,0\right\}$ dimer the coupling is so strong that the state crossing occurs at a field much larger than 9 T, not accessible in our experimental setup. For the other two cases, $\left\{0,2\right\}$ and $\left\{3/2,1/2\right\}$, the crossings are at approximately 3 T and 5 T, respectively. In Figs. \[Figure4\]a–b we show a series of ${\rm d}I/{\rm d}V$ spectra measured around these crossing points for the two dimers, in the voltage range from $-3$ mV to $3$ mV. Here, we clearly observe the emergence of Kondo screening concomitant with a zero-voltage peak in the ${\rm d}I/{\rm d}V$ spectrum, which disappears quickly at fields larger or smaller than $B_{\rm c}$. Note that this two-impurity Kondo state differs significantly from that of a single atom at zero field (Fig. \[Figure3\]a): it reflects the combined screening of the dimer by the substrate electrons. In addition to the central Kondo peak the spectra show low-energy features near the crossing. These features, that cannot be captured by our current model, may contain additional information about the two-impurity Kondo screening phase. Due to some of these features being close to the central peak, it is difficult to extract the exact Kondo line width. However, in both cases the Kondo peak is clearly wider than the thermal energy of our measurement, which is approximately 30 $\mu$eV. Fig. \[Figure4\]c shows the locations of all presented measurements in the phase diagram of Fig. \[Figure1\]. All possible phases are found in our data. The ferromagnetic $\left\{1,1\right\}$ dimer is always in the two-impurity triplet state, while the antiferromagnetic $\left\{2,0\right\}$ dimer is always a two-impurity singlet; for those two instances, no Kondo screening is observed. The weakly FM-coupled $\left\{3/2,3/2\right\}$ dimer is very similar to the single Co: at zero field it behaves like a single Kondo-screened impurity, whereas for large fields the two atoms are separate paramagnetic moments. Lastly, the $\left\{0,2\right\}$ and the $\left\{3/2,1/2\right\}$ both undergo a phase transition from the singlet-phase to being separate moments, through the two-impurity Kondo screening phase. In summary, we have been able to experimentally explore the complete phase space of a system of two coupled Kondo atoms in a magnetic field. We showed that for AFM coupled dimers the phase can be continuously tuned from the two-impurity singlet to the single impurity local moment phase solely by an external magnetic field, leading to the formation of the correlated two-impurity Kondo screening phase at the critical crossover field. This transition, which is fundamentally different from the previously investigated zero-field transition between the single impurity Kondo screening and two-impurity singlet phases [@Bork2011], technically constitutes a quantum phase transition, even though our experiments do not allow us to reveal the associated quantum critical behavior [@Jones1989]. Using a third-order perturbative transport model, we have demonstrated that the spectra can be well reproduced except in the vicinity of the ground state crossings, where additional weight around zero bias voltage occurs. These results may form the basis for future work on one and two-dimensional engineered Kondo lattices. Methods ======= Experimental setup ------------------ The experiments were performed in a commercial STM system (Unisoku USM 1300S), at low-temperature (330 mK) and in ultra-high vacuum ($<2\times 10^{-10}$ mbar). Magnetic fields up to 8 T were applied perpendicular to the sample surface and the hard uniaxial anisotropy of the Co atoms. The Cu$_2$N/Cu(100) sample was prepared in situ by N$_2$ sputtering on the clean Cu crystal [@Leibsle1993]; Cu atoms were evaporated on the precooled Cu$_2$N. The STM tip, made of PtIr, was prepared by indentation in the bare Cu surface. Co dimers were built via vertical atom manipulation [@Hirjibehedin2006]. ${\rm d}I/{\rm d}V$ spectra were recorded with a non-polarized tip, using a lock-in technique with excitation voltage amplitude of 70-100 $\mu$V$_{\rm RMS}$ at 928 Hz. The tunnel current for spectroscopy was set between 1.5 nA and 2 nA at -10 mV or -15 mV sample bias for different types of dimer. Variation of the distance between the tip and the atoms up to 150 pm (corresponding to a factor $\sim30$ variation in tunnel current) on the same atom did not lead to any significant modification of the spectral lineshapes. The resolution of the spectral features is limited mostly by the temperature and the lock-in modulation. Simulations ----------- The simulation of the differential conductance curves was performed using a perturbative approach established by Appelbaum, Anderson, and Kondo [@Kondo1964; @Appelbaum1966; @Anderson1966; @Appelbaum1967] which accounts for spin-flip scattering processes up to 3rd order in the matrix elements. The transition probability $W_{i\rightarrow f}$ for an electron to tunnel between tip and sample or vice versa and simultaneously changing the quantum state of the dimer system between the initial ($i$) and final ($f$) state is in this model given by: $$W_{i\rightarrow f}\propto\left( |M_{i\rightarrow f}|^2+J_{\rm K}\rho_s\sum_{m} \left( \frac{M_{i\rightarrow m}M_{m\rightarrow f}M_{f\rightarrow i}}{\varepsilon_i-\varepsilon_m} +\mbox{c. c.}\right) \right) \delta(\varepsilon_i-\varepsilon_f). \label{equ:W}$$ In this expression, $M_{i\rightarrow j}=\langle \psi_j,\sigma_j|(\mathbf{S}\cdot$$\sigma$$+u)|\psi_i, \sigma_i\rangle$ is the scattering matrix element from the combined state vector $|\psi_i,\sigma_i\rangle$ to $|\psi_j,\sigma_j\rangle$, with $\psi$ as the wavevector of the interacting electron and $\sigma$ as the eigenstate of the localized spin system. Energy conservation between initial energy $\varepsilon_i$ and final energy $\varepsilon_f$ is obeyed by the delta distribution $\delta(\varepsilon_i-\varepsilon_f)$ in equation (\[equ:W\]). The first term of equation (\[equ:W\]) is responsible for the conductance steps observed in the spectra [@Hirjibehedin2007], while the second term leads to logarithmic peaks at the intermediate energy $\varepsilon_m$ and scales with the coupling $J_{\rm K}\rho_s$ to the substrate [@Zhang2013]. Note, that this model cannot cover strong correlations because it neglects higher order effects. Acknowledgements ---------------- This work was supported by the Dutch Organizations FOM and NWO (VIDI) and by the Kavli Foundation. M.T. acknowledges support by the SFB 767. Supplementary Note 1 ==================== Magnetic anisotropy ------------------- Within each type of dimer except for the $\left\{2,0\right\}$, extracted anisotropy parameters were found to be consistent within $\pm5\%$, which is similar to the typical variation found in individual Co atoms. This variation may be attributed to local strain caused by subsurface defects [@Otte2008]. However, as shown in Fig. S1, the $\left\{2,0\right\}$ dimer was found to present much larger variations in the spectra. Interestingly, we could reproduce all the measured curves by varying only one fitting parameter, the anisotropy along the nitrogen direction $\Lambda_{zz}$, and keeping all the other parameters constant. We note that in the conventional notation of $D$ and $E$, two parameters would have to be varied in order to model the observed behaviour. No correlation was found between the $\Lambda_{zz}$ values and either the distance of the dimer from the edge of the Cu$_2$N island, or the size of the supporting island. In Fig. S1c two examples of $\left\{2,0\right\}$ dimers are shown, that exhibit two extreme values of $\Lambda_{zz}$, but that were built on the same island and at very similar distances from the edges. The large variation in the anisotropy of the $\{2,0\}$ dimer can be accounted for since the lattice distortion is in a critical regime. It was noted for Fe dimers in this configuration [@Bryant2013] that the anisotropy is increased compared to the single atom case, due to an increase in the N-Co-N angle. For Co $\{2,0\}$ dimers this angle approaches $180^{\circ}$: at this critical angle, the orbital levels become degenerate. Since $\Lambda_{zz}$ is inversely proportional to the orbital splitting, close to the critical angle small strain-induced modifications of the crystal field can give rise to major changes in the anisotropy parameters. While the atoms in all other dimers are described as easy-plane systems with the hard axis along the vacancy direction (i.e. $\Lambda_{xx}< \Lambda_{yy}\sim\Lambda_{zz}$), for some instances of the $\left\{2,0\right\}$ dimer $\Lambda_{zz}$ has increased to such an extent that $\Lambda_{zz}-\Lambda_{yy}>\Lambda_{yy}-\Lambda_{xx}$, which classifies as an easy-axis system along the nitrogen direction. When written in terms of conventional phenomenological anisotropy parameters $D$ and $E$, technically dimer IV should be specified as $D=+3.99$ meV, $E=1.26$ meV, and dimer V as $D=-4.23$ meV, $E=1.37$ meV. The smooth variation in the spectroscopy of Fig. S1a illustrates that it is more natural to use the description in terms of $\Lambda_{\mu\mu}$. Supplementary Figure 1 ====================== ![**Anisotropy variation in strained dimers.** (**a**) Zero-field differential conductance spectra (blue dots) and corresponding simulated curves (green lines) recorded on different instances of a $\left\{2,0\right\}$ dimer, showing variation in the spectroscopic features. (**b**) Best fit values for $\Lambda_{zz}$ for all dimers shown in **a**; we found $\Lambda_{xx}=0$, $\Lambda_{yy}=6.2$ eV$^{-1}$, $J=2$ meV and $J_{\rm K}\rho_{\rm s}=0.15$ for all measured curves. (**c**) Topographic STM image of dimers II and V. These data also published in Ref. [@Bryant2015].](FigureS1.png) Supplementary Table 1 =====================  [^1]: a.f.otte@tudelft.nl

--- author: - | Leonard Adolphs, Thomas Hofmann\ Department of Computer Science\ ETH Zurich\ {ladolphs, thomas.hofmann}@inf.ethz.ch bibliography: - 'ftwp\_arxiv.bib' title: | LeDeepChef \ Deep Reinforcement Learning Agent for Families of Text-Based Games ---

--- abstract: 'Starting from the ordered state, we investigate the short-time behaviour of the hard-disk model. For the positional order, we determine the critical exponents $\eta$ and $z$ from the dynamic relaxation of the order parameter and the cumulant with molecular dynamics simulations. The results are compared with previous Monte Carlo (MC) simulations. The bond orientational order is studied with MC dynamics.' address: 'Universität - GH Siegen, D-57068 Siegen, Germany' author: - 'A. Jaster' title: | Short-time behaviour of the two-dimensional\ hard-disk model --- Introduction ============ For a long time it was believed that universal scaling behaviour can be found only in the long-time regime. Therefore, numerical simulations were performed in the thermodynamic equilibrium. Such simulations in vicinity of the critical point are affected by the critical slowing down. However, recently Janssen, Schaub and Schmittmann [@JASCSC] showed that universality exists already in the early time of the evolution. They discovered that a system with non-conserved order parameter and energy (model A) quenched from a high temperature state to the critical temperature shows universal short-time behaviour already after a microscopic time scale $t_{\mathrm{mic}}$. Starting from an unordered state with a small value of the order parameter $m_0$, the order increases with a power law $M(t) \sim m_0 \, t^\theta$, where $\theta$ is a new dynamic exponent. A number of Monte Carlo (MC) investigations [@ZHENG; @LUSCZH; @JAMASCZH] support this short-time behaviour. These simulations can be also used to calculate the conventional (static and dynamic) exponents as well as the critical point [@SCHZHE]. This may eliminate critical slowing down, since the simulations are performed in the short-time regime. First simulations of the dynamic relaxation in the short-time regime started from an unordered state. However, short-time dynamical scaling can be also found starting from the ordered state ($M(t=0)=1$). There exist no analytical calculations for this situation, but several MC simulations were done [@STAUFFER; @LISCZH; @ZHENG]. Also, all critical exponents, except for the new exponent $\theta$, can be calculated starting from the ordered state. Up to now, simulations of the dynamic relaxation have only been performed with MC dynamics. Normally, molecular dynamics (MD) simulations can cause ergodicy problems, since the energy of the system is conserved. However, for the two-dimensional hard-disk model the potential energy of the allowed configurations does not depend on the positions of the particles, but is constant. Therefore, the restriction of energy conservation does not lead to a reduction of possible configurations. The nature of the two-dimensional melting transition is a longstanding puzzle [@STRAND; @GLACLA]. The Kosterlitz-Thouless-Halperin-Nelson-Young (KTHNY) theory [@KTHNY] predicts two continuous phase transitions. The first transition (dislocation unbinding) at the melting temperature $T_{\mathrm{m}}$ transforms the solid with quasi-long-range positional order and long-range orientational order into a new [*hexatic*]{} phase that posses short-range positional order and quasi-long-range orientational order. The disclination unbinding transition at $T_{\mathrm{i}}$ transforms this hexatic phase into an isotropic phase in which the positional and orientational order are short-range. There are several other theoretical scenarios for the melting transition in two dimensions [@STRAND]. Most of them predict a first-order phase transition from the solid to the isotropic phase with a coexistence region instead of a hexatic phase. Even for the simple hard-disk system no consensus about the nature of the melting transition has been established. A large number of simulations of the two-dimensional hard-disk model in the thermodynamic equilibrium have been performed. A melting transition was first seen in a computer study by Alder and Wainwright [@ALDWAI]. They investigated 870 disks with MD methods (constant number of particles $N$, volume $V$ and energy $E$) and found the transition to be first order. However, the results of such a small system are affected by large finite-size effects. Recent simulations used MC techniques either in the $NVT$ ensemble (constant volume) [@ZOLCHE; @WEMABI; @MIWEMA; @JASTER] or the $NpT$ ensemble (constant pressure) [@LEESTR; @FEALST]. Unfortunately, the results of these simulations are not compatible. In this article, we study the short-time behaviour of the two-dimensional hard-disk model starting from the ordered state (perfect crystal). First we examine the positional order parameter $\psi_{\mathrm{pos}}(t)$ and the cumulant $\tilde{U}_{\mathrm{pos}}(t)$ with MD simulations. The power law behaviour of these observables is used to determine the critical exponents $\eta$ and $z$. The results are compared with those of a previous MC simulation [@JASTER2]. In the second part we study the bond orientational order parameter $\psi_6(t)$ and the cumulant $\tilde{U}_6(t)$ with MC dynamics. All simulations are performed in a rectangular box with ratio $2:\sqrt{3}$, which is necessary for the ordered state, and with periodic boundary conditions. The disk diameter is set equal to one. Positional order ================ The positional order parameter $\psi_{\mathrm{pos}}$ can be computed via $$\psi_{\mathrm{pos}}(t) = \left | \frac{1}{N} \sum_{i=1}^N \exp ( {\mathrm{i}} \, \vec{G} \cdot \vec{r}_i(t) ) \right | \ ,$$ where $\vec{G}$ denotes a reciprocal lattice vector and $\vec{r}_i(t)$ is the position of particle $i$ at time $t$. $\vec{G}$ has a magnitude of $2 \pi /a$, where $a=\sqrt{2/(\sqrt{3}\rho)}$ is the average lattice spacing. The direction of $\vec{G}$ is fixed to that of a reciprocal lattice vector of the perfect crystal (which are unique due to the boundary condition of a rectangular box of ratio $2:\sqrt{3}$). The reason for fixing $\vec{G}$ is that large crystal tilting is not possible since we simulate only the short-time behaviour of the system. MC simulations of the dynamic relaxation of systems with quasi-long-range order were performed for the 6-state clock model [@CZERIT], the XY model [@OKSCYAZH; @LUOZHE], the fully frustrated XY model [@LUSCZH; @LUOZHE], the quantum XY model [@YILUSCZH] and the hard-disk model [@JASTER2]. However, no investigations exist for the relaxation with molecular dynamics. Independent of the dynamics, the scaling form of the second moment of the order parameter at or above $\rho_{\mathrm{m}}$ is $$\label{Eqscale} {\psi_{\mathrm{pos}}}^2(t,L) = b^{-\eta} {\psi_{\mathrm{pos}}}^2(b^{-z}t,b^{-1}L) \ ,$$ where $b$ denotes the rescaling factor. This leads for sufficiently large $L$ to a power law time dependence of the form $$\label{EqPLpsi2} {\psi_{\mathrm{pos}}}^2(t) \sim t^{-\eta/z} \ .$$ From a finite-size scaling (FSS) analysis of the time-dependent cumulant $$\tilde{U}_{\mathrm{pos}}(t)= \frac{{\psi_{\mathrm{pos}}}^4(t)}{ \left ( {\psi_{\mathrm{pos}}}^2(t) \right ) ^2} -1$$ one obtains $$\tilde{U}_{\mathrm{pos}}(t) \sim t^{d/z} \ ,$$ where $d$ is the dimension of the system. One can use $\tilde{U}_{\mathrm{pos}}(t)$ to determine the dynamic critical exponent $z$ and then, with $z$ in hand, the static exponent $\eta$ from the behaviour of ${\psi_{\mathrm{pos}}}^2(t)$. In the following we simulate the hard-disk model with molecular dynamics at the melting density $\rho_{\mathrm{m}} \approx 0.933$ [@JASTER2] (the density is given as usual in reduced units) and in the solid phase ($\rho=1.0$) to investigate the time evolution of ${\psi_{\mathrm{pos}}}^2$ and $\tilde{U}_{\mathrm{pos}}$. Since the positional order of the system is quasi-long-range, we expect to find a power law behaviour for $\rho \ge \rho_{\mathrm{m}}$ as in the case of the MC study. Starting from the ordered state ($\psi_{\mathrm{pos}}=1$), i.e. from a perfect crystal with lattice spacing $a$, we release the system to evolve with molecular dynamics. Solving exactly the simultaneous classical equations of motion for hard-disks means in practice to calculate the next collision point of two particles and their new momenta. The initial components of the momenta are chosen randomly with a distribution proportional $\exp (p^2_i/2)$. Obviously, in the case of a hard-core potential a global change in the kinetic energy just leads to a rescaling of time. We use systems of $16^2$, $32^2$, $64^2$ and $128^2$ hard-disks and measure the observables up to $t=6$ in time intervals of $\Delta t=0.04$. The number of independent data sets ranges from ${\cal O}(150)$ for $N=128^2$ to ${\cal O}(30\,000)$ for $N=16^2$. Statistical errors are calculated by dividing the data into different subsamples. Systematic errors are estimated by least square fits of different system sizes and different time intervals, i.e. the exponent $c$ of the power law behaviour is calculated for each system in small intervals $[t_i,t_j]$. For large enough systems, $c$ is independent of the system size between $t_{\mathrm{min}}$ and $t_{\mathrm{max}}$. The microscopic time scale $t_{\mathrm{min}}$ depends on the microscopic details, but is independent from the system size. In contrast to this, $t_{\mathrm{max}}$ — the time when the system starts to show finite-size effects — scales with the number of particles. In Fig. \[fig\_pos\_O2\] we plot the time evolution of ${\psi_{\mathrm{pos}}}^2$ at the melting density $\rho_{\mathrm{m}}$ for different system sizes in a double logarithmic scale. The figure shows that the power law behaviour starts after a microscopic time scale $t_{\mathrm{mic}}$ of approximately $0.16$. For times up to $0.5$ the difference between the systems with $64^2$ and $128^2$ hard-disks is negligible. We use these two systems and time intervals of $t=[0.2 \dots 0.8]$ or smaller for the determination of the critical exponents. Power law fits for the different time intervals and system sizes lead to $\eta/z = 0.201(4)$. The situation in the solid phase at $\rho = 1.0$ is very similar and we get $\eta / z = 0.0695(25)$. To determine $z$ independently, we also measure the time evolution of the cumulant $\tilde{U}_{\mathrm{pos}}$. As before the behaviour can be well described by a power law ansatz. From the slope we get $z=1.04(3)$ at $\rho_{\mathrm{m}}$, while the analysis for $\rho=1.0$ yields $z=1.06(3)$. Thus we get $\eta=0.194(9)$ at $\rho=0.933$ and $\eta=0.0737(49)$ at $\rho=1.0$, respectively. The values of $\eta$ coincide with the results from the dynamic relaxation with MC dynamics [@JASTER2] (using the same methods) within statistical errors, as can be seen from Table \[table1\]. For comparison we also show the results obtained from conventional FSS [@JASTER2]. The value of the dynamic critical exponent $z$ changes from $z \approx 2$ for MC dynamics to $z \approx 1$ for MD simulations. These are the usual values for local MC updating schemes — which can be understood as due to the ‘diffusion’ of the changes induced by the local updating — and MD simulations, respectively. The time when the system starts to shows finite-size effects scales approximately with $L^z$ for MD simulations (as can be estimated from Fig.  \[fig\_pos\_O2\]) as well as for MC investigations. [cc\*[2]{}[r@[.]{}l]{}c\*[2]{}[r@[.]{}l]{}c\*[1]{}[r@[.]{}l]{}]{} & & & & & &\ $\rho$ & & & & & & & &\ 0.933 & & 1&04(3) & 0&194(9) & & 2&01(2) & 0&199(3) & & 0&200(2)\ 1.0 & & 1&06(3) & 0&0737(49) & & 2&06(4) & 0&0794(29) & & 0&0791(6)\ Our results show that the determination of the critical exponents in the short-time region can be done with molecular dynamics at least in the case of hard-core potentials. The disadvantage of the MD simulations compared to the MC investigations with the Metropolis algorithm is the scaling of CPU time with the number of particles. In the Metropolis algorithm the CPU time for a sweep is proportional to the number of particles $N$, i.e. a single step of a disk is independent of $N$. In the MD simulations the number of collisions during a fixed time interval $\Delta t $ scales also with $N$. However, the search for the next two colliding disks is not independent of $N$. The CPU time for a naive algorithm scales with $N^2$, while an improved version yields a factor $N$. Therefore, also the improved MD algorithm leads to a poorer scaling behaviour as the MC version. However, the MD simulations not only enlarge the knowledge of dynamic relaxation but perhaps also offer a possibility to compare the results with experimental investigations. Bond orientational order ======================== The orientational order of the hard-disk system can be described by the bond orientational order parameter $$\psi_6 = \frac{1}{N} \sum_{i=1}^{N} \frac{1}{N_i} \sum_{j=1}^{N_i} \exp ( 6 \, {\mathrm{i}}\, \theta_{ij} ) \ .$$ The sum on $j$ is over the $N_i$ neighbours of the particle $i$ and $\theta_{ij}$ is the angle between the particles $i$ and $j$ and an arbitrary but fixed reference axis. Two particles are defined as neighbours, if the distance is less than $1.4$ times the average lattice spacing $a$. This definition is computationally less expensive than the precise determination with the Voronoi construction [@VORONOI]. We examine the bond orientational order of the hard-disk model in the liquid regime at $\rho=0.885$, in vicinity of the transition point $\rho_{\mathrm{i}} \approx 0.899$ [@JASTER] at $\rho=0.9$ and $\rho=0.905$ and in the solid phase at $\rho=0.94$. We measure the second moment of the order parameter ${\psi_6}^2$ and the cumulant $\tilde{U}_6$ as a function of time, where the MC dynamics is given by the Metropolis algorithm [@METRO]. The new positions of the particles are chosen with equal probability within a circle centered about its original position. As before, we start the relaxation process from the perfect ordered crystal ($\psi_6=\psi_{\mathrm{pos}}=1$). We use systems of $64^2$ and $128^2$ hard-disks and measure up to $10\,000$ MC sweeps. In case of a KTHNY-like scenario $\rho_{\mathrm{i}}$ is the beginning of the hexatic phase, i.e., the lower bond of the critical line which ends at $\rho_{\mathrm{m}}$. Therefore, we expect a power law behaviour similar to Eq. (\[EqPLpsi2\]) between $\rho_{\mathrm{i}}$ and $\rho_{\mathrm{m}}$ assuming a scaling behaviour of the form (\[Eqscale\]) for the positional order parameter. The value of the critical exponent $\eta_6$ (at $\rho=\rho_{\mathrm{i}}$) is predicted with $1/4$ [@KTHNY] and was measured with 0.251(36) [@JASTER], while the value of the dynamic critical exponent $z$ for the local Metropolis algorithm is normally about two. For a conventional weak first-order phase transition the behaviour of the positional order parameter should be also approximately power-like. Figure \[fig\_or\_O2\] shows the time evolution of ${\psi_6}^2$ for the different densities. Obviously, the time dependence in vicinity of $\rho_{\mathrm{i}}$ can not be described by a simple power law behaviour as expected. Therefore, the scaling form of the second moment of the order parameter ${\psi_6}^2$ is not given by (\[Eqscale\]). The behaviour of ${\psi_6}^2(t)$ at $\rho=0.9$ and $0.905$ is also not consistent with a simple weak first-order phase transition. The behaviour in the fluid phase ($\rho = 0.885$) is incompatible with both scenarios, since it is not of the form $t^{-\eta_6/z} \exp (-t/\xi_{\mathrm{t}})$ [@JASTER3]. To rule out that the non power law behaviour is just an effect coming from our method determing neighbours, we also make simulations using the precise Voronoi definition. The result for ${\psi_6}^2(t)$ at $\rho=0.9$ is visualized in Fig. \[fig\_or\_voro\]. For small times the difference between both definitions is negligible. At larger times the number of disclinations increase and causes errors in case of using the distance for the determination of neighbours. Therefore, the deviation between the different values of ${\psi_6}^2$ grows if the time increases. However, also in case of using the Voronoi construction we find no power law dependence. A possible explanation for the behaviour of ${\psi_6}^2(t)$ could be a mixing of the order parameters[^1]. In the simplest case we expect that the behaviour should be described by the sum of two power law functions. However, we find that the the curves at $\rho=0.9$ and $\rho=0.905$ could not be fitted very well by such an ansatz. Nevertheless, we use the almost linear behaviour in the intervals $t_1=[10,100]$ and $t_2=[1000,10\,000]$ to determine the values of the exponents, i.e. we examine if the behaviour in one of these intervals is consistent with a power law coming from the bond orientational order parameter. At $\rho=0.9$ we get from ${\psi_6}^2(t)$ the exponent $c_1=0.124$ and $0.0264$, respectively. The slope of the cumulant $\tilde{U}_6(t)$ (which is shown in Fig. \[fig\_or\_Ut\]) gives the exponent $c_{\mathrm{U}}$. This yields $z=2.12$, $\eta=0.262$ (for $t_1=[10,100]$) and $z=4.95$, $\eta=0.130$ (for $t_2=[1000,10\,000]$). The first value of $\eta$ is consistent with previous measurements of $\eta_6$ in equilibrium [@JASTER], while the value determined in the interval $t_2$ is too small. However, also the behaviour of ${\psi_6}^2(t)$ in the interval $t_1$ is inconsistent with (\[Eqscale\]) since the value is not constant in the solid phase. Thus, neither the behaviour in the time interval $t_1$ nor in the interval $t_2$ is compatible with the scaling form (\[Eqscale\]). Additionally, we try to estimate the dynamic critical exponent $z$ from the exponential behaviour the the fluid phase. But these measurements give also inconsistent results, i.e. we get a value $z\approx 1$. Conclusions =========== We investigated the short-time behaviour of the hard-disk model for the positional and bond orientational order. The positional order parameter was studied with MD simulations at $\rho_{\mathrm{m}}$ and in the solid phase. We determined the critical exponent $\eta$ as well as the dynamic critical exponent $z$ from the power law behaviour of ${\psi_{\mathrm{pos}}}^2(t)$ and $\tilde{U}_{\mathrm{pos}}(t)$. The values of $\eta$ are in agreement with previous measurements. Our results show that MD simulations can be used for the determination of critical exponents from the short-time behaviour. The dynamic critical exponent changes from $z \approx 2$ for MC to $z \approx 1$ for MD simulations. The bond orientational order was studied with conventional MC dynamics. The time evolution of the order parameter and the cumulant is not given by a simple power law behaviour, i.e.  the scaling behaviour is not of the form (\[Eqscale\]). This phenomena could not be explained by a weak first-order phase transition or by the way of determining neighbours. A possible mixing of order parameter was discussed. Acknowledgement {#acknowledgement .unnumbered} =============== Critical comments by Lothar Schülke are gratefully acknowledged. Especially I benefitted from discussions with Conny Deiters. This work was supported in part by the Deutsche Forschungsgemeinschaft under Grant No. DFG Schu 95/9-1. [99]{} H.K. Janssen, B. Schaub and B. Schmittmann, Z. Phys. B 73 (1989) 539. A review of short-time dynamics is given in: B. Zheng, Int. J. Mod. Phys. B 12 (1998) 1419. H.J. Luo, L. Schülke and B. Zheng, Phys. Rev. E 57 (1998) 1327; Phys. Rev. Lett. 81 (1998) 180. A. Jaster, J. Mainville, L. Schülke, B. Zheng, J. Phys. A: Math. Gen. 32 (1999) 1395. L. Schülke and B. Zheng, Phys. Lett. A 215 (1996) 81. D. Stauffer, Physica A 186 (1992) 197. Z.B. Li, L. Schülke and B. Zheng, Phys. Rev. E 53 (1996) 2940. K.J. Strandburg, Rev. Mod. Phys. 60 (1988) 161. M.A. Glaser and N.A. Clark, Adv. Chem. Phys.  83 (1993) 543. J.M. Kosterlitz and D.J. Thouless, J. Phys. C 6 (1973) 1181; J.M. Kosterlitz, J. Phys. C 7 (1974) 1046; B.I. Halperin and D.R. Nelson, Phys. Rev. Lett. 41 (1978) 121; A.P. Young, Phys. Rev. B 19 (1979) 1855. B.J. Alder and T.E. Wainwright, Phys. Rev. 127 (1962) 359. J.A. Zollweg and G.V. Chester, Phys. Rev. B 46 (1992) 11186. H. Weber and D. Marx, Europhys. Lett. 27 (1994) 593; H. Weber, D. Marx and K. Binder, Phys. Rev. B 51 (1995) 14636. A.C. Mitus, H. Weber and D. Marx, Phys. Rev. E 55 (1997) 6855. A. Jaster, Europhys. Lett. 42 (1998) 277; Phys. Rev. E 59 (1999) 2594. J. Lee and K.J. Strandburg, Phys. Rev. B 46 (1992) 11190. J.F. Fernández, J.J. Alonso and J. Stankiewicz, Phys. Rev. Lett. 75 (1995) 3477; Phys. Rev. E 55 (1997) 750. A. Jaster, Phys. Lett. A 258 (1999) 59. P. Czerner and U. Ritschel, Phys. Rev. E 53 (1996) 3333. K. Okano, L. Schülke, K. Yamagishi and B. Zheng, J. Phys. A: Math. Gen. 30 (1997) 4527. H.J. Luo and B. Zheng, Mod. Phys. Lett. B 11 (1997) 615. H.P. Ying, H.J. Luo, L. Schülke and B. Zheng, Mod. Phys. Lett. B 12 (1999) 1237. For a definition of Voronoi cell see: D.P. Fraser, M.J. Zuckerman and O.G. Mouritzen, Phys. Rev. A 42 (1990) 3186. N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller and E. Teller, J. Chem. Phys. 21 (1953) 1087. A. Jaster, Phys. Lett. A 258 (1999) 177. Z.B. Li, X.W. Liu, Z.G. Pan and L. Schülke, ’Initial order mixing in the kinetic Ashkin-Teller model’, to be published in Communications in Theoretical Physics, International Academic Publishers, Beijing China. [^1]: A mixing of order parameters was also found for the Ashkin-Teller model [@LILIPASC].

--- abstract: 'We present a set of probabilistic models applied to binary classification as defined in the DEFT’05 challenge. The challenge consisted a mixture of two differents problems in Natural Language Processing : identification of author (a sequence of François Mitterrand’s sentences might have been inserted into a speech of Jacques Chirac) and thematic break detection (the subjects addressed by the two authors are supposed to be different). Markov chains, Bayes models and an adaptative process have been used to identify the paternity of these sequences. A probabilistic model of the internal coherence of speeches which has been employed to identify thematic breaks. Adding this model has shown to improve the quality results. A comparison with different approaches demostrates the superiority of a strategy tha combines learning, coherence and adaptation. Applied to the DEFT’05 data test the results in terms of precision (0.890), recall (0.955) and [*Fscore*]{} (0.925) measure are very promising.' author: - 'Marc El-Bèze' - 'Juan-Manuel Torres-Moreno' - Frédéric Béchet bibliography: - 'biblio\_rnti.bib' title: | Un duel probabiliste pour départager\ deux présidents (LIA @ DEFT’2005) --- Introduction ============ Dans le cadre des conférences TALN[^1] et RECITAL[^2] tenues en juin 2005 à Dourdan (France), un atelier a été organisé autour du défi de fouille textuelle proposé par [@aze:2005]. Ce défi portait le nom de DEFT’05 (DÉfi Fouille de Textes). Il a été motivé par le besoin de mettre en place des techniques de fouille de textes permettant soit d’identifier des phrases non pertinentes dans des textes, soit d’identifier des phrases particulièrement singulières dans des textes apparemment sans réel intérêt. Concrètement, il s’agissait de supprimer les phrases non pertinentes dans un corpus de discours politiques en français. Cette tâche est proche de la piste [*Novelty*]{} du challenge TREC [@soboro:2004] qui dans sa première partie consiste à identifier les phrases pertinentes puis, parmi celles-ci, les phrases nouvelles d’un corpus d’articles journalistiques. Pour mieux comprendre à quoi correspondait dans DEFT’05 la suppression des phrases non pertinentes d’un corpus de discours politiques [@alphonse:2005], une brève description du but général[^3] s’impose. Un corpus de textes, allocutions officielles issues de la Présidence (1995-2005) de Jacques Chirac a été fourni. Dans ce corpus, des passages issus d’un corpus d’allocutions (1981-1995) du Président François Mitterrand ont été insérés. Les passages d’allocutions de F. Mitterrand insérés sont composés d’au moins deux phrases successives et ils sont censés traiter une thématique différente[^4]. Un corpus, formé de discours de J. Chirac entrecoupés d’extraits de ceux de F. Mitterrand, est ainsi constitué. Certaines informations sont supprimées de ce corpus afin de constituer les trois corpus ci-dessous : - Corpus C1 : aucune présence d’années ni de noms de personnes : ils ont été remplacés par les balises <span style="font-variant:small-caps;">&lt;date&gt;</span> et <span style="font-variant:small-caps;">&lt;nom&gt;</span> ; - Corpus C2 : pas d’années : elles ont été remplacées par la balise <span style="font-variant:small-caps;">&lt;date&gt;</span> ; - Corpus C3 : présence des années et des noms de personnes. Le but du défi consistait à déterminer les phrases issues du corpus de F. Mitterrand introduites dans le corpus composé d’allocutions de J. Chirac. Ce but est commun aux trois tâches T1, T2 et T3 relatives aux trois corpus C1, C2 et C3 ont ainsi été définies. Intuitivement, la tâche T1 est la plus difficile des trois car le corpus afférent C1 contient moins d’informations que les deux autres. Les résultats (calculés uniquement sur les phrases de F. Mitterrand extraites) peuvent être évalués sur un corpus de test (T) avec des caractéristiques semblables à celui de développement (D) (cf. tableau \[tab:stats\]), en calculant le [*Fscore*]{} : $$Fscore(\beta)= \frac{(\beta^2+1)\times Pr\acute{e}cision \times Rappel}{\beta^2\times Pr\acute{e}cision + Rappel}$$ Dans le cadre de DEFT’05, le calcul du [*Fscore*]{} retenu par les organisateurs a été effectué uniquement sur les phrases de Mitterrand, et il a été modifié [@alphonse:2005] comme suit (cette réécriture suppose évidemment que $\beta$ soit égal à 1 de façon à ne privilégier ni précision ni rappel) : $$Fscore(\beta = 1)= \frac{2 \times \textrm{Nb\_phrases\_correctes\_extraites}}{\textrm{Nb\_total\_extraites} + \textrm{Nb\_total\_pertinentes}}$$ - [Nb\_phrases\_correctes\_extraites]{} : nombre de phrases qui appartiennent réellement au corpus de Mitterrand dans le fichier résultat fourni par le système ; - [ Nb\_total\_extraites]{} : nombre de phrases données dans le fichier résultat (que le système considère comme étant Mitterrand); - [ Nb\_total\_pertinentes]{} : nombre total de phrases de Mitterrand dans le corpus de test. = 1.6 Discours Phrases (D) Mots (D) Phrases (T) Mots (T) ------------ ------------- ----------- ------------- ---------- Chirac 52 936 1 148 208 30 148 638 547 Mitterrand 8 027 218 124 5 027 134 111 : Statistiques sur les corpus de développement (D) et de test (T).[]{data-label="tab:stats"} On pourrait être tenté de traiter chacune des trois tâches en appliquant les méthodes employées habituellement en classification. [*A priori*]{}, un problème de classification à deux classes (ici Chirac et Mitterrand[^5]) paraît simple. Or, de nombreuses raisons font que la question est complexe. Au terme d’une étude portant sur 68 interventions télévisées composées de 305 124 mots, [@labbe:l990] distingue quatre périodes dans les discours de Mitterrand. L’une d’elles dénommée « Le président et le premier ministre » (octobre 1986 - mars 1988) n’est probablement pas la plus facile à traiter sous l’angle particulier proposé par le défi DEFT’05. Dans d’autres conditions, c’eût été loin d’être évident. Ici, on peut s’attendre à des difficultés accrues pour différencier deux orateurs qui se sont exprimés dans maints débats sur les mêmes sujets. Facteur aggravant : on ne dispose que d’un petit corpus déséquilibré. Pour la tâche T1, 109 279 mots pleins pour un président et 582 595 pour le second répartis dans 587 discours (dont la date n’est pas fournie). Pour donner une idée de la difficulté de ce défi, notons qu’une classification supervisée binaire avec un perceptron optimal à recuit simulé [@torres-moreno:2002] appliqué sur la catégorie grammaticale de mots (l’utilisation de tous les mots générant une matrice trop volumineuse) donne un taux d’extraction des segments Mitterrand décevant avec un [*Fscore*]{} $\approx$ 0,43 ; la méthode classique [*K-means*]{} sur les mêmes données, conduit à un [*Fscore*]{} $\approx$ 0,4. En comparaison, avec des classifieurs à large marge réputés performants tels que [*AdaBoost*]{} [@freund:1997] avec [*BoosTexter*]{} [@schapire:2000] et [*Support Vector Machines*]{} avec SVM-Torch [@collobert:2001], on plafonne à un [*Fscore*]{} $\approx$ 0,5. Enfin, avec une méthode de type [*base-line*]{} vraiment simpliste, où on classerait tout segment comme appartenant à la catégorie [*M*]{}, on obtiendrait un [*Fscore*]{} $\approx$ 0,23 sur l’ensemble de développement et de 0,25 pour le test. Comme ces résultats se sont avérés décevants, nous avons décidé d’explorer des voies totalement différentes. Nous présentons en section \[sec:modeles\] une approche reposant sur des modèles bayésiens, une chaîne de Markov, des adaptations statiques et dynamiques et un réseau sémantique de noms propres. Des approches probabilistes avec ou sans filtrage et lemmatisation sont utilisées. En section \[sec:cohérence\], nous développons une approche de la cohérence interne des discours qui permet encore d’augmenter le [*Fscore*]{}. La section \[sec:expériences\] est consacrée aux expériences et à leurs résultats. Nous comparons et tentons de fusionner ces différentes approches avant de conclure et d’envisager quelques perspectives. L’annexe présente une analyse détaillée d’un discours de la classe [*C*]{}, où l’utilisation de sa cohérence interne permet de mieux le classer. Modélisation {#sec:modeles} ============ La chaîne de traitement que nous allons décrire dans les sous-sections suivantes est constituée de quatre composants : un ensemble de modèles bayésiens (cf. \[sec:bayes\]), un automate de Markov (cf. \[subsec:prise\]), un modèle d’adaptation (cf. \[subsec:adaptation\]) et un réseau sémantique (cf. \[subsec:reseau\]). Le seul composant totalement dédié à la tâche est l’automate. Le réseau sémantique dépend du domaine. Sur un Pentium portable cadencé à 1,7 GHz et doté d’une RAM de 384 Mo, l’intégralité de la chaîne d’adaptation s’exécute en 20’ qui se décomposent en 5’ pour l’apprentissage, et 15’ pour le test, soit une minute par itération du couple adaptation-étiquetage. Le calcul de la cohérence interne demande un temps supplémentaire de 7’ qui reste très raisonnable. Modèles bayésiens {#sec:bayes} ----------------- Guidée par une certaine intuition que nous aurions pu avoir des caractéristiques de la langue et du style de chacun des deux orateurs, une analyse des données d’apprentissage aurait pu nous pousser à retenir certaines de leurs caractéristiques plutôt que d’autres. En premier lieu, il aurait été naturel de tabler sur une caractérisation s’appuyant sur les différences de vocabulaire. Des études anciennes comme celles de [@cotteret:1969] sur le vocabulaire du Général de Gaulle, ou d’autres plus récentes [@labbe:l990] partent du même présupposé. Pour plusieurs raisons, cette approche semble prometteuse mais comme on en rencontre tôt ou tard les limites, on est amené naturellement à ne pas s’en contenter. En effet, la couverture des thématiques abordées par les différents présidents est très large. Il est inévitable que les trajets politiques de deux présidents consécutifs se soient à maintes reprises recoupés. En conséquence, on observe de nombreux points communs dans leurs interventions. On suppose que ces recouvrements viennent s’ajouter les reproductions conscientes ou inconscientes (citations ou effets de mimétisme). ### Modélisation avec lemmatisation Au travers d’une modélisation classique [@manning:2000], nous avons testé quelques points d’appuis comme la longueur des phrases (LL), le pourcentage de conjonctions de subordination (Pcos), d’adverbes (Padv) ou d’adjectifs (Padj) et la longueur moyenne des mots pleins (Plm). Cinq de ces variables (Pcos, Padv, Padj, LL, et Plm) ont été modélisées par des gaussiennes $p_i$ dont les paramètres ont été estimés sur le seul corpus d’apprentissage. En ce qui concerne le vocabulaire lui-même, qu’il s’agisse de lemmes ou de mots, nous avons entraîné sur ce même corpus des modèles [*n*]{}-grammes et [*n*]{}-lemmes (P\#M et P\#L), avec $n < 3$. La probabilité de l’étiquette $t$ (Chirac ou Mitterrand) résulte de la combinaison suivante : $$\label{eq:lambda} P(t)= \sum_{i=1}^r \lambda_i \times p_i(t)$$ avec $\sum_{i=1}^r \lambda_i = 1$. Les valeurs des coefficients $\lambda_i$ que nous avons attribuées de façon empirique à chacune de ces 9 variables $i$ figurent dans le tableau \[tab:exemple\]. L’estimation de ces valeurs a bien entendu, été réalisée sur le corpus d’apprentissage. Comme le montre le tableau \[tab:exemple\], le poids accordé aux lemmes est deux fois plus important que celui accordé aux mots. = 1.3 $i$ P1L P1M Padj LL P2L P2M Pcos Plm Padv ------------- ------ ------ ------ ------ ------ ------ ------ ------ ------ $\lambda_i$ 0,30 0,15 0,15 0,15 0,30 0,15 0,05 0,02 0,01 : Caractères employés pour la modélisation bayésienne et coefficients associés.[]{data-label="tab:exemple"} Lorsqu’on utilise des chaînes de Markov en traitement automatique de la langue naturelle (TALN), on est toujours confronté au problème de la couverture des modèles. Le taux de couverture décroît quand augmente l’ordre du modèle. Le problème est bien connu et des solutions de type lissage ou [*Back-off*]{} [@manning:2000] ; [@katz:1987] sont une réponse classique au fait que le corpus d’apprentissage ne suffit pas à garantir une estimation fiable des probabilités. Le problème devenant critique lorsqu’il y a un déséquilibre flagrant entre les deux classes, il nous a semblé inutile, voire contre-productif de calculer des tri-grammes. En nous inspirant des travaux menés en lexicologie sur les discours de Mitterrand, nous avons essayé de prendre en compte certains des traits qualifiés de dominants chez Mitterrand par [@illouz:2000] : adverbe négatif, pronom personnel à la première personne du singulier, point d’interrogation, ou des expressions comme « c’est », « il y a », « on peut », « il faut » (dans les quatre cas, à l’indicatif présent). Ceux-ci ont été traités de la même façon que les autres caractères de la modélisation bayésienne. Après vérification de la validité statistique de ces traits sur le corpus DEFT’05, nous les avons intégrés dans la modélisation mais dans un second temps, nous les avons retirés car même s’ils entraînaient une légère amélioration sur les données de développement, rien ne garantissait qu’il ne s’agissait pas, là, de tics de langage liés à une période potentiellement différente de celle du corpus de test. Par ailleurs, en cas de portage de l’application à un autre domaine ou une autre langue, nous ne voulions pas être dépendants d’études lourdes. En tous les cas, nous avons préféré faire confiance aux modèles de Markov pour capturer automatiquement une grande partie de ces tournures. ### Modélisation sans lemmatisation Parallèlement et à l’inverse de nos préoccupations de la sous-section précédente, nous avons souhaité faire fonctionner nos modèles sur le texte à l’état brut, sans enrichissement ou annotation. Pour aller dans ce sens, nous nous sommes demandé à quel point la recherche automatisée des caractéristiques propres à un auteur pourrait être facilitée ou perturbée par le fait de ne pas filtrer ni éliminer quoi que ce soit des discours. Ainsi, nous avons fait l’hypothèse que l’utilisation répétée, voire exagérée de certains termes ne servant qu’à assurer le bâti de la phrase, pouvait prétendre au statut d’indicateur fiable. Pour ce deuxième modèle[^6], nous sommes partis du principe que les techniques de [*n*]{}-grammes appliquées à des tâches de classification, pourraient se passer d’une phase préalable de lemmatisation ou de [*stemming*]{}, du rejet des mots-outils et de la ponctuation. Les systèmes [*n*]{}-grammes, [@jalam:2002] ; [@sahami:1999] ont montré que leurs performances ne s’améliorent pas après [*stemming*]{} ou élimination des mots-outils. Dans cet esprit, nous avons laissé les textes dans leur état originel. Aucun prétraitement n’a été effectué, même si cette démarche a ses limites : par exemple, « Gasperi » et « Gaspéri » comptent pour des mots différents, qu’il y ait ou non erreur d’accent ; « premier » et « première » sont aussi comptabilisés séparément en absence de lemmatisation. Malgré cela, nous avons voulu donner au modèle un maximum de chances de capturer des particularités de style (manies de ponctuer le texte par l’emploi de telle ou telle personne, de subjonctifs, gérondifs,…) qui sont gommées après application de certains prétraitements comme la lemmatisation. Une classification naïve et un calcul d’entropie ont déjà été rapportés lors de l’atelier DEFT’05 avec un automate légèrement différent [@el-beze:2005]. Seule variante, l’ajout d’une contrainte : tout mot de longueur $\leq$ 5 n’est pas pris en compte afin d’alléger les calculs. Ceci correspond à un « filtrage » relativement indépendant de la langue. Automate de Markov {#subsec:prise} ------------------ Comme cela était dit en introduction, un discours de Chirac peut avoir fait l’objet de l’insertion d’au plus une séquence de phrases. La séquence [*M*]{}, si elle existe, est d’une longueur supérieure ou égale à deux. Pour prendre en compte cette contrainte particulière, nous avions, initialement, pensé écrire des règles, même si une telle façon de faire s’accorde généralement peu avec les méthodes probabilistes. Dans le cas présent, que faut-il faire si une phrase détachée de la séquence [*M*]{} a été étiquetée [*M*]{}, avec une probabilité plus ou moins élevée (certainement au dessus de 0,5, sinon elle aurait reçu l’étiquette [*C*]{}) ? Renverser la décision, ou la maintenir ? Si l’on opte pour la seconde solution, il serait logique d’extraire également toutes les phrases qui la séparent de la séquence [*M*]{}, bien qu’elles aient été étiquetées [*C*]{}. Mais, dans ce cas, un gain aléatoire en rappel risque de se faire au prix d’une chute de précision. Pour pouvoir trouver, parmi les chemins allant du début à la fin du discours, celui qui optimise la production globale du discours, nous avons exploité un automate probabiliste à cinq états (dont un initial [*I*]{} et trois terminaux, $C_1$, $C_2$, et $M_2$. Comme on peut le voir sur la figure \[fig:markov\], vers les états dénommés $C_1$ et $C_2$ (respectivement $M_1$ et $M_2$) n’aboutissent que des transitions étiquetées [*C*]{} (respectivement [*M*]{}). À une transition étiquetée [*C*]{} (respectivement [*M*]{}), est associée la probabilité d’émission combinant pour [*C*]{} (respectivement [*M*]{}) les modèles probabilistes définis en section \[sec:bayes\]. Avant de décrire les étapes ultérieures du processus de catégorisation segmentation, notons que c’est ce composant qui a permis de faire un saut conséquent (plus de 25% en absolu) au niveau des performances et a ouvert ainsi la voie à la mise en place de procédures d’adaptation décrites en section suivante. S’il s’avère qu’étiqueter un bloc de plusieurs segments est plus fiable qu’étiqueter individuellement chaque phrase, il est naturel que cela ait un impact positif sur les performances. Remarquons par ailleurs que la question aurait pu être gérée autrement, par exemple en utilisant, pour chaque discours, la partie triangulaire supérieure d’une matrice carrée $\Psi[d,d]$ ($d$ étant le nombre de phrases contenues dans le discours en question, voir les figures \[fig:decoupage\] et \[fig:matrice\]). Dans chaque case $\Psi[i,j]$, on calcule la probabilité que la séquence soit étiquetée [*M*]{} entre $i$ et $j$, et [*C*]{} du début jusqu’au $i-1$ et de $j+1$ à $d$. Déterminer les bornes optimales de la séquence Mitterrand revient alors à rechercher un maximum sur toutes les valeurs $\Psi[i,j]$ telles que $i > j$. Si cette valeur optimale est inférieure à celle qu’on aurait obtenue en produisant toute la chaîne avec le modèle associé à Chirac, on se doit de supprimer la séquence [*M*]{}. Sauf si on factorise les calculs pour remplir les différentes cases, la complexité de cette seconde méthode est supérieure à celle de l’algorithme de Viterbi [@manning:2000]. Il nous a paru néanmoins intéressant d’en faire état dès à présent, car elle offre la possibilité de combiner aisément des contraintes globales plus élaborées que celles que nous prenons en compte dans l’adaptation. Elle peut aussi permettre de mixer des modèles issus de l’apprentissage et d’autres optimisant des variables dédiées à la modélisation de la cohésion interne des séquences qui se trouvent dans le discours traité, et n’ont fait l’objet d’aucun apprentissage préalable, comme nous le montrerons en section \[sec:cohérence\]. Adaptation statique et dynamique {#subsec:adaptation} -------------------------------- La contrainte de ne pouvoir enrichir le corpus d’apprentissage, sous peine de disqualification[^7], nous a poussé à tirer un parti intégral des données mises à notre disposition. Or, en dehors du corpus d’apprentissage, il ne restait plus qu’une issue : intégrer dans l’apprentissage (bien entendu, sans les étiquettes de référence) une partie des données de test. C’est sur ces données que l’adaptation a été pratiquée. Dériver un modèle à partir de l’intégralité des discours de test correspond à ce que nous appelons ici [*adaptation statique*]{}. L’[*adaptation dynamique*]{}, quant à elle, repose sur un modèle découlant seulement du discours en train d’être testé. Évidemment, il n’est pas interdit de conjuguer les deux approches. Dans un premier temps, nous avions envisagé de pratiquer un étiquetage des données de test, l’objectif étant à l’itération $i+1$ de n’adjoindre au corpus d’apprentissage[^8] de $X$ que les phrases $s$ ayant reçu au pas $i$ une probabilité $P_i(X|s)$ supérieure à un certain seuil $T_i$. Un apprentissage de type maximum de vraisemblance effectué sur les données ainsi collectées peut autant rapprocher qu’éloigner du point optimal. Pour pallier cette difficulté, nous avons opté pour un apprentissage d’[*Expectation-Maximisation*]{} (EM), consistant à ne compter pour chaque couple { élément $= e, X \}$ observé dans les données d’adaptation que la fraction d’unité égale à la probabilité de l’orateur $X$ sachant la phrase qui contient $e$. La prise de décision repose sur une formule analogue à celle de la formule (\[eq:lambda\]). La variable en position 0 est la probabilité de l’étiquette sachant la phrase qui lui a été attribuée à l’itération $i$. Nous avons fait décroître le poids $\lambda_0$ qui lui est associé, de façon progressive, d’une itération à l’autre par pas de 0,1. Les quatre modèles employés sont, pour les deux premiers, lemmes et mots issus de l’adaptation locale (dynamique), pour les deux derniers, lemmes et mots issus de l’adaptation globale (statique). La pondération entre les différentes probabilités est restée la même durant toutes les itérations : Dynamique { lemmes = 0,4 ; mots = 0,1 } / Statique { lemmes = 0,4 ; mots = 0,1 }. Les procédures d’adaptation statique et dynamique mises en œuvre durant cette étape ont permis de gagner entre 3 et 4 points de [*Fscore*]{}. Réseau de Noms Propres {#subsec:reseau} ---------------------- À partir de la tâche T2, l’ensemble des noms propres était dévoilé aux participants. Établir un lien entre différents éléments apparaissant dans des phrases même éloignées d’un discours donné, nous a paru être un bon moyen pour mettre en évidence une sorte de réseau sémantique permettant aux segments de s’auto-regrouper autour d’un lieu, de personnes et de façon implicite d’une époque. Dans le cas de données bien séparables, plusieurs ensembles de noms ancrés dans une Histoire et une Géographie commune devraient former des composantes connexes (idéalement deux) sur lesquelles il suffirait ensuite de mettre l’étiquette [*M*]{} ou [*C*]{}. Bien que cela ne soit pas tout à fait la démarche que nous avons adoptée, ces remarques aident à en comprendre l’esprit. 2 171 termes ont été regroupés dans 314 « concepts »[^9] qui pour épouser la richesse des discours traités dépassent largement un cadre restreint aux seules considérations géopolitiques (le Sport et la Culture sont souvent abordés lors de cérémonies de remises de médailles). Un terme peut se retrouver dans plusieurs classes, comme par exemple «  Miguel Angel Asturias », qui a été placé aussi bien dans la classe des Guatémaltèques que dans celle des écrivains étrangers. Afin de mixer les relations entretenues entre les noms de pays, leurs habitants, les capitales, le pouvoir exécutif, nous avons complété un réseau fourni par le Centre de Recherche de Xerox[^10], en y rajoutant quelques relations issues des Bases de Connaissance que l’équipe TALN du LIA utilise pour faire fonctionner son système de Questions / Réponses [@bellot:2003]. En table \[tab:extrait\], figure un petit extrait de ce réseau non structuré. = 1.6 CONCEPT TERMES --------------- -------------------------------------------------------------------- Argentin Argentine Alfonsin Carlos\_Menem Bioy\_Casares Buenos\_Aires Alfredo\_Arias Jorge\_Remes Guatemalteque Guatemala\_Ciudad Guatemala Guatémaltèque Guatémaltèques Guatémaltais Guatémaltaise Guatémaltaises Ciudad\_Vieja Permedo **Miguel\_Angel\_Asturias** Alvaro Arzu Irigoyen Rigoberta\_Menchu Rigoberta Mexicain Mexique Mexico Zedillo Zédillo Benito\_Juarez Carlos\_Fuentes Octavio\_Paz FOX Fox Cancun Monterrey Mexicain Mexicaine Mexicaines Mexicains Ecrivains\_e Gao Saramago Virgilio\_Ferreira Fernando\_Pessoa Ionesco Cioran Fukuzawa\_Yukichi Nadia\_Tueni Amin\_Maalouf Tahar\_Ben\_Jelloun Senghor Rachid\_Boujedra Bioy\_Casares Boubou\_Hama Hector\_Bianciotti **Miguel\_Angel\_Asturias** Dostoïevsky : Extrait du réseau de noms propres.[]{data-label="tab:extrait"} On enrichit les segments en leur ajoutant les concepts auxquels appartiennent les termes qui les composent. Deux segments qui ont en commun plus d’un certain nombre d’éléments (termes ou concepts) sont considérés 2 à 2 comme liés thématiquement et mis dans une même classe de segments. Les classes sont élargies par une itération de ce processus. La probabilité de chaque segment est combinée avec la probabilité de la classe à laquelle il appartient. Après quatre itérations, sur 57 301 phrases valides que comptait le corpus de développement (test : 27 120), 6 285 ont été regroupées en 942 groupes (test : 456 groupes de 3 127). Un peu plus de 11% des segments se retrouvent donc dans des groupes, dont le cardinal moyen est d’environ 7 phrases. Le plus grand groupe contient 50 segments (test : 66). Seuls 16 groupes (test : 12) regroupent, de façon confuse, des étiquettes [*M*]{} et [*C*]{}. C’est le cas du discours 38, où la phrase 30 étiquetée [*M*]{} possède en commun « Casablanca MAGHREB » (en fait, il s’agissait du sommet de Casablanca) avec la phrase 173 étiquetée [*C*]{}, où Chirac fait état de ses récents voyages au Maroc. L’avantage d’un réseau probabiliste est que cette erreur n’est pas rédhibitoire. En effet, dans notre soumission, la phrase 30 a été correctement extraite et non la phrase 173. Cela ne fonctionne pas toujours aussi bien ! Dans le cas du discours 739, la séquence [*C*]{} et la séquence [*M*]{} ont en commun deux « termes-concepts » (« Espagne-Espagnol » et « Méditerranée-Méditerranéen »). Il se trouve que la seconde confusion aurait pu être évitée si le TGV Paris-Lyon-Méditerranée dont parle Mitterrand n’avait pas fait l’objet d’une sur-découpe au moment de la [*tokenisation*]{}. Mais cela n’aurait pas suffi, car avec l’aide de l’autre terme (Espagne) quatre phrases [*M*]{} (30, 35, 36 et 37) ont été regroupées par transitivité avec 12 phrases étiquetées [*C*]{} (1, 3, 6-17, 20-25, 27, 47). De fait, aucun segment du discours 739 n’a été extrait. Il est clair que nous sommes encore loin d’une représentation élaborée des relations entretenues entre des concepts et leur expression au travers de textes. Néanmoins, le réseau que nous avons élaboré à peu de frais est un premier pas dans cette direction. Cohésion thématique des discours {#sec:cohérence} ================================ En section \[subsec:prise\], nous avions avancé l’hypothèse qu’étiqueter un bloc est plus fiable qu’étiqueter chaque phrase de façon indépendante l’une de l’autre. Cela se discute en fait si on se borne à rechercher la suite de segments qui optimise la cohésion thématique[^11] de chacun des deux blocs, il est indispensable de conjuguer cette approche thématique avec un étiquetage d’auteur. Cette étape est motivée par une idée simple découlant des présupposés de DEFT’05 : « *Les passages de F. Mitterrand introduits traitent d’une thématique différente. Par exemple, dans les allocutions de J. Chirac évoquant la politique internationale, les phrases de F. Mitterrand introduites sont issues de discours traitant de politique nationale. Ainsi, la rupture thématique peut être une des manières de détecter les phrases issues du corpus de F. Mitterrand.* »[^12] Dans cette optique, on peut vouloir trouver un découpage de chaque discours soit en un bloc ([*C…C*]{}), soit en deux blocs ([*C …C-M …M*]{}) ou ([*M…M-C…C*]{}) soit en trois blocs ([*C…C-M…M-C…C*]{}) tels que le bloc des segments étiquetés [*M*]{} ou les blocs (1 ou 2) étiquetés [*C*]{} présentent tous les deux une cohérence thématique interne optimale. Pour cela, nous proposons de formaliser le problème comme suit : la probabilité de production d’une phrase est évaluée au moyen d’un modèle appris sur toutes les phrases du bloc auquel elle appartient sauf elle. En maximisant le produit des probabilités d’émission de toutes les phrases du discours, on a toutes les chances de bien identifier des ruptures thématiques. Mais rien ne garantit qu’elles correspondent à des changements d’orateurs. En effet, supposons qu’il n’y ait pas dans un discours donné, d’insertion de phrases de Mitterrand, et que dans les discours de Chirac se trouve une longue digression de 20 phrases. Notre approche risque de reconnaître à tort ces 20 phrases comme attribuables à la classe [*M*]{}. Pour éviter ce travers, nous proposons une optimisation mettant en œuvre conjointement les modèles de cohérence interne et ceux issus de l’apprentissage. En annexe, nous donnons en exemple le discours 520 pour lequel ce phénomène se produit. Nous voyons comment la cohérence interne réussit à renverser presque totalement la situation : ainsi, un gros bloc étiqueté [*M*]{} par l’adaptation seule, au sein d’un discours dont la classe est [*C*]{} a été étiqueté correctement par la cohérence interne, à l’exception de deux phrases dont les probabilités penchaient trop fortement vers la classe [*M*]{}. Le modèle de cohérence interne cherche donc à maximiser la probabilité d’appartenance des phrases proches aux frontières de segments. Il peut utiliser *a)* le réseau de noms propres et *b)* la probabilité issue de l’apprentissage par Markov. Pour un discours $S^{d}_1$ donné, de longueur [*d*]{}, nous cherchons un découpage optimal $\widetilde{D}$ (cf. figure \[fig:decoupage\]) et un étiquetage $\widetilde{E}$ tels que : $$(\widetilde{D},\widetilde{E}) = \textrm{Arg}\max_{D,E} \left\{P_{I}(D,E|S^{d}_{1}) \times P'(D,E|S^{d}_{1})\right\}$$ où $P'(D,E|S^{d}_{1})$ est la probabilité issue de l’apprentissage et $P_{I}(D,E|S^{d}_{1})$ la probabilité de cohérence interne (à l’intérieur d’un discours). La conjugaison des modèles d’apprentissage et de cohérence interne est réalisée par le produit entre $P'$ et $P_{I}$, qu’il semble légitime de considérer indépendants l’un de l’autre. Comme le découpage ne peut être déduit de l’apprentissage, nous faisons l’hypothèse que $P'(D,E|S^{d}_{1}) \cong P'(E|S^{d}_{1})$. Donc : $$\label{eq:DE} (\widetilde{D},\widetilde{E}) = \textrm{Arg}\max_{D,E} \left\{P_{I}(D,E|S^{d}_{1}) \times P'(E|S^{d}_{1})\right\}$$ Or, d’après le théorème de Bayes : $$P_{I}(D,E|S^{d}_{1}) = \frac{P(S^{d}_{1}|D,E)P(D|E)}{P(S^{d}_{1})} \mathrm{\ \ et \ \ } P'(E|S^{d}_{1}) = \frac{P'(S^{d}_{1}|E)P'(E)}{P'(S^{d}_{1})}$$ De ce fait, l’équation (\[eq:DE\]) devient : $$(\widetilde{D},\widetilde{E}) \cong \textrm{Arg}\max_{D,E} \left\{\frac{P(S^{d}_{1}|D,E)P(D|E)}{P(S^{d}_{1})} \times \frac{P'(S^{d}_{1}|E)P'(E)}{P'(S^{d}_{1})}\right\}$$ Nous savons que $P(D|E)$ prend toujours des valeurs {0, 1} car le découpage est toujours déterminé par les étiquettes (mais pas vice-versa). La probabilité $P'(E)$ ne peut pas être déduite de l’apprentissage (le choix de $D$ peut être considéré comme aléatoire) et $P(S)$ et $P'(S)$ ne dépendent pas de $D$ ou de $E$. Alors : $$(\widetilde{D},\widetilde{E}) \cong \textrm{Arg}\max_{D,E} \left\{ P_{I}(S^{d}_{1}|D,E) \times P'(S^{d}_{1}|E)\right\}$$ ![Schéma de découpage des discours.[]{data-label="fig:decoupage"}](decoupage){width="7cm"} Nous avons choisi de représenter un couple $(D,E)$ par un couple de deux indices $i$ et $j$ dont la signification est donnée par la figure \[fig:decoupage\]. Ces deux indices correspondent aux bornes du bloc des segments étiquetés $M$ et à la ligne et la colonne de la matrice $\Psi$ évoquée en section \[subsec:prise\]. Pour un discours donné, on aura donc : $$\Psi[i,j] = P(S_{1 \ldots i-1,j+1 \ldots d}|C) \times P(S_{i \ldots j}|M) \times P'(S_{1 \ldots i-1,j+1 \ldots d}|C) \times P'(S_{i \ldots j}|M)$$ En faisant l’hypothèse[^13] que les segments sont indépendants, nous introduisons le produit sur toutes les phrases du discours en distinguant celles qui sont à l’intérieur du bloc $S_{i \rightarrow j} (k=i \ldots j)$ de celles qui sont à l’extérieur $(k=1 \ldots i-1, j+1 \ldots d)$ : $$\Psi[i,j] = \prod\limits_{ k=1 \ldots i-1,j+1 \ldots d }[P(S_{k}|\chi) \times P'(S_{k}|C)] \times \prod\limits_{k=i \ldots j}[P(S_{k}|\mu) \times P'(S_{k}|M)]$$ où $d$ est la longueur du discours et $\chi = C \setminus S_{k}$ et $\mu = M \setminus S_{k}$. Ceci revient à exclure le segment $S_k$ des données qui servent à estimer les paramètres utilisés pour calculer la probabilité de production de ce même segment $S_k$. Notons que, si les probabilités $P(S_{k}|\chi)=1$ et $P(S_{k}|\mu)=1$, alors la valeur de $\Psi[i,j]$ est réduite au cas de Markov (adaptation simple). Nous avons exploité la matrice $\Psi[i,j]$ (cf. figure \[fig:matrice\]) en nous limitant à sa partie triangulaire supérieure. Le fait d’exclure la diagonale principale dans les calculs illustre l’exploitation de la contrainte respectée par les fournisseurs du corpus DEFT’05. S’il y a des segments de la classe [*M*]{} insérés, il y a en au moins deux. Le cas des discours étiquetés uniquement [*C*]{} n’est pas représenté dans la figure, mais il a été pris en considération, même s’il n’est pas intégré dans la matrice. = 2.6 1 ... $i$ ... $d$ ---------- ----------- ----------- ----------- -------------- ----------- ----------- $\vdots$ $\bullet$ $\vdots$ $j$ $\bullet$ $\bullet$ $P(S_i,S_j)$ $\bullet$ $\bullet$ $\bullet$ $\vdots$ $\vdots$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\vdots$ $d$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ $\bullet$ Expériences {#sec:expériences} =========== Pour la Modélisation I, tous les corpus (apprentissage et test) ont été traités par l’ensemble d’outils LIA\_TAGG[^14]. Ces outils contiennent les modules suivants : - un module de formatage de texte permettant de découper un texte en unités (ou [*tokens*]{}) en accord avec un lexique de référence ; - un module de segmentation insérant des balises de début et fin de phrase dans un flot de texte, en accord avec un certain nombre d’heuristiques ; - un étiqueteur morphosyntaxique, basé sur l’étiqueteur ECSTA [@spriet:1998] ; - un module de traitement des mots inconnus permettant d’attribuer une étiquette morphosyntaxique à une forme inconnue du lexique de l’étiqueteur en fonction du suffixe du mot et de son contexte d’occurrence. Ce module est basé sur le système DEVIN présenté dans [@spriet:1996]. - un lemmatiseur associant à chaque couple mot/étiquette morphosyntaxique un lemme en fonction d’un lexique de référence. Dans la phase de développement, le corpus d’apprentissage a été découpé en cinq sous-corpus de telle sorte que pour chacune des cinq partitions, un discours appartient dans son intégralité soit au test soit à l’apprentissage. À tour de rôle, chacun de ces sous-corpus est considéré comme corpus de test tandis que les quatre autres font office de corpus d’apprentissage. Cette répartition a été préférée à un tirage aléatoire des phrases tolérant le morcellement des discours. En effet, un tel tirage au sort présente deux inconvénients majeurs. Le premier provient du fait qu’un tirage aléatoire peut placer dans le corpus de test des segments très proches de segments voisins qui eux ont été placés dans le corpus d’apprentissage. Le second inconvénient (le plus gênant des deux), tient au fait qu’une telle découpe ne permet pas de respecter le schéma d’insertion défini dans les spécificités de DEFT’05. Résultats de l’adaptation ------------------------- Des résultats de nos modèles uniquement avec adaptation ont été publiés dans les actes du colloque TALN 2005. Nous reproduisons ici les observations majeures qui pouvaient être faites sur ces résultats. Le [*Fscore*]{} s’améliore de façon notable au cours des cinq premières itérations de l’adaptation. Au-delà, il n’y a pas à proprement parler de détérioration mais une stagnation qui peut être vue comme la captation par un maximum local. L’apport des réseaux bâtis autour des noms propres est indéniable [@el-beze:2005]. Nous montrons au tableau \[tab:resultats\] et sur la figure \[fig:officiels\] les meilleurs [*Fscore*]{} officiels soumis pour l’ensemble de participants pour les trois tâches. On peut voir que le système du LIA senior est positionné, dans les trois cas, en première place. La méthode de [@rigouste:2005] en deuxième position, utilise quelques méthodes probabilistes semblables aux nôtres, mais ils partent de l’hypothèse où la segmentation thématique est faite au niveau des orateurs (pas au niveau du discours), ils ont besoin de pondérer empiriquement les noms et les dates (tâches T2 et T3), leurs machines de Markov sont plus complexes et il ne font pas d’adaptation, entre autres. Le dévoilement des dates (tâche T3) permet d’améliorer très légèrement les résultats du modèle II, mais entraîne une dégradation sur le modèle I. En ce qui concerne la précision et le rappel au fil des itérations sur l’ensemble de Test(T) ainsi que sur le Développement(D), c’est le gain en précision qui explique l’amélioration due aux Noms Propres. Ce gain allant de pair avec un rappel quasi identique (légèrement inférieur pour le test), il apparaît que le composant Noms Propres fonctionne comme un filtre prévenant quelques mauvaises extractions (mais pas toutes). = 1.6 Equipes T1 T2 T3 ------------------------------------------------------- ---------- ---------- ---------- **1 El-Bèze, Torres, Bechet : LIA senior** **0,87** **0,88** **0,88** 2 Rigouste, Cappé, Yvon : ENST 0,86 0,85 0,87 3 Pierron, Durkal, Freydigier : LORIA/UHP 0,82 0,82 0,82 4 Labadie, Romero, Sitbon : LIA junior 0,76 0,74 0,75 5 Maisonnasse, Tambellini : CLIPS 0,75 0,75 0,76 6 Kerloch, Gallinari : LIP6 0,73 0,79 0,79 7 Hernandez : LIMSI 0,56 0,56 0,57 8 Plantié, Dray, Montmain, Meimouni, Poncelet : LGI2P 0,49 0,52 0,51 9 Hurault-Plantet, Jardino, Illouz : LIMSI 0,49 0,56 0,56 10 Chauche : LIRMM 0,32 0,31 0,31 11 Henry, Marley, Amblard, Moot : LABRI 0,18 0,18 0,42 : Résultats officiels de l’atelier DEFT’05 du [*Fscore*]{} sur les trois tâches de test {T1, T2, T3} pour les meilleures soumissions de l’ensemble des 11 équipes.[]{data-label="tab:resultats"} ![[*Fscore*]{} officiels pour les trois tâches (T1 : pas de noms, pas de dates ; T2 : pas de noms et T3 : avec noms et dates) pour l’ensemble de participants DEFT’05. Les membres des équipes sont cités au tableau \[tab:resultats\].[]{data-label="fig:officiels"}](officiels.pdf){width="6.5cm"} Résultats avec la cohérence interne ----------------------------------- Les résultats ont été améliorés grâce à la recherche d’une cohérence interne des discours. Cette étape intervient après application de l’automate markovien et avant la phase d’adaptation. Nous montrons, sur les figures \[fig:tache1I\], \[fig:tache2I\] et \[fig:tache3I\], le [*Fscore*]{} obtenu pour les trois tâches à l’aide d’une adaptation plus la cohérence interne pour les corpus de Développement (D) et de Test (T). Dans tous les cas, l’axe horizontal représente les itérations de l’adaptation. Sur les graphiques, la ligne pointillée correspond aux valeurs du [*Fscore*]{} obtenues avec l’adaptation seule et les lignes continues à celles de la cohérence interne (une itération : ligne grosse ; deux itérations : ligne fine). Pour les trois tâches, on observe une amélioration notable du modèle de cohérence interne par rapport à celui de l’adaptation seule. Enfin, la valeur la plus élevée de [*Fscore*]{} est à présent obtenue pour la tâche T3 (figure \[fig:tache3I\]), à un niveau de 0,925. Ce score dépasse largement le meilleur résultat ([*Fscore*]{} = 0,88) atteint lors du défi DEFT’05. Les valeurs précises des courbes sont rapportées dans les tableaux \[tab:modeleI\] et \[tab:test1I\]. = .6 ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- It Adaptation Cohérence puis Adaptation Cohérence puis Adaptation Cohérence puis seule adaptation seule adaptation seule adaptation 0 0,841 0,850 0,853 0,852 0,855 0,857 1 0,845 0,871 0,863 0,874 0,864 0,879 2 0,853 0,872 0,863 0,875 0,863 0,880 3 0,857 0,875 0,867 0,877 0,868 0,882 4 0,860 0,877 0,868 0,880 0,868 0,885 5 **0,862** **0,877** **0,868** **0,881** **0,870** **0,887** ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- : Modèle I [*Fscore*]{} Développement : Adaptation seule et Cohérence interne.[]{data-label="tab:modeleI"} ![[*Fscore*]{} tâche T1 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache1I"}](tache1D.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T1 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache1I"}](tache1T.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T2 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache2I"}](tache2D.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T2 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache2I"}](tache2T.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T3 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache3I"}](tache3D.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T3 Modèle I / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache3I"}](tache3T.pdf "fig:"){width="6cm"} = .6 ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- It Adaptation Cohérence puis Adaptation Cohérence puis Adaptation Cohérence puis seule adaptation seule adaptation seule adaptation 0 0,856 0,870 0,863 0,887 0,864 0,887 1 0,862 0,885 0,868 0,902 0,884 0,909 2 0,865 0,888 0,868 0,906 0,884 0,915 3 0,867 0,892 0,866 0,912 0,887 0,920 4 0,869 0,894 0,869 0,914 0,890 0,923 5 **0,870** **0,899** **0,874** **0,917** **0,897** **0,925** ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- : Modèle I $Fscore$ Test : Adaptation seule et Cohérence interne.[]{data-label="tab:test1I"} Les figures \[fig:tache1II\], \[fig:tache2II\] et \[fig:tache3II\] montrent, avec la même convention que les figures précédentes, le [*Fscore*]{} pour le modèle II, où n’ont été appliqués ni filtrage ni lemmatisation. Ici encore, les valeurs les plus élevées sont obtenues pour la tâche T3, avec un [*Fscore*]{} = 0,873. La comparaison avec les performances ([*Fscore*]{} = 0,801 pour la tâche T3) de ce même modèle que nous avons employé lors du défi DEFT’05, est avantageuse : sept points de plus[^15]. Cette amélioration est due essentiellement à la cohérence interne et permet d’approcher la meilleure valeur ([*Fscore*]{} = 0,881 rapporté dans [@el-beze:2005]) qui avait été obtenue avec l’adaptation seule et un filtrage et lemmatisation préalables. Bien que l’utilisation de ce modèle soit un peu moins performante (et de ce fait contestée), nous pensons qu’il peut être utile d’y recourir, si l’on veut éviter la lourdeur des certains processus de prétraitement. = .6 ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- It Adaptation Cohérence puis Adaptation Cohérence puis Adaptation Cohérence puis seule adaptation seule adaptation seule adaptation 0 0,834 0,867 0,842 0,868 0,844 0,867 1 0,844 0,881 0,845 0,881 0,846 0,881 2 0,847 0,881 0,848 0,882 0,849 0,883 3 0,850 0,881 0,851 0,882 0,851 0,883 4 0,854 0,881 0,855 **0,883** 0,854 **0,883** 5 **0,857** **0,882** **0,856** 0,882 **0,857** 0,882 ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- : Modèle II $Fscore$ Développement : Adaptation seule et Cohérence interne.[]{data-label="tab:modeleII"} ![[*Fscore*]{} tâche T1 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache1II"}](tache1D-II.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T1 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache1II"}](tache1T-II.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T2 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache2II"}](tache2D-II.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T2 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache2II"}](tache2T-II.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T3 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache3II"}](tache3D-II.pdf "fig:"){width="6cm"} ![[*Fscore*]{} tâche T3 Modèle II / Adaptation vs Cohérence interne / corpus D et T.[]{data-label="fig:tache3II"}](tache3T-II.pdf "fig:"){width="6cm"} = .6 ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- It Adaptation Cohérence puis Adaptation Cohérence puis Adaptation Cohérence puis seule adaptation seule adaptation seule adaptation 0 0,856 0,857 0,853 0,860 0,852 0,859 1 0,855 0,873 0,853 0,877 0,852 0,872 2 0,854 0,873 0,853 0,877 0,853 0,873 3 0,856 **0,873** 0,855 0,877 0,854 0,874 4 0,860 0,872 **0,859** 0,877 0,856 0,874 5 **0,860** 0,871 0,858 **0,877** **0,856** **0,874** ---- ------------ ---------------- ------------ ---------------- ------------ ---------------- -- -- : Modèle II $Fscore$ Test : Adaptation seule et Cohérence interne.[]{data-label="tab:testII"} Fusion de méthodes ------------------ Le dernier test que nous avons réalisé fait appel à une fusion de l’ensemble des modèles. Nous avons appliqué un algorithme de vote sur presque toutes les hypothèses issues des modèles I et II. Les hypothèses (qui vont faire office de juges) proviennent des différentes itérations de l’adaptation, avec ou sans cohérence interne. Nous avons tenu compte des avis d’un nombre de juges donné ([*NbJ*]{}), en pondérant l’avis de chaque juge [*j*]{} par un poids $\alpha _{j}$ de telle sorte que le critère de décision final est le suivant : $$\theta_i = signe \left( \sum^{NbJ}_{j=1} \alpha_{j} \xi_{i,j} - \delta \right)$$ Si $\theta_i$ est négatif alors l’étiquette du segment $i$ sera [*C*]{} ; [*M*]{} autrement. Avec $\alpha_j \in \Re$, $\beta_{i,j} \in \{0,1\}$ et la convention 0 = $C$ et 1 = $M$. La stratégie est la suivante : afin d’avoir un degré de confiance suffisant, il faut retenir les segments auxquels une majorité de juges attribue l’étiquette $M$. Les paramètres $\alpha_j$ et $\delta$ ont été ajustés pour minimiser le nombre d’erreurs sur l’ensemble de développement (D). Nous nous sommes proposés de voir cette estimation comme un problème de classification à $NbJ$ entrées et une sortie, c’est-à-dire, comme un problème d’apprentissage supervisé. Nous avons ainsi défini un exemple d’apprentissage comme le vecteur binaire $\xi_j = \{0,1\}, j=1,\cdots,NbJ$. La sortie (classe de référence) de cet exemple est un scalaire $\tau=\{-1,1\}$ ($-1$ pour la classe $C$, $+1$ pour la classe $M$). L’ensemble d’apprentissage est donc constitué de $S$ segments et $NbJ$ juges, et nous le dénoterons par $\aleph = \{\vec\xi_i,\tau_i\}; i=1,\cdots,S$. Trouver les poids $\alpha_j$ correspond donc à trouver les $j$ poids d’un perceptron entraîné sur l’ensemble $\aleph$. Nous avons utilisé un perceptron optimal à recuit déterministe[^16] entraîné par l’algorithme Minimerror [@Minimerror; @Monoplane; @torres-moreno:2002], où l’apprentissage garantit que si l’ensemble $\aleph$ est linéairement séparable, l’algorithme trouve la solution optimale (marge maximale de séparation) et s’il ne l’est pas (comme cela semble être le cas ici), il trouve une solution qui minimise le nombre de fautes commises. Ainsi, nous avons trouvé un seuil $\delta$ = 8,834 et les poids $\alpha_j$ avec un nombre de juges $NbJ$ = 89. Pour l’ensemble de développement, les résultats obtenus au moyen de cette fusion sont encore meilleurs qu’avec les autres méthodes. Nous obtenons, dans ce cas, un [*Fscore*]{} = 0,914 avec une précision de 0,916 et un rappel de 0,911. Cependant, pour l’ensemble de test, la fusion ne dépasse pas le meilleur résultat obtenu jusqu’à présent. En effet, on atteint un [*Fscore*]{} = 0,914, avec une précision de 0,892 et un rappel de 0,937. Il est connu que les perceptrons (et les réseaux de neurones en général) trouvent parfois des valeurs de poids trop bien adaptées à l’ensemble d’apprentissage (phénomène de sur-apprentissage). Le fait de n’avoir pas eu de meilleurs résultats sur le test le confirme. Cependant, nous pensons que si les $\vec\xi_{i}$ étaient des probabilités au lieu d’être des 0 et des 1, on aurait pu observer un meilleur comportement. Analyse des erreurs ------------------- Nous avons analysé les erreurs commises par notre système. Sur un total de 27 163 phrases de l’ensemble de Test de la tâche T3, le Modèle I avec la méthode d’adaptation et la cohérence interne des discours, a fait un total de 578 erreurs ([*Fscore*]{} = 0,925) : - 233 erreurs de la classe [*C*]{} (faux négatifs assimilés au rappel), dont : - 37 phrases [*C*]{} à la frontière inversée ($\approx$ 16%) ; - 113 phrases [*C*]{} en blocs ($\approx$ 49%) ; - 83 phrases [*C*]{} entre blocs [*C*]{} ($\approx$ 36%) ; - 345 erreurs de la classe [*M*]{} (faux positifs assimilés à la précision), dont : - 35 phrases [*M*]{} à la frontière inversée ($\approx$ 10%) ; - 126 phrases [*M*]{} en blocs ($\approx$ 37%) ; - 184 phrases [*M*]{} insérées dans 21 discours de classe [*C*]{} exclusive ($\approx$ 53%). Le problème le plus grave concerne la précision (59% du total des erreurs), et ici, la plus grande majorité (53% de faux positifs) est due aux insertions des phrases [*M*]{} dans des discours de classe [*C*]{}[^17]. L’autre problème se présente dans les 126 phrases en blocs inversés (37%). Ces problèmes sont peut-être dus à l’utilisation de la cohérence interne : sur le tableau \[tab:precisionT3MI\], on voit qu’en adaptation seule, la précision est toujours plus élevée que le rappel (en D comme en T). Pour la cohérence interne, la situation est inversée : le rappel est bien meilleur que la précision. Le même comportement a été retrouvé dans le Modèle II. Un autre pourcentage important d’erreurs (49% de faux positifs) a lieu dans l’inversion d’un nombre important de blocs (113 phrases). Enfin, une autre partie non négligeable (10% de faux positifs, 16% de faux négatifs) correspond à l’inversion de catégorie d’une phrase unique à la frontière des découpages (soit $i$ ou $j$, voir figure \[fig:decoupage\]). La détection de cette frontière, reste un sujet très délicat avec nos approches. = 1.1 --------------- ----------- -------- ----------- ----------- Précision Rappel Précision Rappel Développement **0,918** 0,826 0,882 **0,892** Test **0,931** 0,866 0,912 **0,939** --------------- ----------- -------- ----------- ----------- : Précision et Rappel pour la tâche T3, Modèle I, adaptation et cohérence interne à la dernière itération de l’adaptation.[]{data-label="tab:precisionT3MI"} La figure \[fig:precisionM3T3\] montre les courbes de Précision et de Rappel. Pour rester concis, nous allons présenter ici seulement les résultats correspondant à la tâche T3 du modèle I. La cohérence interne est affichée uniquement sur la première itération. Dans les deux cas, nous montrons des résultats sur les corpus de Développement (D) et Test (T). On voit que sur l’ensemble de test T, la précision de la cohérence interne puis adaptation est moins élevée que celle de l’adaptation seule. La même situation se produit pour le rappel. Néanmoins, un phénomène d’inversion se présente en développement : en rappel on est plus performant avec la cohérence qu’avec l’adaptation seule, et en précision avec l’adaptation seule que avec la cohérence. ![Précision-Rappel Modèle I / tâche T3 : Adaptation vs Cohérence interne.[]{data-label="fig:precisionM3T3"}](precisionM3T3 "fig:"){width="5.7cm"} ![Précision-Rappel Modèle I / tâche T3 : Adaptation vs Cohérence interne.[]{data-label="fig:precisionM3T3"}](rappelM3T3 "fig:"){width="5.7cm"} Nous avons aussi calculé la longueur moyenne (en mots) des segments mal classés (tâche T3 / modèle I) suivant le type d’erreur (cf. figure \[fig:erreurs\]) : discours exclusifs de la classe [*C*]{}, longueur moyenne $\approx$ 21. Erreurs de début du bloc : type 1 longueur moyenne $\approx$ 21 ; type 2 longueur moyenne $\approx$ 22. Erreurs de fin du bloc : type 3 longueur moyenne $\approx$ 21 ; type 4 longueur moyenne $\approx$ 26. Cependant, il est difficile de tirer de conclusions à partir de cette information. ![Types d’erreurs frontière de bloc.[]{data-label="fig:erreurs"}](erreurs.pdf){width="10cm"} Enfin, de façon comparative les résultats du [*Fscore*]{} sur les trois tâches sans adaptation, montrés au tableau \[tab:fscoreSA\], confirment l’importance de l’utilisation de l’automate de Markov : il fait un gain global de $\approx$ 25% avec les deux modèles proposés. = 1.1 ----------- ------- ------- ----------- ------- ------- ----------- T1 T2 T3 T1 T2 T3 Modèle I 0,570 0,570 **0,574** 0,593 0,595 **0,596** Modèle II 0,549 0,551 **0,555** 0,581 0,582 **0,585** ----------- ------- ------- ----------- ------- ------- ----------- : [*Fscore*]{} pour les trois tâches sans adaptation.[]{data-label="tab:fscoreSA"} Conclusions et perspectives {#sec:conclusions} =========================== Nous avons introduit une formalisation de la cohérence interne des discours qui a beaucoup amélioré nos résultats rapportés en [@el-beze:2005]. Cette cohérence ainsi que l’adaptation ont été combinées conjointement avec les modèles d’apprentissage préalablement développés, comme la modélisation bayésienne qui semble déterminant, l’automate de Markov et des processus d’adaptation. Les résultats que nous obtenons pour la tâche T3 avec la cohérence interne en terme de [*Fscore*]{} = 0,925 sont très encourageants. Cependant, l’utilisation de la cohérence interne présente un risque : quelques phrases avec une thématique différente, peuvent faire basculer tout un bloc vers l’autre étiquette. Ce type de comportement local entraîne des instabilités globales (semblables à ce qui se produit dans le jeu « Reversi »), dont la prévision reste très difficile, ayant comme conséquence une baisse générale des performances. Ne pas lemmatiser et ne rien filtrer dégrade un peu les performances ([*Fscore*]{} = 0,874 avec le modèle II) mais permet d’éviter l’application d’un processus additionnel de prétraitement qui pour certaines langues peut être relativement lourd. La fusion des hypothèses vue comme un vote de plusieurs juges pondérés par un perceptron optimal a permis de surpasser les résultats précédents en développement ([*Fscore*]{} = 0,914). Cependant nous pensons qu’il reste encore du travail pour améliorer cette stratégie afin d’obtenir de meilleures performances en test. Des études comme celle de [@rigouste:2005] sur le même corpus confirment que l’utilisation de méthodes probabilistes est la mieux adaptée à ce type de segmentation thématique. Le recours à un réseau de Noms Propres est utile et nous encourage par la suite à employer une ressource lexicale comme [@eurowordnet:1998] pour tirer parti de réseaux sur les noms communs. Pour s’affranchir des contraintes liées à la constitution d’une ressource sémantique, il serait judicieux de recourir à des approches telles que *Latent Semantic Analysis* [@deerwester:1990] ou PLSA [@hofmann:1999]. D’autres perspectives d’application, comme celle de la séparation de thèmes sont aussi envisageables. Il faut reconnaître, cependant, que s’il s’était agi de traiter un texte composite moins artificiel que celui proposé par DEFT, par exemple un dialogue, la difficulté aurait été accrue. Des frontières thématiques ne coïncident pas forcément avec des débuts de phrase. Les thèmes peuvent s’entremêler et composer un tissu discursif où les fils sont enchevêtrés de façon subtile. Beaucoup reste à faire pour pouvoir différencier plusieurs orateurs ou plusieurs thèmes comme envisagé dans le cadre du Projet Carmel [@chen:2005]. Annexe {#annexe .unnumbered} ====== Nous souhaitons illustrer le fonctionnement du modèle de cohérence interne en rapportant ici une étude sur le discours 520 du corpus de test, dont toutes les phrases appartiennent à la classe $C$ (Chirac). Le tableau \[tab:Annexe\] montre des extraits des phrases de ce discours, ainsi que leurs probabilités $p(C)$ d’appartenance à la classe $C$ (probabilité $p(M) = 1-p(C)$) calculées par adaptation et celles $p_I$ obtenues avec la cohérence interne du discours (cf. section \[sec:cohérence\]), puis adaptation. L’étiquette associée est aussi indiquée. En gras, nous affichons les mots pleins ([**Afrique, africains, Congo, développ(er/ment)**]{},...) des phrases étiquetées $C$. En souligné les mots pleins communs (, ) aux deux classes, et en petites majuscules les mots pleins (<span style="font-variant:small-caps;">accord</span>) présents uniquement à l’intérieur du bloc $M$. Les adjectifs, adverbes, pronoms, conjonctions de subordination ainsi que les noms propres trop fréquents dans le corpus DEFT’05 (tels que [*France, français(e), Paris, international, pays,*]{}...) ont été supprimés pour ne pas fausser les calculs. ![Probabilités $p(C)$ de Markov (ligne pointillée) et de la cohérence interne puis adaptation (trait gras) du discours 520 (dont tous les segments appartiennent à la classe C).[]{data-label="fig:probabilites"}](analyse520.pdf){width="10cm"} Les phrases 39 et 40 (en italiques) méritent d’être analysées en détail. Leurs probabilités d’appartenance à la classe $M$, calculées uniquement avec l’adaptation seule, étaient déjà très élevées : 0,671 ($p(C)=0,329$) et 0,853 ($p(C)=0,147$) respectivement (et sont d’ailleurs les plus élevées de tout ce discours), donc difficiles à renverser. Elles faisaient partie d’un gros bloc (phrases 31-40) étiqueté $M$ par Markov. Après le calcul de la cohérence, les phrases 39 et 40 seront encore étiquetées comme $M$, avec des valeurs 0,91 ($p(C)=0,09$) et 0,89 ($p(C)=0,11$) respectivement. Cependant, la méthode de la cohérence interne a fait basculer vers la classe $C$ sept phrases (31-38) du bloc original, étiquetées $M$ à tort dans un premier temps. Ce renversement n’est en rien négligeable comme en atteste l’augmentation du Fscore due au mécanisme de la cohérence. La figure \[fig:probabilites\] montre les probabilités $p(C)$ en fonction du numéro de la phrase du discours. En pointillé, nous affichons les probabilités de l’adaptation seule et en gras continu celles de la cohérence interne, puis adaptation. On voit que, sauf pour les deux phrases en question, toutes les autres ont été réarrangées de façon à avoir des probabilités $p(C)$ bien au-dessus de leurs valeurs précédentes (plusieurs phrases ont maintenant des probabilités $p(C)=1$). La rupture de cohérence interne du discours reflète ici plutôt un changement thématique : tout le discours 520 s’inscrit fortement dans une vision politique centrée sur l’Afrique, alors que les phrases 39 et 40 parlent soudain de politique internationale dans un sens beaucoup plus large ([*New York, Washington, institutions internationales*]{}). Ceci illustre la tendance (toujours difficile à modéliser) qu’a parfois un locuteur d’introduire subitement des changements (quasi aléatoires) dans son discours. Cet honneur revient dans le cas présent, à Jacques Chirac. = .5 --------------------- ----- ------- Discours 520 (Test) Phrase $p$ $p_I$ --------------------- ----- ------- : Exemple de découpage de la cohérence interne. Discours 520 du Test.[]{data-label="tab:Annexe"} [^1]: Traitement Automatique des Langues Naturelles. [^2]: Rencontre des Etudiants Chercheurs en Informatique pour le Traitement Automatique des Langues. [^3]: Le lecteur intéressé trouvera une description plus détaillée ainsi que les données et résultats dans le site officiel du défi : http://www.lri.fr/ia/fdt/DEFT05 [^4]: Par exemple, dans les allocutions de Jacques Chirac évoquant la politique internationale, les phrases de François Mitterrand introduites sont issues de discours traitant de politique nationale. [^5]: Pour des facilités d’écriture, nous prenons dorénavant la liberté de désigner les deux derniers présidents de la République, par leur nom de famille, sans les faire précéder d’un titre, ou d’un prénom, et pour plus de concision, il nous arrivera de nous contenter de remplacer « Mitterrand » et « Chirac » par les étiquettes [*M*]{} et [*C*]{}. [^6]: Qui sera appelé par la suite Modèle II et par opposition celui avec lemmatisation sera appelé Modèle I. [^7]: « Les équipes utilisant dans leur méthode des corpus de J. Chirac et de F. Mitterrand autres que ceux fournis par les organisateurs seront disqualifiées. Par exemple, la méthode consistant à acquérir un corpus de F. Mitterrand et/ou de J. Chirac par des ressources extérieures pour identifier les phrases de F. Mitterrand présentes dans le corpus fournis par les organisateurs sera considérée comme non valide. » Source : http://www.lri.fr/ia/fdt/DEFT05/resultats.html [^8]: $X$ pouvant prendre ici les valeurs [*M*]{} ou [*C*]{}. [^9]: Les termes ont été regroupés de façon manuelle pour former les concepts du réseau. [^10]: http://www.xrce.xerox.com [^11]: La cohésion thématique étant un des éléments permettant d’apprécier la cohérence interne d’un discours, nous emploierons de préférence l’expression « cohérence interne » dans la suite de l’article. [^12]: Source : http://www.lri.fr/ia/fdt/DEFT05 [^13]: Cette hypothèse va quelque peu à l’encontre de l’objectif recherché, à savoir considérer les segments d’un même bloc comme un tout, mais nous ne savons pas comment faire autrement. [^14]: Téléchargeable à l’adresse : http://www.lia.univ-avignon.fr [^15]: Plus de détails sont rapportés aux tableaux \[tab:modeleII\] et \[tab:testII\]. [^16]: La position de l’hyperplan séparateur des classes se fait par une modification progressive des poids (descente en gradient) contrôlés au moyen d’une température de recuit lors de l’apprentissage. [^17]: Voir en annexe l’analyse du discours 520, concernant cette situation problématique.

--- abstract: 'Neutrinos escaping from a core collapse supernova a few seconds after bounce pass through the shock wave, where they may encounter one or more resonances corresponding to $\Delta m^2_{\rm atm}$. The neutrino mass eigenstates in matter may stay coherent between these multiple resonances, giving rise to oscillations in the survival probabilities of neutrino species. We provide an analytical approximation to these inevitable phase effects, that relates the density profile of the shock wave to the oscillation pattern. The phase effects are present only if the multiple resonances encountered by neutrinos are semi-adiabatic, which typically happens for $10^{-5} \lsim \sin^2 \theta_{13} \lsim 10^{-3}$. The observability of these oscillations is severely limited by the inability of the detectors to reconstruct the neutrino energy faithfully. For typical shock wave profiles, the detection of these phase effects seems rather unlikely. However, if the effects are indeed identified in the $\nuebar$ spectra, they would establish inverted hierarchy and a nonzero value of $\theta_{13}$.' author: - Basudeb Dasgupta and Amol Dighe title: Phase effects in neutrino conversions during a supernova shock wave --- Introduction {#intro} ============ The neutrino fluxes emitted from a core collapse supernova (SN) contain information about the primary fluxes produced inside the star, the neutrino mixing pattern as well as the matter densities that the neutrinos have passed through. The high statistics neutrino signal that one expects from a future galactic SN needs to be decoded in order to extract this information. The neutrinos and antineutrinos produced inside the SN pass through the core, mantle and envelope of the progenitor star before escaping. They encounter matter densities ranging from nuclear densities deep inside the core all the way to vanishingly small densities in the interstellar space. Neutrino masses and mixing angles, and hence the extent of their flavor conversions, depend crucially on the density of surrounding matter, hence it is important to study these matter effects in detail. The matter effects [@Wolfenstein:1977ue] give rise to MSW resonances [@Mikheev:gs] when the matter density corresponds to \_R = 2M\_N / (2G\_[F]{} Y\_e E) . \[msw\] Here $Y_e$ is the electron fraction, $M_N$ the average nucleon mass, and the plus and minus signs correspond to neutrinos and antineutrinos respectively. For neutrinos of energy $E$, resonances are possible at two densities, the $H$ resonance corresponding to $(\dmsq,\theta) \approx (\msqa,\theta_{13})$ and the $L$ resonance corresponding to $(\dmsq,\theta) \approx (\msqs,\theta_{12})$. The energies of SN neutrinos are typically in the range $5$–$50$ MeV. For these energies, the $H$ resonance takes place around $\rho_H \sim 500$–$5000$ g/cc. It occurs in neutrinos for normal hierarchy and in antineutrinos for inverted hierarchy. The $L$ resonance that takes place around $\rho_L \sim 20$–$200$ g/cc always occurs in neutrinos. However, since $\theta_{12}$ is large, significant flavor conversions of antineutrinos also take place at the $L$ resonance. The adiabaticities at the $H$ and $L$ resonances determine the net neutrino conversion probabilities. The $L$ resonance is always adiabatic, given the values of $\msqs, \theta_{12}$ and the typical density profile of the progenitor star around $\rho_L$. The adiabaticity at the $H$ resonance is very sensitive to the value of $\theta_{13}$ and the density profile of the star in the resonance region. Indeed the neutrino conversion rates are crucially dependent on the value of $\theta_{13}$, and whether the $H$ resonance is in neutrinos or antineutrinos. The SN signal can therefore be an extremely sensitive probe of $\theta_{13}$ and whether the mass hierarchy is normal or inverted [@ds]. In addition to divulging the neutrino mixing scenario, SN neutrino fluxes can also allow us to have a peek at the propagation of the shock wave while it is still inside the mantle of the star. The violent density fluctuations caused by the SN shock wave can change the adiabaticity at the $H$ resonance in a time dependent manner, thus leaving their imprint on the time dependent neutrino spectra [@fuller; @takahashi; @lunardini; @fogli]. In particular, the observations of the time dependent neutrino spectra can confirm the presence of forward as well as reverse shock wave through the “double dip” feature [@revshock], and in addition can track the positions of the shocks as they pass through the $H$ resonance region. The feasibility of such a tracking at a water Cherenkov detector has been explored in [@fogli2; @stochastic]. Our understanding of the SN explosion mechanism is still unsatisfactory [@janka; @buras2], which makes it very important to extract as much information about the shock wave as possible. In this paper, we demonstrate how the neutrinos that pass through the shock wave near the $H$ resonance carry information about the density profile of the shock wave. In addition to making the $H$ resonance temporarily nonadiabatic, the shock wave also forces the neutrinos to encounter multiple $H$ resonances. The relative phases that the neutrino mass eigenstates gain between two or more of such $H$ resonances manifest themselves as oscillations in the neutrino flavor conversion probabilities as a function of the neutrino energy. These oscillations are related to the shock wave density profile, and in principle carry information about it. We therefore study the relation between the oscillations and the shock wave profile in detail. Most of the analyses of SN neutrino conversions till now calculate the [*probabilities*]{} of conversions between neutrino mass eigenstates in matter in each resonance region, and combine the results at different resonances, assuming them to be independent. This includes the implicit assumption that the information on the relative phase between the neutrino mass eigenstates is lost between successive resonances. For neutrinos that undergo a single $H$ and a single $L$ resonance, this is a valid assumption since the two resonance regions are well separated [@Keranen:2006gd]. However, when the neutrinos encounter multiple $H$ resonances that are close to each other, coherence between the neutrino mass eigenstates is maintained, and one has to compute the [*amplitudes*]{} of neutrino flavor conversions at the resonances, keeping track of the relative phases. This gives rise to the phase effects that we explore in this paper. Such effects were pointed out in the context of solar neutrinos in [@Haxton:1986bc] and also observed in the numerical studies of supernova neutrinos in [@fogli2; @Kneller:2005hf]. We provide an analytical approximation to study these effects, relating them to the the density profile of the medium. We also provide criteria to decide when decoherence or finite energy resolution of detectors renders these effects unimportant. If the multiple resonances are semi-adiabatic, the phase effects may be strong and if the oscillations are indeed observed in the neutrino spectra, they can help us infer about the density profile of the shock wave, which in turn can provide us important clues about the shock wave propagation and the SN explosion. The observability of these effects is dependent on factors like neutrino fluxes at the source, possible coherent development of neutrinos due to neutrino-neutrino interactions, the different density profiles encountered by neutrinos coming from different regions of the neutrinosphere, stochastic fluctuations of density, effective luminosity of the supernova, detector capabilities, et cetera. We comment on different model predictions of source fluxes and point out the model independent features of the phase effect. We investigate the number of events and detector capabilities required for the effect to be observable and find them to be demanding. Our emphasis in this paper is to first ascertain the origin of these phase effects, their dependence on various parameters in the problem, and to check whether they are important at least in a simplified analysis where effects like coherent development of neutrinos [@coll-duan1; @coll-raff], anisotropy [@revshock] and turbulence [@stochastic; @noise] of the medium are neglected. We find that even in a simplified setting, the observation of these effects is challenging. The paper is organised as follows. In Sec. \[analytic\], we give an analytical approximation for calculating the neutrino conversion probability when multiple $H$ resonances are taken into account in a two-neutrino framework. The results of this section are general and can be applied to any situation where multiple resonances are involved. In Sec. \[actual\], we apply the results to the case of a SN shock wave, where the third neutrino and the $L$ resonance are included. We show the neutrino conversion probabilities obtained by using a realistic shock wave profile and study the feasibility of detecting the phase effects. Sec. \[concl\] concludes. Phase effects from multiple resonances {#analytic} ====================================== In this section, we calculate the survival probability of $\nu_e$ when they pass through multiple resonances, keeping track of the relative phases between the mass eigenstates. The calculations are performed in the $2$-$\nu$ framework. The results are readily extended to the $3$-$\nu$ framework in the case of a SN shock wave, as will be shown in Sec. \[actual\]. Although all the arguments in this section are given for neutrinos, they are valid for antineutrinos just by changing the sign of the matter potential $A(x)$. However, the flavor conversion analysis at the $H$ resonance is applicable to neutrinos in the normal mass hierarchy and for antineutrinos in the inverted mass hierarchy. Therefore, while analyzing antineutrinos, the value of $\dmsq$ should be taken to be negative. Survival probability of $\nue$ for a small mixing angle {#anal} ------------------------------------------------------- The relevant mixing angle at the resonance $H$ is $\theta =\theta_{13}$, on which we currently have a strong bound: $\sin^2 \theta_{13} < 0.05$ [@t13-bound]. Therefore, we try to solve the problem using a small angle approximation. We follow the notation and framework outlined in [@balantekin] and work in the flavor basis. Let $\nu_\beta$ be the relevant linear combination of $\nu_\mu$ and $\nu_\tau$. When neutrinos pass through matter, their propagation is described by the Schrödinger equation i ( [c]{} \_[e]{}\ \_\ )=H ( [c]{} \_[e]{}\ \_\ ). \[coupledeqns\] Upto a matrix proportional to the unit matrix,the Hamiltonian $H$ is given by, ( [cc]{} A(x)- 2& 2\ 2& -A(x)+ 2\ ), \[hamiltonian\] where $A(x) \equiv 2 E V(x) \equiv 2 \sqrt{2} G_F Y_e \rho E / M_N$. These two coupled first order equations give rise to the second order equation -\_[e]{} - (\^2 + i’ )\_[e]{} = \^2 \_[e]{} \[nu\_eeqn\]\ \[order2\] where (x) = (A(x) - 2) ,   = 2and prime ($'$) denotes derivative with respect to $x$. In order to find the survival probability of $\nu_e$, we solve for the $\nu_{e}$ wavefunction with the initial conditions $\nu_{e}(0) = 1$, $\nu_{\beta}(0) = 0$. These conditions are equivalent to \_[e]{}(0) = 1 ,   i.|\_[0]{} = (0) . \[initial\] The “logarithmic perturbation” approximation [@balantekin] solves the differential equation (\[nu\_eeqn\]) for small mixing angles by choosing g 1 - 2\[g-def\] as the small expansion parameter. Denoting \_[e]{} = e\^[S(x)]{} , S’(x) = c\_0(x) + gc\_1(x) + O(g\^2) \[sx-def\] so that \_[e]{}(x) &=& (\_0\^x[dx\_1 c\_0(x\_1)]{}.\ & & .+g\_0\^x[dx\_1 c\_1(x\_1)]{}+ O(g\^2)) , the solution becomes \_[e]{}(x) &=&\ &&+ [O]{}(g\^2) . \[og-soln\] Here we have defined the “accumulated phase” Q(x) \_0\^x dx\_1 ( A(x\_1)-) . \[Q-def\] The survival probability $P_{ee}(x) \equiv P(\nu_{e}\rightarrow\nu_{e})$ at $x=X$ then becomes P\_[ee]{}(X) &=&\ &&+ O(g\^2) . \[Pee\] The integral in the above expression can be evaluated using the stationary phase approximation. The integral oscillates rapidly unless $Q'(x) \approx 0$. So the entire contribution to the integral can be taken to be from the saddle point $x_s$, which is the point where $Q'(x_s)=0$, i.e. A(x\_s) = . Note that this is also the resonance point in the small angle limit. For a monotonic density profile, there is only one saddle point $x_s$ and the survival probability is P\_[ee]{} , \[pee-mono\] which agrees with the Landau-Zener jump probability [@lz1; @lz2] in the limit of small mixing angle, and hence small $g$, even when $P_{ee} \sim 1$. For a non-monotonic density profile, neutrinos can experience more than one resonance at the same density but at different positions. In that case $Q'(x) =0$ at more than one point. If the resonances are sufficiently far apart, the contributions from each of them may be added independently of each other. Their total contribution to the integral in (\[Pee\]) is \_0\^X dx e\^[iQ(x)]{} \_ie\^[i\_i]{}e\^[iQ(x\_[i]{})]{} ( )\^ [[1/2]{}]{} where $i$ runs over all the saddle points. Note that $\alpha = \pi /4$ if $A'(x_{s}) < 0$ and $\alpha = 3\pi /4$ if $A'(x_{s}) > 0$. The probability calculated using (\[Pee\]) now has terms which depend on the differences between the integrated phases \_[ij]{} & & Q(x\_[j]{})-Q(x\_[i]{})+\_j-\_i\ & = & \_[x\_[i]{}]{}\^[x\_[j]{}]{} (A(x)-) dx + \_[j]{}-\_[i]{} . \[phi-ij-def\] In general, P\_[ee]{}(X) &=&\ &&+ [O]{}(g\^2) \[Pee-general\] where a\_i ( ) \^[1/2]{} . \[a-def\] For example, when there are only two saddle points the survival probability is given by P\_[ee]{} &=& (-g a\_1\^2) (-g a\_2\^2) (-2 g a\_1 a\_2 \_[12]{})\ & &+ [O]{}(g\^2) . \[master-eqn\] The first two factors in (\[master-eqn\]) are the individual Landau-Zener jump probabilities for the two level crossings. The last factor gives rise to oscillations in $P_{ee}$ as a function of energy. The oscillation pattern has its maxima at $\Phi_{12} = (2 n + 1) \pi$ and minima at $\Phi_{12} = 2 n \pi$ where $n$ is an integer. This expression is valid as long as $g a_{i}^2 \lsim 1$, so that the ${\cal O}(g^2)$ terms remain small, and none of the resonances overlap. We illustrate the validity and limitations of the small angle approximation with a toy density profile (x) = { [ll]{} a + b\_1 x\^2 & (x&lt;0)\ a + b\_2 x\^2 & (x&gt;0)\ . , \[toy-ne\] as shown in figure \[densityplot\]. We take $Y_e=0.5$ and $\dmsq=0.002~{\rm eV}^2$. Neutrinos are produced at $x \to -\infty$ and we calculate $P_{ee}$ at $x \to \infty$. We also show the positions of resonance densities for various energies, which are given by \_R\[[g/cc]{}\] . \[msw2\] Figure \[nonmonoplot\] shows the survival probability $P_{ee}$ as a function of energy, both the exact numerical result and the result of our analytical approximation for small angles. It can be seen that at such small angles ($\theta = 0.02$ rad $\approx 1.1^\circ$), the approximation works extremely well. Note that the amplitude of the oscillations is comparable to the deviation of the average survival probability from unity. That is, the oscillation effect is not a small effect. Indeed, the oscillation term is of the same order as the averaged effect, as can be seen from (\[master-eqn\]). Figure \[nonmonoplot\] also shows the average value of $P_{ee}$ that one would have obtained if one naively combined the jump probabilities at the two resonances. Our analysis gives additional oscillations in the survival probability as a function of neutrino energy about this average value. This effect is what we call as the phase effect, and is clearly significant as can be seen from the figure. An important feature of the oscillations is that the “wavelengths,” i.e. the distances between the consecutive maxima or minima, are larger at larger $E$. The resonances start overlapping at $E \approx E_{R(max)}$ (figure \[densityplot\]), which is where our approximation starts breaking down, as can be seen in figure \[nonmonoplot\]. For $E > E_{R(max)}$, the neutrinos no longer encounter a strict resonance, and our approximation gives $P_{ee}=0$ identically. However, the resonances have finite widths which may affect the conversion probabilities of neutrinos with $E \approx E_{R(max)}$. The sharp jump observed in figure \[nonmonoplot\] at $E \approx 52$ MeV is therefore not a real effect, but a limitation of our technique. The small angle approximation starts failing for larger angles and lower energies, where $ga_{i}^2 \gsim {\cal O}(1)$. Figure \[pee-large\] shows that the amplitude at low energies is not calculated correctly for $\theta = 0.1$ rad $\approx 5.7^\circ$. However, note that the positions of maxima and minima of $P_{ee}$ are still predicted to a good accuracy. We shall argue in the next subsection that these can be computed accurately for the whole allowed range of $\theta_{13}$, given the non-monotonic density profile between the two resonances. The oscillations in $P_{ee}(E)$ {#freq} ------------------------------- Let us consider a density profile with a “dip” as in the toy model in the previous section. A neutrino with energy $E$ encounters two resonances $R_1$ and $R_2$ at $x=x_1$ and $x=x_2$ respectively, so that \_R (x\_1) = (x\_2) . We assume $Y_e$ to be a constant throughout the region of interest. We also assume that the propagation of neutrino mass eigenstates is adiabatic everywhere except in the resonance regions $(x_{1-}, x_{1+})$ and $(x_{2-},x_{2+})$ around the resonance points $x_1$ and $x_2$ respectively. In the limit of a small mixing angle, the widths of the resonances are small: 2 . Therefore, $x_{1-}\approx x_1 \approx x_{1+}$ and $x_{2-}\approx x_2 \approx x_{2+}$. We shall work in this approximation, and shall use the notation $x_{i\pm}$ only for the sake of clarity wherever needed. At $x \ll x_1$, the density $\rho(x) \gg \rho_R$, so that the heavier mass eigenstate $\nu_H$ is approximately equal to the flavor eigenstate $\nu_e$. Let us start with $\nu_e$ as the initial state: \_e (x x\_1) \_H . The mass eigenstate $\nu_H$ propagates adiabatically till it reaches the resonance region $x \approx x_1$: \_e (x\_[1-]{}) \_H . While passing through the resonance, unless the resonance is completely adiabatic, the state becomes a linear combination of $\nu_H$ and $\nu_L$, the lighter mass eigenstate. Note that the phases of $\nu_H$ and $\nu_L$ can be defined to make their relative phase vanish at $x=x_{1+}$. \_e(x\_[1+]{}) = \_1 \_H + \_1 \_L , where $P_1 \equiv \sin^2 \chi_1$ is the “jump probability” at $R_1$ if it were an isolated resonance. The two mass eigenstates $\nu_H$ and $\nu_L$ propagate to the other resonance $R_2$, gaining a relative phase in the process (the overall phase of the state is irrelevant): \_e(x\_[2-]{}) = \_1 \_H +\_1 ( i \_[x\_[1]{}]{}\^[x\_[2]{}]{} dx ) \_L , where $\Delta \tilde m^2$ is the mass squared difference between $\nu_H$ and $\nu_L$ in matter: m\^2 = ((m\^2 2- 2 E V(x))\^2 + (2))\^[1/2]{} . \[dmtilde\] The effect of the resonance $R_2$ may be parametrised in general as ( [c]{} \_H(x\_[2+]{})\ \_L(x\_[2+]{})\ ) = ( [cc]{} \_2 & \_2 e\^[i]{}\ -\_2 e\^[-i]{} & \_2\ ) ( [c]{} \_H(x\_[2-]{})\ \_L(x\_[2-]{})\ ). where $P_2 \equiv \sin^2 \chi_2$ is the “jump probability” at $R_2$ if it were an isolated resonance. From (\[og-soln\]), one can deduce that in the limit $x_{2-} \approx x_{2+}$, we have $\varphi \approx Q(x_{2+}-x_{2-}) \approx 0$. The state $\nue(x_{2+})$ can then be written as \_e(x\_[2+]{}) &=& \_H\ &&+ \_L . For $x > x_{2+}$, the mass eigenstates travel independently and over sufficiently large distances, decohere from one another. At $x \gg x_2$, since $\rho(x) \gg \rho_R$, the heavier mass eigenstate $\nu_H$ again coincides with $\nu_e$ and we get the $\nu_e$ survival probability as P\_[ee]{} & = & \^2(\_1-\_2)\ &&- 2\_1 2\_2 \^2 ( \_[x\_[1]{}]{}\^[x\_[2]{}]{} dx ) . \[eq-freq\] If the phase information were lost, either due to decoherence or due to finite energy resolution of the detectors [@decoherence] the survival probability would have been P\_[ee ]{} &=& P\_1 P\_2 + (1-P\_1) (1-P\_2)\ &=& \^2 \_1 \^2 \_2 + \^2 \_1 \^2 \_2, \[prob-combo\] which matches with (\[eq-freq\]) when the $\sin^2(\int ..)$ term is averaged out to $1/2$. The $\sin^2(\int ..)$ term in (\[eq-freq\]) gives rise to the oscillations in $P_{ee}(E)$. If two consecutive maxima of $P_{ee}$ are at energies $E_k$ and $E_{k+1}$ such that $E_k > E_{k+1}$, then the condition \_[x\_1(E\_[k+1]{})]{}\^[x\_2(E\_[k+1]{})]{}  dx &-& \_[x\_1(E\_k)]{}\^[x\_2(E\_k)]{}  dx\ &=& 2 \[integrals\] is satisfied. The quantity $(E_k - E_{k+1})$ is the “wavelength” of the oscillations. Note that $\Delta \tilde m^2(x,E)$ is equal to $|A(x) -\Delta m^2|$ in the small angle limit. Moreover, this quantity is rather insensitive to $\theta$ in the allowed range of $\theta_{13}$. Therefore, it is not a surprise that the predictions of the positions of maxima and minima in the small angle approximation (Sec. \[anal\]) are accurate and robust in the whole range $\theta < 13^\circ$. Since $\theta$ is small, the left hand side of (\[integrals\]) is approximately equal to the area of the region in the density profile plot enclosed by the densities $\rho_{E_k}$ and $\rho_{E_{k+1}}$: 2 . The distance between the two resonances in the region $\rho_{E_k} < \rho < \rho_{E_{k+1}}$ is then r\_k / (\_[E\_[k+1]{}]{} - \_[E\_[k]{}]{}) . \[res-dis\] This procedure may be repeated for various $k$ values to estimate the separation between the resonances at the corresponding densities, and hence to constrain the form of the density profile. Although this seems straightforward in our simplified analysis, the effect of density variations due to convection, turbulence and anisotropies would greatly complicate such a reconstruction in practice. Nonmonotonic density profiles are encountered by neutrinos escaping from a core collapse supernova during the shock wave propagation. If the phase effects are observable at neutrino detectors, the above procedure may help us reconstruct the shock wave partially. Of course the procedure is somewhat crude, and the information obtained would only be on the density profile along the line of sight. We have also assumed that neutrinos coming from different parts of the neutrinosphere encounter nearly the same density profiles. However, since neutrinos are the only particles that can even in principle carry information from so deep inside the exploding star, it is important to check whether the detection of these phase effects is feasible even with these simplifying assumptions. We shall do this in the next section. Oscillations during the SN shock wave propagation {#actual} ================================================= In this section, we apply the results in the last section to the neutrinos travelling through a supernova shock wave. Though we have to consider three-neutrino mixing in this case, the separation of $H$ and $L$ resonances [@kuo] means that we can calculate the transition probabilities at these resonances separately. Each of these resonances can then be treated as an effective two-neutrino level crossing. The $L$ resonance that takes place in neutrinos is always adiabatic [@ds]. If the net survival probability after passing through all the $H$ resonances is denoted by $P_H$, the survival probability of $\nue$ after passing through all the $H$ and $L$ resonances is p = P\_H \^2 \_ p = \^2 \_ \[p-def\] where NH and IH stand for normal and inverted mass hierarchy respectively. Here $\theta_\odot$ is the solar mixing angle. Similarly, the survival probability of $\nuebar$ after passing through all the $H$ and $L$ resonances is |[p]{} = P\_H \^2 \_ |[p]{} = \^2 \_ . \[pbar-def\] Clearly, since the phase effects appear through $P_H$, they will be visible only in $\nu_e$ for normal hierarchy and only in $\nuebar$ for inverted hierarchy. In present and planned water Cherenkov [@hk] and scintillation [@lena] detectors, the main neutrino detection channel is the inverse beta decay reaction $\bar\nu_e p\to ne^+$ that allows the reconstruction of $\nuebar$ energies. Therefore we consider only the $\bar\nu_e$ spectrum in our analysis. However an analogous analysis can be easily performed in the neutrino channel for a detector able to measure the $\nue$ spectrum, for example using liquid argon [@liq-ar]. In order to illustrate the phase effects on $p$ or $\bar{p}$, we consider a typical snapshot of the density profile of a SN during a shock wave [@revshock], as shown in figure \[snapshot\]. The forward shock F and the reverse shock R are sharp density discontinuities, the density change of a factor of two or more taking place over a distance of much less than a km. The density variation in the “contact discontinuity” C, which is the transition region between the shock-accelerated and neutrino-heated SN ejecta, takes place more slowly, over a distance of more than $100$ km [@revshock]. The mass accretion region A behind the forward shock wave, and the low density bubble B have gradually changing densities. The region T is the tail of the shock wave. The neutrinos, while passing through these regions, may undergo multiple level crossings. The extent of flavor conversion in each region will depend on the value of $\theta_{13}$ and the steepness of the density profile in that region. It is found that for $\theta \sim 0.01$ rad or higher, the density variations in the mass accretion region A, the low density bubble B and the contact discontinuity C are too gradual for any non-adiabaticity. We therefore concentrate on the forward shock F, the reverse shock R and the tail T. The coherence between mass eigenstates and the oscillations in the survival probability may be lost due to two major sources: the separation of mass eigenstate wavepackets and the finite energy resolution of the detector. The coherence length $L_{coh}$, defined as the distance over which the wavepackets separate, is given by [@Nussinov] L\_[coh]{} \~4 E\^[2]{}/  , \[coherence-crit\] where $\sigma$ is the width of the wavepacket at source. Taking $\sigma \sim 10^{-9}$ cm near the neutrinosphere  [@Anada:1989fk] in the relevant energy range of $5$–$80$ MeV, the coherence length for SN neutrinos is $L_{coh} \sim 10^{3}$–$10^{5}$ km. Resonances separated by distances larger than $L_{coh}$ may be taken to be incoherent. Since the distances involved are ${\cal O}(10^4~{\rm km})$ (See figure \[snapshot\]), a definite conclusion about decoherence due to wavepacket separation cannot be reached with this simple estimate. However, for observability the oscillations in survival probability must also occur over energy intervals much larger than $\delta E_{det}$, the uncertainity in energy measurement at the detector. This turns out to be the dominant factor in smearing out the oscillations. We can estimate from (\[integrals\]) the difference in energies at which successive maxima of the survival probability occur to be \_[E\_[k]{}]{} E\_[k+1]{}-E\_[k]{} \~4 E\_[k]{}\^2/ r\_[k]{}, \[ener-crit\] where $r_{k}$ is the distance between the two resonances encountered by a neutrino of energy $E_{k}$. For the energy range of $5$–$80$ MeV, this gives $\lambda_{E_{k}}$ of $1$–$10$ MeV for the T-R resonance pair with $r_{k}\sim10^3$–$10^4$ km. The oscillations are faster for the T-F and R-F pairs that occur about $10^5$–$10^6$ km apart. Typically the energy resolution is $1$–$10$ MeV for a water Cherenkov detector and $0.1$–$1$ MeV for a scintillation detector over this energy range. Moreover, the charged lepton energy is not of the same as the energy of the neutrino, which introduces additional smearing. Indeed as we shall see, the fast oscillations due to the T-F or R-F resonance pairs will be smeared and therefore clearly unobservable even at a scintillation detector. Note that the above arguments are only qualitative; we have neglected the matter effects while estimating $\L_{coh}$ and $\lambda_{E_k}$. Figure \[pbarplot\] shows the value of $\bar{p}$ as a function of energy for $\theta=0.02$ rad $\approx 1.1^\circ$. The rapid oscillations correspond to the relative phase $\Phi_{RF} (\approx \Phi_{TF})$ that is accumulated by the mass eigenstates between resonance regions R and F (T and F). Such high frequency oscillations are virtually impossible to observe, given the practical limits on the energy resolutions of neutrino detectors. We smear the high frequency oscillations by taking a “running average” over the energy range corresponding to the typical energy resolution of a scintillation detector. The low frequency oscillations that survive are found to correspond to the relative phase $\Phi_{TR}$ accumulated by the mass eigenstates between resonances in regions T and R. Since these two resonances are closer compared to the resonance pairs R-F or T-F, the frequency of oscillations is smaller. The same oscillation pattern is observed if the survival probability is computed by assuming that the resonance in region F is completely non-adiabatic, which confirms that the pattern is indeed due to the level crossings in regions T and R. Note that $\bar{p}$ goes to $\cos^2\theta_\odot$ at its maximum where $P_H$ goes to unity, whereas at $\bar{p}~\approx 0$ at high energies and low energies where $P_H$ goes to zero. The oscillations in the low energy region ($E<~20$ MeV) are too rapid to be observable. The fluctuations observed in the running average at low energies are not robust: they depend partly on the details of the density profile and are partly numerical artifacts. In figure \[analpbarplot\], we plot the value of $\bar{p}$ for the same parameter values, but calculated using (\[Pee-general\]) at the T, R, F resonances. The smeared probability curve is calculated by dropping the oscillatory terms due to the resonance at F and combining the survival probability at the T-R pair with that at F using (\[prob-combo\]). We see that the analytical expression agrees quite well with the numerical result of figure \[pbarplot\]. In particular, the oscillation frequency matches quite well. Moreover, the slope of $A(x)$ at T is about $0.1$ times that at R, therefore (\[a-def\]) and (\[master-eqn\]) predict the amplitude of oscillations to be $\sqrt {A'(T)/A'(R)}~\approx 0.3$ times the mean value of $\bar{p}$. This estimate is also in good agreement with our numerical results in the figure. The analytical approximation in (\[Pee-general\]) breaks down when $g a_{i}^2 > {\cal O}(1)$. This happens below a certain value of energy, that is higher for larger mixing angles. Comparing figure \[pbarplot\] and figure \[analpbarplot\] we see that for $\theta=0.01$ rad the approximation is resonably accurate for neutrino energies above $20$ MeV. Figure \[pbarlargetheta\] shows the smeared $\bar{p}$ calculated numerically and analytically for $\theta=0.05$ rad. It is clear that the approximation is valid above $25$ MeV. The approximation fails at low energies where $\bar{p}$ calculated analytically does not tend to zero. For higher energies, the oscillations due to the phase effect are predicted quite accurately. In order to explore the observability of the phase effects, we use the parametrization for the primary fluxes given by [@Keil:2002in]: F\^0\_[\_i]{}= N()()\^ , where $N(\alpha)~=~(1+\alpha)^{1+\alpha}/\Gamma(1+\alpha)$. For illustration, we choose two models of neutrino fluxes, the Garching model [@garching] that uses the parameters \_ = \_[|\_x]{} =3 ,   E\_0() = 15  ,\ E\_0(|\_x) = 18  ,  \_0()/\_0(|\_x) = 0.8 ,and the Livermore model [@livermore] that uses \_ = \_[|\_x]{} =3 ,   E\_0() = 15  ,\   E\_0(|\_x) = 24  ,  \_0()/\_0(|\_x) = 1.6 . We plot in figure \[fig-flux\](a) the quantity dN\_/dE\_& = & (E\_) F\_\ & = & (E\_) \[F\^0\_[|\_x]{} + |[p]{} (F\^0\_-F\^0\_[|\_x]{})\] \[sigma-e\] as a function of the neutrino energy $E_\nuebar$, where we have normalized the spectrum such that the total number of events is $10^5$. Here we use the differential cross section as computed in [@strumia]. The phase effect is most prominent around $4$–$5$ seconds postbounce when the tail T and reverse shock R pass through the $H$ resonance and the T-R resonance pair is about $10^3$ km apart. We expect from (\[ener-crit\]) that the extrema in the survival probability will occur at energies separated by about $1$–$10$ MeV. This is clearly visible in figure \[fig-flux\](a). Moreover, the positions of the extrema are independent of the primary fluxes. However, the spectrum shown in figure \[fig-flux\](a) is not directly observable: one can only observe the energy spectrum of positrons produced by the inverse beta reaction. Assuming quasielastic scattering, the positron energy is given by [@strumia] E\_e= \[E-positron\] and p\_e= , \[e-positron\] where $\vartheta$ is the angle of scattering, $M_p$ the proton mass, $\epsilon \equiv E_\nuebar/M_p$, and $\kappa \equiv (1+\epsilon)^2 - (\epsilon \cos\vartheta)^2$. For $E \approx 40$ MeV, we have $\epsilon \approx 1/25$, so that the positron energy is spread over a range of $\approx 4$ MeV depending on the scattering angle. Given that the successive maxima of the oscillation pattern are separated by only about $2$–$8$ MeV in this energy range, the oscillation pattern is significantly smeared out. A further smearing of the oscillation pattern is caused by the finite energy resolution of the detector. The energy resolution of a water Cherenkov detector is typically $\Delta E_{\rm CH}{\rm (MeV)} \approx 1.6 \sqrt{E/10~{\rm MeV}}$ and washes off the oscillations completely. For a scintillation detector, the resolution is much better, $\Delta E_{\rm SC} {\rm (MeV)} \approx 0.2 \sqrt{E/10~{\rm MeV}}$. We show in figure \[fig-flux\](b) the spectrum of the observed positron energy $E'_e$ after taking (\[E-positron\]) and (\[e-positron\]) into account and using the energy resolution of a scintillation detector. It is observed that one or two extrema at high energies ($E \approx$ $40$–$60$ MeV) may still survive for the Livermore model where the spectrum extends to higher energies, but their clean identification would require $\sim 10^5$ events at a scintillation detector in a single time bin. Figure \[fig-flux\](b) also shows the positron energy spectrum for $10^5$ events, binned in $0.5$ MeV energy intervals, which is approximately the energy resolution of a scintillation detector near $E = 40$ MeV. The total number of events expected even at a future 50 kt scintillation detector is $\sim 10^4$ for a SN at a distance of $10$ kpc [@lena]. Even with 1 sec time bins, the number of events in each bin will be $\sim 10^3$. This is a number too small for our purposes. Thus, the identification of the phase effects seems very unlikely, unless the SN is as close as a kpc. The density profiles in the shock wave are quite uncertain, and one may expect that some possible profiles give rise to oscillations with larger amplitudes and wavelengths, which would be easier to observe. The amplitude of oscillations is proportional to the ratio $|A'(T)/A'(R)|$ \[See (\[a-def\]) and (\[master-eqn\])\]. Therefore, larger amplitudes need a larger ratio $|A'(T)/A'(R)|$ whereas larger wavelengths need a sharp tail, i.e. a larger $|A'(T)|$. At the same time, the adiabaticity parameter at the two resonances has to be $0.1 \lsim \gamma \lsim 2$ for $E \approx 40$ MeV. However, we find that even with such finely tuned density profiles, the improvement in observability is not significant. Concluding remarks {#concl} ================== When the neutrinos escaping from the core of a core collapse SN pass through the shock wave, they may encounter multiple “$H$” resonances corresponding to $\dmsq_{\rm atm}$ and $\theta_{13}$, when the shock wave is in the regions with densities around $500$–$5000$ g/cc. We have shown that this necessarily gives rise to oscillations in the neutrino survival probabilities, which we have calculated as a function of energy. We present an analytical approximation for small mixing angles and show that the oscillations are a significant effect: they can be of the same order as the non-oscillating terms. The typical values of $\theta_{13}$ that gives rise to the oscillation features are in the “transition region” of the neutrino mixing parameter space [@ds] that is significant in range ($10^{-5} \lsim \sin^2 \theta_{13} \lsim 10^{-3}$) but is usually neglected in the SN analysis for the sake of simplicity. The local maxima and minima in the survival probabilities of $\nue$ or $\nuebar$ are determined by the relative phase accumulated by the neutrino mass eigenstates between the multiple resonances. The positions of these extrema depend on the density profile and are independent of the primary neutrino spectra. If these extrema were identified they would reveal information on the propagation of the shock wave: its location as well as the density variation present around it. Since neutrinos are the only particles that can carry information about the shock wave while it is still deep inside the exploding star, it is important to explore the observability of these phase effects. Moreover, the mere identification of these effects in the $\nuebar$ spectrum would establish the inverted hierarchy and nonzero $\theta_{13}$, which are two of the most important quantities in neutrino phenomenology, and even faint chances of their determination should be explored. It is interesting that for typical shock wave profiles, oscillations with “wavelengths” of $2$–$8$ MeV are indeed present in the neutrino survival probabilities in the energy range $E \approx$ $30$–$60$ MeV. These wavelengths are tantalizingly close to the resolving power of the neutrino detectors. However, the inability of the detectors in reconstructing the energy of the incoming neutrino tends to wash out the oscillation pattern. For typical shock wave density profiles, the energy resolution of water Cherenkov detectors is insufficient to detect the oscillations. Even a scintillation detector with a superior energy resolution will need $\sim 10^5$ events in the relevant time bin of $4$–$5$ sec postbounce for identifying one or two extrema in the most optimistic scenario. Therefore, we expect that neglecting the phase effects is a valid practical approximation except under extreme cases. Note that it is still possible to observe other robust signatures of shock wave propagation like the dips in the time spectrum of number of events [@fogli] or the double dip feature in $\la E \ra$ [@revshock]. Some of these features survive even in the presence of somewhat extreme stochastic density fluctuations, as discussed in [@stochastic]. The oscillatory effects might contribute additional scatter to these signals but will not spoil the signatures. We must qualify the above conclusions by stating the effects ignored in our analysis to obtain the present results. We have assumed a smooth spherically symmetric density profile and ignored anisotropies that are likely to be present [@revshock]. This is a justified assumption only as long as the deviations from this assumed profile occur only on transverse distances larger than the size of the neutrinosphere so that the neutrinos in our line of sight do not experience the anisotropy. We have not included the recently discussed collective effects of coherent flavor development [@coll-duan1; @coll-raff], which may be important in the inverted hierarchy. However, the extent and nature of its impact on flavor conversion has not been worked out in detail, which makes it difficult to be included in the present analysis. Similarly, the effects of a realistic spectrum of stochastic fluctuations in the medium density or turbulent convections behind the shock wave are yet to be calculated for SN neutrinos [@stochastic; @noise]. We think that with a better understanding of the collective effects and density fluctuations one could include their effect on the realistic observability of the phase effects. The phase effects pointed out here result from the interference between two or more MSW resonances. This phenomenon is not restricted to the SN alone, but may occur whenever neutrinos pass through nonmonotonic matter densities and the resonances are semiadiabatic. The technique developed in this paper for treating coherent multiple resonances is applicable to such cases. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank R. Buras and L. Scheck for help in understanding features of the shock wave. AD would like to thank the Max Planck Institute for Physics for hospitality during the initial part of the work. We would like to especially thank B. Bhattacharya for useful discussions. We are grateful to S. Choubey, H. -Th. Janka, E. Lisi, G. G. Raffelt and R. Tomàs for comments on the manuscript. This work was partly supported through the Partner Group program between the Max Planck Institute for Physics and Tata Institute of Fundamental Research. [99]{} L. Wolfenstein, Phys. Rev. D [**17**]{} (1978) 2369. S. P. Mikheev and A. Yu. Smirnov, Sov. J. Nucl. Phys.  [**42**]{} (1985) 913 \[Yad. Fiz.  [**42**]{} (1985) 1441\]. A. S. Dighe and A. Yu. Smirnov, Phys. Rev. D [**62**]{} (2000) 033007 \[hep-ph/9907423\]. R. C. Schirato and G. M. Fuller, astro-ph/0205390. K. Takahashi, K. Sato, H. E. Dalhed and J. R. Wilson, Astropart. Phys.  [**20**]{} (2003) 189 \[astro-ph/0212195\]. C. Lunardini and A. Yu. Smirnov, J. Cosmol. Astropart. Phys. JCAP06 (2003) 009 \[hep-ph/0302033\]. G. L. Fogli, E. Lisi, D. Montanino and A. Mirizzi, Phys. Rev. D [**68**]{} (2003) 033005 \[hep-ph/0304056\]. R. Tomàs, M. Kachelrieß, G. Raffelt, A. Dighe, H. -Th. Janka and L. Scheck, J. Cosmol. Astropart. Phys. JCAP09 (2004) 015 \[astro-ph/0407132\]. G. L. Fogli, E. Lisi, A. Mirizzi and D. Montanino, J. Cosmol. Astropart. Phys. JCAP04 (2005) 002 \[hep-ph/0412046\]. G. L. Fogli, E. Lisi, A. Mirizzi and D. Montanino, \[hep-ph/0603033\]. R. Buras, M. Rampp, H. -Th. Janka and K. Kifonidis, Phys. Rev. Lett.  [**90**]{} (2003) 241101 \[astro-ph/0303171\]. R. Buras, M. Rampp, H. -Th. Janka and K. Kifonidis, \[astro-ph/0507135\]. P. Keranen, J. Maalampi, M. Myyrylainen and J. Riittinen, Phys. Rev.  D [**75**]{}, 033006 (2007) \[hep-ph/0611037\]. W. C. Haxton, Phys. Rev. D [**35**]{},(1987) 2352. J. P. Kneller and G. C. McLaughlin, Phys. Rev. D [**73**]{}, (2006) 056003 \[hep-ph/0509356\]. H. Duan, G. M. Fuller and Y. Z. Qian, \[astro-ph/0511275\]. S. Hannestad, G. G. Raffelt, G. Sigl and Y. Y. Y. Wong, \[astro-ph/0608695\]. A. Friedland and A. Gruzinov, \[astro-ph/0607244\]. S. Goswami, A. Bandyopadhyay and S. Choubey, Nucl. Phys. Proc. Suppl.  [**143**]{} (2005) 121 \[hep-ph/0409224\]. A. B. Balantekin, S. H. Fricke and P. J. Hatchell, Phys. Rev. D [**38**]{} (1988) 935. L. Landau, Phys. Z. Sowjetunion [**2**]{} (1932) 46. C. Zener, Proc. Roy. Soc. Lond. A [**137**]{} (1932) 696. A. S. Dighe, Q. Y. Liu and A. Yu. Smirnov, \[hep-ph/9903329\]. T. K. Kuo and J. T. Pantaleone, Rev. Mod. Phys.  [**61**]{} (1989) 937. K. Nakamura, Int. J. Mod. Phys. A [**18**]{} (2003) 4053. L. Oberauer, F. von Feilitzsch and W. Potzel, Nucl. Phys. Proc. Suppl.  [**138**]{} (2005) 108. A. Bueno, I. Gil-Botella and A. Rubbia, hep-ph/0307222. S. Nussinov, Phys. Lett. B [**63**]{}, (1976) 201. H. Anada and H. Nishimura, Phys. Rev. D [**41**]{}, (1990) 2379. M. T. Keil, G. G. Raffelt and H. T. Janka, Astrophys. J.  [**590**]{} (2003) 971 \[astro-ph/0208035\]. R. Buras, H. -Th. Janka, M. T. Keil, G. G. Raffelt and M. Rampp, Astrophys. J.  [**587**]{} (2003) 320 \[astro-ph/0205006\]. T. Totani, K. Sato, H. E. Dalhed and J. R. Wilson, Astrophys. J.  [**496**]{} (1998) 216 \[astro-ph/9710203\]. A. Strumia and F. Vissani, Phys. Lett. B [**564**]{} (2003) 42 \[astro-ph/0302055\].

--- abstract: 'We study observational implications of the stochastic gravitational wave background and a non-Gaussian feature of scalar perturbations on the curvaton mechanism of the generation of density/curvature fluctuations, and show that they can determine the properties of the curvaton in a complementary manner to each other. Therefore even if Planck could not detect any non-Gaussianity, future space-based laser interferometers such as DECIGO or BBO could practically exhaust its parameter space.' --- ICRR-Report-551\ RESCEU-26/09 1.35cm [**Gravitational Wave Background and Non-Gaussianity\ as a Probe of the Curvaton Scenario** ]{} 1.2cm Kazunori Nakayama$^{(a)}$ and Jun’ichi Yokoyama$^{(b,c)}$ 0.4cm [ *$^a$Institute for Cosmic Ray Research (ICRR), University of Tokyo, Kashiwa, Chiba 277-8582, Japan*]{} [ *$^b$Research Center for the Early Universe (RESCEU), Graduate School of Science, The University of Tokyo, Bunkyo-ku, Tokyo 113-0033, Japan* ]{} [*$^c$Institute for the Physics and Mathematics of the Universe (IPMU), University of Tokyo, Kashiwa 277-8568, Japan*]{} Introduction ============ The idea of inflation [@lindebook] has become a standard paradigm since it was proposed in the early 1980’s. Inflation is an accelerated expansion epoch in the very early Universe, which solves the flatness, the horizon and also the monopole problems naturally. Furthermore quantum fluctuation of the inflaton $\phi$, which is a scalar field responsible for the accelerated expansion, can provide the seed of cosmic density/curvature fluctuations [@yuragi] observed through the cosmic microwave background (CMB) anisotropy [@Komatsu:2008hk], galaxy clustering, etc. Another prediction of inflation, which is in fact more generic, is the generation of the stochastic gravitational wave background or the tensor perturbation, whose spectrum is nearly scale-invariant [@staro; @Allen:1987bk; @Turner:1990rc; @Sahni:1990tx; @Turner:1993vb; @Turner:1993xz; @Liddle:1993zj; @Turner:1995ge; @Turner:1996ck; @Allen:1997ad; @Seto:2003kc; @Weinberg:2003ur; @Tashiro:2003qp; @Ungarelli:2005qb; @Smith:2005mm; @Seto:2005qy; @Kudoh:2005as; @Boyle:2005se; @Smith:2006xf; @Chongchitnan:2006pe; @Friedman:2006zt; @Zhao:2006is; @Watanabe:2006qe; @Chiba:2007kz; @Boyle:2007zx; @jyessay; @Smith:2008pf; @Nakayama:2008ip; @Nakayama:2008wy; @Kuroyanagi:2008ye; @Mangilli:2008bw] (see Ref. [@Maggiore:1999vm] for a review). The amplitude of the gravitational wave is simply proportional to the Hubble parameter, $H_{\rm inf}$, during inflation and many inflation models predict detectable amplitude of gravitational waves. In particular, future space-based gravitational wave detectors such as DECIGO [@Seto:2001qf] and/or BBO have a chance to detect inflationary gravitational wave background. Recently, it was pointed out that thermal history of the early universe is imprinted in the spectral shape of the gravitational wave background [@Seto:2003kc; @Boyle:2005se; @Boyle:2007zx; @jyessay; @Nakayama:2008ip; @Nakayama:2008wy]. In particular, the reheating temperature of the Universe after inflation can be determined or constrained from future gravitational wave experiments [@jyessay; @Nakayama:2008ip; @Nakayama:2008wy]. Thus any detection of primordial gravitational wave background gives useful information on the early Universe. Traditionally, the spectrum of the comoving curvature perturbation is parametrized as $$\Delta^2_{{{{\cal R}_c}}}(k)=\Delta^2_{{{{\cal R}_c}}}(k_*)\left ( \frac{k}{k_*} \right )^{n_s-1}.$$ Here $k_\ast$ is the pivot scale which we take $k_*=0.002$ Mpc$^{-1}$ and $n_s$ is the scalar spectral index. Inflation predicts nearly scale-invariant spectrum with $n_s\cong 1$, with its precise value determined by the shape of the potential [@yuragi], and it agrees well with observations so far [@Komatsu:2008hk]. If $\phi$ has a single component with a canonical kinetic term, the resultant curvature fluctuation is Gaussian distributed, which is again in agreement with the observation today [@Komatsu:2008hk]. However, the origin of the density perturbation is not limited to the quantum fluctuation of the inflaton. Another scalar field, called curvaton, may be responsible for the generation of the observed density perturbation [@Mollerach:1989hu; @Lyth:2001nq]. It is undoubtedly an important task to distinguish them in order to understand the physics of the early Universe. One such signature may come from the non-Gaussianity in the CMB anisotropy [@Bartolo:2004if], since the curvaton scenario can produce large enough non-Gaussian feature to be detected [@Lyth:2002my; @Enqvist:2005pg], while standard inflation models predict negligible non-Gaussianity [@Acquaviva:2002ud; @Maldacena:2002vr; @Seery:2005wm; @Yokoyama:2007uu]. However, the situation where the curvaton generates large non-Gaussianity is somewhat limited and it is possible that the curvaton accounts for the observed density perturbation without generating large non-Gaussianity. In this paper, we point out that the observation of stochastic gravitational wave background plays a very important role to probe the physics of the curvaton scenario. In the case that the curvaton generates negligible non-Gaussianity, there must be an entropy production process by the curvaton decay itself, and such a non-standard thermal history is imprinted in the spectrum of the inflationary gravitational wave background. Future space-based gravitational wave detectors, such as DECIGO and/or BBO may be able to do this job. This opens up a possibility to find an evidence of the curvaton scenario. Thus measurement of the non-linearity parameter $f_{\rm NL}$ which is a simplified measure of non-Gaussianity and that of tensor perturbations by DECIGO/BBO play complementary roles to each other. The rest of the paper is organized as follows. In Sec. \[sec:curvaton\], the gravitational wave background spectrum and non-Gaussianity in the curvaton scenario are summarized. In Sec. \[sec:probe\] the detection possibility is discussed. Sec. \[sec:conc\] is devoted to the conclusion. Features of the Curvaton Scenario {#sec:curvaton} ================================= In this section, we summarize the cosmological consequences of the curvaton scenario, namely, the spectra of gravitational wave background and the density/curvature perturbation, and non-Gaussianity. All these ingredients are essential for probing the curvaton scenario and constraining the parameter space, as will be discussed in Sec. \[sec:probe\]. Gravitational wave background spectrum in the curvaton scenario --------------------------------------------------------------- In the inflationary era, quantum fluctuations of the RMS amplitude of $H_{\rm inf}/(2\pi)$ are induced on all massless fields in each Hubble time. Here “massless” means that the mass is much smaller than $H_{\rm inf}$. The tensor perturbation of the metric consists of two free massless scalar components, which can be quantized in de-Sitter space-time. After a mode left the horizon during inflation, it actually becomes a classical fluctuation which can be viewed as a stochastic gravitational wave. Thus inflation necessarily generates a gravitational wave background from a cosmological scale to a sub-kilometer scale The former is a target of detection by the B-mode polarization of CMB anisotropy and direct detection by space-based laser interferometer experiments. Tensor perturbation of the metric is defined by the following line element, $$ds^2=a(\tau)^2\left[ -d\tau^2 +(\delta_{ij}+h_{ij})dx^i dx^j \right],$$ where $a(\tau)$ is the scale factor and $h_{ij}$ denotes the metric perturbation satisfying the transverse-traceless conditions $\partial^i h_{ij}=0$ and $h^i_i=0$. Thus $h_{ij}$ has two physical degrees of freedom, which are denoted as $h^\lambda$ with $\lambda=+,\times$. In the inflationary era, the tensor perturbation has a quantum fluctuation whose spectrum is given by $$\Delta_h^{(\rm p) 2}(k)=64\pi G\left ( \frac{H_{\rm inf}}{2\pi} \right )^2 \left ( \frac{k}{k_*} \right )^{n_t},$$ where the tensor spectral index is given by $n_t=-2\epsilon$. Here $\epsilon$ is one of the slow-roll parameters during inflation defined by $$\epsilon = \frac{M_{\rm Pl}^2}{2} \left ( \frac{V'}{V} \right )^2,~~~~~ \eta = M_{\rm Pl}^2 \frac{V''}{V},$$ where $V$ is the inflaton potential and the prime denotes derivative with respective to the inflaton field $\phi$, and $M_{\rm Pl}=(8\pi G)^{-1/2}$ is the reduced Planck scale. After the production of the gravitational waves during inflation, the amplitude of each Fourier mode remains constant when the corresponding mode lies outside the Hubble radius. However, once it enters the horizon, its amplitude decreases as $\propto a^{-1}$. Thus the present energy density of the gravitational wave background per logarithmic frequency interval is written as $$\frac{d\rho_{\rm gw}}{d\ln k}=\frac{k^2}{32\pi Ga_0^2} \Delta_h^{(\rm p) 2}(k) \left ( \frac{a_{\rm in}(k)}{a_0} \right )^2,$$ where $a_0$ is the present scale factor, and $a_{\rm in}(k)$ denotes the scale factor at which the corresponding mode with wave number $k$ enters the horizon. It behaves as $\propto k^{-2} (k^0)$ for the mode which enters the horizon at matter (radiation) dominated era. Therefore thermal history of the Universe is imprinted in the spectrum of the gravitational wave background, and this is the reason why the observations of the gravitational wave are expected to have great impacts on cosmology [@Seto:2003kc; @Boyle:2005se; @Boyle:2007zx; @Nakayama:2008ip; @Nakayama:2008wy]. In terms of the density parameter, it can be rewritten as $$\Omega_{\rm gw}(k)=\frac{k^2}{12a_0^2H_0^2}\Delta_h^2(k),$$ where $H_0$ is the present Hubble parameter, and $$\Delta_h^2(k) = \Delta_h^{(\rm p)2}(k) \Omega_{\rm m} ^{2} \left ( \frac{3j_1(z_k)}{z_k} \right )^2 \left ( \frac{g_*(T_{\rm in})}{g_{*0}} \right ) \left ( \frac{g_{*s0}}{g_{*s}(T_{\rm in})} \right )^{4/3} T_1^2(x_{\rm eq}) T_2^2(x_{\rm R}), \label{Deltah}$$ where $g_*(T_{\rm in})$ denotes the effective relativistic degrees of freedom at the temperature $T_{\rm in}$ when $k$-mode enters the horizon, and $j_1(z)$ is the spherical Bessel function of the first rank with $z_k\equiv 2k/(a_0H_0)$.[^1] The transfer functions $T_1(x)$ and $T_2(x)$ are given by [@Turner:1993vb; @Nakayama:2008wy] $$\begin{gathered} T_1^2(x) = 1+1.57 x +3.42 x^2,\\ T_2^2(x) = [1-0.32 x +0.99 x^2]^{-1}.\end{gathered}$$ The former connects the gravitational wave spectrum of the mode entering the horizon before ($x_{\rm eq}\equiv k/k_{\rm eq} >1$) and after ($x_{\rm eq}<1$) the matter-radiation equality, where $k_{\rm eq}\equiv a(t_{\rm eq})H(t_{\rm eq})=7.3\times 10^{-2} \Omega_{\rm m}h^2~ {\rm Mpc}^{-1}$.[^2] The latter transfer function connects the mode entering the horizon before ($x_{\rm R}\equiv k/k_{\rm R} >1$) and after ($x_{\rm R}<1$) the reheating subsequent to inflation. Here $k_{\rm R}$ is the comoving wavenumber corresponding to the horizon scale at the reheating epoch when the Universe became radiation dominant. Without any significant entropy production after the reheating, it is given by $$\begin{split} \frac{k_{\rm R}}{a_0} &=1.7\times 10^{13}~{\rm Mpc}^{-1} \left ( \frac{g_{*s}(T_{\rm R})}{106.75} \right )^{1/6} \left ( \frac{T_{\rm R}}{10^6~{\rm GeV}} \right ),\\ &T_{\rm R}=\left(\frac{10}{\pi^2 g_*(T_{\rm R})}\right)^{1/4} \sqrt{\Gamma_\phi M_{\rm Pl}}, \label{kR} \end{split}$$ with $T_{\rm R}$ and $\Gamma_\phi$ being the reheating temperature after inflation and the decay rate of the inflaton, respectively. This corresponds to the frequency $f=$0.026Hz for $T_{\rm R}=10^6$GeV, which is close to the most sensitive frequency range of the planned future space-based laser interferometer experiments, DECIGO or BBO. Thus by observing the spectral shape of the gravitational wave background, the reheating temperature of the Universe can be determined or constrained. As mentioned above, the above correspondence assumes the standard thermal history with no significant entropy production after reheating. In the curvaton scenario, however, its coherent oscillation may once dominate the Universe, which introduces an additional matter-dominated era, and then decays releasing huge amount of entropy. In this case the gravitational wave spectrum receives an additional suppression [@Seto:2003kc; @Nakayama:2008ip; @Nakayama:2008wy] which can be quantified by the dilution factor, $F$, defined by $$F = \frac{s(T_\sigma)a^3(T_\sigma)}{s(T_{\rm R})a^3(T_{\rm R})} =\left \{ \begin{array}{ll} \displaystyle \frac{\sigma_i^2}{6M_{\rm Pl}^2}\frac{T_{\rm R}}{T_\sigma} ~~&{\rm for}~~~m_\sigma > \Gamma_\phi \\ \displaystyle \frac{\sigma_i^2}{6M_{\rm Pl}^2}\frac{T_{\rm osc}}{T_\sigma} ~~&{\rm for}~~~m_\sigma < \Gamma_\phi \end{array} \right., \label{F}$$ where $\Gamma_\phi$ is the decay rate of the inflaton, $T_\sigma$ is the radiation temperature just after the curvaton decay, and $T_{\rm osc}$ is the temperature at which the curvaton begins to oscillate. This expression is valid for $F \gg 1$. The resultant gravitational wave spectrum is given by $$\begin{split} \Delta_h^2(k) = &\Delta_h^{(\rm p)2}(k) \Omega_{\rm m}^2 \left ( \frac{3j_1(z_k)}{z_k} \right )^2 \left ( \frac{g_*(T_{\rm in})}{g_{*0}} \right ) \left ( \frac{g_{*s0}}{g_{*s}(T_{\rm in})} \right )^{4/3} \\ &\times T_1^2(x_{\rm eq}) T_2^2(x_{\sigma}) T_1^2(x_{\sigma \rm R}) T_2^2(x_{\rm R}(F)), \end{split}$$ where $x_\sigma=k/k_\sigma$ with $k_\sigma$ given in an analogous way with (\[kR\]) after replacing $T_{\rm R}$ with $T_\sigma$, and $x_{\sigma \rm R} = k/k_{\sigma \rm R}$ with $k_{\sigma \rm R} = k_\sigma F^{2/3}$ and $x_{\rm R}(F) = k/k_{\rm R}(F)$ with $k_{\rm R}(F) = k_{\rm R} F^{-1/3}$. In Fig. \[fig:GWspec\] we show spectra of the gravitational wave background as a function of its present frequency. In the top panel, the spectra for $H_{\rm inf}=10^{14}$ GeV and $10^{13}$ GeV with the reheating temperature $T_{\rm R}=10^6$ GeV are shown. Also plotted are sensitivities of DECIGO with a correlation analysis (blue dashed line), ultimate-DECIGO (red dotted line), and correlation of analysis of ultimate-DECIGO (purple dot-dashed line) [@Kudoh:2005as]. In the bottom panel, the gravitational wave background spectra in the presence of entropy production is shown, for $F=10$ and $T_\sigma$=10 GeV and $T_{\rm R}=10^7$ GeV. ![(Top) Spectra of the gravitational wave background for inflationary scale $H_{\rm inf}=10^{14}$ GeV and $10^{13}$ GeV. Here we have taken $T_{\rm R}=10^7$ GeV. Also shown are sensitivities of planned space-based gravitational wave detectors, DECIGO with a correlation analysis (blue dashed line), ultimate-DECIGO (purple dotted line), and correlation of analysis of ultimate-DECIGO (red dot-dashed line). (Bottom) Same as the top panel for the dilution factor $F=10$ for $T_\sigma$=10 GeV and $T_{\rm R}=10^7$ GeV. []{data-label="fig:GWspec"}](GWspec.eps "fig:"){width="0.7\linewidth"} ![(Top) Spectra of the gravitational wave background for inflationary scale $H_{\rm inf}=10^{14}$ GeV and $10^{13}$ GeV. Here we have taken $T_{\rm R}=10^7$ GeV. Also shown are sensitivities of planned space-based gravitational wave detectors, DECIGO with a correlation analysis (blue dashed line), ultimate-DECIGO (purple dotted line), and correlation of analysis of ultimate-DECIGO (red dot-dashed line). (Bottom) Same as the top panel for the dilution factor $F=10$ for $T_\sigma$=10 GeV and $T_{\rm R}=10^7$ GeV. []{data-label="fig:GWspec"}](GW_D10.eps "fig:"){width="0.7\linewidth"} Scalar perturbation in the curvaton scenario -------------------------------------------- The curvaton is a scalar field other than the inflaton, which remains light during inflation and has a quantum fluctuation $\delta \sigma \sim H_{\rm inf}/(2\pi)$. Then, according to the $\delta N$-formalism [@Starobinsky:1986fxa], the comoving curvature perturbation ${{{\cal R}_c}}$ is given by $${{{\cal R}_c}}= N_\sigma \delta \sigma +\frac{1}{2}N_{\sigma \sigma}(\delta \sigma)^2,$$ up to the second order in $\delta \sigma$, where $N$ is the local number of $e$-folds, given by the integral of the local expansion from an initial spatially flat hypersurface to a final uniform density hypersurface, and $N_\sigma$ is its derivative with respective to $\sigma$. Note that we assume the curvaton behaves as free scalar field and hence $\delta \sigma$ can be regarded as a random Gaussian variable. Then we obtain [@Lyth:2005fi] $${{{\cal R}_c}}= \frac{2R}{3}\left ( \frac{\delta \sigma}{\sigma_i} \right ) +\left( \frac{R}{3}-\frac{4R^2}{9}-\frac{2R^3}{9} \right) \left ( \frac{\delta \sigma}{\sigma_i} \right )^2, \label{zeta}$$ where $\sigma_i$ is the initial amplitude of the curvaton during inflation, and $$R=\left. \frac{3\rho_\sigma}{4\rho_r + 3\rho_\sigma} \right|_{\sigma~{\rm decay}}$$ roughly denotes the fraction of the curvaton energy density to the total energy density at the epoch of curvaton decay. Here $\rho_\sigma$ and $\rho_r$ are the energy densities of the curvaton and that of radiation, respectively. In order to reproduce the observed density perturbation of the Universe, we must have $$\sqrt{ \Delta_{{{\cal R}_c}}^2(k_*)} =\frac{R}{3} \left(\frac{H_{\rm inf}}{\pi \sigma_i}\right) = 5\times 10^{-5}. \label{normal}$$ Notice that the inflaton also generates curvature perturbation whose magnitude is $\Delta_{{{{\cal R}_c}}\phi}^2 = H_{\rm inf}^2/(8\pi^2 \epsilon M_{\rm Pl}^2)$. It is much smaller than curvaton’s contribution (\[normal\]) by assumption. This leads to a constraint $$\epsilon > \frac{9}{8R^2}\left ( \frac{\sigma_i}{M_{\rm Pl}} \right )^2 = 9\times 10^{-3} \left ( \frac{H_{\rm inf}}{10^{14} ~{\rm GeV}} \right )^2, \label{epsconst}$$ where we have used the WMAP normalization (\[normal\]). We can calculate the tensor-to-scalar ratio $r$ in the curvaton scenario, as $$r \equiv \frac{\Delta_h^2(k_*)}{\Delta_{{{\cal R}_c}}^2(k_*)} =\frac{18}{R^2}\left ( \frac{\sigma_i}{M_{\rm Pl}} \right )^2 =0.14\left(\frac{H_{\rm inf}}{10^{14}\mathrm{GeV}}\right)^2. \label{r}$$ It can be checked that this is smaller than $16\epsilon$ using Eq. (\[epsconst\]), which is the prediction of the standard inflation scenario, once we assume that the curvature perturbation from the inflaton should be smaller than that from the curvaton. The scalar spectral index is given by [@Lyth:2001nq] $$n_s = 1-2\epsilon +\frac{2m_\sigma^2}{3H_{\rm inf}^2},$$ where $m_\sigma$ is the curvaton mass. We can determine it through the relation $$m_\sigma^2=\frac{3}{2}(n_s-n_t-1)H_{\rm inf}^2 =11(n_s-n_t-1)r\times(10^{14}\mathrm{GeV})^2,$$ in principle. But it is generically much smaller than $H_{\rm inf}^2$, so that it would be more practical to use the above equality as a consistency relation, $$n_s-n_t\cong 1, \label{consistency}$$ for the curvaton scenario. Thus in the curvaton scenario, both the scalar and tensor spectral tilt are predicted to be $-2\epsilon$, and hence a measurement of the tensor spectral index by future space-based gravitational wave detectors [@Seto:2005qy] will give an evidence of the curvaton scenario. Non-Gaussianity in the curvaton scenario ---------------------------------------- Statistics of the observed CMB anisotropy is currently consistent with Gaussian distribution. The deviation from Gaussianity is parameterized by the non-linearity parameter $f_{\rm NL}$, whose definition is given by $${{{\cal R}_c}}= {{{\cal R}_c}}^{(\rm g)}+\frac{3}{5}f_{\rm NL}{{{\cal R}_c}}^{(\rm g)2},$$ where ${{{\cal R}_c}}^{(\rm g)}$ denotes the Gaussian part of the curvature perturbation. WMAP5 result gives a constraint on it as $-9<f_{\rm NL}<111$ at 95% C.L. [@Komatsu:2008hk]. From Eq. (\[zeta\]), we can estimate the non-linearity parameter as[^3] $$f_{\rm NL}=\frac{5}{4R}\left (1- \frac{4}{3}R-\frac{2}{3}R^2 \right ). \label{fNL-R}$$ Thus $f_{\rm NL}$ can be significantly large for small $R$, and this may provide an observational hint of the curvaton scenario if large $f_{\rm NL}$ is detected, because standard inflation models predict a small non-linearity parameter, $f_{\rm NL}\sim \mathcal O (\epsilon, \eta)$. Here we have assumed that the curvaton potential is quadratic, $V=(1/2)m_\sigma^2 \sigma^2$. If it deviates from the quadratic one, the prediction of $f_{\rm NL}$ changes [@Enqvist:2005pg] and in this case the coefficient of the trispectrum, $g_{\rm NL}$, may be useful to distinguish curvaton models. We do not go into the detail on this point. Notice that in the case where a large non-Gaussianity ($f_{\rm NL} \gtrsim 10$) is obtained, the curvaton decay does not increase entropy because it must be a subdominant component at the instance of its decay ([*i.e.*]{}, $R\ll 1$), so that the gravitational wave background spectrum is not modified. In this sense, the detections of $f_{\rm NL}$ by the CMB observation and entropy production process by the gravitational wave background as a probe of the curvaton scenario are complementary to each other. To summarize, we have five possible observable quantities: scalar spectral index $n_s$, tensor-to-scalar ratio $r$, tensor spectral index $n_t$, the dilution factor $F$ and non-linearity parameter $f_{\rm NL}$. Among them, $r$ will be accurately determined by future B-mode polarization measurements. Direct detection of gravitational waves will observe $n_t$, and comparing it with $n_s$ will confirm the curvaton scenario. Then, either large enough $F$ or $f_{\rm NL}$ will be observed depending on whether the curvaton once dominated the Universe or not. In the former case, $F$ can be determined by comparing $r$ and directly observed magnitude of the gravitational waves. In the latter case, $f_{\rm NL}$ will be determined by CMB measurements such as Planck. Then these observables may be used to pin down a curvaton model. In the next section we investigate the possibility to detect either signature of the curvaton scenario and how they can fix properties of the curvaton. Probing the Curvaton Scenario {#sec:probe} ============================= We have seen that a curvaton scenario may leave distinct signatures on either a shape of the gravitational wave background or primordial non-Gaussianity. Here we show parameter regions where a curvaton scenario has characteristic features on either of them. Curvaton models are characterized by two parameters, the initial amplitude of the curvaton $\sigma_i$ and its decay temperature $T_\sigma$. The relevant quantity is the abundance of the curvaton coherent oscillation at the time of its decay, which depends on these two parameters and the reheating temperature after inflation, $T_{\rm R}$, but does not depend on the curvaton mass.[^4] Once these parameters are fixed, we can estimate the inflation scale $H_{\rm inf}$ in order to reproduce the observed magnitude of the density perturbation (see Eq. (\[normal\])). The inflation scale $H_{\rm inf}$ also gives overall normalization of the gravitational wave background spectrum, which can be directly measured independently using B-mode polarization of CMB. The spectral shape is determined by $T_{\rm R}$ and the curvaton abundance at its decay. Thus we can uniquely predict the gravitational wave signal at an arbitrary frequency for each parameter set ($T_{\rm R}, T_{\sigma}, \sigma_i$). Similarly, the level of non-Gaussianity depends only on the curvaton abundance at its decay, as given in Eq. (\[fNL-R\]), and hence is also uniquely predicted. Figures 2 depict the parameter region of the curvaton which can be probed by observation of DECIGO/BBO. The upper panel represents the region accesible with a single ultimate-DECIGO and B-mode experiments such as EPIC [@Bock:2008ww], CMBPol [@Baumann:2008aq], and LiteBIRD [@LiteBIRD] which are expected to reach down to $r \simeq 10^{-3}$. The lower panel shows the ideal case of correlation analysis of ultimate-DECIGO together with low-noise delensed CMB map which would hopefully reach $r\sim 2\times 10^{-6}$ [@Marian:2007sr] (see also [@Knox:2002pe]). Sensitivities of these direct detection experiments for the low frequency are limited by stochastic noise from white dwarf binaries [@Farmer:2003pa], and hence we have cut the sensitivities below 0.1 Hz. We have set $\epsilon$ to satisfy the constraint (\[epsconst\]), but the precise value of $\epsilon$ does not affect the results as long as it is sufficiently small. The region above the purple wedge corresponds to $r>0.2$ which is excluded by WMAP. Also plotted there are contours of the non-linearity parameter $f_{\rm NL}$ and the dilution factor $F$. The region with $f_{\rm NL}> 100$ is also disfavored by WMAP. On the other hand, Planck can measure it if it lies in the range $10< f_{\rm NL}$. As argued before, it occupies a detached domain from the region with $F>1$. Therefore below we study how the curvaton parameters are determined in each domain separately. First let us consider the case $f_{\rm NL}\gg 1$ is confirmed by, say, Planck experiment. Then from (\[fNL-R\]) we find $R \cong 5/(4f_{\rm NL})$ and therefore (\[r\]) determines the initial amplitude of the curvaton in terms of the observable quantities alone as $$\frac{\sigma_i}{M_{\rm Pl}}=\frac{5}{12f_{\rm NL}} \left(\frac{r}{2}\right)^{1/2}. \label{sigmai}$$ In this case the curvaton decays before dominating the cosmic energy density. Hence we find $$\left. \frac{\rho_\sigma}{\rho_r}\right|_{\sigma~{\rm decay}} =\frac{1}{6}\left(\frac{\sigma_i}{M_{\rm Pl}}\right)^2 \frac{a(T_\sigma)}{a(T_{\rm R})} =\frac{1}{6}\left(\frac{\sigma_i}{M_{\rm Pl}}\right)^2 \frac{T_{\rm R}}{T_\sigma}\cong \frac{4}{3}R.$$ These two equalities lead to $$\frac{T_\sigma}{T_{\rm R}}=\frac{5r}{576f_{\rm NL}}.$$ Thus both of the curvaton parameters are fixed by the observable quantities in this case. Finally we note that in this case there is a simple relation between $T_\sigma/T_{\rm R}$ and $\sigma_i/M_{\rm Pl}$ as $$\frac{T_\sigma}{T_{\rm R}}=\frac{f_{\rm NL}}{10} \left(\frac{\sigma_i}{M_{\rm Pl}}\right)^2,$$ which was used to draw the contours of $f_{\rm NL}$ in Figures 2. Next we consider the case the curvaton dominates the energy density of the universe when it decays releasing significant amount of entropy with $R=1$ and $F > 1$. Then from (\[r\]) we find $$\frac{\sigma_i}{M_{\rm Pl}}=\left(\frac{r}{18}\right)^{1/2}$$ and from (\[F\]) we obtain $$\frac{T_\sigma}{T_{\rm R}}=\frac{r}{108F}, \label{tsigma}$$ for $T_{\rm osc}>T_{\rm R}$. We can determine $F$ if we can measure tensor perturbations both by the B-mode polarization of CMB and DECIGO/BBO, provided there is no other source of entropy production mechanism after reheating besides the curvaton. Then all the relevant parameters are again fixed by the observable quantities. The corresponding parameter space is shown by the orange regions in Figs. \[fig:DEC\]. In case there is other entropy production besides curvaton, the right-hand-side of (\[tsigma\]) will give a lower bound on $T_\sigma/T_{\rm R}$. Finally, if B-mode polarization measurements fail to find the tensor mode and if it is detected only by DECIGO/BBO, which is shown by the yellow regions in Figs. \[fig:DEC\], the values of $F$ and $r$ cannot be determined independently. Only the combination $rF^{-4/3}$ is determined in this case. Results are summarized in Fig. \[fig:DEC\]. Contours of the non-linearity parameter $f_{\rm NL}$ are plotted on the $\sigma_i/M_{\rm Pl}$- $T_\sigma/T_{\rm R}$ plane. Regions where $f_{\rm NL}$ is much larger than 100 is inconsistent with WMAP observation [@Komatsu:2008hk]. On the other hand, if it lies in the range $10<f_{\rm NL}<100$, it can be confirmed by the Planck satellite. Region above the purple line corresponds to $H_{\rm inf} > 10^{14}$ GeV, which is excluded since the tensor mode contribution to the CMB anisotropy becomes too large. Future satellite experiments dedicated to detect a B-mode polarization of CMB will find a primordial tensor mode for $H_{\rm inf}\gtrsim 10^{13}~$GeV [@Bock:2006yf]. Also we show a region where space-laser interferometer experiments can detect a stochastic gravitational wave background by the blue line. Top and bottom panels correspond to single ultimate-DECIGO and correlation analysis of ultimate-DECIGO, respectively. In the region $f_{\rm NL} \ll 10$ and above the detectability lines of gravitational waves, a significant amount of entropy production occurs, which can be confirmed by these gravitational wave detectors. It can be seen that such a confirmation is complementary to the detection of non-Gaussianity from these figures, as already stated. The region below all of these lines is also allowed, although no observable signature may arise. ![ Range of the curvaton parameters $\sigma_i/M_{\rm Pl}$- $T_\sigma/T_{\rm R}$ which can be probed by space-based laser interferometers. The upper panel represents the case with single ultimate-DECIGO and B-mode measurements down to $r=10^{-3}$, while the lower panel shows an ideal case with correlation analysis of ultimate-DECIGO and B-mode measurements accesible to $r=2\times 10^{-6}$. Region above the red wedge is excluded since the tensor mode contribution to the CMB anisotropy becomes too large. Also shown there are contours of the non-linearity parameter $f_{\rm NL}$. Upper left region above the solid line is disfavored by WMAP. In the green region all the curvaton parameters can be determined in terms of $f_{\rm NL}$ and $r$, while in the orange region they can be determined by $F$ and $r$ provided other sources of entropy production is absent. In the yellow region, $F$ and $r$ are not determined independently, and only the combination $rF^{-4/3}$ is determined. []{data-label="fig:DEC"}](ulDEC.eps "fig:"){width="0.7\linewidth"} ![ Range of the curvaton parameters $\sigma_i/M_{\rm Pl}$- $T_\sigma/T_{\rm R}$ which can be probed by space-based laser interferometers. The upper panel represents the case with single ultimate-DECIGO and B-mode measurements down to $r=10^{-3}$, while the lower panel shows an ideal case with correlation analysis of ultimate-DECIGO and B-mode measurements accesible to $r=2\times 10^{-6}$. Region above the red wedge is excluded since the tensor mode contribution to the CMB anisotropy becomes too large. Also shown there are contours of the non-linearity parameter $f_{\rm NL}$. Upper left region above the solid line is disfavored by WMAP. In the green region all the curvaton parameters can be determined in terms of $f_{\rm NL}$ and $r$, while in the orange region they can be determined by $F$ and $r$ provided other sources of entropy production is absent. In the yellow region, $F$ and $r$ are not determined independently, and only the combination $rF^{-4/3}$ is determined. []{data-label="fig:DEC"}](ulDECcor.eps "fig:"){width="0.7\linewidth"} Conclusion {#sec:conc} ========== In this paper we have investigated a possible observable signatures from curvaton scenarios, including non-Gaussianity, detection of tensor modes in CMB and direct detection of gravitational wave background. First the measurement of the tensor power spectrum by CMB is very important to identify the curvaton scenario through the consistency relation (\[consistency\]). If the curvaton once dominates the Universe, an entropy production process by its decay is imprinted in the gravitational wave background spectrum and can be confirmed by future space-based laser interferometer experiments. For the opposite case where the curvaton is a subdominant component before it decays, a large non-Gaussianity is predicted, which may be confirmed by Planck experiment. In this sense, the gravitational wave signal and non-Gaussian signal are complementary to each other. This allows us to probe the curvaton scenario for large parameter spaces, giving information on the properties of the curvaton such as its decay rate and amplitude, etc. It will in turn provide important information on high-energy physics beyond the reach of terrestrial experiments. Acknowledgements {#acknowledgements .unnumbered} ================ K.N. would like to thank S. Kuroyanagi for valuable discussion. He would also like to thank the Japan Society for the Promotion of Science for financial support. This work was supported in part by JSPS Grant-in-Aid for Scientific Research No. 19340054(JY) and by Global COE Program “the Physical Sciences Frontier", MEXT, Japan. A.H. Guth, [ Phys. Rev.]{} [**D23**]{}, 347 (1981); K. Sato, Mon. Not. Roy. Astron. Soc. [**195**]{}, 467 (1981); A.A. Starobinsky Phys. Lett. [**91B**]{}, 99 (1980); For a review, see, for example, A.D. Linde, arXiv:0705.0164 \[hep-th\]. S.W. Hawking, [ Phys. Lett.]{} [**115B**]{}, 295 (1982); A.A. Starobinsky, Phys. Lett. [**117B**]{}, 175 (1982);  A.H. Guth and S-Y. Pi, [ Phys. Rev. Lett.]{} [**49**]{}, 1110 (1982). E. Komatsu [*et al.*]{} \[WMAP Collaboration\], arXiv:0803.0547 \[astro-ph\]. A.A. Starobinsky, JETP Lett. [**30**]{}, 682 (1979); V.A. Rubakov, M.V. Sazin, and A.V. Veryaskin, Phys. Lett. [**115B**]{}, 189(1982); L.F. Abbott and M.B. Wise, Nucl. Phys. B [**244**]{}, 541 (1984). B. Allen, Phys. Rev.  D [**37**]{}, 2078 (1988). M. S. Turner and F. Wilczek, Phys. Rev. Lett.  [**65**]{}, 3080 (1990). V. Sahni, Phys. Rev.  D [**42**]{}, 453 (1990). M. S. Turner, M. J. White and J. E. Lidsey, Phys. Rev.  D [**48**]{}, 4613 (1993) \[arXiv:astro-ph/9306029\]. M. S. Turner, Phys. Rev.  D [**48**]{}, 3502 (1993) \[arXiv:astro-ph/9302013\]. A. R. Liddle, Phys. Rev.  D [**49**]{}, 3805 (1994) \[Erratum-ibid.  D [**51**]{}, 4603 (1995)\] \[arXiv:gr-qc/9307036\]. M. S. Turner and M. J. White, Phys. Rev.  D [**53**]{}, 6822 (1996) \[arXiv:astro-ph/9512155\]. M. S. Turner, Phys. Rev.  D [**55**]{}, 435 (1997) \[arXiv:astro-ph/9607066\]. B. Allen and J. D. Romano, Phys. Rev.  D [**59**]{}, 102001 (1999) \[arXiv:gr-qc/9710117\]. N. Seto and J. Yokoyama, J. Phys. Soc. Jap.  [**72**]{}, 3082 (2003) \[arXiv:gr-qc/0305096\]. S. Weinberg, Phys. Rev.  D [**69**]{}, 023503 (2004) \[arXiv:astro-ph/0306304\]. H. Tashiro, T. Chiba and M. Sasaki, Class. Quant. Grav.  [**21**]{}, 1761 (2004) \[arXiv:gr-qc/0307068\]. C. Ungarelli, P. Corasaniti, R. A. Mercer and A. Vecchio, Class. Quant. Grav.  [**22**]{}, S955 (2005) \[arXiv:astro-ph/0504294\]. T. L. Smith, M. Kamionkowski and A. Cooray, Phys. Rev.  D [**73**]{}, 023504 (2006) \[arXiv:astro-ph/0506422\]. N. Seto, Phys. Rev.  D [**73**]{}, 063001 (2006) \[arXiv:gr-qc/0510067\]. H. Kudoh, A. Taruya, T. Hiramatsu and Y. Himemoto, Phys. Rev.  D [**73**]{}, 064006 (2006) \[arXiv:gr-qc/0511145\]. L. A. Boyle and P. J. Steinhardt, Phys. Rev.  D [**77**]{}, 063504 (2008) \[arXiv:astro-ph/0512014\]. T. L. Smith, H. V. Peiris and A. Cooray, Phys. Rev.  D [**73**]{}, 123503 (2006) \[arXiv:astro-ph/0602137\]. S. Chongchitnan and G. Efstathiou, Phys. Rev.  D [**73**]{}, 083511 (2006) \[arXiv:astro-ph/0602594\]; G. Efstathiou and S. Chongchitnan, Prog. Theor. Phys. Suppl.  [**163**]{}, 204 (2006) \[arXiv:astro-ph/0603118\]. B. C. Friedman, A. Cooray and A. Melchiorri, Phys. Rev.  D [**74**]{}, 123509 (2006) \[arXiv:astro-ph/0610220\]. W. Zhao and Y. Zhang, Phys. Rev.  D [**74**]{}, 043503 (2006). Y. Watanabe and E. Komatsu, Phys. Rev.  D [**73**]{}, 123515 (2006) \[arXiv:astro-ph/0604176\]. T. Chiba, Y. Himemoto, M. Yamaguchi and J. Yokoyama, Phys. Rev.  D [**76**]{}, 043516 (2007) \[arXiv:0705.2851 \[gr-qc\]\]. L. A. Boyle and A. Buonanno, Phys. Rev.  D [**78**]{}, 043531 (2008) \[arXiv:0708.2279 \[astro-ph\]\]. J. Yokoyama, An essay submitted to Gravity Research Foundation and received Honorble Mention (2007). T. L. Smith, M. Kamionkowski and A. Cooray, Phys. Rev.  D [**78**]{}, 083525 (2008) \[arXiv:0802.1530 \[astro-ph\]\]. K. Nakayama, S. Saito, Y. Suwa and J. Yokoyama, Phys. Rev.  D [**77**]{}, 124001 (2008) \[arXiv:0802.2452 \[hep-ph\]\]. K. Nakayama, S. Saito, Y. Suwa and J. Yokoyama, JCAP [**0806**]{}, 020 (2008) \[arXiv:0804.1827 \[astro-ph\]\]. S. Kuroyanagi, T. Chiba and N. Sugiyama, Phys. Rev.  D [**79**]{}, 103501 (2009) \[arXiv:0804.3249 \[astro-ph\]\]. A. Mangilli, N. Bartolo, S. Matarrese and A. Riotto, arXiv:0805.3234 \[astro-ph\]. M. Maggiore, Phys. Rept.  [**331**]{}, 283 (2000) \[arXiv:gr-qc/9909001\]. N. Seto, S. Kawamura and T. Nakamura, Phys. Rev. Lett.  [**87**]{}, 221103 (2001) \[arXiv:astro-ph/0108011\]; S. Kawamura [*et al.*]{}, Class. Quant. Grav.  [**23**]{} (2006) S125. S. Mollerach, Phys. Rev.  D [**42**]{}, 313 (1990); A. D. Linde and V. F. Mukhanov, Phys. Rev.  D [**56**]{}, 535 (1997) \[arXiv:astro-ph/9610219\]. D. H. Lyth and D. Wands, Phys. Lett.  B [**524**]{}, 5 (2002) \[arXiv:hep-ph/0110002\]; T. Moroi and T. Takahashi, Phys. Lett.  B [**522**]{}, 215 (2001) \[Erratum-ibid.  B [**539**]{}, 303 (2002)\] \[arXiv:hep-ph/0110096\]; K. Enqvist and M. S. Sloth, Nucl. Phys.  B [**626**]{}, 395 (2002) \[arXiv:hep-ph/0109214\]. For a review, see N. Bartolo, E. Komatsu, S. Matarrese and A. Riotto, Phys. Rept.  [**402**]{}, 103 (2004) \[arXiv:astro-ph/0406398\]. D. H. Lyth, C. Ungarelli and D. Wands, Phys. Rev.  D [**67**]{}, 023503 (2003) \[arXiv:astro-ph/0208055\]; N. Bartolo, S. Matarrese and A. Riotto, Phys. Rev.  D [**69**]{}, 043503 (2004) \[arXiv:hep-ph/0309033\]; K. A. Malik and D. H. Lyth, JCAP [**0609**]{}, 008 (2006) \[arXiv:astro-ph/0604387\]; M. Sasaki, J. Valiviita and D. Wands, Phys. Rev.  D [**74**]{}, 103003 (2006) \[arXiv:astro-ph/0607627\]; Q. G. Huang, Phys. Lett.  B [**669**]{}, 260 (2008) \[arXiv:0801.0467 \[hep-th\]\]; JCAP [**0809**]{}, 017 (2008) \[arXiv:0807.1567 \[hep-th\]\]; JCAP [**0811**]{}, 005 (2008) \[arXiv:0808.1793 \[hep-th\]\]; K. Ichikawa, T. Suyama, T. Takahashi and M. Yamaguchi, Phys. Rev.  D [**78**]{}, 023513 (2008) \[arXiv:0802.4138 \[astro-ph\]\]; M. Beltran, Phys. Rev.  D [**78**]{}, 023530 (2008) \[arXiv:0804.1097 \[astro-ph\]\]; T. Takahashi, M. Yamaguchi, J. Yokoyama and S. Yokoyama, Phys. Lett.  B [**678**]{}, 15 (2009) \[arXiv:0905.0240 \[astro-ph.CO\]\]. K. Enqvist and S. Nurmi, JCAP [**0510**]{}, 013 (2005) \[arXiv:astro-ph/0508573\]; K. Enqvist and T. Takahashi, JCAP [**0809**]{}, 012 (2008) \[arXiv:0807.3069 \[astro-ph\]\]; arXiv:0909.5362 \[astro-ph.CO\]; Q. G. Huang and Y. Wang, JCAP [**0809**]{}, 025 (2008) \[arXiv:0808.1168 \[hep-th\]\]; M. Kawasaki, K. Nakayama and F. Takahashi, JCAP [**0901**]{}, 026 (2009) \[arXiv:0810.1585 \[hep-ph\]\]; P. Chingangbam and Q. G. Huang, JCAP [**0904**]{}, 031 (2009) \[arXiv:0902.2619 \[astro-ph.CO\]\]. V. Acquaviva, N. Bartolo, S. Matarrese and A. Riotto, Nucl. Phys.  B [**667**]{}, 119 (2003) \[arXiv:astro-ph/0209156\]. J. M. Maldacena, JHEP [**0305**]{}, 013 (2003) \[arXiv:astro-ph/0210603\]. D. Seery and J. E. Lidsey, JCAP [**0506**]{}, 003 (2005) \[arXiv:astro-ph/0503692\]; JCAP [**0509**]{}, 011 (2005) \[arXiv:astro-ph/0506056\]; D. Seery, J. E. Lidsey and M. S. Sloth, JCAP [**0701**]{}, 027 (2007) \[arXiv:astro-ph/0610210\]. S. Yokoyama, T. Suyama and T. Tanaka, JCAP [**0707**]{}, 013 (2007) \[arXiv:0705.3178 \[astro-ph\]\]; Phys. Rev.  D [**77**]{}, 083511 (2008) \[arXiv:0711.2920 \[astro-ph\]\]. A. A. Starobinsky, JETP Lett.  [**42**]{} (1985) 152 \[Pisma Zh. Eksp. Teor. Fiz.  [**42**]{} (1985) 124\]. Y. Nambu and A. Taruya, Class. Quant. Grav.  [**13**]{}, 705 (1996) \[arXiv:astro-ph/9411013\]. M. Sasaki and E. D. Stewart, Prog. Theor. Phys.  [**95**]{}, 71 (1996) \[arXiv:astro-ph/9507001\]; M. Sasaki and T. Tanaka, Prog. Theor. Phys.  [**99**]{}, 763 (1998) \[arXiv:gr-qc/9801017\]. H. Kodama and T. Hamazaki, Phys. Rev.  D [**57**]{}, 7177 (1998) \[arXiv:gr-qc/9712045\]; D. H. Lyth, K. A. Malik and M. Sasaki, JCAP [**0505**]{}, 004 (2005) \[arXiv:astro-ph/0411220\]; T. Hamazaki, Phys. Rev.  D [**78**]{}, 103513 (2008) \[arXiv:0811.2366 \[astro-ph\]\]. D. H. Lyth and Y. Rodriguez, Phys. Rev. Lett.  [**95**]{}, 121302 (2005) \[arXiv:astro-ph/0504045\]. M. Kawasaki, K. Nakayama, T. Sekiguchi, T. Suyama and F. Takahashi, JCAP [**0811**]{}, 019 (2008) \[arXiv:0808.0009 \[astro-ph\]\]; JCAP [**0901**]{}, 042 (2009) \[arXiv:0810.0208 \[astro-ph\]\]; M. Kawasaki, K. Nakayama and F. Takahashi, JCAP [**0901**]{}, 002 (2009) \[arXiv:0809.2242 \[hep-ph\]\]; E. Kawakami, M. Kawasaki, K. Nakayama and F. Takahashi, JCAP [**0909**]{}, 002 (2009) \[arXiv:0905.1552 \[astro-ph.CO\]\]; K. Nakayama and F. Takahashi, Phys. Lett.  B [**679**]{}, 436 (2009) \[arXiv:0907.0834 \[hep-ph\]\]. D. Langlois, F. Vernizzi and D. Wands, JCAP [**0812**]{}, 004 (2008) \[arXiv:0809.4646 \[astro-ph\]\]. C. Hikage, K. Koyama, T. Matsubara, T. Takahashi and M. Yamaguchi, arXiv:0812.3500 \[astro-ph\]; C. Hikage, D. Munshi, A. Heavens and P. Coles, arXiv:0907.0261 \[astro-ph.CO\]. J. Bock [*et al.*]{}, arXiv:0805.4207 \[astro-ph\]. D. Baumann [*et al.*]{} \[CMBPol Study Team Collaboration\], AIP Conf. Proc.  [**1141**]{} (2009) 10 \[arXiv:0811.3919 \[astro-ph\]\]. http://cmbpol.kek.jp/litebird/index.html L. Marian and G. M. Bernstein, Phys. Rev.  D [**76**]{}, 123009 (2007) \[arXiv:0710.2538 \[astro-ph\]\]. L. Knox and Y. S. Song, Phys. Rev. Lett.  [**89**]{}, 011303 (2002) \[arXiv:astro-ph/0202286\]; M. Kesden, A. Cooray and M. Kamionkowski, Phys. Rev. Lett.  [**89**]{}, 011304 (2002) \[arXiv:astro-ph/0202434\]. A. J. Farmer and E. S. Phinney, Mon. Not. Roy. Astron. Soc.  [**346**]{}, 1197 (2003) \[arXiv:astro-ph/0304393\]. [^1]: There was an error in Eq. (14) of Ref. [@Nakayama:2008wy]. Eq. (\[Deltah\]) is the correct one. [^2]: The effect of neutrino free streaming is known to lead a suppression on the gravitational wave background spectrum around the frequency $\sim 10^{-9}$ Hz [@Weinberg:2003ur; @Watanabe:2006qe]. But our concern is around 1 Hz, and hence we simply neglect this effect. [^3]: Here we assume that there are no CDM/baryonic isocurvature perturbations, which might cause other types of non-Gaussianity [@Kawasaki:2008sn; @Langlois:2008vk; @Hikage:2008sk]. [^4]: If the curvaton begins to oscillate after the inflaton decays, the mass dependence appears. In this case, however, it is sufficient to replace $T_{\rm R}$ with $T_{\rm osc}$, the temperature at which the curvaton begins to oscillate, as in Eq. (\[F\]).

--- author: - | [Yu. Arestov$^{1,2}$, R.M. Santilli$^2$ and V. Solovianov$^1$]{}\ [*$^1$ Institute for High Energy Physics, 142284, Protvino, Russia* ]{}\ [*$^2$ Institute for Basic Research, P. O. Box 1577, Palm Harbor, FL 34682*]{} title: | **Insufficiency of Available Data on the Behaviour\ of the Meanlives of Unstable Hadrons with Energy\ ** --- =-1.5cm =-1.5cm \ #### Data on local speeds of light. {#data-on-local-speeds-of-light. .unnumbered} Strictly speaking, the speed of electromagnetic (elm) waves is not a “universal constant”, but rather a quantity $c = c_o/n$ depending on local physical conditions representable via the index of refraction n, where $c_o$ is the speed in vacuum. When experimentally established, deviations from $c_o$ are then rather forceful evidence of deviations from the conventional Minkowskian spacetime of the vacuum \[1a\]. Speeds $c = c_o/n < c_o$ are known in our Newtonian environment. Lesser known is the fact that one of the first studies on the implications of speeds $c < c_o$ were first studied by Lorentz \[1b\] (see the related mention in Pauli’s book \[1c\]). Speeds $c = c_o / n > c_o$ have been apparently measured by A. Enders and G. Nimtz \[1d\] in the tunneling of photons between certain guides (see review \[1e\] for additional references and details). Apparent speeds $c = c_o/n > c_o$ have also been identified in certain astrophysical events \[1f-1h\] (see also the recent data \[1i\]). Note that the hopes of regaining the exact Minkowskian spacetime by reducing light to photons scattering among molecules, even though valid as a first approximation, is no longer viable because: 1) the reduction to second quantization is questionable for elm waves in our atmosphere, say, with one meter wavelength; 2) the reduction does not permit quantitative studies of superluminal speeds; and 3) the reduction eliminates the representation of the inhomogeneity and anisotropy of physical media, which have apparent, experimentally measurable effects (see below). Recall that hadrons are not ideal spheres with isolated points in them, but rather some of the densest media measured in laboratory until now. If spacetime anomalies are established for media of relatively low density, the hypothesis that the Minkowskian spacetime can be [*exact*]{} within hadrons in its conventional realization has little scientific credibility (see below for the exact character of an axiom-preserving covering spacetime). Also, deviations are expected from the complete mutual penetration of the wavepackets of the constituents, thus resulting in the historical open legacy of the existence of nonlinear, nonlocal and nonpotential effects in the interior of hadronic. One of the first quantitative studies of the above legacy was done by D. L. Blokhintsev \[2a\] in 1964, followed by L. B. Redei \[2b\], D. Y. Kim \[2c\] and others. Note that the exact validity of the Minkowskian geometry for the [*center-of-mass behavior*]{} of a hadron in a particle accelerator is beyond scientific doubts. The authors of Refs. \[2a-2c\] then argued that a possibility for internal anomalies due to nonlocal and other effects to manifest themselves in the outside is [*via deviations from the conventional Minkowskian behavior of the meanlives of unstable hadrons with the speed*]{} v (or energy E). Note that the Minkowski metric can be written $\eta = {\rm Diag}. (1, 1, 1, -c_o^2)$. Therefore, [*any deviation*]{} $\hat{\eta}$ [*from*]{} $\eta$ [*necessarily implies a deviation from*]{} $c_o$, as one can see by altering any component of the metric and then using Lorentz transforms. Along these lines, R. M. Santilli \[2d\] submitted in 1982 the hypothesis that [*contact-nonpotential interactions (thus including the strong interactions as per the above legacy) can accelerate ordinary (positive) masses at speed bigger than the speed of light in vacuun* ]{} much along the subsequent astrophysical measures \[1f-1h\]. The above hypothesis implies that [*photons travel inside the hyperdense hadrons at speeds bigger than that in vacuum*]{}. V. de Sabbata and M. Gasperini \[2e\] conducted the first phenomenological verification within the context of the conventional gauge theories supporting the hypothesis of Ref. \[2d\], and actually reaching limit speeds up to $75 c_o$ for superheavy hadrons. The above hypothesis is also supported by the phenomenological calculations conducted by H. B. Nielsen and I. Picek \[2f\] via the spontaneous symmetry breaking in the Higgs sector of conventional gauge theories, which have resulted in the anomalous Minkowskian metrics (here written in the notation above) $$\begin{aligned} \pi:~~ \hat {\eta} = {\rm Diag}.[(1 + 1.2\cdot 10^{-3}),~(1 + 1.2\cdot 10^{-3}),~(1 + 1.2\cdot 10^{-3}),~ - c_o^2(1 - 3.79\cdot 10^{-3})],\\ K:~~ \hat{\eta} = {\rm Diag}. [(1 - 2.0\cdot 10^{-4}),~ (1 - 2.0\cdot 10^{-4}),~ (1 - 2.0\cdot 10^{-4}),~ - c_o^2(1 + 6.00\cdot 10^{-4})]. \label{(2)}\end{aligned}$$ As one can see, calculations \[2f\] confirm speeds of photons $c = c_o/n > c_o$ for the interior of kaons, as conjectured in Ref. \[2d\]. Recall that: spacetime anomalies are expected to increase with the density; all hadrons have approximately the same size; and hadrons have densities increasing with mass. Therefore, results similar to (2) are expected for all hadrons [*heavier*]{} than kaons, as supported by phenomenological studies \[2e\]. The first direct experimental measures on the behavior of the meanlife of $K_S^o$ with energy, $\tau (E)$, were done by S. H. Aronson [*et al.*]{} \[3a\] at Fermilab and they suggested [*deviations*]{} from the Minkowskian spacetime in the energy range of 30 to 100 GeV. Subsequent direct measures also for $K_S^o$ were done by S. H. Aronson [*et al.*]{} \[3b\] also at Fermilab, suggesting instead [*no deviations*]{} of $\tau (E)$ from the Minkowskian form in the [*different*]{} energy range of 100 to 400 GeV. More recently, a test of the decay law at short decay times was made by the OPAL group at LEP \[3c\]. In the latter experiment the ratio of number of events $Z^0 \rightarrow \tau^{+} \tau^{-}$ with deviations of $\tau$ from the conventional law to number of ”normal” events was $(1.1 \pm 1.4 \pm 3.5)\%$. #### Isominkowskian geometrization of physical media. {#isominkowskian-geometrization-of-physical-media. .unnumbered} A geometrization of all deviations from the Minkowskian spacetime was submitted by Santilli \[4a\] in 1983 under the name of [*isominkowskian geometry*]{} (see \[4b\] for the latest account) and resulted to be: “axiom-preserving” (in the sense that the isominkowskian geometry is isomorphic to the conventional one, a property denoted with the prefix “iso”); “invariant” (in the sense of admitting a symmetry isomorphic to the Poincaré symmetry \[4a-4d\]; and “universal” (in the sense of admitting all infinitely possible, well behaved, signature-preserving and symmetric modifications of the Minkowski metric \[4e\]). Moreover, the isominkowskian geometry has permitted the [*exact*]{} reconstruction of the special relativity under [*arbitrary*]{} local speeds of light \[4f\]. Refs. \[4\] have therefore established that, contrary to a popular belief (see, e.g., the “Lorentz asymmetry” of Ref. \[2f\]), the Minkowskian axioms, the Lorentz and Poincaré symmetry and the special relativity remain [*exact*]{} under all the above [*spacetime anomalies*]{}, of course, when properly formulated. The isominkowskian geometry is essentially characterized by the lifting of the Minkowskian metric $\eta \rightarrow \hat{\eta} = \hat T\times \eta$, where $\hat T(x, v, E, \mu, \tau, \omega, ...)$ is a positive-definite $4\times 4$ matrix with an arbitrary local dependence on coordinates x, speeds v, energies E, ensities $\mu$, temperatures $\tau$, frequencies $\omega$, and any other needed variable. . Jointly, the basic unit of the Minkowski space, $I$ = Diag. (1, 1, 1, 1), is lifted by an amount which is the [*inverse*]{} of the deformation of the metric, $I \rightarrow \hat I = 1/\hat T$. The dual lifting $\eta \rightarrow \hat{\eta} = \hat T\times \eta$ and $I \rightarrow \hat I = 1/\hat T$ then implies the preservation of all original spacetime axioms \[4\] (see Ref. \[41-4k\] for mathematical studies and \[4l\] for physical profiles). The isominkowskian geometry provides a geometrization of physical media at bopth the classical and operator levels \[4l\]. Since $\hat T$ is positive-definite, $\hat{\eta}$ can always be diagonalized in the form $\hat{\eta}$ = Diag. $(1/n_1^2, 1/n_2^2, 1/n_3^2, -c_o^2/n_4^2)$, thus providing a geometrization of: the local [*inhomogeneity*]{} (e.g., via a dependence of the n’s from the density); the local [*anisotropy*]{} (e.g., via a differentiation between the space and time n’s); as well as [*arbitrary local speeds of elm waves*]{} (via the expression $c(x, \mu, \omega, ...) = c_o/n_4(x, \mu, \omega, ...)$ first proposed in \[4a\]). The isotopic behavior of the meanlife with speed (or energy) for isotropic space with $n_1 = n_2 = n_3 = n_s(x, \mu, \omega, ...)$ (yet with general spacetime anisotropy $n_s \not = n_4$) is given by \[4a-4c\] $$\hat{\tau} = \tau_{o}\hat\gamma, ~\hat\gamma = (1-\hat\beta^2)^{-1/2}, ~\hat\beta = (v/n_s) / (c_{o}/n_4), \label{(3)}$$ and includes all existing or otherwise possible laws \[2\] via different power series expansions in terms of different parameters with different truncations \[4e\]. This eliminates the ambiguity of individually testing the several different laws of Refs. \[2\]. Note that, when a hadron is studied from the outside, one evidently can only use the [*average of the n-quantities to constants*]{}, called ”characteristic constants” of the medium considered. Note also that a possible anysotropy of the medium implies a deviation from the conventional Doppler shift studied by Mignani \[5a\] and others which will be studied elsewhere as a possible complement to measures \[2,3\]. Note finally that the latter anomalies are eliminated by the reduction of of light to photons [*moving in vacuum*]{} and scattering among molecules. Isotopic law (3) was applied by F. Cardone, R. Mignani and R. M. Santilli \[5b\] to the experimental data of Refs. \[3a,3b\] resulting in the single fit of both experiments, $$1/n_1^2 = 1/n_2^2 = 1/n_3^2 = 0.909080 \pm 0.004,~~ 1/n_4^2 = 1.003 \pm 0.002. \label{(4)}$$ Therefore, even under the assumption of the correct character of measures \[3b\], they do not establish the validity of the Minkowskian geometry inside hadrons because of the above isominkowskian fit. Note also that fit (4) confirms the superluminal character of the propagation of light within the hyperdense hadronic media, a property that appears to be confirmed by other studies (see the outline in \[4b\]). We should finally mention that nonlinear and nonlocal effects at short distances have been recently studied in Refs. \[5d,5e,4l\]. #### An alternative data elaboration. {#an-alternative-data-elaboration. .unnumbered} In this note we focus the attention on the range-energy selection rule which can be applied to re-elaborate the initial data on $K_S$ decays from the experiment \[3b\]. By taking into account the results as they were done, we performed Monte Carlo simulations of the main features of experiment \[3b\] and made our own fits for $K_S^o$. Our conclusions and recommendations are the following:\ 1) We agree that the parameters in the full formula $dN/dt$ for the proper time evolution are strongly correlated. This may cause a generally non-relevant regular dependence of the parameters on entities which are not present in the formula, such as number of runs, energy, etc., apart from the systematic uncertainties. Therefore, the above dependence may shadow the weak energy dependence we are interested in, as can be seen from the large values of the correlation elements. 2\) The authors of Ref. \[3b\] solved the problem of non-correlated fit by selecting the $K_S^o$ momenta greater than 100 GeV/c. By means of that energy cut, they obtained the data sample in which the CP violating terms contribute up to 1.6%. However, it seems unrealistic to look for the deviations from the Minkowskian decay law of the order of some percent. More realistic is to test the decay law on the level of $10^{-3}$, as suggested by studies \[2\]. In fact, the assumption of 1.6% contribution from PC violation in the data elaboration of Ref. \[3b\] implies looking for the energy dependence of $\tau_s$ at the level $k\cdot 10^{-2}$, thus rendering meaningless to look for more realistic deviations of the order of $10^{-3}$ and smaller. 3\) we propose to suppress the CP violating terms significantly using selection rule for the ratio $R/E$, where $R$ and $E$ are $K_S^o$ range and energy. In the experiment, $R/E$ ranges from 2.3 to 36.1 cm/GeV. The $R/E$ interval should be selected to make the contribution of the CP violating terms less than a desirable value, say $k\cdot 10^{-3}$. An effective ($R,E$) plot can then be calculated via Monte Carlo methods applied to the real decay volume.\ [r]{}[8.0cm]{} [ Fig. 1. Comparison of the various fitting functions (curves 1,2 and 3) applied to the simulated lifetime $\tau$(E) dependence in ref. 3b under the energy-range selection rule (see the text). ]{} The price we pay for more accurate data handling due to the range-energy selection rule will be [*lower statistics.*]{} In fact, under the above new assumptions, 60-70% events will be rejected, i.e., only 63K - 84K events of the total 220K will be useful. Apart from the loss of a major part of the data, 1/3 of the decay volume in the experiment turns out to be also useless. The large inefficiency of the experiment occurred because it had not been optimized for the problem. Basicly, the experimental design and data selection rules followed that of conventional $K_S,\, K_L$ studies. We illustrate the above arguments with two fits shown in Fig. 1. 220 000 $K_S$ decays at six energy values (from 125 to 375 GeV) were generated in the decay volume with the ranges from 9.3 m to 25.3 m. The energy dependence of the lifetime was assumed in the form $\tau(E)=\tau_S(1+\epsilon E)$ with $\tau_S$ = 0.8927, the world average of the mean lifetime, and $\epsilon$ = 4$\cdot 10^{-5}$. After applying the range-energy selection rule, a sample of 64K events was chosen for which the contribution of the CP violating terms was less then 0.008. Namely we deal with the following distribution for the proper lifetime: $$\begin{aligned} \frac{dN}{dx} = N\lbrace \exp{(-x)}~+~{\rm CPV} \rbrace,\end{aligned}$$ where $N$ is a normalization constant, $x=t/\tau(E)$ and CP violating terms are equal to $$\begin{aligned} \nonumber {\rm CPV}~=~\mid \eta_{+-} \mid ^2 {\rm exp}(-xy)~+~2D\mid \eta_{+-}\mid ~{\rm cos}(\Delta m~t - \phi_{+-}){\rm exp}(-x(1+y)/2)\end{aligned}$$ where $y$ stands for $\tau_S (E)/\tau_L$. The values of other parameters are taken as the world average values. These are $\mid \eta_{+-} \mid$ = 2.284$\cdot 10^{-3}$, the magnitude of the CP-nonconservation parameter in $ K_{L}^{o}\to \pi^+ \pi^-$ decay, $\phi_{+-}$=43.7$^o$, and $\Delta m=0.5333\cdot 10^{10}$ $\hbar sec^{-1}$ is the mass difference of $K_L^o - K_S^o$. The dilution factor $D$ is defined as the ratio $(N-\bar{N})/(N+\bar{N})$ where $N~(\bar{N})$ is the number of $K^o ~(\bar {K^o})$ produced by the proton beam on the target. We accepted the value $D$=0.75. In Fig. 1 the sequence of the mean proper lifetimes is plotted versus $E$, $K_S^o$ laboratory energies. The dependence was obtained by simulations of $K_S^o$ decays in the experimental volume under the conditions described above. The figure presents also results of three fitting procedures:\ a) one-parameter fit by a constant function $\tau (E) =c$ with $c$=0.90$\pm$0.01 and $\chi^2$/ndf = 0.7/5 (dashed line 1);\ b) one-parameter fit by the energy-dependent formula of the type $\tau (E)=0.8927 (1~+~p_1 E)$ with the obtained value of the parameter $p_1$ = (4 $\pm$ 5)$\cdot 10^{-5}$ and $\chi^2$/ndf = 0.38/5 (solid line 2);\ c) two-parameter fit to the formula of Ref. \[3b\], $\tau (E)~=~p_2 (1~+~p_1 E) $. In this case, the value of the crucial parameter $p_1$ is equal to (4 $\pm$ 23)$\cdot 10^{-5}$ with $\chi^2$/ndf = 0.38/4 (dotted line 3 which coinsides practically with solid line 2). There is a difference in interpretation of parameters in the two fitting formulae with the energy dependence. The parameter $p_2$ in the fit from the cited paper was interpreted as the zero-energy mean value of the proper lifetime. We think that it is difficult to extrapolate the fitting formulae from the energy interval 100-400 GeV to zero. Instead, we try to find the energy dependence in the limited energy interval by fit starting from a definite point. This difference in interpretation is important because, in general, various approaches in fitting procedures may lead to crucially different numerical results. Thus, in the selected amount of the events, both fits dig up well the mean value of the hidden parameter $\epsilon$ determining the energy dependence in the simulated $K_S^o$ decays, however the error bars differ strongly. Though both results for fitting values of $p_1$ are still insignificant statistically even in the selected sample of events, the 100% error bar in our fit being rather promising. It opens the door for new manipulations with the selection procedure aiming to improve the result. So we encourage the re-elaboration of the original data of ref. \[3b\] under the modified selection rules to obtain possible hopeful estimations of $\tau (E)$ instead of previous hopeless ones. We finally note that no firm spacetime anomalies can be established via the above re-elaboration for PC violating contributions smaller than $1.6\%$ because said anomalies are visually within the errorbars (Fig. 1) due to insufficient statistuics and other reasons. Corresponding deviations cannot be considered for PC violating contribution larger than $1.6\%$ because the latter are experimentally known to be excluded for the energy range of measures \[3b\]. Despite that, the analysis of this note establishes the insufficiencies of both measures \[3a,3b\] and the need for novel more accurate measures. [99]{} H. Minkowski, Nachr. Ges. Wiss. Goettingen [**[43]{}**]{} (1908) \[1a\]. H. A. Lorentz, [*Versuch einer Theorie der Elektrishen und Magnetishen Erscheinungen in bewenglen Korpern*]{}, Leyda \[1895 \[1b\]. W. Pauli, [*Theory of Relativity*]{}, Pergamon Press, New York (1958) \[1c\]. A. Enders and G. Nimtz, J. Phys. France [**[2]{}**]{}, 1693 (1992) \[1d\]. G. Nimtz and W. Heitmann, Progr. Quantum Electr. [**[21]{}**]{}, 81 (1997) \[1e\]. F.Mirabel and F. Rodriguez, Nature [**[371]{}**]{}, 464 (1994) \[1f\]. J. Tingay et al., Nature [**[374]{}**]{}, 141 (1995) \[1g\]. D. Baylin et al., IAU Comm. 6173 (1995) \[1h\]. P. Saari and K. Reivelt, Phys. Rev. Letters [**79**]{} (1997), in press \[1i\]. D.I. Blochintsev, Phys Rev Lett [**12**]{}, 272 (1964) \[2a\]. L.B. Redei, Phys.Rev. [**145**]{}, 999 (1966) \[2b\]. D.Y. Kim, Hadronic J.,[**1**]{}, 1343 (1978) \[2c\]. R. M. Santilli, Lett. Nuovo Cimento [**33**]{}, 145 (1982) \[2d\]. V. de Sabbata and M. Gasperini, Lett. Nuovo Cimento [**34**]{}, 337 (1982) \[2e\]. H.B. Nielsen and I. Picek, Nucl.Phys. [**B211**]{}, 269 (1983) \[2f\]. S.H. Aronson et al., Phys.Rev. [**D28**]{}, 495 (1983) \[3a\]. N. Grossman et al., Phys.Rev.Lett. [**59**]{}, 18 (1987) \[3b\]. G. Alexander et al., Phys. Lett. B [**368**]{}, 244 (1996) \[3c\]. R.M. Santilli, Lett. Nuovo Cimento,[**37**]{}, 545 (1983) \[4a\]. R. M. Santilli, Intern. J. Modern Phys., in press \[4b\]. R. M. Santilli, J. Moscow Phys. Soc. [**[3]{}**]{}, 255 (1993) \[4c\]. J. V. Kadeisvili, Math. Methods in Applied Sciences, [****]{}, 1349 (1996) \[4d\]. A. K. Aringazin, Hadronic J. [****]{}, 71 (1989); A. K. Aringazin et al., in [*Frontiers of Fundamental Physics*]{}, M. Barone and F. Selleri, Editors, Plenum, New York (1995), p. 153 \[4e\]. R. M. Santilli, in [*Proceedings of the International Conference on Modern Modified Theories of Gravitation and Cosmology*]{}, E. I. Guendelman, Ed., Hadronic J. [**21**]{} (1998), in press \[4f\]. R. M. Santilli, Geometries [**[10]{}**]{}, 273 (1993) \[4g\]. J. V. Kadeisvili, Algebras, Groups and Geometries [****]{}, 283 and 319 (1992) \[4h\]. Gr. T. Tsagas and D. S. Sourlas, Algebras, Groups and Geometries [****]{}, 1 and 67 (1995) \[4i\]. P. Vacaru, Algebras, Groups and Geometries [****]{}, 211 (1997) \[4j\]. R. M. Santill, Rendiconti Circolo Matematico Palermo, Suppl. [**[42]{}**]{}, 7 (1996) \[4k\]. R. M. Santilli, Found. Phys. [**27**]{}, 625 (1997) \[4l\]. R. Mignani, Physics Essays [**5**]{}, 531 (1992) \[5a\]. F. Cardone, R. Mignani and R. M. Santilli, J. Phys G. [****]{}, L61 and L141 (1992) \[5b\]. D. Schuch, Phys. Rev A [**55**]{}, 955 (1997) \[5c\]. C. A. C. Dreismann [*et al.*]{}, Phys. Review Letters [**79**]{}, 2390 (1997) \[5d\].

--- author: - Sven Maerivoet - Steven Logghe - Ben Immers - Bart De Moor bibliography: - 'paper.bib' date: 'Received: June 28, 2005 / Revised version: ' title: ' Relating the dynamics of road traffic in a stochastic cellular automaton to a macroscopic first-order model ' --- Introduction ============ Considering the existing relations between the stochastic TCA model of Nagel and Schreckenberg [@NAGEL:92; @NAGEL:95c] and the macroscopic first-order model of Lighthill, Whitham, and Richards (LWR) [@LIGHTHILL:55; @RICHARDS:56], from traffic flow theory, there are already numerous links between both modeling approaches. An example is the so-called totally asymmetric simple exclusion process (TASEP) [@DERRIDA:92], which corresponds to the LWR model with a noisy and diffusive conservation law if a random sequential update is assumed [@NAGEL:95c; @NAGEL:96]. Another example is the STCA which can be approximated by a so-called mean field theory (MFT), and its successive refinements, such as the car-oriented mean-field theory (COMF), and the recently developed site-oriented cluster-theoretic approach [@SCHRECKENBERG:95; @SCHADSCHNEIDER:97; @SCHADSCHNEIDER:98; @SANTEN:99; @SCHADSCHNEIDER:99b; @CHOWDHURY:00; @SCHADSCHNEIDER:02]. A more elaborate discussion on these analytical treatments can be found in [@MAERIVOET:05]. In summary, we can say that there already exist several methods for bridging both the microscopic STCA and the macroscopic LWR model (note that we do not consider the class of hybrid models, as we are only interested in direct analogies between both microscopic and macroscopic models, and not in pure combinations of these model classes). In this paper, we reconsider the STCA and LWR models, but we take a different approach at studying their relation: we consider a TCA model as a particle-based discretization scheme for macroscopic traffic flow models. It is from this latter point of view that our work addresses the common structure between both models. Our main goal is therefore to provide a means for implicitly incorporating the STCA’s stochasticity into the LWR model, which is in fact deterministic in nature [@MAERIVOET:03g]. Note that we use the term *implicit* to denote the fact that the STCA’s stochasticity is not introduced in the equations by means of *explicit* noise terms. Rather, our methodology implies that the stochasticity is introduced through the shape of the LWR’s fundamental diagram.\ This paper is organized as follows: in section \[sec:Recapitulation\], we briefly recapitulate both the approaches taken by traffic cellular automata models (with the STCA in particular), and the macroscopic first-order hydrodynamic LWR model. In section \[sec:InitialMethodology\], we then consider a methodology for implicitly incorporating the STCA’s stochasticity into the LWR’s triangular fundamental diagram. Continuing, we apply this technique to a small case study in section \[sec:ApplicationToACaseStudy\], which points us to some discrepancies between both modeling approaches. Highlighting some of the resulting artifacts, and investigating the main reason for the difference in behavior, we move on to section \[sec:AlternateDerivation\] where we present an alternate derivation of the fundamental diagrams. Finally, the paper concludes with section \[sec:Conclusions\], stating a summary of our findings. Recapitulating the STCA and LWR models {#sec:Recapitulation} ====================================== With respect to the modeling of traffic flows, there are largely two model classes possible, i.e., the microscopic and macroscopic approach, respectively. In the former class, interactions between vehicles in a traffic stream are explicitly modeled, giving rise to car-following and lane-changing submodels. With respect to the latter class, traffic streams are mostly treated as inviscid but compressible fluids. In this section, we briefly recapitulate a special class of microscopic models, i.e., traffic cellular automata (TCA) models. We then describe a stochastic TCA model, called the stochastic traffic cellular automaton (STCA) of Nagel and Schreckenberg, after which we conclude with an overview of the most prominent features of the macroscopic first-order hydrodynamic model of Lighthill, Whitham, and Richards (LWR). Traffic cellular automata (TCA) models {#sec:TCAModels} -------------------------------------- In the field of traffic flow modeling, microscopic traffic simulation has always been regarded as a time consuming, complex process involving detailed models that describe the behavior of individual vehicles. Approximately a decade ago, however, new microscopic models were being developed, based on the cellular automata programming paradigm from statistical physics. Let us first describe the operation of a single-lane traffic cellular automaton as depicted in [Fig. \[fig:TCABasics\]]{}. We assume $N$ vehicles are driving on a circular lattice containing $K$ cells, i.e., periodic boundary conditions (each cell can be occupied by at most one vehicle at a time). Time and space are discretized, with $\Delta T = 1$ s and $\Delta X = 7.5$ m, leading to a velocity discretization of $\Delta V = 27$ km/h. Furthermore, the velocity $v_{i}$ of a vehicle $i$ is constrained to an integer in the range $\lbrace 0, \ldots, v_{\mbox{max}} \rbrace$, with $v_{\mbox{max}}$ typically 5 cells/s (corresponding to 135 km/h). \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] ![ Schematic diagram of the operation of a single-lane traffic cellular automaton (TCA); here, the time axis is oriented downwards, the space axis extends to the right. The TCA’s configuration is shown for two consecutive time steps $t$ and $t + 1$, during which two vehicles $i$ and $j$ propagate through the lattice. Without loss of generality, we denote the number of empty cells in front of vehicle $i$ as its space gap $g_{s_{i}}$. []{data-label="fig:TCABasics"}](figures/tca-classic-lattice "fig:"){width="\figurewidth"} Each vehicle $i$ has a space headway $h_{s_{i}}$ and a time headway $h_{t_{i}}$, defined as follows: $$\begin{aligned} h_{s_{i}} & = & g_{s_{i}} + l_{i}, \label{eq:SpaceHeadway}\\ h_{t_{i}} & = & g_{t_{i}} + \rho_{i} \label{eq:TimeHeadway}.\end{aligned}$$ In these definitions, $g_{s_{i}}$ and $g_{t_{i}}$ denote the space and time gaps respectively; $l_{i}$ is the length of a vehicle and $\rho_{i}$ is the occupancy time of the vehicle (i.e., the time it ‘spends’ in one cell). Note that in a traffic cellular automaton the space headway of a vehicle is always an integer number, representing a multiple of the spatial discretization $\Delta X$ in real world measurement units. So in a jam, it is taken to be equal to the space the vehicle occupies, i.e., $h_{s_{i}} = {\ensuremath{\mbox{1}}}$ cell. Local interactions between individual vehicles in a traffic stream are modeled by means of a rule set. In this paper, we assume that all vehicles have the same physical characteristics. The system’s state is changed through synchronous position updates of all the vehicles, based on a rule set that reflects the car-following behavior. Note that most rule sets of TCA models do not use the space headway $h_{s_{i}}$ or the space gap $g_{s_{i}}$, but are instead based on the number of empty cells $d_{i}$ in front of a vehicle $i$. Keeping equation [(\[eq:SpaceHeadway\])]{} in mind, we therefore adopt the convention that, for a vehicle $i$ its length $l_{i} = {\ensuremath{\mbox{1}}}$ cell; the resulting TCA models are called *single-cell models*. This means that when the vehicle is residing in a compact jam, its space headway $h_{s_{i}} = {\ensuremath{\mbox{1}}}$ cell and its space gap is consequently $g_{s_{i}} = {\ensuremath{\mbox{0}}}$ cells. This abstraction gives us a rigorous justification to formulate the TCA’s update rules more intuitively using space gaps. The stochastic traffic cellular automaton (STCA) model of Nagel and Schreckenberg --------------------------------------------------------------------------------- In 1992, Nagel and Schreckenberg proposed a TCA model that was able to reproduce several characteristics of real-life traffic flows, e.g., the spontaneous emergence of traffic jams [@NAGEL:92; @NAGEL:95c]. Their model is called the *NaSch TCA*, but is more commonly known as the *stochastic traffic cellular automaton* (STCA). It explicitly includes a stochastic noise term in one of its rules. The STCA then comprises the following three rules (note that in Nagel and Schreckenberg’s original formulation, they decoupled acceleration and braking, resulting in four rules): > **R1**: *acceleration and braking*\ > $$\label{eq:STCAR1} > v_{i}(t) \leftarrow \mbox{min} \lbrace v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}},g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),v_{\mbox{max}} \rbrace,$$ > > **R2**: *randomization*\ > $$\label{eq:STCAR2} > \xi(t) < p \Longrightarrow v_{i}(t) \leftarrow \max \lbrace {\ensuremath{\mbox{0}}},v_{i}(t) - {\ensuremath{\mbox{1}}}\rbrace,$$ > > **R3**: *vehicle movement*\ > $$\label{eq:STCAR3} > x_{i}(t) \leftarrow x_{i}(t - {\ensuremath{\mbox{1}}}) + v_{i}(t).$$ The STCA contains a rule for increasing the speed of a vehicle and braking to avoid collisions, i.e., rule R1, equation [(\[eq:STCAR1\])]{}. It furthermore also contains rule R2, equation [(\[eq:STCAR2\])]{}, which introduces stochasticity in the system. At each discrete time step $t$, a random number $\xi(t) \in [{\ensuremath{\mbox{0}}},{\ensuremath{\mbox{1}}}[$ is drawn from a uniform distribution. This number is then compared with a stochastic noise parameter $p \in [{\ensuremath{\mbox{0}}},{\ensuremath{\mbox{1}}}]$ (called the *slowdown probability*); as a result, there is a probability of $p$ that a vehicle will slow down to $v_{i}(t) - {\ensuremath{\mbox{1}}}$ cells/time step. According to Nagel and Schreckenberg, the randomization of rule R2 captures natural speed fluctuations due to human behavior or varying external conditions. The rule introduces overreactions of drivers when braking, providing the key to the formation of spontaneously emerging jams. Finally, rule R3, equation [(\[eq:STCAR3\])]{}, allows for the actual movement of vehicles in the system. The STCA model is called a *minimal model*, in the sense that all these rules are a necessity for mimicking the basic features of real-life traffic flows. To get an intuitive feeling for the STCA’s system dynamics, we have provided two time-space diagrams in [Fig. \[fig:STCATimeSpaceDiagrams\]]{}. Both diagrams show the evolution for a global density of $k =$ 0.2 vehicles/cell, but with $p$ set to 0.1 for the left diagram, and $p = $ 0.5 for the right diagram. As can be seen in both diagrams, the randomization in the model gives rise to many unstable artificial phantom mini-jams. The downstream fronts of these jams smear out, forming *unstable interfaces* [@NAGEL:03]. This is a direct result of the fact that the intrinsic noise (as embodied by $p$) in the STCA model is too strong: a jam can always form at *any* density, meaning that breakdown can (and will) occur, even in the free-flow traffic regime. For low enough densities however, these jams can vanish as they are absorbed by vehicles with sufficient space headways, or by new jams in the system [@KRAUSS:99]. It has been experimentally shown that below the critical density, these jams have finite life times with a cut-off that is about ${\ensuremath{\mbox{5}}}\times {\ensuremath{\mbox{10}}}^{{\ensuremath{\mbox{5}}}}$ time steps and independent of the lattice size. When the critical density is crossed, these long-lived jams evolve into jams with an infinite life time, i.e., they will survive for an infinitely long time [@NAGEL:94b; @NAGEL:95c; @SCHADSCHNEIDER:99b]. ![ Typical time-space diagrams of the STCA model (the time and space axes are oriented left to right, and bottom to top, respectively). The two shown closed-loop lattices each contain 300 cells, with a visible period of 580 time steps (each vehicle is represented as a single colored dot). Both diagrams have a global density of $k =$ 0.2 vehicles/cell. *Left:* the evolution of the system for $p =$ 0.1. *Right:* the evolution of the system, but now for $p = $ 0.5. The effects of the randomization rule R2 are clearly visible in both diagrams, as there occur many unstable artificial phantom mini-jams. Furthermore, the speed $w$ of the backward propagating kinematic waves decreases with an increasing $p$. []{data-label="fig:STCATimeSpaceDiagrams"}](figures/stca-tx-diagram-k20-p010 "fig:"){width="\figurehalfwidth"} ![ Typical time-space diagrams of the STCA model (the time and space axes are oriented left to right, and bottom to top, respectively). The two shown closed-loop lattices each contain 300 cells, with a visible period of 580 time steps (each vehicle is represented as a single colored dot). Both diagrams have a global density of $k =$ 0.2 vehicles/cell. *Left:* the evolution of the system for $p =$ 0.1. *Right:* the evolution of the system, but now for $p = $ 0.5. The effects of the randomization rule R2 are clearly visible in both diagrams, as there occur many unstable artificial phantom mini-jams. Furthermore, the speed $w$ of the backward propagating kinematic waves decreases with an increasing $p$. []{data-label="fig:STCATimeSpaceDiagrams"}](figures/stca-tx-diagram-k20-p050 "fig:"){width="\figurehalfwidth"} The macroscopic first-order traffic flow model of Lighthill, Whitham, and Richards (LWR) {#sec:LWRModelDescription} ---------------------------------------------------------------------------------------- Considering traffic as an inviscid but compressible fluid, it follows from this assumption, that densities $k$, space-mean speeds $\overline v_{s}$, and flows $q$ are defined as continuous variables, in each point in time and space, hence leading to the names of *continuum models*, *fluid-dynamic models*, or *macroscopic models*. The first aspect of such a fluid-dynamic description of traffic flow, consists of a *scalar conservation law*; ‘scalar’ because it is a first-order partial differential equation (PDE). A typical derivation can be found in [@GARTNER:97] and [@JUNGEL:02]: the derivation is based on considering a road segment with a finite length on which no vehicles appear or disappear other than the ones that enter and exit it. After taking the infinitesimal limit, this will result in a PDE that expresses the interplay between continuous densities and flows on a local scale: $$\label{eq:FirstOrderConservationLaw} \frac{\partial k(t,x)}{\partial t} + \frac{\partial q(t,x)}{\partial x} = {\ensuremath{\mbox{0}}},$$ with the density $k$ and flow $q$ dynamically (i.e., time varying) defined over a single spatial dimension. Lighthill and Whitham were among the first to develop such a traffic flow model in 1955 [@LIGHTHILL:55]. One year later, Richards independently derived the same fluid-dynamic model [@RICHARDS:56], albeit in a slightly different form. Because of the nearly simultaneous and independent development of the theory, the model has become known as the *LWR model*, after the initials of its inventors who receive the credit. In some texts, the model is also referred to as the *hydrodynamic model*, or the *kinematic wave model* (KWM), attributed to the fact that the model’s solution is based on characteristics, which are called kinematic waves (e.g., shock waves). Crucial to their approach, was a fundamental hypothesis, essentially stating that flow is a function of density, i.e., there exists a $q_{e}(k(t,x))$ equilibrium relationship, more commonly known as the *fundamental diagram* [@HAIGHT:63]. Central to their theory, Lighthill and Whitham assumed that the fundamental hypothesis holds at all traffic densities, not just for light-density traffic but also for congested traffic conditions. In order to solve equation [(\[eq:FirstOrderConservationLaw\])]{}, we also need the *fundamental relation of traffic flow theory*, which relates the macroscopic traffic flow characteristics density $k$ (vehicles/kilometre), space-mean speed $\overline v_{s}$ (kilometres/hour), and flow $q$ (vehicles/hour) to each other as follows: $$\label{eq:FundamentalRelation} q = k~\overline v_{s}.$$ In general however, there are two restrictions, i.e., the relation is only valid for (1) continuous variables, or smooth approximations of them, and (2) traffic composed of substreams (e.g., slow and fast vehicles) which comply to the following two assumptions: > **Homogeneous traffic** > > :   > > There is a homogeneous composition of the traffic substream (i.e., the same type of vehicles).\ > > **Stationary traffic** > > :   > > When observing the traffic substream at different times and locations, it ‘looks the same’. Putting it a bit more quantitatively, all the vehicles’ trajectories in a time-space diagram should be parallel and equidistant [@DAGANZO:97]. > The latter of the above two conditions, is also referred to as traffic operating in a *steady state* or at *equilibrium*. Employing the fundamental diagram $q_{e}(k)$, relates the two dependent variables in equation [(\[eq:FirstOrderConservationLaw\])]{} to each other, thereby making it possible to solve the partial differential equation. Thus, reconsidering equation [(\[eq:FirstOrderConservationLaw\])]{}, taking into account the fundamental diagram, the conservation law is now expressed as: $$\label{eq:FirstOrderPDEConservationLaw} k_{t} + q_{e}(k)_{x} = {\ensuremath{\mbox{0}}},$$ in which we introduced the standard differential calculus notation for PDEs. Recognizing the fundamental relation of traffic flow theory [(\[eq:FundamentalRelation\])]{}, the conservation law [(\[eq:FirstOrderPDEConservationLaw\])]{} can also be cast in a non-linear wave equation, using the chain rule for differentiation [@GARTNER:97]: $$\label{eq:FirstOrderPDEIVP} k_{t} + \frac{dq_{e}(k)}{dk} k_{x} = {\ensuremath{\mbox{0}}}.$$ Analytically solving the previous equation using the method of characteristics, results in shock waves that travel with speeds equal to: $$w = \frac{dq_{e}(k)}{dk},$$ i.e., the tangent to the $q_{e}(k)$ fundamental diagram. As a consequence, solutions, being the characteristics, of equation [(\[eq:FirstOrderPDEIVP\])]{} have the following form: $$k(t,x) = k(x - wt),$$ with the observation that the density is constant along such a characteristic. Whenever in the solution of the conservation equation, two of its characteristics intersect, the density takes on two different values (each one belonging to a single characteristic). As this mathematical quirk is physically impossible, the so-called entropy solution states that both characteristics terminate and breed a *shock wave*; as such, these shock waves form boundaries that discontinuously separate densities, flows, and space-mean speeds [@GARTNER:97]. The speed of such a shock wave is related to the following ratio [@PIPES:64]: $$\label{eq:FirstOrderShockWaveSpeed} w_{\mbox{shock}} = \frac{\Delta q}{\Delta k},$$ with $\Delta q = q_{u} - q_{d}$ and $\Delta k = k_{u} - k_{d}$ the relative difference in flows, respectively densities, up- and downstream of the shock wave. Note that going from a low to a high density regime typically results in a shock wave, whereas the reverse transition is accompanied by an emanation of a *fan of characteristics* (also called *expansion*, *acceleration*, or *rarefaction waves*). In shock wave theory, the densities on either side of a shock are well defined (i.e., unique solutions exist); along the shock wave however, the density jumps discontinuously from one value to another. Because of the well-defined properties of the LWR model, it is possible to derive analytical solutions to certain types of problems. These solutions can even be drawn graphically in a time-space diagram, thereby clearly showing the evolution of first-order macroscopic traffic flow characteristics (e.g., the speed of backward-travelling jams, …). Besides the previous analytic derivation of a solution to the conservation law expressed as a PDE, it is also possible to treat the problem numerically. Converting the PDEs into finite difference equations (FDEs), and solving them numerically stable, can be done by casting the LWR model in the context of *Godunov FDE methods*, allowing for arbitrary $q_{e}(k)$ fundamental diagrams [@DAGANZO:95b; @LEBACQUE:96]. As such, the LWR model is very appealing because it provides an attractive means for gaining insight into the primary characteristics of the evolution of traffic flows. As a surplus, the model has the benefits of the ability to efficiently construct analytical solutions, as well as the existence of thoroughly-developed and understood numerical schemes. Implicitly incorporating the STCA’s stochasticity {#sec:InitialMethodology} ================================================= As mentioned in the introduction, we reconsider the STCA and LWR models, taking a different approach at studying their relation. Our main goal is to provide a means for implicitly incorporating the STCA’s stochasticity into the LWR model. To this end, we provide a practical methodology for specifying the fundamental diagram to the LWR model. Assuming that a *stationarity condition* holds on the STCA’s rules, we incorporate the STCA’s stochasticity directly into the LWR’s fundamental diagram. Relating both the STCA and the LWR models is now done using a simple two-step approach, in which we first rewrite the STCA’s rules into a single rule, leading to a set of *linear inequalities*. These constraints can be considered as a $\overline v_{s_{e}}(\overline h_{s})$ fundamental diagram (see e.g., the left part in [Fig. \[fig:DerivingStationaryFundamentalDiagrams\]]{}). This latter diagram can then be converted into an equivalent triangular flow versus density $q_{e}(k)$ fundamental diagram. Rewriting the STCA’s rule set ----------------------------- Considering a vehicle’s average speed, the STCA’s rules R1 and R2, equations [(\[eq:STCAR1\])]{} and [(\[eq:STCAR2\])]{} respectively, state that a vehicle slows down with probability $p$, and that it does not slow down with probability ${\ensuremath{\mbox{1}}}- p$. As such, they can be rewritten into the following single rule that is expressed in *continuous* speeds and space gaps: $$\begin{aligned} v_{i}(t) & \leftarrow & p \cdot \mbox{min} \lbrace v_{i}(t - {\ensuremath{\mbox{1}}})~\cancel{+ {\ensuremath{\mbox{1}}}}~~\cancel{- {\ensuremath{\mbox{1}}}}, g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}},\nonumber\\ & & v_{\mbox{max}} - {\ensuremath{\mbox{1}}}\rbrace + ({\ensuremath{\mbox{1}}}- p) \cdot \mbox{min} \lbrace v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}},\nonumber\\ & & g_{s_{i}}(t - {\ensuremath{\mbox{1}}}), v_{\mbox{max}} \rbrace,\label{eq:STCASingleRule}\end{aligned}$$ with $v_{i}(t) \leftarrow \mbox{max} \lbrace 0, v_{i}(t) \rbrace$. Furthermore, the following two algebraic relations always hold: $$\begin{aligned} a \cdot \mbox{min} \lbrace b, c \rbrace & = & \mbox{min} \lbrace ab, ac \rbrace,\label{eq:MinAlgebraRule1}\\ \mbox{min} \lbrace a, b \rbrace + \mbox{min} \lbrace c, d \rbrace & = & \mbox{min} \lbrace a + c, a + d,\nonumber\\ & & \qquad b + c, b + d \rbrace.\label{eq:MinAlgebraRule2}\end{aligned}$$ Applying relation [(\[eq:MinAlgebraRule1\])]{} to our rule [(\[eq:STCASingleRule\])]{}, yields the following result: $$\begin{aligned} v_{i}(t) & \leftarrow & \mbox{min} \lbrace p v_{i}(t - {\ensuremath{\mbox{1}}}), p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}}), p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) \rbrace \nonumber\\ & & + \mbox{min} \lbrace ({\ensuremath{\mbox{1}}}- p) (v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}), ({\ensuremath{\mbox{1}}}- p) g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & \qquad~~ ({\ensuremath{\mbox{1}}}- p) v_{\mbox{max}} \rbrace.\end{aligned}$$ Using relation [(\[eq:MinAlgebraRule2\])]{} to this result, allows us to obtain a formulation with a single minimum-operator: $$\begin{aligned} v_{i}(t) \leftarrow \mbox{min} \lbrace & & p v_{i}(t - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) (v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}),\nonumber\\ & & p v_{i}(t - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & p v_{i}(t - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) v_{\mbox{max}},\nonumber\\ & & p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) (v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}),\nonumber\\ & & p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) v_{\mbox{max}},\nonumber\\ & & p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) (v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}),\nonumber\\ & & p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + ({\ensuremath{\mbox{1}}}- p) v_{\mbox{max}} \rbrace.\end{aligned}$$ Expanding all the terms between parentheses gives the following result: $$\begin{aligned} v_{i}(t) \leftarrow \mbox{min} \lbrace & & \cancel{p v_{i}(t - {\ensuremath{\mbox{1}}})} + v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}- \cancel{p v_{i}(t - {\ensuremath{\mbox{1}}})} - p,\nonumber\\ & & p v_{i}(t - {\ensuremath{\mbox{1}}}) + g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - p g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & p v_{i}(t - {\ensuremath{\mbox{1}}}) + v_{\mbox{max}} - p v_{\mbox{max}},\nonumber\\ & & p g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - p + v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}- p v_{i}(t - {\ensuremath{\mbox{1}}}) - p,\nonumber\\ & & \cancel{p g_{s_{i}}(t - {\ensuremath{\mbox{1}}})} - p + g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - \cancel{p g_{s_{i}}(t - {\ensuremath{\mbox{1}}})},\nonumber\\ & & p g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - p + v_{\mbox{max}} - p v_{\mbox{max}},\nonumber\\ & & p v_{\mbox{max}} - p + v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}- p v_{i}(t - {\ensuremath{\mbox{1}}}) - p,\nonumber\\ & & p v_{\mbox{max}} - p + g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - p g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & \cancel{p v_{\mbox{max}}} - p + v_{\mbox{max}} - \cancel{p v_{\mbox{max}}} \rbrace.\end{aligned}$$ And finally, regrouping for $p$ yields: $$\begin{aligned} v_{i}(t) \leftarrow \mbox{min} \lbrace & & v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}}- p,\nonumber\\ & & p (v_{i}(t - {\ensuremath{\mbox{1}}}) - g_{s_{i}}(t - {\ensuremath{\mbox{1}}})) + g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & p (v_{i}(t - {\ensuremath{\mbox{1}}}) - v_{\mbox{max}}) + v_{\mbox{max}},\nonumber\\ & & p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - v_{i}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{2}}}) + v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}},\nonumber\\ & & g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - p,\nonumber\\ & & p (g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + v_{\mbox{max}},\nonumber\\ & & p (v_{\mbox{max}} - v_{i}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{2}}}) + v_{i}(t - {\ensuremath{\mbox{1}}}) + {\ensuremath{\mbox{1}}},\nonumber\\ & & p (v_{\mbox{max}} - g_{s_{i}}(t - {\ensuremath{\mbox{1}}}) - {\ensuremath{\mbox{1}}}) + g_{s_{i}}(t - {\ensuremath{\mbox{1}}}),\nonumber\\ & & v_{\mbox{max}} - p \rbrace.\label{eq:STCASingleRuleRegrouped}\end{aligned}$$ If we now assume traffic is *stationary* (see e.g., Daganzo’s description of stationary traffic in section \[sec:LWRModelDescription\]), then we can assert that the state of a vehicle at time $t$ is the same as its state at time $t - {\ensuremath{\mbox{1}}}$, i.e., $v_{i}(t) = v_{i}(t - {\ensuremath{\mbox{1}}})$ and $g_{s_{i}}(t) = g_{s_{i}}(t - {\ensuremath{\mbox{1}}})$. As a result, equation [(\[eq:STCASingleRuleRegrouped\])]{} gets transformed into the following set of *linear inequalities* that express constraints on the relations between $v_{i}(t)$, $g_{s_{i}}(t)$, $p$, and $v_{\mbox{max}}$: $$\begin{aligned} \cancel{v_{i}(t)} + {\ensuremath{\mbox{1}}}- p \geq \cancel{v_{i}(t)} & & \quad \mbox{(C1)},\nonumber\\ p (v_{i}(t) - g_{s_{i}}(t)) + g_{s_{i}}(t) \geq v_{i}(t) & & \quad \mbox{(C2)},\nonumber\\ p (v_{i}(t) - v_{\mbox{max}}) + v_{\mbox{max}} \geq v_{i}(t) & & \quad \mbox{(C3)},\nonumber\\ p (g_{s_{i}}(t) - v_{i}(t) - {\ensuremath{\mbox{2}}}) + \cancel{v_{i}(t)} + {\ensuremath{\mbox{1}}}\geq \cancel{v_{i}(t)} & & \quad \mbox{(C4)},\nonumber\\ g_{s_{i}}(t) - p \geq v_{i}(t) & & \quad \mbox{(C5)},\nonumber\\ p (g_{s_{i}}(t) - v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + v_{\mbox{max}} \geq v_{i}(t) & & \quad \mbox{(C6)},\nonumber\\ p (v_{\mbox{max}} - v_{i}(t) - {\ensuremath{\mbox{2}}}) + \cancel{v_{i}(t)} + {\ensuremath{\mbox{1}}}\geq \cancel{v_{i}(t)} & & \quad \mbox{(C7)},\nonumber\\ p (v_{\mbox{max}} - g_{s_{i}}(t) - {\ensuremath{\mbox{1}}}) + g_{s_{i}}(t) \geq v_{i}(t) & & \quad \mbox{(C8)},\nonumber\\ v_{\mbox{max}} - p \geq v_{i}(t) & & \quad \mbox{(C9)}.\nonumber\end{aligned}$$ Let us now examine each of these nine constraints C1 through C9. - Constraint C1 states that ${\ensuremath{\mbox{1}}}- p \geq {\ensuremath{\mbox{0}}}$, i.e., $p \leq {\ensuremath{\mbox{1}}}$. This logically follows from the STCA’s condition that $p \in [{\ensuremath{\mbox{0}}},{\ensuremath{\mbox{1}}}]$. - Constraint C2 states that $p (v_{i}(t) - g_{s_{i}}(t)) + g_{s_{i}}(t) \geq v_{i}(t)$, i.e., $g_{s_{i}}(t) \cancel{({\ensuremath{\mbox{1}}}- p)} \geq v_{i}(t) \cancel{({\ensuremath{\mbox{1}}}- p)}$. This corresponds to $v_{i}(t) \leq g_{s_{i}}(t)$, which states that vehicles strive for collision-free driving. - Constraint C3 states that $p (v_{i}(t) - v_{\mbox{max}}) + v_{\mbox{max}} \geq v_{i}(t)$, i.e., $v_{\mbox{max}} \cancel{({\ensuremath{\mbox{1}}}- p)} \geq v_{i}(t) \cancel{({\ensuremath{\mbox{1}}}- p)}$. This corresponds to $v_{i}(t) \leq v_{\mbox{max}}$, which logically follows from the STCA’s condition that $v_{i}(t) \in \lbrace {\ensuremath{\mbox{0}}}, \ldots, v_{\mbox{max}} \rbrace$. - Constraint C4 states that $p (g_{s_{i}}(t) - v_{i}(t) - {\ensuremath{\mbox{2}}}) + {\ensuremath{\mbox{1}}}\geq {\ensuremath{\mbox{0}}}$, i.e., $v_{i}(t) \leq g_{s_{i}}(t) - {\ensuremath{\mbox{2}}}+ \frac{{\ensuremath{\mbox{1}}}}{p}$ (for $p \neq {\ensuremath{\mbox{0}}}$). - Constraint C5 states that $g_{s_{i}}(t) - p \geq v_{i}(t)$, i.e., $v_{i}(t) \leq g_{s_{i}}(t) - p$, which is a more stringent constraint than C2 and C4. - Constraint C6 states that $p (g_{s_{i}}(t) - v_{\mbox{max}} - {\ensuremath{\mbox{1}}}) + v_{\mbox{max}} \geq v_{i}(t)$, i.e., $v_{i}(t) \leq v_{\mbox{max}} ({\ensuremath{\mbox{1}}}- p) + p (g_{s_{i}}(t) - {\ensuremath{\mbox{1}}})$. - Constraint C7 states that $p (v_{\mbox{max}} - v_{i}(t) - {\ensuremath{\mbox{2}}}) + {\ensuremath{\mbox{1}}}\geq {\ensuremath{\mbox{0}}}$, i.e., $v_{i}(t) \leq v_{\mbox{max}} - {\ensuremath{\mbox{2}}}+ \frac{{\ensuremath{\mbox{1}}}}{p}$ (for $p \neq {\ensuremath{\mbox{0}}}$). - Constraint C8 states that $p (v_{\mbox{max}} - g_{s_{i}}(t) - {\ensuremath{\mbox{1}}}) + g_{s_{i}}(t) \geq v_{i}(t)$, i.e., $v_{i}(t) \leq g_{s_{i}}(t) ({\ensuremath{\mbox{1}}}- p) + p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}})$. - Constraint C9 states that $v_{\mbox{max}} - p \geq v_{i}(t)$, i.e., $v_{i}(t) \leq v_{\mbox{max}} - p$, which is a more stringent constraint than C3 and C7. Taking the previous considerations into account, we can see that constraints C1, C2, and C3 are always satisfied. The remaining three pairs of similar constraints on the relations between $v_{i}(t)$, $g_{s_{i}}(t)$, $p$, and $v_{\mbox{max}}$, are the following: constraints C5 and C9, C4 and C7, and C6 and C8. In order to gain insight into the more difficult constraints C6 and C8, we first rewrite them as follows: $$\begin{aligned} \mbox{(C6)} & & v_{i}(t) \leq \underbrace{p}_{\mbox{slope}} g_{s_{i}}(t) + \underbrace{({\ensuremath{\mbox{1}}}- p) v_{\mbox{max}} - p}_{\mbox{intercept}},\nonumber\\ \mbox{(C8)} & & v_{i}(t) \leq \underbrace{({\ensuremath{\mbox{1}}}- p)}_{\mbox{slope}} g_{s_{i}}(t) + \underbrace{p (v_{\mbox{max}} - {\ensuremath{\mbox{1}}})}_{\mbox{intercept}},\nonumber\end{aligned}$$ where we have separated the terms containing $g_{s_{i}}(t)$. Plotting the speed $v_{i}(t)$ versus the space gap $g_{s_{i}}(t)$ in [Fig. \[fig:Constraints6And8\]]{}, allows us to more easily interpret the combined effects of these two constraints. On the one hand, if we continuously change $p = {\ensuremath{\mbox{0}}}\rightarrow {\ensuremath{\mbox{1}}}$, then constraint C6 goes from a horizontal line at $v_{i}(t) = v_{\mbox{max}}$, to a slanted line with a slope of $+{\ensuremath{\mbox{1}}}$, intercepting the horizontal and vertical axes at $+{\ensuremath{\mbox{1}}}$ and $-{\ensuremath{\mbox{1}}}$, respectively. In all cases, the point at ($v_{\mbox{max}} + {\ensuremath{\mbox{1}}}$,$v_{\mbox{max}}$) remains invariant. On the other hand, changing $p = {\ensuremath{\mbox{0}}}\rightarrow {\ensuremath{\mbox{1}}}$ turns constraint C8 from a slanted line with a slope of $+{\ensuremath{\mbox{1}}}$, passing through the origin, into a horizontal line at $v_{i}(t) = v_{\mbox{max}} - {\ensuremath{\mbox{1}}}$. In all cases, the point at ($v_{\mbox{max}} - {\ensuremath{\mbox{1}}}$,$v_{\mbox{max}} - {\ensuremath{\mbox{1}}}$) remains invariant. \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[l\] \[\]\[l\] \[\]\[\] \[\]\[\] ![ A visual representation of the constraints C6 and C8. *Left:* as $p = {\ensuremath{\mbox{0}}}\rightarrow {\ensuremath{\mbox{1}}}$, C6 changes from a horizontal line at $v_{i}(t) = v_{\mbox{max}}$, to a slanted line with a slope of $+{\ensuremath{\mbox{1}}}$, intercepting the horizontal and vertical axes at $+{\ensuremath{\mbox{1}}}$ and $-{\ensuremath{\mbox{1}}}$, respectively. *Right:* at the same time, constraint C8 changes from a slanted line with a slope of $+{\ensuremath{\mbox{1}}}$, passing through the origin, into a horizontal line at $v_{i}(t) = v_{\mbox{max}} - {\ensuremath{\mbox{1}}}$. []{data-label="fig:Constraints6And8"}](figures/lwr-stca-constraints-6-8 "fig:"){width="\figurewidth"} Deriving the fundamental diagram -------------------------------- The next step of our approach, considers the most determining linear inequalities C5, C6, C8, and C9 as boundaries in a $\overline v_{s_{e}}(\overline g_{s})$ fundamental diagram. As such, we note the following observations: - **Increasing the slowdown probability $p$, holding $v_{\mbox{max}}$ constant**: - The average speed $\overline v_{\mbox{ff}}$ in the free-flow regime decreases towards $v_{\mbox{max}} - p$. - The transition point at the critical space gap $g_{s_{c}}$ remains invariant. - The space gap $g_{s_{j}}$, corresponding to the jam density, increases. - **Decreasing the maximum speed $v_{\mbox{max}}$, holding $p$ constant**: - The average speed $\overline v_{\mbox{ff}}$ in the free-flow regime decreases towards $v_{\mbox{max}} - p$. - The transition point at the critical space gap $g_{s_{c}}$ decreases. - The space gap $g_{s_{j}}$, corresponding to the jam density, remains invariant. From equation [(\[eq:SpaceHeadway\])]{} from traffic flow theory, it follows that a vehicle’s space headway $h_{s}$ is equal to its space gap $g_{s}$ (i.e., the distance between the vehicle’s frontal bumper and the one of its direct frontal leader), plus the vehicle’s own length $l$. As such, the derived $\overline v_{s_{e}}(\overline g_{s})$ fundamental diagram can be converted into a $\overline v_{s_{e}}(\overline h_{s})$ fundamental diagram. Because we originally started from a single-cell TCA model (i.e., the STCA model with all vehicles having the same unit length), we can use our convention which states that a vehicle’s length $l_{i} \geq {\ensuremath{\mbox{1}}}~\mbox{cell} \propto \Delta X$ (see our discussion at the end of section \[sec:TCAModels\] for more details). Let us now consider the relation between the macroscopic traffic flow characteristic density $k$ and the microscopic characteristic average space headway $\overline h_{s}$, i.e., $\overline h_{s} = k^{-{\ensuremath{\mbox{1}}}}$ [@WARDROP:52; @DAGANZO:97]. This allows us to effectively transform the $\overline v_{s_{e}}(\overline h_{s})$ fundamental diagram into a $\overline v_{s_{e}}(k)$ fundamental diagram. As can be seen in the left part of [Fig. \[fig:DerivingStationaryFundamentalDiagrams\]]{}, increasing the stochasticity leads to the same observations that we previously mentioned. Finally, using the fundamental relation of traffic flow theory [(\[eq:FundamentalRelation\])]{} (see section \[sec:LWRModelDescription\]), our constraints are transformed into an equivalent *triangular* $q_{e}(k)$ fundamental diagram. Applying this technique results in the following analytical expressions for the parameters of the LWR’s fundamental diagram: $$\begin{aligned} \overline v_{\mbox{ff}} & = & (v_{\mbox{max}} - p) ~ \frac{\Delta X}{\Delta T} ~ \mbox{3.6},\label{eq:LWRFDExpressionVFF}\\ k_{\mbox{crit}} & = & \frac{\mbox{1000}}{(v_{\mbox{max}} + l) ~ \Delta X},\\ k_{\mbox{jam}} & = & \frac{\mbox{1000}}{(l + p) ~ \Delta X},\end{aligned}$$ with $l = {\ensuremath{\mbox{1}}}$ cell as the length of all vehicles in the single-cell STCA model. The capacity flow is calculated using the fundamental relation [(\[eq:FundamentalRelation\])]{}, resulting in the following expression: $$\label{eq:LWRFDExpressionQCAP} q_{\mbox{cap}} = k_{\mbox{crit}}~\overline v_{\mbox{ff}}.$$ As is visible in the right part of [Fig. \[fig:DerivingStationaryFundamentalDiagrams\]]{}, an increase of the stochasticity leads to a lower capacity flow $q_{\mbox{cap}}$, an invariant critical density $k_{c}$, and a smaller jam density $k_{j}$. \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] ![ *Left:* deriving a stationary $\overline v_{s_{e}}(\overline h_{s})$ fundamental diagram from the STCA’s constraints C1 – C9. The stochastic diagram has a higher inverse jam density, but the same inverse critical density as its deterministic counterpart (for the same $v_{\mbox{max}}$). *Right:* an equivalent triangular $q_{e}(k)$ fundamental diagram. []{data-label="fig:DerivingStationaryFundamentalDiagrams"}](figures/lwr-stca-hsv-kq-fds "fig:"){width="\figurewidth"} In conclusion, we note how rewriting the STCA’s rule set allowed us to obtain a stationary triangular $q_{e}(k)$ fundamental diagram. This fundamental diagram, which implicitly incorporates the STCA’s stochasticity, can then be specified as a parameter to the macroscopic first-order LWR model of section \[sec:LWRModelDescription\]. Finally note that the shape of the derived fundamental diagram is dictated by the inequality constraints C1 – C9. As such, it actually represents an ‘outer envelope’, that is to say, all possible fundamental diagrams lie beneath this envelope. This includes curved fundamental diagrams, more generally piecewise linear fundamental diagrams, et cetera. Application to an illustrative case study {#sec:ApplicationToACaseStudy} ========================================= After deriving a relation between the STCA and LWR models by means of the process explained in the previous section \[sec:InitialMethodology\], we now apply our methodology to a small case study. We first describe the setup of the test scenario, after which we interpret and discuss our obtained results. Description of the case study ----------------------------- The case study we consider, is modeled as a single-lane road that has a middle section with a reduced maximum speed (corresponding to e.g., an elevation, a speed limit, …). This road consists of three consecutive segments $A$, $B$, and $C$, as depicted in [Fig. \[fig:TestScenarioDescription\]]{}, whereby vehicles enter the road at segment $A$, travel through segment $B$, and exit it at the end of segment $C$. For the STCA, we assume a temporal and spatial discretization of $\Delta T = {\ensuremath{\mbox{1}}}$ s and $\Delta X =$ 7.5 m, respectively. The first road segment $A$ then consists of 1500 cells (11.25 km), while the second and third segments $B$ and $C$ each consist of 750 cells (i.e., each approximately 5.6 km long). The maximum speed for segments $A$ and $C$ is $v_{\mbox{max}}^{A,C} = {\ensuremath{\mbox{5}}}$ cells/time step, whereas it is $v_{\mbox{max}}^{B} = {\ensuremath{\mbox{1}}}$ cell/time step for segment $B$. The capacity flows for all three segments are denoted as $q_{\mbox{cap}}^{A,C}$ and $q_{\mbox{cap}}^{B}$. \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] \[\]\[\] ![ The single-lane road of the case study we consider, consisting of three consecutive segments $A$, $B$, and $C$. Assuming temporal and spatial discretizations of $\Delta T = {\ensuremath{\mbox{1}}}$ s and $\Delta X =$ 7.5 m, respectively, segment $A$ is composed of 1500 cells, while segments $B$ and $C$ are each composed of 750 cells. The maximum speed for segments $A$ and $C$ is $v_{\mbox{max}} = {\ensuremath{\mbox{5}}}$ cells/time step, whereas it is $v_{\mbox{max}} = {\ensuremath{\mbox{1}}}$ cell/time step for segment $B$. []{data-label="fig:TestScenarioDescription"}](figures/lwr-stca-test-scenario-description "fig:"){width="\figurewidth"} This road is simulated using both the STCA and the LWR model, each time for 3000 time steps. As for the boundary conditions, we assume an overall inflow of $q_{\mbox{cap}}^{B} / {\ensuremath{\mbox{2}}}$, except from time step 200 until time step 600, where we have created a short *traffic burst* of increased demand, with an inflow of $(q_{\mbox{cap}}^{A,C} + q_{\mbox{cap}}^{B}) / {\ensuremath{\mbox{2}}}$. [Fig. \[fig:TCATrajectoriesCloseup\]]{} shows a close up of the individual vehicle trajectories for the STCA in a time-space diagram, near the border between segments $A$ and $B$. As can be seen from the trajectories, heavy congestion sets in and flows upstream into segment $A$, where it starts to dissolve at the end of the traffic burst. The result is a typical triangular-shaped region that contains a queue of slow-moving vehicles (the backward propagating waves are clearly distinguished as the pattern of parallel black and white stripes). Applying our previously discussed methodology, we construct a stationary triangular $q_{e}(k)$ fundamental diagram. Its parameters are calculated by means of equations [(\[eq:LWRFDExpressionVFF\])]{} – [(\[eq:LWRFDExpressionQCAP\])]{}. The results are listed in [Table \[table:LWRCalculatedFDParameters\]]{}, with the TCA’s parameters expressed in cells/time step, vehicles/cell, and vehicles/time step, respectively, and the LWR’s parameters expressed in kilometres/hour, vehicles/kilometre, and vehicles/hour, respectively. [|l|r|r||l|r|r|]{} &\ & TCA & LWR & & TCA & LWR\ $\overline v_{\mbox{ff}}$ & 0.90 & 24.30 & $\overline v_{\mbox{ff}}$ & 4.90 & 132.30\ $k_{\mbox{crit}}$ & 0.50 & 66.67 & $k_{\mbox{crit}}$ & 0.17 & 22.22\ $k_{\mbox{jam}}$ & 0.91 & 121.20 & $k_{\mbox{jam}}$ & 0.91 & 121.20\ $q_{\mbox{cap}}$ & 0.45 & 1620.08 & $q_{\mbox{cap}}$ & 0.83 & 2939.71\ \ &\ & TCA & LWR & & TCA & LWR\ $\overline v_{\mbox{ff}}$ & 0.50 & 13.50 & $\overline v_{\mbox{ff}}$ & 4.50 & 121.50\ $k_{\mbox{crit}}$ & 0.50 & 66.67 & $k_{\mbox{crit}}$ & 0.17 & 22.22\ $k_{\mbox{jam}}$ & 0.67 & 88.89 & $k_{\mbox{jam}}$ & 0.67 & 88.89\ $q_{\mbox{cap}}$ & 0.25 & 900.05 & $q_{\mbox{cap}}$ & 0.77 & 2699.73\ Results and discussion ---------------------- The result of numerically solving the LWR model for the case of $p =$ 0.1 using the Godunov method [@DAGANZO:95b; @LEBACQUE:96], is depicted in the right part of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{}. Note that for the LWR model, each cell in the Godunov scheme corresponds to 5 (i.e., $v_{\mbox{max}}^{A,C}$) consecutive cells of the STCA model. Comparing the tempo-spatial behavior of the LWR model to that of the microscopic system dynamics of the STCA model (i.e., the left part of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{}), we find a good *qualitative* agreement between the two approaches. With respect to the first-order traffic flow characteristics, we note that the buildup and dissolution of congestion queues are fairly analogous for both techniques. ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the microscopic system dynamics of the STCA model. *Right:* the results for the LWR model. In both cases, $p =$ 0.1, with darker regions corresponding to more congested traffic conditions. There is a qualitatively good agreement between the two approaches on the level of first-order traffic flow characteristics: the buildup and dissolution of congestion queues are fairly analogous for both techniques. []{data-label="fig:STCAAndLWRTXDensityDiagramsP01"}](figures/lwr-stca-tx-k-stca-vac5-vb1-p01-qac2400-qb1220 "fig:"){width="\figurehalfwidth"} ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the microscopic system dynamics of the STCA model. *Right:* the results for the LWR model. In both cases, $p =$ 0.1, with darker regions corresponding to more congested traffic conditions. There is a qualitatively good agreement between the two approaches on the level of first-order traffic flow characteristics: the buildup and dissolution of congestion queues are fairly analogous for both techniques. []{data-label="fig:STCAAndLWRTXDensityDiagramsP01"}](figures/lwr-stca-tx-k-lwr-vac5-vb1-p01-qac2400-qb1220 "fig:"){width="\figurehalfwidth"} In [Fig. \[fig:STCAAndLWRTXDensityDiagramsP05\]]{}, we show the results when repeating the same experiment, but this time with the stochastic noise $p$ set to 0.5 for all three segments. As revealed by the shape of the dark triangular region in the LWR model (right part), the buildup and dissolution of congestion queues seems to be *exaggerated*, especially in the upstream flowing queue of segment $A$. ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the microscopic system dynamics of the STCA model. *Right:* the results for the LWR model. In both cases, $p =$ 0.5, with darker regions corresponding to more congested traffic conditions. As revealed by the shape of the dark triangular region in the LWR model (right part), the buildup and dissolution of congestion queues is exaggerated, especially in the upstream flowing queue of segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsP05"}](figures/lwr-stca-tx-k-stca-vac5-vb1-p05-qac1220-qb540 "fig:"){width="\figurehalfwidth"} ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the microscopic system dynamics of the STCA model. *Right:* the results for the LWR model. In both cases, $p =$ 0.5, with darker regions corresponding to more congested traffic conditions. As revealed by the shape of the dark triangular region in the LWR model (right part), the buildup and dissolution of congestion queues is exaggerated, especially in the upstream flowing queue of segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsP05"}](figures/lwr-stca-tx-k-lwr-vac5-vb1-p05-qac1220-qb540 "fig:"){width="\figurehalfwidth"} It is interesting to note that the STCA model reveals a *higher-order effect* that is not visible in the LWR model: there exists a fan of forward propagating density waves in segment $B$ (see the left parts of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{} and [Fig. \[fig:STCAAndLWRTXDensityDiagramsP05\]]{}). As such, in its tempo-spatial diagram, the STCA seems to be able to visualize the characteristics that constitute the solution of the LWR model. In order to more rigourously quantify the discrepancies between the time-space diagrams of both STCA and LWR models, we provide their *absolute differences* in [Fig. \[fig:STCAAndLWRTXDensityDiagramsDifferences\]]{}. The left part shows the differences for $p =$ 0.1, whereas the right part shows the differences for $p =$ 0.5. The most important features to look at, are the dark colored regions which indicate larger differences between both modeling approaches. We can clearly see that there is a problem with respect to a *quantitative* agreement between both STCA and LWR models. It appears as though the LWR model *overestimates* the STCA’s capacity flows. As a result, it dissolves its jams more quickly in segment $B$, and it predicts a more severe onset of congestion in segment $A$ (i.e., the triangular-shaped region containing the spill-back queue is more pronounced in the LWR’s case). The sharply pronounced darker regions in the tempo-spatial left part of segment $B$, are due to the fact that the LWR model does not visualize the characteristics of its solution, in contrast to the STCA model which is able to give a clear indication of them. ![ Time-space diagrams showing the differences in densities for the STCA and LWR models, during 3000 time steps for the road in the case study. Darker regions indicate large differences between both modeling approaches. *Left:* the differences for $p =$ 0.1 are less pronounced, showing only a dark edge at the bottom triangular-shaped region in segment $A$. *Right:* the differences for $p =$ 0.5, showing significant discrepancies in the bottom of the triangular-shaped region in segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsDifferences"}](figures/lwr-stca-lwr-difference-vac5-vb1-p01-qac2400-qb1220 "fig:"){width="\figurehalfwidth"} ![ Time-space diagrams showing the differences in densities for the STCA and LWR models, during 3000 time steps for the road in the case study. Darker regions indicate large differences between both modeling approaches. *Left:* the differences for $p =$ 0.1 are less pronounced, showing only a dark edge at the bottom triangular-shaped region in segment $A$. *Right:* the differences for $p =$ 0.5, showing significant discrepancies in the bottom of the triangular-shaped region in segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsDifferences"}](figures/lwr-stca-lwr-difference-vac5-vb1-p05-qac1220-qb540 "fig:"){width="\figurehalfwidth"} One of the main reasons for this discrepancy between both modeling approaches, lies in the derivation of a triangular $q_{e}(k)$ fundamental diagram for the LWR model, as was explained in section \[sec:InitialMethodology\]. Because we assumed a stationarity condition on the STCA’s rule set, the resulting constraints implied an invariant critical density, and always overestimated the STCA’s capacity flows. In our opinion, the different behavior of both models, mainly stems from this artifact. As a result, the discrepancies will become more articulated when increasing the stochastic noise $p$. Alternate derivation of the fundamental diagram {#sec:AlternateDerivation} =============================================== Considering the results of the previous approach, i.e., deriving the LWR’s fundamental diagram based on the STCA’s rule set, and the problems related to it, the next step is to specify the fundamental diagram *directly*, based on the empirically observed behavior of the STCA model. In the following two sections, we first discuss the effects of explicitly adding noise to the LWR’s fundamental diagram, after which we discuss our obtained results when specifying the fundamental diagram directly. The effect of adding noise to the LWR’s fundamental diagram ----------------------------------------------------------- Adding noise to the LWR model can mainly be accomplished via two ways: either by explicitly incorporating noise terms in the LWR equations (e.g., the conservation equation), or as a noise term in the $q_{e}(k)$ relation (i.e., the fundamental diagram). We refrain from changing the LWR’s conservation equation, because this amounts to introducing some form of numerical diffusion, similar to the viscosity terms that are originally encountered in the conservation equation’s [(\[eq:FirstOrderConservationLaw\])]{} right-hand side (which we set to zero in order to obtain the inviscid LWR model). In [Fig. \[fig:LWRTXDensityDiagramsWithNoise\]]{}, we show the results of supplying uniformly distributed additive noise of 0.1 (left part) and 0.5 (right part). As can be seen, the introduction of noise in the fundamental diagram, leads to a ‘spreading’ of the solution. For small noise levels, some of the characteristics are revealed; for larger noise levels, the characteristics are clearly pronounced, including long jam dissolution times. Specifying the fundamental diagram directly ------------------------------------------- Instead of deriving the fundamental diagram based on the approach taken in section \[sec:InitialMethodology\], we now try to obtain the values for the critical densities and capacity flows directly, by looking at the STCA’s ($k$,$q$) diagrams in [Fig. \[fig:STCAKQDiagrams\]]{} (an equivalent procedure would be to measure the capacity flows directly from the time-space diagram in [Fig. \[fig:TCATrajectoriesCloseup\]]{}). Considering the STCA’s ($k$,$q$) diagrams in [Fig. \[fig:STCAKQDiagrams\]]{}, we can estimate its capacities at approximately $q_{\mbox{cap}}^{B} = $ 0.34 vehicles/time step $\approx$ 1220 vehicles/hour, and $q_{\mbox{cap}}^{A,C} = $ 0.67 vehicles/time step $\approx$ 2400 vehicles/hour for $v_{\mbox{max}}^{B} = {\ensuremath{\mbox{1}}}$ and $v_{\mbox{max}}^{A,C} = {\ensuremath{\mbox{5}}}$ cells/time step, respectively. The stochastic noise $p$ was set to 0.1 for all three segments. Changing $p$ to 0.5 for these segments, we can estimate the capacities at approximately $q_{\mbox{cap}}^{B} =$ 0.15 vehicles/time step $\approx$ 540 vehicles/hour, and $q_{\mbox{cap}}^{A,C} = $ 0.34 vehicles/time step $\approx$ 1220 vehicles/hour for $v_{\mbox{max}} = {\ensuremath{\mbox{1}}}$ and $v_{\mbox{max}} = {\ensuremath{\mbox{5}}}$ cells/time step, respectively. ![ The ($k$,$q$) fundamental diagrams for the STCA model. *Left:* two diagrams for $v_{\mbox{max}}^{B} = {\ensuremath{\mbox{1}}}$ cells/time step. *Right:* two diagrams for $v_{\mbox{max}}^{A,C} = {\ensuremath{\mbox{5}}}$ cells/time step. Each time, the slowdown probability $p \in \lbrace \mbox{0.1}, \mbox{0.5} \rbrace$. Note how a slower maximum speed makes the diagrams more curved, and how an increasing slowdown probability leads to both a lower critical density and capacity flow. []{data-label="fig:STCAKQDiagrams"}](figures/lwr-stca-fundamental-diagrams-stca-vmax1-k-q "fig:"){width="\figurehalfwidth"} ![ The ($k$,$q$) fundamental diagrams for the STCA model. *Left:* two diagrams for $v_{\mbox{max}}^{B} = {\ensuremath{\mbox{1}}}$ cells/time step. *Right:* two diagrams for $v_{\mbox{max}}^{A,C} = {\ensuremath{\mbox{5}}}$ cells/time step. Each time, the slowdown probability $p \in \lbrace \mbox{0.1}, \mbox{0.5} \rbrace$. Note how a slower maximum speed makes the diagrams more curved, and how an increasing slowdown probability leads to both a lower critical density and capacity flow. []{data-label="fig:STCAKQDiagrams"}](figures/lwr-stca-fundamental-diagrams-stca-vmax5-k-q "fig:"){width="\figurehalfwidth"} Instead of calculating the capacity flows from the average free-flow speeds and the critical densities, as was done by means of equation [(\[eq:LWRFDExpressionQCAP\])]{}, we now specify these capacity flows directly to the LWR’s fundamental diagrams and calculate the critical densities from them. The results we obtained, are visualized in the time-space diagrams of [Fig. \[fig:LWRTXDensityDiagramsP01P05\]]{}. Because the STCA’s capacity flows are now better approximated, and not overestimated as with the previous methodology, there seems to be a better qualitative agreement for both noise levels with the STCA’s time-space diagrams in the left parts of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{} and [Fig. \[fig:STCAAndLWRTXDensityDiagramsP05\]]{}. ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the results for the LWR model with $p =$ 0.1. *Right:* the results for the LWR model with $p =$ 0.5. In both diagrams, the fundamental diagram was specified directly to the LWR model, by explicitly stipulating the capacity flows of the STCA model. As a result, there seems to be a better qualitative agreement for both noise levels with the STCA’s time-space diagrams in the left parts of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{} and [Fig. \[fig:STCAAndLWRTXDensityDiagramsP05\]]{}. []{data-label="fig:LWRTXDensityDiagramsP01P05"}](figures/lwr-stca-tx-k-lwr-vac5-vb1-p01-qac2400-qb1220-direct "fig:"){width="\figurehalfwidth"} ![ Time-space diagrams showing the propagation of densities during 3000 time steps for the road in the case study. *Left:* the results for the LWR model with $p =$ 0.1. *Right:* the results for the LWR model with $p =$ 0.5. In both diagrams, the fundamental diagram was specified directly to the LWR model, by explicitly stipulating the capacity flows of the STCA model. As a result, there seems to be a better qualitative agreement for both noise levels with the STCA’s time-space diagrams in the left parts of [Fig. \[fig:STCAAndLWRTXDensityDiagramsP01\]]{} and [Fig. \[fig:STCAAndLWRTXDensityDiagramsP05\]]{}. []{data-label="fig:LWRTXDensityDiagramsP01P05"}](figures/lwr-stca-tx-k-lwr-vac5-vb1-p05-qac1220-qb540-direct "fig:"){width="\figurehalfwidth"} In [Fig. \[fig:STCAAndLWRTXDensityDiagramsDifferencesDirect\]]{} we have depicted the absolute differences between this approach and the STCA’s time-space diagrams. Comparing this to the previous results of [Fig. \[fig:STCAAndLWRTXDensityDiagramsDifferences\]]{}, we can see that in both cases the buildup and dissolution of congestion queues is in good qualitative agreement for both noise levels. *As such, we come the conclusion that it is vital to correctly capture the capacity flows of the STCA model.* Neglecting this property, can result in severe distortion of the system dynamics for higher noise levels. ![ Time-space diagrams showing the differences in densities for the STCA and LWR models, during 3000 time steps for the road in the case study. Darker regions indicate large differences between both modeling approaches. *Left:* the differences for $p =$ 0.1. *Right:* the differences for $p =$ 0.5. In both cases, the differences are less pronounced, showing only dark edges at the bottom of the triangular-shaped region in segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsDifferencesDirect"}](figures/lwr-stca-lwr-difference-vac5-vb1-p01-qac2400-qb1220-direct "fig:"){width="\figurehalfwidth"} ![ Time-space diagrams showing the differences in densities for the STCA and LWR models, during 3000 time steps for the road in the case study. Darker regions indicate large differences between both modeling approaches. *Left:* the differences for $p =$ 0.1. *Right:* the differences for $p =$ 0.5. In both cases, the differences are less pronounced, showing only dark edges at the bottom of the triangular-shaped region in segment $A$. []{data-label="fig:STCAAndLWRTXDensityDiagramsDifferencesDirect"}](figures/lwr-stca-lwr-difference-vac5-vb1-p05-qac1220-qb540-direct "fig:"){width="\figurehalfwidth"} Note that the LWR model is able to correctly capture the first-order effects of jam buildup and dissolution, and that, due to its microscopic treatment, the STCA model allows us to visualize the higher-order effects inside jam. However, as is evidenced by this and the previous section, it is very important to correctly capture the capacity flows in the STCA model, otherwise a growing discrepancy between the LWR and STCA model is introduced with higher noise levels. Conclusions {#sec:Conclusions} =========== In this paper, we presented an alternate methodology for implicitly incorporating the STCA’s stochasticity into the macroscopic first-order LWR model. The innovative aspect of our approach, is that we derive the LWR’s fundamental diagram directly from the STCA’s rule set, by assuming a stationarity condition that converts the STCA’s rules into a set of linear inequalities. In turn, these constraints define the shape of the fundamental diagram that is then specified to the LWR model. For noise-free systems, our method is exact. In the presence of noise, however, the capacity flows in the derived fundamental diagram are overestimates of those of the STCA model. This discrepancy can be explained as follows: the underlying assumption for the LWR model, is that the fundamental diagram is assumed to be exact, and implicitly obeyed, i.e., the existing equilibrium relation is representative for the real traffic situation. In the original LWR formulation, this relation was also assumed to hold also for non-stationary traffic (which is a more or less reasonable assumption if we consider long and crowded roads). Our calculations have shown that a direct translation of the STCA’s rule set into the LWR’s fundamental diagram, does not always result in a valid fundamental diagram, especially for higher noise levels. As such, there can be a significant difference between an *average* fundamental diagram (STCA) and a *stationary* fundamental diagram (LWR). As a result, the STCA model is able to temporarily operate under larger flows and densities than those possible for the LWR’s stationary fundamental diagram. A logical course of action would to better approximate the STCA’s fundamental diagram. By doing so however, we lose the advantage gained through an explicit derivation of the fundamental diagrams’ outer envelope, almost certainly leading to extra conditions that need to make further assumptions about its shape. Directly specifying the STCA’s capacity flows to the triangular LWR fundamental diagram, effectively remedies most of the mismatches between both STCA and LWR models. Our methodology sees the STCA complementary to the LWR model and vice versa, so the results can be of great assistance when interpreting the traffic dynamics in both models. Especially appealing, is the fact that the STCA can visualise the higher-order characteristics of traffic stream dynamics, e.g., the fans of rarefaction waves. Nevertheless, because the LWR model is only a coarse representation of reality, there are still some mismatches between the two approaches. One of the main concerns the authors discovered, is as hinted at earlier, the fact that using a stationary fundamental diagram (i.e., an equilibrium relation between density and flow), always overestimates the practical capacity of a stochastic cellular automaton model. As such, it is vital to correctly capture the capacity flows in both STCA and LWR models, a remark that we feel is valid for all case studies. Acknowledgements {#acknowledgements .unnumbered} ================ Dr. Bart De Moor is a full professor at the Katholieke Universiteit Leuven, Belgium. Our research is supported by: **Research Council KUL**: GOA AMBioRICS, several PhD/postdoc & fellow grants, **Flemish Government**: **FWO**: PhD/postdoc grants, projects, G.0407.02 (support vector machines), G.0197.02 (power islands), G.0141.03 (identification and cryptography), G.0491.03 (control for intensive care glycemia), G.0120.03 (QIT), G.0452.04 (new quantum algorithms), G.0499.04 (statistics), G.0211.05 (Nonlinear), research communities (ICCoS, ANMMM, MLDM), **IWT**: PhD Grants, GBOU (McKnow), **Belgian Federal Science Policy Office**: IUAP P5/22 (‘Dynamical Systems and Control: Computation, Identification and Modelling’, 2002-2006), PODO-II (CP/40: TMS and Sustainability), **EU**: FP5-Quprodis, ERNSI, **Contract Research/agreements**: ISMC/IPCOS, Data4s,TML, Elia, LMS, Mastercard.

--- abstract: 'We apply the recently developed adaptive non-harmonic model based on the wave-shape function, as well as the time-frequency analysis tool called synchrosqueezing transform (SST) to model and analyze oscillatory physiological signals. To demonstrate how the model and algorithm work, we apply them to study the pulse wave signal. By extracting features called the spectral pulse signature, [and]{} based on functional regression, we characterize the hemodynamics from the radial pulse wave signals recorded by the sphygmomanometer. Analysis results suggest the potential of the proposed signal processing approach to extract health-related hemodynamics features.' address: - '(1) Mathematics, University of Toronto, Toronto, Ontario, Canada.' - '(2) Institute of Traditional Medicine, School of Medicine, National Yang-Ming University, Taipei, Taiwan.' - '(3) Department of Chinese Medicine, Taiwan Landseed Hospital, Tao-Yuan, Taiwan.' - '(4) Cardiovascular Division, Department of Internal Medicine, Chang Gung Memorial Hospital, Taoyuan, Taiwan. College of Medicine, Chang Gung University, Taiyuan, Taiwan.' - '(5) Center for Traditional Chinese Medicine, Keelung Chang Gung Memorial Hospital, Taiwan.' - '(6) Department of Healthcare Management, Yuanpei University of Medical Technology, Hsinchu, Taiwan.' - '(7) School of Post-Baccalaureate Traditional Chinese Medicine, and Research Center for Chinese Medicine and Acupuncture, China Medical University, Taiwan.' - '(8) Departments of Chinese Medicine, China Medical University Hospital, Taiwan.' author: - 'Hau-Tieng Wu^1^, Han-Kuei Wu^2,3^, Chun-Li Wang^4^, Yueh-Lung Yang^5^, Wen-Hsiang Wu^6^, Tung-Hu Tsai^2^ and Hen-Hong Chang^7,8^' bibliography: - '../bib/Optimization.bib' - '../../tex20130723/bib/MathBooks.bib' - '../bib/MedicalBooks.bib' - '../bib/PulseDiagnosis.bib' - '../bib/PhysiologicalVariability.bib' - '../bib/noisyManifold.bib' - '../bib/TFanalysis.bib' - '../bib/Thesis.bib' - 'MultivariateTimeSeries.bib' title: 'Modeling the pulse signal by wave-shape function and analyzing by synchrosqueezing transform' --- Introduction ============ Hemodynamics is an essential ingredient of cardiovascular physiology which not only reflects the forces the heart needs to pump blood through the cardiovascular system, but also reflects the integrity of [an individual’s]{} physiological system. While there are [many]{} aspects [to]{} hemodynamics [@Milnor:1989; @Nicholas_ORourke:2005; @Nicholas_ORourke_Vlachopoulos:2011; @Caro_Pedley_Schroter_Seed:2011; @Salvi:2012], it is common to evaluate hemodynamics by assessing the pulse wave signals recorded [at]{} different locations. For example, the brachial cuff systolic and diastolic blood pressures derived from the pulse wave [have important clinical applications, long established beginning]{} 1870 [@Lewington_Clarke_Qizilbash_Peto_Collins:2002]. In addition to the blood pressure, the pulse wave signal itself contains [an abundance of]{} clinical information. For example, [in]{} subjects with dysfunctional left ventricle, different carotid pulse patterns [can be observed, such as]{} hyperkinetic pulse, pulsus alternans, pulse bisferiens, pulsus parvus et tardus, etc [@Morris:1990; @Braunwald_Perloff:2014]. [^1] However, it is difficult to [obtain these]{} carotid pulse wave signals non-invasively [@Chen_Ting_Nussbacher_Nevo_Kass_Pak_Wang_Chang_Win:1996]. On the other hand, the non-invasive pulse diagnosis (pulsology) based on, for example, the brachial or radial pulse signal, provides several [aspects of]{} physiological information, [such as]{} the central pressure wave [@Scuteri_Chen_Yin_Chih_Spurgeon_Lakatta:2001; @ORourke_Pauca_Jiang:2001]. Other indices associated with the pulse wave signals recorded from different locations, [including]{} the augmentation index [@Salvi:2012 Section 6.1][^2] [@Chen_Ting_Nussbacher_Nevo_Kass_Pak_Wang_Chang_Win:1996; @Vlachopoulos_Aznaouridis_ORourke_Safar_Baou_Stefanadis:2010], are [also]{} available for clinical [applications. Since]{} there is a [great amount of]{} information [within]{} the pulse signal, a [deeper and more extensive understanding]{} of the pulse wave is undoubtedly important to [better]{} assess not only the cardiovascular but also the physiological integrity. The common [approaches to analyzing]{} the pulse wave signals [can]{} be classified into two categories – time domain analysis [and]{} frequency domain analysis. In both categories, [it is necessary to have either]{} a representative pulse wave cycle, which reflects the interaction between the one time heart pump driving force and the impedance to blood flow, or a sequence of pulse wave cycles over a period of time, during which the heart rate is relatively constant. In the time domain analysis, different landmarks or features are identified from the pulse signal. For example, [with the]{} radial pulse signal [is characterized by]{} the height of the percussive wave, the height of the dicrotic notch, the length of the cardiac cycle, the length of acute ejection, etc. Different features with different physiological meanings [can]{} be obtained from the pulse signals recorded from other areas, and these features have been widely used in clinics for health evaluation [@Braunwald_Perloff:2014]. [Furthermore, new mathematical methods have been applied to investigate additional physiological information. For example, the central pulse wave could be reconstructed from the radial pulse wave by a generalized transfer function, and then be separated into inflection and reflection waves in order to evaluate markers of early vascular aging [@Melenovsky_Borlaug_Fetics_Kessler_Shively_Kass:2007].]{} In the frequency domain analysis, the energy of different harmonic modes estimated by power spectrum analysis are [also]{} potential features for clinical [applications]{}. According to the resonance theory [@Wang_Chang_Wu_Hsu_Wang:1991], different harmonic modes are associated with the integrity of different organs [@Hsu_Chao_Hsiu_Wang_Li_Wang:2006]. Several existing works show the potential of this spectral approach, [such as]{} the relationship between the internal organ disorders and the spectral content [@Wei_Lee_Chow:1984; @Huang_Wei_Liao_Chang_Kao_Li:2011]. [Although]{} this approach is commonly applied, [it has]{} two significant limitations. The first limitation is the over-simplification of the model underlying the analysis techniques. For example, heart rate variability (HRV) and pulsus alternans are often [not used]{} in the analysis. [Though]{} the time domain landmark features reflect the hemodynamics from different aspects, they might be confounded by HRV and pulsus alternans. Since HRV and pulsus alternans are inevitable physiological dynamics, [without considering them in the model and analysis, both time domain and frequency domain analysis results may be inaccurate and the results might lead to incorrect interpretation.]{} The second limitation is that it is not always straightforward to determine a representative pulse cycle or a period of time [when]{} the pulse wave signal is suitable for [these]{} analysis techniques, [so that]{} human intervention and subjective decision making are required. Indeed, to determine a pulse cycle, [it is necessary]{} to define what a pulse cycle means or, at least, [to determine]{} landmarks within the pattern to separate one pulse [cycle from]{} another. For example, [with an]{} electrocardiogram (ECG) signal, the repeating basic pattern is related to the electrophysiological activity of a normal heart beat, and a common chosen landmark [for this]{} is the R peak. As the R peak is usually dominant and significant, determining the oscillatory pattern for an ECG signal is not difficult. However, for pulse signal [such as those]{} acquired [in this study]{}, it is not always easy to define an oscillatory pattern or a landmark, in particular when the signal is recorded from an abnormal subject. Even [more seriously]{}, due to the [possibility of a]{} noisy record, the pulse cycle morphology [can]{} dramatically change from one pulse to another or [from]{} one subject to another (See Figure \[fig:PulseExample\]). Although this difficulty can be mitigated to some extent by noise reduction algorithms, due to the non-linear nature of the signal, [their]{} performance is not guaranteed. ![An example of the radial pulse signal recorded from a subject with congestive heart failure. Note that at the $3^{rd}$ second there is an obvious artifact, indicated by the blue arrow, [and]{} at the $5.5^{th}$ second there are two shock noise events. The relatively small pulse wave at the $3^{rd}$ second is related to a [premature ]{}atrial contraction. [This is significant since patients]{} with heart failure may have more premature beats than a normal control.[]{data-label="fig:PulseExample"}](Figure1.png){width=".75\textwidth"} To address these limitations, in this paper we apply a recently proposed and analyzed descriptive model, called the [*adaptive non-harmonic model*]{}, to describe the pulse wave signals. [This]{} model is characterized by [what is referred to as the]{} [*wave-shape function*]{} [@Wu:2013], which is defined [in order]{} to capture [what]{} a pulse wave cycle should look like, [including]{} the features in the time domain and frequency domain mentioned above. The main characteristic of the proposed model is that the HRV and pulsus alternans are decoupled from the wave-shape function. A companion algorithm, referred to as the [*synchrosqueezing transform*]{} (SST), is applied to provide an accurate estimation of the wave-shape function. SST is a recently proposed time-frequency analysis technique [that has a]{} mathematical foundation [and robustness to different kinds of noise and different practical issues have]{} been well established in the literature [@Daubechies_Lu_Wu:2011; @Chen_Cheng_Wu:2014]. [By the adaptive non-harmonic model and SST, the above-mentioned limitations can be alleviated.]{} By a suitable manipulation of the estimated pulse wave cycle, a vector [can be obtained]{} as an associated feature, which further leads to our final index, [referred to as the]{} [*global pulse signature (GPS)*]{} [@Wu:2013; @Chui_Lin_Wu:2016]. This paper is organized [as follows]{}. In Section \[Section:methodology\], we systematically model the pulse signal [using]{} the notion of wave-shape function and the adaptive non-harmonic model. In Section \[Section:MaterialandMethod\], we summarize the SST algorithm and discuss how the wave-shape function [can be estimated]{} based on the SST and functional regression technique and [then]{} generate the GPS index. In Section \[Section:materialresult\], we apply the proposed algorithm to study the radial pulse signal. The data collection procedure, the physiological background of pulsology and the analysis result are reported. In Section \[Section:discussion\], we discuss our findings [with their]{} clinical interpretations and indicate [directions for]{} future [research]{}. Adaptive Non-Harmonic Model {#Section:methodology} =========================== Adaptive non-harmonic model for the pulse signal {#Section:AdaptiveHarmonicModel} ------------------------------------------------ In this section, we propose [a]{} mathematical model to describe the pulse wave signal. It is well known that hemodynamics is a [highly complex]{} subject [@Nicholas_ORourke:2005; @Nicholas_ORourke_Vlachopoulos:2011; @Caro_Pedley_Schroter_Seed:2011; @McEniery_Cockcroft_Roman_Franklin_Wilkinson:2014; @Westerhof_vanderWijngaard_Murgo_Westerhof:2008]. Different signals recorded from different areas of the body by different instruments [can shed light on]{} different aspects of hemodynamics. The question of modeling hemodynamics has been discussed by several authors, [who have proposed]{} different models based on different physiological factors, [such as]{} pressure gradient, resonance, vibration, etc, have been proposed; see [@Milnor:1989; @Wang_Hsu_Jan_Wang:2010; @Nicholas_ORourke_Vlachopoulos:2011; @Caro_Pedley_Schroter_Seed:2011] and the literatures cited [therein]{} for detailed discussions [of]{} these models. While these models shade a light [on]{} the understanding of the pulse wave signal, [having]{} a signal model to [effectively]{} quantify the pulse behavior [remains a distant goal]{}. In particular, inevitable physiological [aspects, such as]{} HRV and pulsus alternans, are commonly missed or [disregarded]{} in the existing models. Ignoring them in the model and analysis, [however]{}, might not only limit the analysis but also lead to [incorrect]{} findings. In this paper, we apply a different model and approach to [resolve]{} these issues. The pulse wave signal is oscillatory in nature, and there are several well-known features [that we consider here]{}: how fast the signal oscillates, how much the oscillatory rate changes, how [differently]{} the signal behaves inside an oscillation, etc. Understanding these features plays a significant role in [the]{} data analysis, which [can help]{} to obtain more information about the [physiological]{} system. Based on these observations, we consider the following adaptive non-harmonic model[@Wu:2013] to model the pulse signal. This model [differs]{} from most existing models in that it is purely [*phenomenological*]{}; that is, the parameters in the model are solely driven by observations of the physiological signal, [rather than by]{} explicit and quantitative underlying mechanisms. ### Wave shape function Based on the periodic nature of an oscillatory function, we introduce the [*wave shape function*]{}, which is [used]{} to model how a signal oscillates over the period. This idea has been [previously]{} studied in [@Wu:2013; @Yang:2014; @Hou_Shi:2016]. [To resolve the difficulty of defining a period]{}, we shall call a function $f$ [*$\tau-$periodic*]{}, where $\tau>0$, if (1) the periodic condition is satisfied; that is, for all $t \in {\mathbb{R}}$ and all $k \in {\mathbb{Z}}$, $f(t+k\tau)=f(t)$; and (2) for all $0<\tau'<\tau$ the periodic condition is not satisfied. Take two positive values $\delta$ and $\theta$ and a positive integer $D$. A $1$-periodic function $s$ is called a [*wave shape function*]{} with parameters $\delta$, $D$ and $\theta$ if the following conditions are satisfied. First, [the function]{} is differentiable and its derivative is a Holder continuous function with the Holder coefficient $\alpha>1/2$. Denote $\widehat{s}$ to be the Fourier transform of the function $s$. The second condition [that]{} a wave shape function should satisfy is that $s$ has the unit energy (unit $L^2$-norm) and all the Fourier modes $\widehat{s}(k)$, $k\neq 1$ are dominated by the product of $\delta$ and the first mode coefficient; that is, $$\label{definition:shape:1} \forall k \in {\mathbb{N}}, \, \mbox{ with }k\neq 1, \, \left| \widehat{s}(k) \right| \leq \delta \, \left| \widehat{s}(1) \right|.$$ Furthermore, [it is desirable for]{} $\widehat{s}$ [to be]{} mostly concentrated in the low frequency region, which is quantified by $\theta$ by: $$\label{definition:shape:2} \sum_{n>D}|n\widehat{s}(n)|\leq \theta.$$ [Conditions]{} (\[definition:shape:1\]) and (\[definition:shape:2\]) [reflect]{} the practical finding of the spectral analysis of the pulse signals [@Taylor:1966; @Wei_Chow:1985; @LinWang_Jan_Shyu_Chiang_Wang:2004], [and]{} it [can be]{} observed that the amplitudes of tenth and higher order harmonics are negligible. Note that the commonly selected landmarks and lengths considered in the time domain analysis could be understood as morphological features describing the wave-shape function [that is]{} modeling how a pulse repeats itself. While there is a one-to-one relationship between these landmarks and Fourier coefficients $\widehat{s}$, this relationship is nonlinear in nature. Thus, while we could view the Fourier coefficients of the wave shape function as the “features” of the pulse wave, their physiological interpretations are not directly related to the [hemodynamic]{} interpretations of these landmarks. See Figure \[fig:ShapeExample\] for an example of the wave-shape function describing the pulse wave signal. ### Adaptive non-harmonic model {#adaptive-non-harmonic-model} [Using]{} the notion of wave shape function, we now describe our phenomenological model to capture the recorded pulse signal. Fix parameters $\delta,\,\theta,\,D$ for a wave shape function $s$. We consider the following *Intrinsic Mode Type (IMT) functions* to model the pulse wave signal. An IMF function, $f$, is a bounded and continuous function with a continuous derivative [that]{} satisfies the following format: $$\label{definition:IMT} f(t)=A(t)s(2\pi \phi(t)),$$ where $A$ is a positive differentiable function and $\phi$ is a monotonically differentiable function. Intuitively, $A(t)$ describes how large the oscillation is at time $t$, and the positive function $\phi'(t)$, the first derivative of $\phi$, describes how fast the oscillation is at time $t$. To see how it is interpreted in this way, consider a constant positive function $A$ and a linear function with a positive slope as $\phi$. Suppose $\phi(t)=\xi_0 t$, where $\xi_0>0$. In this case, we know that $f$ is a harmonic function with the frequency $\xi_0$ and the amplitude $A$. We could thus view (\[definition:IMT\]) as a generalization of the harmonic function. [Though]{} the heart rate is not constant, normally it does not change suddenly. [Therefore]{} we need following conditions to better quantify the pulse signal. Fix a small positive constant $\epsilon$. Then we assume that $A(t)$ does not change too fast in the sense that its derivative is bounded by $\epsilon \phi'(t)$ and that $\phi'(t)$ does not change too fast in the sense that $\phi''(t)$ exists and is bounded by $\epsilon\phi'(t)$; that is, we have $$\label{definition:IMT2} |A'(t)|\leq \epsilon\phi'(t)\mbox{ and }|\phi''(t)|\leq \epsilon\phi'(t)\mbox{ for all time }t.$$ We would thus model the pulse wave signal by (\[definition:IMT\]) with the condition (\[definition:IMT2\]), and call this model [the]{} [*adaptive non-harmonic model*]{}. We would call the monotonically increasing function $\phi(t)$ the [*phase function*]{}, $\phi'(t)$ the [*instantaneous frequency*]{} (IF), and $A(t)$ the [*amplitude modulation*]{} (AM). An important issue regarding the identifiability issue of the phase function, the IF and the AM [cannot]{} be ignored if the [discussion is to be]{} rigorous. Indeed, [there might be]{} infinitely many different ways to represent a cosine function $g_0(t)=\cos(2\pi t)$ in the format $a(t)\cos(2\pi\phi(t))$ so that $a>0$ and $\phi'>0$, even though it is well known that $g_0(t)$ is a harmonic function with amplitude $1$ and frequency $1$. Indeed, we could find infinitely many smooth functions $\alpha$ and $\beta$ so that $g_0(t)=\cos(t)=(1+\alpha(t))\cos(2\pi(t+\beta(t)))$, and in general there is no reason to favor $\alpha(t)=\beta(t)=0$. Clearly, before resolving this issue, [amplitude $1$ and frequency $1$ cannot be taken as]{} reliable features to quantify the signal $g_0$. This identifiability issue has been well studied in [@Chen_Cheng_Wu:2014; @Kowalski_Meynard_Wu:2015], [finding]{} that if $g_0(t)=A(t)s(\phi(t))$ and $g_0(t)=(A(t)+\alpha(t))s(\phi(t)+\beta(t))$ also holds, and both satisfy the condition (\[definition:IMT2\]), then $|\alpha(t)|$ and $|\beta'(t)|$ are both bounded by $\epsilon$ for all time $t\in{\mathbb{R}}$. Note that the IF and AM are always positive, but usually not constant. Clearly, when $\phi$ is a linear function with a positive slope and $a$ is a positive constant, then the model is reduced to the harmonic model and the IF is equivalent to the notion frequency in the ordinary Fourier transform sense. The conditions $|a'(t)|\leq \epsilon \phi'(t)$ and $|\phi''(t)|\leq \epsilon\phi'(t)$ force the signal to [locally oscillate “regularly”]{}, that is, around time $t_0$, the signal $A(t)s(\phi(t))$ oscillates like $A(t_0)s(\phi(t_0)-t_0\phi'(t_0)+\phi'(t_0)t)$, and hence the nominations of IF and AM. We mention that [this]{} model is a special case of a wider class of model composed of multiple oscillatory components in the sense that we only have one oscillatory component in the model. For details about a more general model, see [@Daubechies_Lu_Wu:2011; @Wu:2013; @Chen_Cheng_Wu:2014; @Kowalski_Meynard_Wu:2015]. ### Physiological interpretation The main reason we consider a time-varying frequency and amplitude model like (\[definition:IMT\]) is to capture the physiological facts [of]{} HRV and pulsus alternans. [Since]{} the heart does not beat at a constant rate [@Malik_Camm:1995], the pulse signal should not oscillate with a constant frequency. The non-constant heart rate is modeled by IF and hence the HRV could be evaluated from analyzing IF. In the past decade, due to the health information [embedding]{} in HRV and the trend [towards]{} personalized health care [requirements]{}, estimating the HRV from the pulse signal extracted from different resources has attracted [significant research]{} [@Gil_Orini_Bailon_Vergara_Mainardi_Laguna:2010; @Davila:2012Thesis; @Wu_Lewis_Davila_Daubechies_Porges:2015]. However, the existence of HRV is commonly ignored in the pulse analysis literature. For example, in the traditional spectral analysis approach to the radial pulse, [researchers need]{} to analyze the pulse signal over the interval where the heart rate [closely resembles]{} a constant [@LinWang_Jan_Shyu_Chiang_Wang:2004; @Hsu_Chao_Hsiu_Wang_Li_Wang:2006]. The possible pulsus alternans is captured by the AM. This model is particularly important in [patients susceptible to]{} heart failure [since]{} pulsus alternans is a manifestation of severe impairment of the left ventricular systolic function [@Morris:1990]. An example of a wave-shape function and [an]{} IMT function representing the pulse signal [are]{} shown in Figure \[fig:ShapeExample\]. ![Left: an illustrative wave-shape function $s$ with $D=5$, $\delta=0.59$ and $\theta=0$. Right: an IMT function with the wave-shape function $s$ and time varying amplitude and frequency. Here the frequency is about $1.2$ Hz.[]{data-label="fig:ShapeExample"}](Figure2){width=".75\textwidth"} ### Recorded pulse signal In practice, noise is inevitable [when recording]{} signals. Thus, we model the recorded pulse signal by $$\begin{aligned} \label{decompShape} Y(t)=A(t)s(\phi(t))+\sigma(t)\Phi(t),\end{aligned}$$ where $A(t)s(\phi(t))$ is the adaptive non-hamornic model for the pulse signal, $\sigma$ is a slowly varying smooth function quantifying the possible non-stationarity in the data collection process, and $\Phi$ is a stationary random process with unit standard deviation describing the noise. See, for example, [@Chen_Cheng_Wu:2014] for a discussion of the noise. Here, we do not assume that the noise is Gaussian or white, while the commonly encountered Gaussian white noise is a specific example. Since in general the noise might be complicated with time-dependence and [be]{} far from Gaussian, we take this noise model into account. While the pulse signal could be recorded by different ways, like tonometer, photoplethymography, video [@Davila:2012Thesis], etc, the wave-shape function could be different for different types of pulse signals. Methodology {#Section:MaterialandMethod} =========== Synchrosqueezing transform -------------------------- It is a common practice to apply [the]{} Fourier transform to study oscillatory signals, in particular the pulse wave signal [in which we are interested]{} [@Wei_Lee_Chow:1984; @Hsu_Chao_Hsiu_Wang_Li_Wang:2006; @Huang_Wei_Liao_Chang_Kao_Li:2011]. As useful as the spectral analysis is, however, it [is]{} well known that when the signal is not composed of harmonic functions, the power spectrum determined by the Fourier transform might be misleading. [For]{} spectral pulse wave analysis, since HRV and pulsus alternans are inevitable, in practice analysts [must]{} carefully choose the signal. [But]{} there is [still]{} no guarantee that the HRV and pulsus alternans influence could be eliminated, and [this could]{} become a confounder in the analysis. While the oscillatory signals with time-varying AM and IF are ubiquitous, [much effort has been extended over]{} the past few decades to address this problem, [and several approaches [have been]{} proposed, ranging from empirical mode decomposition and its variations [@Huang_Shen_Long_Wu_Shih_Zheng_Yen_Tung_Liu:1998; @Tavallali_Hou_Shi:2014; @Cicone_Liu_Zhou:2014], [to]{} optimized window and the approximation theory approach [@Ricaud_Stempfel_Torresani:2014; @Chui_Mhaskar:2016], to name but a few. [Moreover, research interest]{} has [recently]{} been extended to multivariate time series analysis [@Gao_Jin:2012; @Gao_Yang_Fang_Jin_Xia_Hu:2015; @Gao_Fang_Ding_Jin:2015; @Ahrabian_Looney_Stankovic_Mandic:2015; @Gao_Yang_Zhai_Ding_Jin:2016] to further take the spatial information into account. Among [the]{} different approaches, one]{} active field regarding this issue is time-frequency (TF) analysis. Based on different [principles]{}, several TF analysis techniques have been proposed. Typical examples include linear [methods such as]{} the short time Fourier transform (STFT) and [the]{} continuous wavelet transform (CWT)[, or]{} the quadratic [methods such as]{} the Wigner-Ville distribution [or the Cohen class. We refer the reader to [a]{} standard textbook [@Flandrin:1999] for more information]{}. While these methods have attracted a [great deal]{} of attention in different fields other than signal processing, they are limited by [either]{} the Heisenberg [uncertainty]{} principal or the mixing issue. [We demonstrate this idea by discussing STFT]{}. The main idea [forming the basis for]{} STFT is dividing the signal into overlapping small pieces, and [then studying]{} the spectral behaviors over these small pieces. Mathematically, for a function $f$, its STFT associated with a window function $h(t)$ can be defined as $$V_f^{(h)}(t,\eta):=\int f(s)h(t-s)e^{-i2\pi \eta (t-s)}{\textup{d}}s$$ where $t\in{\mathbb{R}}$ is the time, $\eta\in {\mathbb{R}}^+$ is the frequency, $h$ is the window function determined by the user – [for which]{} a common choice is the Gaussian function with kernel bandwidth $\sigma>0$, i.e. $h(t)=(2\pi\sigma)^{-1/2}e^{-t^2/\sigma^2}$. However, the Heisenberg uncertainty principal limits how well the spectrum [could be estimated]{} over each piece, [thereby limiting]{} the STFT. Similar discussions hold for CWT and other [linear]{} TF analysis techniques, and we refer the [reader]{} to [@Flandrin:1999] for details. [In the current work, in order to capture the hemodynamics, which have oscillations on the order of 1 second, we run STFT with $\sigma=0.5$ so that the window is not too long to lose the dynamic information, and is not too short to cause a numerical artifact.]{} To handle these fundamental issues, [a nonlinear TF analysis technique, synchrosqueezing transform (SST) [@Daubechies_Lu_Wu:2011; @Chen_Cheng_Wu:2014], which is a special reassignment technique (RM) [@Auger_Flandrin:1995], is proposed to obtain a sharper time-frequency representation.]{} Since [these techniques]{} were introduced, they have been widely applied in different fields; see [@Daubechies_Wang_Wu:2016] for a summary of its applications in different fields. Here we [briefly]{} summarize SST, and refer the reader to [@Auger_Flandrin:1995] for RM. [We mention that the SST could be applied to different linear TF analysis, like STFT, CWT, wave packet transform [@Yang:2014] or S-transform [@Huang_Zhang_Zhao_Sun:2015] and theoretically the results do not depend [significantly]{} on the chosen linear TF analysis. But to simplify the discussion, we choose to do the analysis with STFT.]{} The Matlab code for the SST algorithm based on both CWT and STFT [can]{} be downloaded from <https://sites.google.com/site/hautiengwu/home/download>. In [brief]{}, the SST sharpens the TF representation determined by STFT by reallocating the STFT coefficients along the frequency axis to the “correct” frequency slot which represents the IF of an oscillatory component. Mathematically, The SST of $f$ is defined as $$S_f^{(h)}(t,\xi):=\lim_{\alpha \rightarrow 0} \int V_f^{(h)}(t,\eta)\, g_\alpha(|\xi-\omega^{(h)}_f(t,\eta)|)\, {\textup{d}}\eta,$$ where $g_\alpha$ is an approximate $\delta$-function in the sense that $g$ is [a]{} fast decaying smooth function with $\int g(x){\textup{d}}x =1$, so that $g_\alpha(t):=\frac{1}{\alpha}g(\frac{t}{\alpha})$ tends to the Dirac delta measure $\delta$ supported at $0$ weakly as $\alpha\to 0$, and with $\omega^{(h)}_f$ defined by $$\omega^{(h)}_f(t,\eta):= \left\{ \begin{array}{ll} \frac{-i\partial_t V_f^{h}(t,\eta)}{2\pi V_f^{(h)}(t,\eta)} & \mbox{ if }\, V_f^{(h)}(t,\eta)\neq 0\\ -\infty & \mbox{ otherwise}\nonumber. \end{array}\right.$$ In plain language, by reading $\omega^{(h)}_f(t,\eta)$, we collect all STFT coefficients indicating the existence of an oscillatory component with frequency $\xi$ to the slot $\xi$. Note that [compared with RM,]{} in SST, the coefficients are reallocated along the frequency axis, so the causality is preserved; second, in SST we reallocate the STFT coefficient instead of the spectrogram coefficient. These two facts allow us to reconstruct the oscillatory components [of]{} interest, in particular when the signal is noisy or is composed of several oscillatory components. It has been [clearly found]{} that for the phenomenological model of the pulse wave signal [of]{} interest, the IF $\phi'(t)$ and the AM $A(t)$ [can be accurately estimated]{} from the recorded pulse signal [@Daubechies_Lu_Wu:2011; @Chen_Cheng_Wu:2014]. Precisely, [we]{} could prove that at time $t$, the coefficients in $S_f^{(h)}(t,\xi)$ are dominant when $\xi\approx\phi'(t)$. This property allows us an accurate estimate of the IF $\phi'$ by, for example, the curve extraction technique. Denote the estimated $\phi'$ by $\widetilde{\phi}'$. We can then estimate the amplitude modulation $A(t)$ and the phase function $\phi(t)$ by building $$\begin{aligned} & \widetilde{R}(t):=h(0)^{-1}\int_{\{\xi:~|\widetilde{\phi}'(t)-\xi|\leq \epsilon^{1/3}\}}S_f^{(h)}(t,\xi){\textup{d}}\xi. \label{alogithm:sst:reconstruction}\end{aligned}$$ The estimator of $A(t)$ is thus defined as $\widetilde{A}(t):=|\widetilde{R}(t)|$, and hence an estimator for $\phi(t)$, denoted as $\widetilde{\phi}(t)$, can be obtained by unwrapping the phase of the complex-valued signal $\widetilde{R}(t)/\widetilde{A}(t)$. We refer [readers interested]{} in SST to [@Daubechies_Lu_Wu:2011; @Chen_Cheng_Wu:2014] for the [detailed]{} numerical [algorithms and]{} the theory beyond [them]{}. Here, we show the results of pulse analysis by applying the SST in Figure \[fig:SSTExample\]. To demonstrate its benefit on [a]{} noisy signal, we artificially add noise on the second half of the signal, and show the result in Figure \[fig:SSTExampleNoisy\]. In this example, the noise is an ARMA(1,1), where ARMA stands for autoregressive and moving averaging, time series determined by the autoregression polynomial $a(z)=0.5z+1$ and the moving averaging polynomial $b(z)=-0.3z+1$, with the innovation process taken as i.i.d. student $t_3$ random variables so that the signal to noise ratio, defined as $20\log\frac{\text{std}(\text{signal})}{\text{std}(\text{noise})}$, is $0$ dB. It is clear that [though]{} the signal to noise is low, the noise is non-stationary since it exists only over a finite interval, and the noise has the fat tail behavior described by the student $t_3$ random variables. [Thus,]{} the SST algorithm could reliably extract the instantaneous frequency, so that we could obtain reliable HRV information. ![(a) The radial pulse signal $R(t)$ recorded from a normal subject. It is also clear that the IF and amplitude modulation are not constant due to the inevitable HRV and pulsus alternans. The peaks are marked by gray circles, and the lengths between two consecutive peaks are shown. Note that there are two peaks around the second second, and we [choose]{} the one with the larger value as the peak. (b) The time-frequency (TF) representation of $R(t)$ determined by the synchrosqueezing transform. The dominant curve in the TF representation is associated with the IF induced by the heart rate variability (HRV). It is clear that the pulse around [the]{} $5^{th}$ second takes a longer time to finish, which leads to the slower instantaneous frequency. Indeed, the y-axis of the dominant curve around time $5.3$-th second is $1/1.15=0.87$, which reflects how fast the signal oscillates at that moment. Note that the artifacts around [the]{} $5.5$ second and [the]{} $7.5$ second do not play a major role in the analysis [results]{}.[]{data-label="fig:SSTExample"}](Fig3arxiv.pdf){width=".6\textwidth"} ![(a) The radial pulse signal $R(t)$ recorded from a normal subject [that is]{} contaminated by the autoregressive and moving averaged noise generated from the student $t_3$ i.i.d. random process from the $2.5$-th second to the $5.5$-th second, which is denoted as $Y(t)$. The radial pulse signal is the same as that shown in Figure \[fig:SSTExample\]. Note that there are several spikes in $Y(t)$, which are generated [by]{} the fat tail natural of the student t 3 random variable. (b) The time-frequency (TF) representation of $Y(t)$ [is]{} determined by the synchrosqueezing transform (SST). The dominant curve in the TF representation is associated with the IF induced by the heart rate variability (HRV). It is clear that the TF representation determined by SST is robust to the noise even if the noise is time-dependent with [a]{} fat tail distribution, while the dominant curve representing the instantaneous frequency is slightly deformed due to the noise.[]{data-label="fig:SSTExampleNoisy"}](Fig4arxiv.pdf){width=".6\textwidth"} Feature extraction by estimation of the wave shape functions ------------------------------------------------------------ As discussed in Section \[Section:AdaptiveHarmonicModel\], the main feature [of]{} interest [for]{} the pulse analysis is the wave-shape function in the adaptive non-harmonic model, in particular in the pulse spectral analysis. In this section, we describe an algorithm to extract the wave shape function. Note that [since]{} the wave-shape function can be expanded by its Fourier coefficients, the pulse signal $f(t)=A(t)s(\phi(t))$ can be represented as $$\label{IMTexpansionVersion1} f(t)=A(t)\sum_{\ell\in{\mathbb{Z}}}\left[\alpha_{\ell}\cos(2\pi \ell\phi(t))+\beta_{\ell}\sin(2\pi\ell\phi(t))\right],$$ where $\alpha_{\ell}\in{\mathbb{R}}$ and $\beta_{\ell}\in{\mathbb{R}}$ are the Fourier coefficients of the shape function $s$. In this study, we estimate the wave shape functions based on (\[IMTexpansionVersion1\]) using the standard functional regression [@Ramsay_Silverman:1997; @Chui_Lin_Wu:2016]. Consider the wave shape functions $s$ with parameters $\delta, D,\theta$ in (\[decompShape\]). To simplify our discussion, we assume that $\theta=0$ and that the noise is stationary, that is $\sigma=1$. We would choose $D=6$ in this study, based on the discussion in Section \[Section:methodology\]. Thus, the pulse signal (\[IMTexpansionVersion1\]) becomes $$\label{IMTexpansionVersion2} f(t)=\, A(t)\alpha_0+A(t)\sum_{\ell=1}^{D} \left[\alpha_{\ell}\cos(2\pi \ell\phi(t))+\beta_{\ell}\sin(2\pi\ell\phi(t))\right]\nonumber.$$ After discretization by the sampling period $\Delta t>0$ over time interval $[\Delta t, N\Delta t]$, the recorded pulse signal is saved digitally as a $N$-dim vector, ${\boldsymbol{Y}}\in{\mathbb{R}}^N$, so that its $l$-th entry is ${\boldsymbol{Y}}_l=f(l\Delta t)+\Phi_l$, where $l=1,\ldots, N$ and $\Phi$ is a random vector satisfying $\textup{var}(\Phi_l)=1$ for all $l$, which might not be Gaussian and the covariant matrix might not be the identity. Denote the discretized estimators of $A(t)$ and $\phi(t)$ by $\widetilde{A}\in{\mathbb{R}}^{N}$ and $\widetilde{\phi}\in{\mathbb{R}}^{N}$. Note that it has been [well established]{} the discretized estimation of $A$ and $\phi$ are accurate with error of order $\epsilon$ [@Chen_Cheng_Wu:2014]. To [simplify]{} the discussion, [we]{} assume that the estimates $\widetilde{A}$ and $\widetilde{\phi}$ are precise without error; that is, $\widetilde{A}(l)=A(l\Delta t)$ and $\widetilde{\phi}(l)=\phi(l\Delta t)$ for all $l=1,\ldots,N$. In the general case, the analysis result will [deviate]{} by an error of order of $\epsilon$. We thus construct the following “functional vectors” $$\boldsymbol{c}=[c_0^T,c_{1}^T,\ldots,c_{D}^T,d_{1}^T,\ldots,d_{D}^T]^T\in {\mathbb{R}}^{2D\times N},$$ where $\ell=0,\ldots,D$ and $c_{\ell}, d_{\ell}$ are $N$-dim vector whose $k$-th entries are $$c_{\ell}(k):=\widetilde{A}(k)\cos(2\pi\ell\widetilde{\phi}(k)),\quad d_{\ell}(k):=\widetilde{A}(k)\sin(2\pi\ell\widetilde{\phi}(k)),\nonumber$$ where $k=1,\ldots, N$. As a result, the recorded pulse signal satisfies $$\label{model:with_shape_and_trend:regression_model} {\boldsymbol{Y}}=\boldsymbol{\gamma}^T\boldsymbol{c}+\Phi$$ where $\boldsymbol{\gamma}=[\alpha_0,\alpha_{1},\ldots,\alpha_{D},\beta_{1},\ldots,\beta_{D}]\in{\mathbb{R}}^{2D+1}$. To estimate the parameters $\alpha_{\ell}$ and $\beta_{\ell}$ from the functional vectors $\boldsymbol{c}$, observe that the $(2D+1)\times (2D+1)$ matrix $\boldsymbol{c}\boldsymbol{c}^T$ is diagonal dominant. Since $\mathbb{E}(\Delta t\Phi\boldsymbol{c}^T)=0$ and $\mbox{var}(\Delta t\Phi\boldsymbol{c}^T)=O(L\Delta t)$. Thus we can estimate $\boldsymbol{\gamma}$ by $$\label{estimator_shapefunction} \widetilde{\boldsymbol{\gamma}}:=({\boldsymbol{Y}}\boldsymbol{c}^T)(\boldsymbol{c}\boldsymbol{c}^T)^{-1},$$ where $\widetilde{\boldsymbol{\gamma}}=[\widetilde{\alpha}_{0},\widetilde{\alpha}_{1},\ldots,\widetilde{\alpha}_{D},\widetilde{\beta}_{1},\ldots,\widetilde{\beta}_{D}]^T\in{\mathbb{R}}^{2D+1}$, and hence estimate the oscillatory signal by $$\widetilde{f}(t):=\widetilde{\boldsymbol{\gamma}}^T\boldsymbol{c}.$$ [When]{} $\widetilde{f}$ contains an accurate estimation of the wave-shape function $s$ with $L^2$ error of order $\epsilon$, we could take the Fourier coefficients $\widetilde{\boldsymbol{\gamma}}$ into account as the feature of the pulse signal. We call the $2D+1$ dimensional vector $\widetilde{\boldsymbol{\gamma}}$ the [*spectral pulse signature (SPS)*]{} for the recorded pulse signal. Note that the $\ell$-th component of the power spectrum of $s$ could be estimated by $(\widetilde{\alpha}^2_{\ell}+\widetilde{\beta}^2_{\ell})/4$. The main benefit of this approach is that the influence of HRV, which is modeled by IF, is eliminated automatically and the wave-shape function is better estimated. [Comparing the present method to previous ones]{} ------------------------------------------------- Here we summarize the difference between our approach and the commonly applied spectral analysis [@Taylor:1966; @Wei_Chow:1985; @LinWang_Jan_Shyu_Chiang_Wang:2004]. Recall that in spectral analysis, the power spectrum of a selected recorded pulse signal is evaluated and the energy of different harmonic modes are considered [as]{} features of the subject, [indicating aspects of]{} physiological health. The selection criteria for the pulse signal interval is that the heart rate is almost a constant [@LinWang_Jan_Shyu_Chiang_Wang:2004; @Hsu_Chao_Hsiu_Wang_Li_Wang:2006][, since the power spectrum approach is sensitive to the non-constant frequency and non-constant amplitude signal]{}. The first difference is that based on the adaptive non-harmonic model and SST, the instantaneous frequency and amplitude modulation [can be]{} accurately [obtained]{}, so we do not need to select an interval from a recorded pulse signal. [Second, since the HRV and pulsus alternans are physiologically]{} inevitable, [the proposed]{} approach is more physiologically feasible. Third, note that the information obtained from the spectral analysis is different from the SPS. Indeed, under the assumption for (\[IMTexpansionVersion2\]), we have the following direct expansion: $$s(t)=\sum_{\ell=-D}^D\hat{s}(\ell)e^{i2\pi \ell t}=\alpha_0+\sum_{\ell=1}^D\left[\alpha_\ell\cos(2\pi \ell t)+\beta_\ell\sin(2\pi \ell t)\right],$$ where $\alpha_0=\hat{s}(0)$ and for $\ell=1,\ldots,D$, $\hat{s}(\ell)=a_\ell e^{i\theta_\ell}$, $a_\ell\geq 0$, $\theta_\ell\in[0, 2\pi)$, $\alpha_\ell=2a_\ell \cos(\theta_\ell)$ and $\beta_\ell=-2a_\ell\sin\theta_\ell$. Here, $\theta_\ell$ reflects the phase of the $\ell$-th harmonic component hidden inside the wave shape signal. Clearly, since power [spectrum]{} analysis [commonly takes]{} $|\hat{s}(\ell)|^2=a_\ell^2$, the phase information of the $\ell$-th harmonic component is missed. [However, in SPS the phase information is preserved and the hemodynamics could be more faithfully captured.]{} Testbed: the radial pulse diagnosis {#Section:materialresult} =================================== To test how well the proposed adaptive non-harmonic model and the algorithm work in practice, in this section we study the radial pulse wave signal on congestive heart failure subjects. The signals are recorded from three landmarks. First, the radial styloid process; second the middle position between the radial styloid process and the palm and the proximal point with the same distance from the first to the second location. We mention that these three locations are commonly recognized in the pulse diagnosis; the first one is called [*guan*]{}, the second one is called [*chun*]{} and the third one is called [*chi*]{}. [We]{} thus use guan, chun and chi to refer to these three locations in this study. [^3] Material -------- All protocols in this study [were]{} approved by the Institutional Review Board of Chang Gung Memorial Hospital, Linkou, Taoyuan, Taiwan (93-6288), Taiwan and written informed consent was obtained from [all]{} patients. Nineteen normal subjects [without history of]{} heart disease are included in the control group, and $17$ subjects with congestive heart failure (CHF) are included in the study group. The diagnosis of CHF subjects is based on the criteria indicated by the Framingham heart study. The participants were invited for pulse examination in a quiet and temperature-controlled room in Chang Gung Memorial Hospital, Linkou branch, Taiwan. Pulse wave signals were recorded from chun, guan and chi positions of both hands by a tonometer (Wang’s sphygmometer, PDS-2000). The sampling rate of the signal is at $100$Hz. For each subject, we collect $10$ seconds signal for each position on both [hands]{}, and repeat $2$ or $3$ times. The pulse wave was recorded in sitting position with the wrist comfortably [resting]{} on a small pillow at the level of [the]{} heart We recruited $17$ patients with CHF for the study group and $19$ normal individuals for the controls. The age of the study and control group are $64.3\pm 23.7$ and $63.2\pm 15.8$ respectively. The male/female ratio were $15/7$ in the study group and $10/10$ in the control group. There were no significant [differences]{} in age [or]{} sex. As a result, we obtain $53$ (respectively $39$) pulse wave signals recorded from chun from the normal (respectively CHF) group, $55$ (respectively $34$) pulse wave signals recorded from guan from the normal (respectively CHF) group and $50$ (respectively $41$) pulse wave signals recorded from chi from the normal (respectively CHF) group. Statistical Analysis and Global pulse signature (GPS) ----------------------------------------------------- To test if the SPS indices of the normal group [differ from]{} those of the CHF group, we apply the currently developed one-way ANOVA for functional data, called the globalized pointwise F (GPF) test [@Zhang_Liang:2013]. For readers having interest in this technique, we refer them to [@Zhang_Liang:2013]. In this study, we [consider]{} p values less than $0.01$ as significant. As the SPS index is a high dimensional vector, it is not easy to visualize. To provide an easy-to-use index for the pulse diagnosis, we consider the following approach to integrate the information in the SPS index. Suppose from a fixed position of the $i$-th subject we obtain a SPS index $\widehat{\boldsymbol{\gamma}}_i$. The associated outcome of this SPS index is denoted as $y_i$, and $y_i=1$ (respectively $y_i=0$) means the subject is in the CHF group (respectively control group). Thus we have the dataset $\{\widehat{\boldsymbol{\gamma}}_i,y_i\}_{i=1}^N$. [When]{} the SPS index is located in the $2D+1$ Euclidean space and the sampling size is limited, we apply the partial least squares (PLS) regression to find a linear regression model by projecting SPS and the response variables to a new space. Here we briefly recall the PLS regression. PLS regression finds components from $\mathcal{X}=\{\widehat{\boldsymbol{\gamma}}_i\}_{i=1}^N$ that are relevant for the outcome $\mathcal{Y}=\{y_i\}_{i=1}^N$ by seeking a set of components that performs a simultaneous decomposition $\mathcal{X}$ and $\mathcal{Y}$ with the constraint that these components explain as much as possible of the covariance between $\mathcal{X}$ and $\mathcal{Y}$. Then the decomposition of $\mathcal{X}$ is applied to predict the group. For details about PLS regression, see [@Rosipal_Kr:2006]. Suppose the PLS regression coefficient is $\beta\in {\mathbb{R}}^{(2D+1)\times 1}$. Then the prediction result under the linear regression model for $\widehat{\boldsymbol{\gamma}}_i$, denoted as $\hat{y}_i:=[1 \,\,\,\widehat{\boldsymbol{\gamma}}_i]\beta\in{\mathbb{R}}$, is referred to [as]{} the [*global pulse signature (GPS)*]{} index. The GPS index is integrated information derived from the SPS index, which reflects the subject’s condition [of]{} interest. For the [purpose of prediction]{} based on GPS, we [can]{} further apply the receiver operating characteristic (ROC) to determine the threshold to classify the subjects into two groups. We report the sensitivity, specificity, accuracy and the area under curve (AUC) to evaluate the classification result. The confidence interval (CI) of AUC is evaluated by $1000$ bootstrap replicas. To assess how the results based on PLS and ROC will generalize to an independent data set, we run leave-one-out cross validation (LOOCV) $200$ times for each position of both hands, and report the accuracy. Results ------- First, we show the synchrosqueezing transform of the pulse wave signal from a subject with CHF and the associated estimated wave shape function in Figure \[fig:ShapeExtractionExample\]. Note that due to the inevitable deviation, for example the one at the $4$-th second, and the HRV, the power spectrum estimated from the pulse wave signal is spreading. Note that estimating the wave shape function could be viewed as the power spectrum analysis of the pulse wave signal after correcting the IF and AM. In other words, as we could estimate IF and AM accurately from the pulse wave signal, we could resample the signal according to the estimated IF and then normalize the signal by the estimated AM. ![(a) The pulse signal recorded from a subject with congestive heart failure (CHF), [with the]{} y-axis [as]{} the arbitrary unit. (b) [The]{} estimated wave shape function for the subject with CHF. (c) [The]{} time-frequency (TF) representation determined by the synchrosqueezing transform. It is clear that the pulse around [the]{} $2.5^{th}$ second takes a longer time to finish, which leads to the slower instantaneous frequency (the dominant curve in the TF representation is around 0.9 Hz at time$ 2.5^{th}$ second). It is also clear that the instantaneous frequency is not constant due to the inevitable heart rate variability (HRV). Note that while there is a significant deviation at the $2.5$-th second, the estimation result catches most of the shape information. (d) [The]{} power spectrum of the estimated wave shape function shown in (b). []{data-label="fig:ShapeExtractionExample"}](Figure5.pdf){width=".78\textwidth"} The set of SPS indices of different groups from different positions [is]{} shown in Figure \[fig:SPSErrorBar\]. We could see that the means of the normal group and the CHF group are different. The GPF functional ANOVA test shows that the SPS indices evaluated from all positions on both hands are significantly different. Except [for]{} the p value of [the]{} chun position on the right hand, which is $4.4\times 10^{-4}$, the p values of other positions are $<10^{-4}$. ![The deviation from the mean of each entry of the spectral pulse signature (SPS) is shown in the error bar plot. The error bar is determined by the standard deviation. In the top row, the distribution of SPS determined from guan, chun and chi [positions]{} of the left hand are shown from left to right. In the bottom row, the distribution of SPS determined from guan, chun and chi of the right hand are shown from left to right. The results of the normal subjects are shown in the gray error bar, and those of the subjects with congestive heart failure are shown in the black error bar. []{data-label="fig:SPSErrorBar"}](Figure6.png){width=".95\textwidth"} Then we apply PLS to obtain the GPS index to distinguish the two groups. The histogram of GPS indices determined from different positions on both hands are shown in Figure \[fig:GPS\]. It is clear that the GPS of subjects in the CHF group is smaller than that of the control group. The ROC analysis results, including the sensitivity and specificity and AUC, from different positions are shown in Figure \[fig:ROC\]. The sensitivity, specificity, accuracy, AUC and the accuracy of LOOCV of the GPS determined from different positions on both hands are summarized in Table \[tab:allROC\]. ![The histograms of GPSs determined from chun, guan and chi positions on the left (respectively right) hand are shown on the left, middle and right subfigures on the top (respectively bottom) row. The normal group is shown in the gray color and the congestive heart failure group is shown in black. [Gray]{} stars represent the determined thresholds by the ROC binary classification.[]{data-label="fig:GPS"}](Figure7.png){width=".9\textwidth"} ![Left (respectively middle and right): the receiver operating curve (ROC) of the global pulse signature (GPS) of the chun (respectively guan and chi) position on the left hand is shown in the black curve and that of the right hand is shown in the gray hand. [Black]{} and gray circles are the optimal operating point of the ROC.[]{data-label="fig:ROC"}](ROC2.png "fig:"){width=".3\textwidth"} ![Left (respectively middle and right): the receiver operating curve (ROC) of the global pulse signature (GPS) of the chun (respectively guan and chi) position on the left hand is shown in the black curve and that of the right hand is shown in the gray hand. [Black]{} and gray circles are the optimal operating point of the ROC.[]{data-label="fig:ROC"}](ROC1.png "fig:"){width=".3\textwidth"} ![Left (respectively middle and right): the receiver operating curve (ROC) of the global pulse signature (GPS) of the chun (respectively guan and chi) position on the left hand is shown in the black curve and that of the right hand is shown in the gray hand. [Black]{} and gray circles are the optimal operating point of the ROC.[]{data-label="fig:ROC"}](ROC3.png "fig:"){width=".3\textwidth"} ------ ------ ------ ------ ------------------ ------- SEN SPE AC AUC LOOCV chun 0.87 0.9 0.88 0.92 (0.81-0.96) 0.8 guan 0.89 0.79 0.85 0.91 (0.85-0.96) 0.74 chi 0.94 0.78 0.87 0.89 (0.8-0.95) 0.74 SEN SPE AC AUC LOOCV chun 0.87 0.79 0.84 0.89 (0.79-0.95) 0.74 guan 0.94 0.73 0.85 0.91 (0.83-0.95) 0.79 chi 0.94 1 0.96 1 (0.97-1) 0.94 ------ ------ ------ ------ ------------------ ------- : [Sensitivity]{} (SEN), specificity (SPE), accuracy (AC), the area under curve (AUC) and the accuracy of the leave-one-out cross validation (LOOCV) of the global pulse signature evaluated from different positions on both hands. The confidence [intervals]{} of the AUC [are]{} reported in [parentheses]{}.[]{data-label="tab:allROC"} Discussion {#Section:discussion} ========== In summary, in this paper we study the physiological signal by the adaptive non-harmonic model, the SST and the functional regression technique. The usefulness of the proposed scheme is supported by an encouraging analysis result of the radial pulse signal. Although we analyze the radial pulse signal as the test [case for]{} this study, [it should be noted]{} that the proposed model and analysis technique could be applied to other pulse signals obtained by different instruments. For example, contact photoplethysmogram (PPG) measurement [@Allen:2007] or PhysioCam non-contact PPG measurement [@Davila:2012Thesis], which represent the changes of blood volume in the vessel obtained through an optical transmission measurement or real-time camera images, and hence reflect the change in vascular blood volume associated with the cardiac beat. While these signals represent different [aspects]{} of the hemodynamics [than]{} the radial pulse signal we study in this paper, we could expect to obtain a broader angle of view about human health if information obtained from these signals [could be combined]{}. We [will]{} report the [research]{} progress in the future work. Since the potential of the SST and other nonlinear TF analysis techniques have been shown in this study and other clinical problems, for example, [@Lin_Wu_Tsao_Yien_Hseu:2014; @Baudin_Wu_Bordessoule_Beck_Jouvet_Frasch_Emeriaud:2014; @Wu_Talmon_Lo:2015], by taking the wave-shape model into account, we could expect a broader application and better analysis [results]{}. We [also point out]{} the relationship of the pulse signal analysis results with the [traditional Chinese medicine]{} (TCM) theory. In short, the results we obtained in this study [could]{} lead to a potential [means]{} to help establish the foundation of TCM theory. See Figure \[fig:ChunGuanChi\] for an illustration of the summary of the pulse diagnosis theory in the TCM theory. ![An illustration of anatomical locations associated with the terminologies used in pulse diagnosis. In the TCM pulse diagnosis, it is stated that the pulse on the right hand manifests the condition of Qi [@Leung:2011; @Wiseman_Ye:2014], while the pulse on the left hand reflects the blood. For different pulse positions, left chun, guan and chi reflect the heart, the liver, and the kidney respectively; the right chun, guan and chi reflect the lung, the spleen, the kidney respectively. The left chi manifests mainly the kidney yin, and the right chi manifests mainly the kidney yang (the Life Gate in TCM). []{data-label="fig:ChunGuanChi"}](Figure9.jpg){width=".6\textwidth"} Note that while TCM has been widely applied in the eastern culture, up to now, a systematical/scientific theory understanding the mechanism beyond pulse diagnosis is not yet well established [@Leung:2011], but its usefulness [has been demonstrated]{} [@Oparil_Joseph:2006]. There are several [studies]{} based on hemodynamics aiming to understand the mechanism; for example, Wang et. al. [@Wang_Hsu_Jan_Wang:2010; @Wang_Wang_Jan_Wang:2012] proposed that the pulse wave consists of numerous harmonic waves, and each harmonic wave is associated with an internal organ and carries information of different meridians over the body. [That study also]{} noted that the harmonic waves of the upper, middle and lower section of the body may correspond to the three sections of pulse in timeline and that the waveform on both hands are not totally the same. On the other hand, our approach is purely from the phenomenological viewpoint and adaptive harmonic analysis. In our results, the classification result of the CHF group and the normal group is better on the right hand side, especially the right chi. Since the right chi position [demonstrates]{} the most significant difference between groups, it may reveal the importance of [the]{} right chi (kidney yang) signal in evaluating the clinical condition of CHF. According to TCM theory, kidney yang is the fundamental support to maintain the human life. In [Western]{} medicine, the hemodynamics of renal artery, such as the blood pressure, plays an important role [in regulating the overall]{} whole cardiovascular function. On the other hand, a possible cause for this observation is that the pulse pressure of the right arm is usually higher than the pulse pressure of the left arm. It is presumed that the pulse wave signal of the right arm may have a higher signal-to-noise ratio than the pulse wave signal of the left arm. The left chun and right guan position are the two second important roles in our analysis. According to TCM theory, [the]{} left chun position (heart) manifests the general condition of the cardiovascular system, hence our results [are consistent with]{} clinical experience. Thus, our [findings]{} partially support the pulse diagnosis theory in TCM that the waveforms on different positions of radial artery contain different information. [A]{} similar finding [was]{} reported [by]{} Young et. al., [who]{} found that the although the augmentation indices determined from the radial pulse waves recorded from chun, guan and chi are not significantly different, the estimated aortic augmentation indices determined from the radial pulse waves recorded from chun, guan and chi are not identical [@Jeon_Kim_Lee_Lee_Ryu_Lee_Kim:2011]. In conclusion, since the study of pulse diagnosis from different aspects is an active field, we would expect our proposed model and method could help to further study the experiences and working practice [of]{} pulse diagnosis, e.g. the nature and dynamic of disease, and its relationship to modern hemodynamics. Limitations of this study should also be mentioned. First, the tonometer (PDS-2000) we applied in this study records only the two-dimensional data (pressure-time) of the pulse. Since [more advanced]{} instruments [now available can]{} obtain three dimensional data [@Luo_Chung_Yeh_Si_Chang_Hu_Chu:2012], further study should be [undertaken]{}. Second, although the phenomenological model we propose is capable of capturing IF and AM as well as the wave shape function, it is clearly not the optimal solution. It is clear that there is a more [complex]{} interaction between IF, AM and the wave shape function than what we consider in the adaptive non-harmonic model. On the one hand, a more general model, like the time-varying wave shape function, could be considered based on the physiological background. On the other hand, we conjecture that this relationship might be better captured by combining the existing hemodynamic models with the proposed phenomenological model. This finer model might [better]{} capture the physiological information hidden inside the pulse wave signal and lead to a better algorithm to better study the recorded pulse wave signal. A systematic study and its application of this issue will be reported in future work. Moreover, from the clinical viewpoint, the sample size in this study is limited, and the interesting clinical problems, like outcome prediction or early CHF diagnosis are not discussed. Also, to simplify the discussion and avoid possible confounders, in this study we limit our analysis to subjects with CHF. Thus, we could not make the final conclusion about the pulse diagnosis. To use the research results in clinics, a larger scale clinical study with CHF and other diseases is needed to conclude the [current findings]{}, and we will report our continuing [research]{} in the future. Acknowledgement =============== H.-T. Wu also thanks Professor Jin-Ting Zhang for sharing the GPF code. Part of this work was done during H.-T. Wu’s visit to National Center for Theoretical Sciences, Taiwan, and he would like to thank NCTS for its hospitality. [^1]: The [*hyperkinetic pulse*]{} is characterized by an increase in the velocity of the upstroke and amplitude. The [*pulsus alternans*]{} is a beat-to-beat variation in the amplitude of the pressure pulse. The [*pulsus bisferiens*]{} arterial pulse is perceived as two narrowly separated positive waves during systole. The [*pulsus pervious et tardus*]{} is a small and delayed arterial pulse. [^2]: The augmentation index indicates the incidence of reflected waves on the total pulse pressure. See [@Salvi:2012 Section 6.1] for a definition. [^3]: In the literature sometimes chun is called [*inch*]{} or [*chon*]{}, guan is called [*bar*]{} or [*gwan*]{} and chi is called [*cubit*]{} or [*cheok*]{}.

--- abstract: | We study the contributions of the $C\!P$  violating anomalous $WW\gamma$ interactions to $b\rightarrow s \ell^+\ell^-$. We obtain cutoff independent results on $\tilde\kappa$ and $\tilde\lambda$, by constructing an asymmetry for the process $B\rightarrow K^{*}\ell^+\ell^{-}$, where the $B$ and $\bar B$ events are added. We show that a sample of $10^4$ $B\rightarrow K^{*}e^+e^{-}$ events can yield a bound, $|\tilde\kappa| < 0.42$ at 90% C.L., which is much tighter than the recent constraint from D0. address: ' Institute of Mathematical Sciences, Taramani, Chennai 600113, India.' author: - 'Nita Sinha and Rahul Sinha [^1] [@author]' date: 22 July 1997 title: '$C\!P$  violating anomalous trilinear gauge couplings from $B\to K^* \ell^+ \ell^-$' --- =8.5in c i ł Ł ø Ø ¶ § u 2[m\_[B\^\*]{}\^[2]{}]{} \#1[Phys. Rev. Lett. [**\#1**]{}]{} \#1[Phys. Rev. [**D\#1**]{}]{} \#1[Phys. Lett. [**B\#1**]{}]{} \#1[Nucl. Phys. [**B\#1**]{}]{} \#1[Mod. Phys. Lett. [**A\#1**]{}]{} \#1[Int. J. Mod. Phys. [**A\#1**]{}]{} \#1[Z. Phys. [**C\#1**]{}]{} Two of the aspects of the Standard Model(SM) which remain to be further explored are the gauge nature of the W, Z bosons and $C\!P$  violation. The vector boson self interactions are uniquely fixed by the $SU(2)_L\times U(1)$ gauge structure of the SM. However, new physics as well as higher order corrections will modify these self interactions. While tests of trilinear gauge couplings are now being done at present colliders, only weak bounds on them exist[@bounds; @CPV; @bounds; @CPC]. $C\!P$  violation on the other hand is only parameterized in the SM via the CKM matrix, and has been seen only in the neutral K mesons. Much effort is now being put into studying $C\!P$  violation in the $B$ mesons. Large number of $B$ mesons are expected to be produced in the near future, at the new asymmetric colliders and upgraded Tevatron, which would enable study of its rare decay modes. Rare decays of $B$ mesons, occur in the SM only through loops and are thus expected to be very sensitive to new physics. In particular the mode $b \rightarrow s \ell^+\ell^-$ has been extensively studied[@Hewett; @Wyler; @Burdman2] for possible contributions from new physics. New physics that affects the gauge nature of the W, Z bosons, can be parameterized in terms of $C\!P$  conserving as well as $C\!P$  violating anomalous trilinear gauge boson couplings. The $C\!P$  conserving couplings have been considered in the rare modes $b \rightarrow s\gamma$, $b \rightarrow s \ell^+\ell^-$[@He; @Drew; @Numata]. Since, the origin of $C\!P$  violation is not understood, any potential source of $C\!P$  violation should be pursued. Here we focus attention on the $C\!P$  violating anomalous couplings, by studying the mode $B \rightarrow K^{*} \ell^+\ell^-$. The couplings of W bosons to the photon and Z can be described by an effective Lagrangian, the most general form of which may be written down with the minimum requirements– that of Lorentz invariance, global $SU(2)$ and local $U(1)$ symmetry [@peccei] as: $$\begin{array}{rcl} \displaystyle {\cal L}_{\it eff}^{V}= &-& g_{V} \left[i g_1^V \left( W^\dagger_{\alpha \beta} W^\alpha - W^{\dagger\alpha} W_{\alpha \beta} \right) V^\beta + i \kappa_V W^\dagger_{\alpha} W_\beta V^{\alpha\beta}\right.\\[3ex] &+&\displaystyle \left . i \frac{\lambda_V}{M_W^2} W^\dagger_{\alpha \beta} {W^\beta}_\sigma V^{\sigma\alpha} + i \tilde{\kappa}_V W^\dagger_{\alpha} W_\beta \tilde{V}^{\alpha\beta} + i\frac{\tilde{\lambda}_V}{M_W^2} W^\dagger_{\alpha \beta} {W^\beta}_\sigma \tilde{V}^{\sigma\alpha} \right.\displaystyle\\[3ex] &+& \left. g_4^V W^\dagger_{\alpha} W_\beta (\partial^\alpha V^\beta+\partial^\beta V^\alpha) + g_5^V \epsilon^{\alpha\beta\mu\nu} W^\dagger_\alpha \stackrel{\leftrightarrow}{\partial_\mu} W_\beta V_\nu \right]. \label{lagrangian} \end{array}$$ In the above equation $V$ represents the neutral gauge bosons either the photon or the Z, $V_{\alpha\beta} = \partial_\alpha V_\beta - \partial_\beta V_\alpha $, $W_{\alpha\beta} = \partial_\alpha W_\beta- \partial_\beta W_\alpha $, $\tilde{V}^{\alpha\beta}=\frac{1}{2}\epsilon^{\alpha\beta\rho\sigma} V_{\rho\sigma}$ and $g_{V}$ is the $WWV$ coupling strength in the SM with $g_{\gamma}=e$ and $g_{Z}=ec/s$, where $c^2 \equiv 1 -s^2 \equiv M_W^{2}/M_Z^2$. In the SM, the couplings $g_1^V$, $\kappa_V$ are unity and all others are zero. New physics may result in a modification of these couplings. The deviation from the SM, the so called anomalous couplings, need to be constrained experimentally. We shall here concentrate on the $C\!P$  violating couplings $\tilde\kappa$ and $\tilde\lambda$. While many theoretical and experimental studies of anomalous triple gauge boson couplings have been done, few bounds exist on the anomalous $C\!P$  violating couplings. Direct bounds have recently been obtained[@bounds; @CPV] on $\tilde\kappa$ and $\tilde\lambda$ from $p\bar p\to W \gamma X$ at the tevatron. Stringent indirect constraints[@Marciano] on these couplings come from electric dipole moment (EDM) of the neutron and electron. However, these constraints assume naturalness and are cutoff dependent. Hence, bounds from $C\!P$  violating asymmetries that we pursue here, would be complementary to the EDM bounds. Possible contributions of the couplings $\tilde\kappa$ and $\tilde\lambda$ to processes at colliders have also been examined for LEPII[@peccei], upgraded tevatron[@Valencia] as well as future linear colliders[@Colliders]. In the chiral Lagrangian approach the dimension 6 operator with coefficient, $\tilde\lambda$, is ignored. For ‘$b$’ penguin modes such as $b\rightarrow s\gamma$ or $b\rightarrow s \ell^+\ell^-$, unitarity of the CKM, results in cutoff independence for contributions from $\tilde\kappa$ and $\tilde\lambda$. Since both $\tilde\kappa$ and $\tilde\lambda$ contributions are finite, requiring no cutoff, we do not neglect $\tilde\lambda$. However, the constraints on $\tilde\lambda$ are expected to be weaker. The effective short distance Hamiltonian relevant to the decay $b\to s\ell^+\ell^-$ [@Buras-Munz; @Grinstein; @ODonnell] leads to the QCD corrected matrix element, $${\cal M} (b\to s \ell^+\ell^-) =\frac{\alpha G_F}{\sqrt{2}\pi} v_t \{ -2 i C_7 m_b \frac{q^\nu}{q^2}\;\bar{s}\sigma_{\mu\nu}b_R\; \bar{\ell}\gamma^\mu\ell +C_8\;\bar{s}\gamma_\mu b_L\;\bar{\ell}\gamma^\mu\ell +C_9\;\bar{s}\gamma_\mu b_L\;\bar{\ell}\gamma^\mu\gamma_5\ell \}, \label{heff}$$ where only the dominant top quark contribution to the loop is retained. $C_{j}$ ($j=7,8,9$) are the Wilson coefficients given in Ref.[@Buras-Munz; @Grinstein], $m_b$ is the mass of the $b$ quark, $q^2$ is the invariant lepton mass squared, $b_{L,R}=(1\mp\gamma_5)/2\,b$ and $v_t=V^*_{ts}V_{tb}$ is the product of the CKM matrix elements. The anomalous WWV couplings, result in a shift in the values of the short distance coefficients $C_{j}$’s at $\mu=M_{W}$, $$C_{j} = C_{j}^{SM}+\imath\;\tilde\kappa C_{j}^{\tilde\kappa}+\imath\;\tilde\lambda C_{j}^{\tilde\lambda},$$ where, $C_{j}^{SM}$ are the SM coefficients and $C_{j}^{\tilde\kappa}$, $C_{j}^{\tilde\lambda}$ are the contributions from the anomalous couplings. These shifted values of $C_j$’s are then evolved down to the $b$ quark scale. The anomalous $WW\gamma$ coupling coefficients have been evaluated to be, $$\begin{aligned} C_{8}^{\tilde\kappa}=C_{8}^{\tilde\lambda}= \frac{m_b^2}{M_W^2}\,\frac{x_t}{12}\, \left( {{2 + 37\,{x_t} + 10\,{x_t^2} - {x_t^3}}\over {6\, {{\left( 1 - x_t \right) }^4}}} + {{{x_t}\,\left( 3 + 5\,{x_t} \right) \,\ln {x_t}}\over { {{\left( 1 - x_t \right) }^5}}} \right),\nonumber \end{aligned}$$ $$\begin{aligned} C_{7}^{\tilde\kappa}= \frac{ {x_t}}{2} \left({ { 1 }\over {{\left( 1 - x_t \right) }^2} } + {{{x_t}\,\left( 3 - {x_t} \right) \, \ln {x_t}}\over {2 {{\left( 1 - x_t \right) }^3} }} \right), \end{aligned}$$ $$\begin{aligned} C_{7}^{\tilde\lambda}= \frac{{x_t}}{2}\,\frac{\left( 1 - {x_t^2} + 2\,{x_t}\,\ln {x_t} \right) } {2\,{{\left( 1 - x_t \right) }^3}},\nonumber\end{aligned}$$ where, $x_t=\displaystyle\frac{m_t^2}{M_W^2}$, $m_t$ is the mass of the top quark. The WWZ coefficients can be related to the corresponding $WW\gamma$ coefficients listed above, by modifying the couplings and propagator. $C_{9}^{\tilde\kappa,\tilde\lambda}$ also contributes to the WWZ vertex, with a functional form in $x_t$, identical to that of $C_{8}^{\tilde\kappa,\tilde\lambda}$. However, the anomalous contributions from the WWZ vertex may be neglected in comparison to that of the photon. Our results for $C_{7}^{\tilde\kappa}$ are in agreement with those of refs.[@He] and [@Numata] but we disagree with ref.[@Numata] on $C_{7}^{\tilde\lambda}$ and on $C_{8}^{\tilde\lambda}$. $C_{8}^{\tilde\kappa}$ has been evaluated for the first time. Note that $C_{8}^{\tilde\kappa}$ turns out to be exactly equal to $C_{8}^{\tilde\lambda}$, perhaps indicating that for the contact term $C_8$, the derivatives from dimension 6 operator and the gauge part of the propagators contributing in case of the dimension 4 operator, behave similarly. Another interesting feature is that $C_8$ is doubly chirally suppressed for the $C\!P$  violating couplings $\tilde\kappa,\tilde\lambda$. Previous studies[@He; @Numata] of $C\!P$  violating anomalous gauge boson couplings in rare B decays, have used the mode $b\to s\gamma$, to constrain $\tilde\kappa$ and $\tilde\lambda$ separately, disallowing any possible cancelations between them. They also used only the decay rate of $b\to s\gamma$ and hence depended quadratically on these couplings. In order to achieve maximum sensitivity to $C\!P$  violating couplings and to have a clear signal of $C\!P$  violation, we need to estimate observables that are explicitly $C\!P$  odd and must therefore depend linearly on the $C\!P$  violating anomalous couplings. $C\!P$  violating asymmetries have also been considered in $b\to s\gamma$, in SM and two Higgs doublet model[@Soares; @Wolfenstein; @chinese]. However, such asymmetries rely on the presence of a large strong phase arising out of final state interactions. $C\!P$  violating asymmetries in rare modes that require flavour tagging are going to be very difficult to detect. We therefore prefer to use a technique that neither needs flavour nor time tagging. Such an asymmetry has been considered in ref.[@Sinha]. We adopt this technique and construct asymmetries for the mode $B \rightarrow K^{*}\ell^{+}\ell^{-}$ ($\ell^+\ell^-$ non resonant), to obtain simultaneous bounds on $\tilde\kappa$ and $\tilde\lambda$. The mode $B \rightarrow K^{*}\ell^+\ell^{-}$ is theoretically clean and can provide additional information due to its richer kinematics. We present here only the case of charged $B$’s. For neutral $B$’s, after time integration, the asymmetries that we construct, have exactly the same form as that for the charged $B$’s [@Sinha]. The transition matrix element for the exclusive process $B(p)\to K^*(k) \ell^+\ell^-\to K(k_1) \pi (k_2)\ell^+(q_1)\ell^-(q_2)$ can be written for each of the operators in Eq.(\[heff\]) as, $$\begin{aligned} \langle K\pi|\bar{s}i\sigma_{\mu\nu}(1\pm\gamma_5)q^\nu b|B\rangle &=& i{\cal A} \epsilon_{\mu\nu\alpha\beta} K^\nu k^\alpha q^\beta\pm {\cal B}K_\mu\pm {\cal C} k_\mu \, ,\nonumber \\ \langle K\pi|\bar{s}\gamma_\mu(1\mp\gamma_5)b|B\rangle &=& i{\cal D}\epsilon_{\mu\nu\alpha\beta} K^\nu k^\alpha q^\beta \pm {\cal E} K_\mu\pm {\cal F} k_\mu \,.\end{aligned}$$ The form-factors ${\cal A,\cdots,F}$ are unknown functions of $q^2=(p-k)^2$ and other dot products involving momentum, $k=k_1+k_2$ and $K=k_1-k_2$ and can be related to those used in [@Wyler], as given in the Table. The variable $\sigma$, in the table, arises due to the decay of $K^*\to K\pi$, evaluated in the zero width approximation. They can be similarly related to those in heavy quark effective theory (HQET)[@Burdman2]. The HQET form factors cannot currently be reliably predicted over the entire dilepton mass range. The results that we shall obtain, do depend on the numerical values of the form factors. In future, with large number of $B$’s available, the form factors can be determined experimentally. The current proportional to $q_\mu$ does not contribute as it couples to light leptons. In our notation $M_B,m_{K^*},m_K$ and $m_\pi$ are the masses of the $B$, $K^*$, $K$ mesons and the pion respectively. The matrix element for the process $B\to K^* \ell^+\ell^-\to K \pi \ell^+\ell^-$ can be written as $${\cal M}( B \displaystyle\to K\pi\ell^+\ell^-)=\frac{\alpha G_F}{\sqrt{2}\pi} \left\{\left(i \alpha_L\,\epsilon_{\mu\nu\alpha\beta}\, K^\nu k^\alpha q^\beta\right.\right. \displaystyle\left. \left.+ \beta_L\, K_\mu+ \rho_L\, k_\mu \right){\bar\ell}\gamma^\mu\,L\,\ell+ L\to R \right\}\,,$$ where $L,R=\displaystyle\frac{(1\mp\gamma_5)}{2}$, $q=q_1+q_2$ and $Q=q_1-q_2$, and the coefficients $\alpha_{L,R},~\beta_{L,R}$ and $\rho_{L,R}$ are given by $$\begin{array}{rclcrcl} \displaystyle\alpha_{R,L} &=&\displaystyle v_t\,\{\frac{(C_8\pm C_9)}{2}\,{\cal D} -\frac{m_b}{q^2}\,C_7\,{\cal A}\}&=&\displaystyle|{\rm a}_{R,L}|\,exp\left(i\delta^{\alpha}_{R,L}\right)\,exp \left(i\phi^{\alpha}_{R,L}\right) \\[2ex] \displaystyle\beta_{R,L}& =&\displaystyle v_t\,\{\frac{(C_8\pm C_9)}{2}\,{\cal E} -\frac{m_b}{q^2}\,C_7\,{\cal B}\}&=&\displaystyle|{\rm b}_{R,L}|\,exp\left(i\delta^{\beta}_{R,L}\right)\, exp\left(i\phi^{\beta}_{R,L}\right) \\[2ex] \displaystyle\rho_{R,L} &=&\displaystyle v_t\,\{\frac{(C_8\pm C_9)}{2}\,{\cal F} -\frac{m_b}{q^2}\,C_7\,{\cal C}\}&=&\displaystyle|{\rm r}_{R,L}|\,exp\left(i\delta^{\rho}_{R,L}\right)\, exp\left(i\phi^{\rho}_{R,L}\right) \,. \end{array} \label{alpha_beta_rho}$$ In the above equation $\alpha$, $\beta$ and $\rho$ are recast in terms of $a$, $b$ and $r$ so as to identify the strong phases $\delta$ and the weak phases $\phi$. Using $CPT$ invariance, the matrix element for the decay ${\bar B}\to\bar K \bar\pi\ell^+\ell^-$ can be obtained from the $B\to K\pi\ell^+\ell^-$ by replacing $\alpha_{L,R}\to -{\bar\alpha}_{L,R}, \beta_{L,R}\to {\bar\beta}_{L,R}, \rho_{L,R}\to{\bar\rho}_{L,R}$ [@valencia2; @Okubo2], where $$\displaystyle{\bar \alpha}_{R,L}=\displaystyle|{\rm a}_{R,L}|\,exp\left(i\delta^{\alpha}_{R,L}\right)\, exp\left(-i\phi^{\alpha}_{R,L}\right) \label{alpha_beta_rho_bar}$$ and similar relations hold for $\bar\beta$ and $\bar\rho$. The matrix element mod. squared for the process is worked out retaining the imaginary parts in $\a,\b$ and $\r$ to be, |[M]{}(B(p).&.K(k\_1)(k\_2)\^+(q\_1)\^-(q\_2))|\^2=\ {. &2\_k\^K\^q\^Q\^ ( KQ[Im]{}(\_[L]{}\_[L]{}\^\* + \_[R]{}\_[R]{}\^\*)+ [Im]{}(\_[R]{}\_[R]{}\^\* -\_[L]{}\_[L]{}\^\*) .\ &. -kQ[Im]{}(\_[L]{}\_[L]{}\^\*+ \_[R]{}\_[R]{}\^\*) ) + 2 [Re]{}(\_[L]{}\_[L]{}\^\* -\_[R]{}\_[R]{}\^\*)\ & ( -kK q\^2kQ + kqkQKq + m\_[K\^\*]{}\^2q\^2KQ - [kq]{}\^2KQ )\ +& 2[Re]{}(\_[L]{}\_[L]{}\^\*+\_[R]{}\_[R]{}\^\*) ( - kK q\^2 + kqKq - kQKQ )\ + & 2[Re]{}(\_[R]{}\_[R]{}\^\* -\_[L]{}\_[L]{}\^\*) ( K\^2q\^2kQ - kQ[Kq]{}\^2 - kKq\^2KQ .\ + &. kqKqKQ ) + ( \_L\^2+\_R\^2 ) ( -m\_[K\^\*]{}\^2q\^2 + kq\^2 -kQ\^2 )\ + & ( [\_L]{}\^2 + [\_R]{}\^2) ( -K\^2q\^2[kQ]{}\^2 + [kQ]{}\^2[Kq]{}\^2+ 2kKq\^2kQKQ .\ &-. 2kqkQKqKQ - m\_[K\^\*]{}\^2q\^2[KQ]{}\^2 + [kq]{}\^2[KQ]{}\^2 )\ +&. ( [[\_[L]{}]{}\^2]{} +[[\_[R]{}]{}\^2]{} ) ( -K\^2q\^2+[Kq]{}\^2 - [KQ]{}\^2 ) } \[Matmodsq\] where kK &=& m\_K\^2-m\_\^2, qQ = 0\ kq &=& (M\_B\^2-m\_[K\^\*]{}\^2-q\^2)\ kQ &=& X M\_B \_e\ Kq &=& ł\_K X M\_B \_[K\^\*]{} + kq\ KQ &=& kQ + ł\_K( kq \_l \_K - \_l \_K ) .Here $X$ is the three momentum of the $\ell^+\ell^-$ or $K\pi$ invariant system in the $B$ meson rest frame and is given by $$X=\frac{(k\cdot q^2-q^2 m_{K^*}^2)^\frac{1}{2}}{M_B}$$ $\l_K$ is related to the $K$ three momentum in the $K^*$ rest frame and is defined as $$\l_K=\left(1-\frac{(m_K+m_\p)^2}{m_{K^*}^2}\right)^\frac{1}{2} \left(1-\frac{(m_K-m_\p)^2}{m_{K^*}^2}\right)^\frac{1}{2}$$ and similarly, $\lambda_e$ is related to the lepton momentum in the $\ell^+\ell^-$ rest frame and is given by, $$\lambda_e=\sqrt{1-\frac{4\,m_e^2}{q^2}}$$ $$\e_{\m\n\r\s} k^\m K^\n q^\r Q^\s = -X M_B \l_K \sqrt{q^2 m_{K^*}^2} \sin\q_l \sin\q_K \sin\vf\;,$$ $\q_l$ ($\q_k$) is the angle between the $\ell^-$ ($K$) three-momentum vector in the $\ell^+\ell^-$ ($K\p$) rest frame and the direction of total $\ell^+\ell^-$ ($K^*$) three-momentum vector defined in the $B$ rest frame. $\vf$ is the angle between the normals to the planes defined by $\ell^+\ell^-$ and the $K\p$, in the $B$ rest frame. The differential decay rate is then given by $$d\Gamma=\displaystyle\frac{1}{2^{14} \pi^6 M_B^2}\int|{\cal M}|^2 X\lambda_K \lambda_e\, dq^2 d\cos\theta_K d\cos\theta_l d\varphi\,,$$ assuming a narrow width approximation for the decay $K^*\to K\pi$. It can easily be seen from eq.(\[Matmodsq\]) that, the only terms proportional to $\sin(\varphi)$ or $\sin(2 \varphi)$ are those that depend on the imaginary parts of the products any two of $\alpha,~\beta$ or $\rho$. For instance only the coefficient of ${\rm Im} (\alpha_L\beta^*_L +\alpha_R\beta^*_R)$ is proportional to $\sin(2 \varphi)$. Hence we can isolate this term by considering the following asymmetric width in terms of the differential decay rates of the $B$ meson with respect to $\varphi$, $$\displaystyle \Delta \Gamma_1=\displaystyle \ (\displaystyle\int_0^{\frac{\pi}{2}} -\int_{\frac{\pi}{2}}^\pi +\int_\pi^{\frac{3\pi}{2}} -\int_{\frac{3\pi}{2}}^{2\pi})\frac{d\Gamma}{d\varphi}d\varphi\,.$$ The imaginary part in the term under consideration can be due to either a strong phase or a weak phase. Such $C\!P$ violating asymmetries can be obtained not by considering the difference of differential rates for $B$ and ${\bar B}$ , but the sum of these rates. It follows trivially from eqn.(\[alpha\_beta\_rho\] and \[alpha\_beta\_rho\_bar\]) that the asymmetric width for $B$(${\bar B}$ ) is, $$\Delta \Gamma_1( \Delta {\bar \Gamma_1}) \propto\pm\displaystyle \sum_{j,k} \{\displaystyle |a^j_L| |b^k_L| \sin(\displaystyle(\delta_L^{jk})\pm(\phi_L^{jk}))+L \to R \}, \label {dgamma}$$ where $\delta_L^{jk}\equiv(\delta_L^{\alpha_j}-\delta_L^{\beta_k})$ and $\phi_L^{jk}\equiv(\phi_L^{\alpha_j}-\delta_L^{\beta_k})$. We define the asymmetry ${\sf A}_1$ as the sum of the asymmetric widths, normalized to the total widths, ${\sf A}_1=\displaystyle\frac{ \Delta \Gamma_1 + \Delta {\bar \Gamma_1}}{\Gamma + \bar \Gamma}$ and from eq.(\[dgamma\]), we have, $${\sf A}_1\propto\displaystyle \sum_{j,k} \{\displaystyle |a^j_L| |b^k_L| \cos(\displaystyle\delta_L^{jk})\,\sin(\phi_L^{jk})+L \to R \} \,,\label{asymm}$$ which is [*nonzero if and only if there is $C\!P$ violation represented by non-zero phases $\phi$*]{}[@Sinha; @valencia2; @Burdman]. $\delta$ can arise from electromagnetic final state interactions, which are negligible and ignored. For top quark in the penguin loop, the case of $\delta$ being nonzero due to intermediate quark on shell, does not arise. It is also possible to construct a different asymmetry that isolates another combination of the imaginary terms. Such an asymmetry [@kramer-palmer] considers the difference distribution of the same hemisphere and opposite hemisphere events, and the asymmetric width in this case can be defined by, $$\displaystyle \Delta \Gamma_2= \displaystyle (\displaystyle\int_0^\pi-\int_\pi^{2\pi}) d\varphi\displaystyle\int_Dd\cos\theta_l\int_Dd\cos\theta_K \displaystyle\frac{d\Gamma}{d\cos\theta_l d\cos\theta_K d\varphi}\;,$$ where $\displaystyle\int_D\equiv\displaystyle\int_{-1}^0-\int_0^{1}$. Analogous to ${\sf A_1}$, we define the asymmetry ${\sf A}_2$, ${\sf A}_2=\displaystyle\frac{\Delta \Gamma_2+ \Delta {\bar \Gamma_2}}{\Gamma + {\bar \Gamma}}$. The asymmetries ${\sf A_1}$ and ${\sf A_2}$ are evaluated to be, $${\sf A}_1=\displaystyle -2\,x\,\Delta\,\int\,dq^2 \,{\sf C}\, ({a_0\,V-A_0\,g})\,X^2\, ,\; {\sf A}_2 =\displaystyle\,x\,\int\,dq^2 \,{\sf C}\,{\sf F} \frac{1}{m_{K^*}\,\sqrt{q^2}}\,X^2\;,$$ where, x&=&,\ [C]{}&=&(C\_7\^+C\_7\^) [C\_8\^[SM]{}]{}, =M\_B\^2-m\_[K\^\*]{}\^2,\ [F]{}&=&{2 X\^2 M\_B\^2 ([a\_[+]{}V-A\_[+]{}g]{})+ ([a\_0V-A\_0g]{})[kq]{}}. Note that these asymmetries are independent of $C_9^{SM}$. Since, $C_{8}^{\tilde\kappa,\tilde\lambda}$ terms are suppressed in comparison with $C_{7}^{\tilde\kappa,\tilde\lambda}$, for all our estimations, we ignore them. For a rate asymmetry [A]{} to provide a $n$ standard deviation signal of $C\!P$  violation, we require that ${\sf A}=n/\sqrt{N}$, where N is the total number of events in the channel. Thus for ${\sf A_2}$, at 1$\sigma$ (1.64 $\sigma$) level, it is possible to place the bounds, $|\tilde\kappa |<0.34$ (0.92) for $\tilde\lambda=0$ and $|\tilde\lambda |< 1.02$ (2.75) for $\tilde\kappa=0$, using 10,000 $(B \rightarrow K^{*}e^{+}e^{-})$ events. This corresponds to 2x$10^9$ $B$’s, however, by adding the $(B \rightarrow K^{*}\mu^{+}\mu^{-})$ mode, $10^9$ $B$’s will be required. Much weaker bounds would be obtained from ${\sf A_1}$. However, we can improve the statistical significance, by looking at the $q^{2}$ distributions, shown in Fig. 1. We bin the asymmetry ${\sf A_2}$ (${\sf A_1}$ being small is ignored, and we shall drop the subscript 2 hereafter) and use a $\chi^{2}$ fit to obtain bounds on the parameters $\tilde\kappa$ and $\tilde\lambda$. The asymmetry as a function of $q^2$ is given by, $$\displaystyle A (q^2)= \frac{\displaystyle (\displaystyle\int_0^\pi-\int_\pi^{2\pi}) d\varphi\displaystyle\int_Dd\cos\theta_l\int_Dd\cos\theta_K \;\Gamma_{sum}(q^2)} {\displaystyle (\displaystyle\int_0^\pi+\int_\pi^{2\pi}) d\varphi\displaystyle\int_{-1}^1 d\cos\theta_l\int_{-1}^1 d\cos\theta_K \;\Gamma_{sum}(q^2)}\;,$$ where $\Gamma_{sum}(q^2)=\displaystyle\frac{d(\Gamma+\bar\Gamma)}{d q^2 d\cos\theta_l d\cos\theta_K d\varphi}$, $d(\Gamma+\bar\Gamma)$ is the sum of the differential widths of $B$ and $\bar B$. The average value of asymmetry in the $i^{th}$ bin is then, $$\displaystyle A_{i} = \frac{N_B}{\Gamma_B\,N_i}\int_{q^2_{min}}^{q^2_{max}} dq^2 A(q^2)\frac{d(\Gamma+\bar\Gamma)}{dq^2}\;,$$ where, $N_B$ is the total number of $B$’s, $\Gamma_B$ is the total $B$ width and $q^2_{min}$,$q^2_{max}$ are the minimum and maximum $q^2$ values in the bin; the number of events in the $i^{th}$ bin are, $$\displaystyle N_i = \frac{N_B}{\Gamma_B}\int_{q^2_{min}}^{q^2_{max}} dq^2 \frac{d(\Gamma+\bar\Gamma)}{dq^2}\;.$$ The value of the asymmetry, in the $i^{th}$ $q^2$ bin, coming from $\tilde\kappa$ and $\tilde\lambda$ contributions, is of the form $$A_i=\frac{a_i \tilde\kappa+b_i \tilde\lambda}{X_i}.$$ In $X_i$, terms quadratic in $\tilde\kappa$ and $\tilde\lambda$ are ignored, which is a reasonable approximation as long as $\tilde\kappa$, $\tilde\lambda$ are small. The observed asymmetry $A_i^{obs}$ in the $i^{th}$ bin is assumed to be chosen from a Gaussian distribution with mean equal to the theoretical asymmetry $A_i^{th}$ and variance $\sigma_i^2$. Then, the method of least squares gives, $$\chi^2=\sum_{i=1}^{nbin}\left(\frac{A_{i}^{obs}-A_{i}^{th}} {\sigma_{i}}\right)^2\;.$$ The statistical errors are taken to be $\sigma_{i}=\sqrt{1/N}$. We evaluate the difference, $\Delta\chi^2=\chi^2_{SM}-\chi^2_{min}$, where $\chi^2_{SM}$ is the $\chi^2$ corresponding to the SM values of the parameters i.e. (0,0) and $\chi_{min}^2$ is evaluated from the values ($\tilde\kappa, \tilde\lambda$) that minimize $\chi^2$. The linear dependence on the parameters $\tilde\kappa,\tilde\lambda$ results in the simple form, $$\Delta\chi^2\displaystyle=\displaystyle\sum_{i=1}^{nbin}\displaystyle \frac{\displaystyle\left(\frac{\displaystyle\tilde\kappa a_i}{X_i} \right)^2+\left(\frac{\displaystyle\tilde\lambda b_i}{X_i} \right)^2+\frac{\displaystyle 2 \tilde\lambda \tilde\kappa a_i b_i}{X_i^2}}{\sigma_{i}^2}\;.$$ Now, $\Delta\chi^2=n^2$ will give the $n$-standard deviation bounds for $\tilde\kappa$ and $\tilde\lambda$. For 20 bins, for 10,000 $(B \rightarrow K^{*}e^{+}e^{-})$ events, it is possible to obtain the improved individual bounds, || &&lt;& 0.25 (0.42) 68.3(90)%\ || &&lt;& 0.76(1.25) 68.3(90)% The possible combined bounds on $\tilde\kappa$ and $\tilde\lambda$ are shown in Fig. 2. To conclude, we have studied $C\!P$  violating anomalous couplings $\tilde\kappa$ and $\tilde\lambda$ contributing to the process $B \rightarrow K^{*}\ell^{+}\ell^{-} \to K\pi\ell^+\ell^-$. These anomalous couplings can be constrained by constructing an asymmetry, requiring the addition of $B$ and $\bar B$ events. No strong phases, nor any flavour/time tagging are required for this technique. Although, EDM of neutron places strong bounds on the couplings $\tilde\kappa$ and $\tilde\lambda$, the $C\!P$  odd asymmetries studied here, depend on a different combination of these anomalous couplings and thus provide useful additional information. $\tilde\lambda$ being the coefficient of a higher dimensional operator, is expected to be better constrained by high energy collider events than from rare B decays. The recent results of D0 provide good bounds on $\tilde\lambda$, but for $\tilde\kappa$ the approach discussed here can provide a much tighter constraint. We would like to thank G.Date and G.Rajasekaran for discussions. [References:]{} D0 Collaboration, S. Abachi et al.,Phys. Rev. Lett.[**78**]{}, 3634 (1997);[**75**]{}, 1034 (1995); CDF Collaboration, F. Abe et al.,Phys. Rev. Lett.[**74**]{}, 1936 (1995). D0 Collaboration, S. Abachi et al.,Phys. Rev. Lett.[**75**]{}, 1023 (1995);[**77**]{}, 3303 (1996); CDF Collaboration, F. Abe et al.,Phys. Rev. Lett.[**75**]{}, 1017 (1995); CLEO Collaboration, M.S. Alam eta al., Phys. Rev. Lett.[**74**]{}, 2885 (1995); UA2 Collaboration, J.Alitti et al., Phys. Lett.[**B277**]{}, 194 (1992); M. Samuel et al., Phys. Lett.[**B280**]{}, 124 (1992). M. Gronau and D. London, Phys. Rev. [**D55**]{}, 2845 (1997); J.L. Hewett and J.D. Wells, Phys. Rev. [**D55**]{}, 5549 (1997); J.L. Hewett, Phys. Rev. [**D53**]{}, 4964 (1996);N. G. Deshpande, K. Panose and J.Trampetic, Phys. Lett. [**B308**]{}, 322 (1993). C. Greub, A. Ioannissian and D. Wyler, Phys. Lett. [**B346**]{}, 149 (1995). G. Burdman, Phys. Rev. [**D52**]{}, 6400 (1995). Xiao-Gang He and Bruce Mc Kellar, Phys. Lett. [**B320**]{}, 165 (1994). K.A. Peterson, Phys. Lett.[**B282**]{}, 207 (1992); T.G. Rizzo, Phys. Lett. [**B315**]{}, 471 (1993). K. Numata Z. Phys. [**C52**]{}, 691 (1991). K. Hagiwara, R.D. Peccei and D. Zeppenfeld, , 253 (1987). W. Marciano an.d A. Queijeiro, Phys. Rev. D [**33**]{}, 3449 (1986); F. Boudjema [*et al.*]{}, Phys. Rev. D [**43**]{}, 2223 (1991). S. Dawson, Xiao-Gang He and G. Valencia, Phys. Lett.[**B390**]{}, 431 (1997). P. Kalyniak, P. Madsen, N. Sinha and R. Sinha, Phys. Rev.[**D52**]{}, 3826 (1995); G. Belanger and G. Couture, Phys. Rev. D [**49**]{}, 5720 (1994). M. Misiak, Nucl. Phys. [**B393**]{}, 23 (1993); A. J. Buras and M. Műnz, Phys. Rev [**D 52**]{}, 186 (1995). B. Grinstein, M. J. Savage and M. B. Wise, Nucl. Phys. [**B319**]{}, 271 (1989). R. Grigjains, P, J. O’Donnell, M. Sutherland and N. Navelet, Phys. Rep.[**228**]{}, 93 (1993). J. M. Soares, Nucl. Phys. [**B367**]{}, 575 (1991). L. Wolfenstein and Y. L. Wu Phys. Rev. Lett. [ **73**]{}, 2809 (1994); Y. L. Wu and L. Wolfenstein, Phys. Rev. Lett. [ **73**]{}, 1762 (1994). Dong-Sheng Du and Mao-Zhi Yang, Phys. Rev. [**D54**]{}, 882 (1996); T. M. Aliev, D. A. Demir, E. Illtan and N. K. Pak, hep-ph/9511352. R. Sinha, hep-ph/9608314. G. Valencia, Phys. Rev. [**D39**]{}, 3339 (1989). S. Okubo, in [*“ A gift of prophecy”– essays in celebration of R. E. Marshak*]{}, edited by E. C. G. Sudarshan, World Scientific (1995). G. Burdman and J. F. Donoghue, Phys. Rev. [**D45**]{}, 187b (1992). G. Kramer and W. F. Palmer, Phys. Lett. [**B279**]{}, 181 (1992). ------------------------------------------------------------------------------------------------------------------------------ [                                ]{} QM [@Wyler] [                    ]{} -------------------------------------- ------------------------------------------------------------ -------------------------- ${\cal A}$ $-2\,g\,\sigma$ \[0ex\] ${\cal B}$ $a_0\,\Delta\,\sigma$ \[2ex\] ${\cal C}$ $\displaystyle 2\,a_{+}\,W\cdot q\,\sigma-\zeta\,{\cal B}$ \[2ex\] ${\cal D}$ $\displaystyle -2\,\frac{V}{M_B+m_{K^*}}\,\sigma $ \[2ex\] ${\cal E}$ $\displaystyle \frac{A_0}{M_B+m_{K^*}}\, \Delta\, \sigma $ \[1ex\] ${\cal F}$ $\displaystyle 2\,\frac{A_{+}}{M_B+m_{K^*}}\,W\cdot q\,\sigma-\zeta \,{\cal E} $ \[2ex\] ------------------------------------------------------------------------------------------------------------------------------ : Relations between the form factors used in this paper, a quark model (QM) that reproduces heavy quark limit . $\displaystyle W_\mu=(K_\mu-\zeta\,k_\mu)$, $\sigma^2=96 \pi^2/(m_{K^*}^2 {\lambda_K}^3)$, $\zeta=\displaystyle\frac{k\cdot K}{m_{K^*}^2}$, $\lambda_K$ and $~\Delta$ are defined in the text. []{data-label="tab1"} [^1]: e-mail: nita,sinha@imsc.ernet.in

--- abstract: 'Within the framework of nonrelativistic QED, we prove that, for small values of the coupling constant, the energy function, $E_{{{\vec{P}}}}$, of a dressed electron is twice differentiable in the momentum ${{\vec{P}}}$ in a neighborhood of ${{\vec{P}}}=0$. Furthermore, $\frac{\partial^2E_{{{\vec{P}}}}}{(\partial |{{\vec{P}}}|)^2}$ is bounded from below by a constant larger than zero. Our results are proven with the help of *iterative analytic perturbation theory*.' author: - | Jürg Fröhlich[^1]\ \ \ - | Alessandro Pizzo\ \ \ date: '[27-June-2008]{}' title: Renormalized Electron Mass in Nonrelativistic QED --- Description of the problem, definition of the model, and outline of the proof ============================================================================= In this paper, we study problems connected with the renormalized electron mass in a model of quantum electrodynamics (QED) with nonrelativistic matter. We are interested in rigorously controlling radiative corrections to the electron mass caused by the interaction of the electron with the *soft* modes of the quantized electromagnetic field. The model describing interactions between nonrelativistic, quantum-mechanical charged matter and the quantized radiation field at low energies (i.e., energies smaller than the rest energy of an electron) is the “standard model”, see [@Fierz-Pauli]. In this paper, we consider a system consisting of a single spinless electron, described as a nonrelativistic particle that is minimally coupled to the quantized radiation field, and photons. Electron spin can easily be included in our description without substantial complications.\ The physical system studied in this paper exhibits space translations invariance. The Hamiltonian, $H$, generating the time evolution, commutes with the vector operator, ${{\vec{P}}}$, representing the total momentum of the system, which generates space translations. If an infrared regularization, e.g., an infrared cutoff $\sigma$ on the photon frequency, is imposed on the interaction Hamiltonian, there exist single-electron or dressed one-electron states, as long as their momentum is smaller than the bare electron mass, $m$, of the electron. This means that a notion of mass shell in the energy momentum spectrum is meaningful for velocities $|{{\vec{P}}}|/m$ smaller than the speed of light $c$; (with $c\equiv m \equiv 1$ in our units). Vectors $\{\Psi^{\sigma}\}$ describing dressed one-electron states are normalizable vectors in the Hilbert space ${{\mathcal{H}}}$ of pure states of the system. They are characterized as solutions of the equation $$\label{eq:I.1} H^{\sigma}\Psi^{\sigma}\,=E^{\sigma}_{{{\vec{P}}}}\Psi^{\sigma}\,,\, |{{\vec{P}}}|<1$$ where $H^{\sigma}$ is the Hamiltonian with an infrared cutoff $\sigma$ in the interaction term and $E^{\sigma}_{{{\vec{P}}}}$, the energy of a dressed electron, is a function of the momentum operator ${{\vec{P}}}$. If in the joint spectrum of the components of ${{\vec{P}}}$ the support of the vector $\Psi^{\sigma}$ is contained in a ball centered at the origin and of radius less than $1\,\equiv mc$ then Eq. (\[eq:I.1\]) has solutions; see [@ChenFroehlichPizzo2]. Since $[H,{{\vec{P}}}]=0$, Eq. (\[eq:I.1\]) can be studied for the fiber vectors, $\Psi_{\bar{{{\vec{P}}}}}^{\sigma}$, corresponding to a value, ${{\vec{P}}}$, of the total momentum; (both the total momentum operator and points in its spectrum will henceforth be denoted by ${{\vec{P}}}$ – without danger of confusion). Thus we consider the equation $$\label{eq:I.2bis} H^{\sigma}_{{{\vec{P}}}}\Psi_{{{\vec{P}}}}^{\sigma}\,=E^{\sigma}_{{{\vec{P}}}}\Psi_{{{\vec{P}}}}^{\sigma}\,,$$ where $H^{\sigma}_{{{\vec{P}}}}$ is the fiber Hamiltonian at fixed total momentum ${{\vec{P}}}$, and $E_{{{\vec{P}}}}^{\sigma}$ is the value of the function $E^{\sigma}_{{{\vec{z}}}}$ at the point ${{\vec{z}}}\equiv {{\vec{P}}}$. Physically, states $\{\Psi^{\sigma}\}$ solving Eq. (\[eq:I.1\]) describe a freely moving electron in the absence of asymptotic photons.\ It is an essential aspect of the “infrared catastrophe” in QED that Eq. (\[eq:I.1\]) does not have any normalizable solution in the limit where the infrared cut-off $\sigma$ tends to zero, and the underlying dynamical picture of a freely moving electron breaks down; see [@ChenFroehlichPizzo1]. Nevertheless, the limiting behavior of the function $E^{\sigma}_{{{\vec{P}}}}$ is of great interest for the following reasons. As long as $\sigma>0$, a natural definition of the renormalized electron mass, $m_r$, is given by the formula $$\label{eq:I.2} m_r(\sigma):=\big[\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2} |_{{{\vec{P}}}=0}\big]^{-1}\,.$$ (Note that $E^{\sigma}_{{{\vec{P}}}}\equiv E^{\sigma}_{|{{\vec{P}}}|}$ is invariant under rotations). Equation (\[eq:I.2\]) is expected to remain meaningful in the limit $\sigma\to0$. In particular, the quantity on the R.H.S. of Eq. (\[eq:I.2\]) is expected to be positive and bounded from above uniformly in the infrared cutoff $\sigma$. More importantly, one aims at mathematical control of the function $$m_r(\sigma,| {{\vec{P}}}|):=\big[\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2} \big]^{-1}$$ in a full neighborhood, ${{\mathcal{S}}}$, of ${{\vec{P}}}=0$, corresponding to a slowly moving electron; (i.e., in the nonrelativistic regime). When combined with a number of other spectral properties of the Hamiltonian of nonrelativistic QED the condition $$\label{eq:I.5} \frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2} > 0\quad,\quad{{\vec{P}}}\in{{\mathcal{S}}}\,,$$ uniformly in $\sigma>0$, suffices to yield a consistent scattering picture in the limit when $\sigma\to0$ in which the electron exhibits *infraparticle* behavior. In fact, (\[eq:I.5\]) is a crucial ingredient in the analysis of Compton scattering presented in [@Pizzo2005], [@ChenFroehlichPizzo1].\ *Main results* Assuming the coupling constant, $\alpha$, small enough, the following results follow. 1\) The function $$\label{eq:I.6} \Sigma_{|{{\vec{P}}}|}:=\lim_{\sigma\to0}\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}$$ is well defined for ${{\vec{P}}}\in{{\mathcal{S}}}:=\{{{\vec{P}}}\,|\,|{{\vec{P}}}|<\frac{1}{3}\}$; furthermore, it is Hölder-continuous in ${{\vec{P}}}$. 2\) The function $$E_{{{\vec{P}}}}:=\lim_{\sigma\to0}E_{{{\vec{P}}}}^{\sigma}$$ is twice differentiable in ${{\vec{P}}}\in{{\mathcal{S}}}$ and $$\label{eq:I.6bis} \frac{\partial^2E_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}= \Sigma_{|{{\vec{P}}}|}\,.$$ 3\) $$\label{eq:I.7} \lim_{\alpha\to0}\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2 } =\frac{1}{m}\quad,\quad{{\vec{P}}}\in{{\mathcal{S}}}\,,$$ uniformly in $\sigma$, where $m$ is the bare electron mass. (Our results can be extended to a region ${{\mathcal{S}}}$ (inside the unit ball) of radius larger than $\frac{1}{3}$.)\ We wish to mention some related earlier results. Using *operator-theoretic renormalization group methods*, results (\[eq:I.6\]) and (\[eq:I.7\]) have been proven in [@ChenBachFroehlichSigal] for the special value ${{\vec{P}}}=0$. The point ${{\vec{P}}}=0$ is exceptional, because the Hamiltonian $H_{{{\vec{P}}}}$ is *infrared regular* at ${{\vec{P}}}=0$; it has a normalizable ground state. Thomas Chen (see [@Chen]) has established the results in (\[eq:I.6bis\]), (\[eq:I.7\]) (using smooth infrared cut-offs) by a highly non-trivial extension of the analysis of [@ChenBachFroehlichSigal] to arbitrary momenta ${{\vec{P}}}\in{{\mathcal{S}}}$. The procedure presented in our paper relies on *iterative analytic perturbation theory* (see Section \[SectII\] where this tool is recalled) that makes our proof substantially different and much shorter in comparison to a renormalization group approach. The main feature is a more transparent treatment of the so called *marginal* terms of the interaction, where an essential role is played by explicit Bogoliubov transformations that transform the infrared representations of the CCR of photon creation- and annihilation operators determined by dressed one-particle states of fixed momentum ${{\vec{P}}}(\neq 0)$ back to the Fock representation. The use of these Bogoliubov transformations is a crucial device in our fight against the infrared problem. The way in which we are using them is new, at least in the context of mathematically rigorous results on the infrared problem in QED. In our paper, the regularity properties of $E_{{{\vec{P}}}}$ come with an explicit control of the asymptotics of the fiber ground state eigenvectors $\Psi^{\sigma}_{{{\vec{P}}}}$ as $\sigma$ tends to zero. (This improves earlier results in [@ChenFroehlichPizzo2].) Along the lines of [@BachFroehlichPizzo], these results are preparatory to developing an infrared finite algorithm for the asymptotic expansion of the renormalized electron mass in powers and, probably, logarithms of the finestructure constant $\alpha$, up to an arbitrarily small remainder term. However, the expansion in the coupling constant $\alpha$ is not studied in this paper. With regard to *ultraviolet* corrections to the electron mass in nonrelativistic QED models, we refer the reader to [@Lieb-Loss], [@Lieb-Loss2], [@Hiroshima-Spohn], and [@Hainzl-Seiringer]. In Section \[Definition\] below, the model is defined rigorously. Then, for the convenience of the reader, in Section \[Outline\] we outline the key ideas of the proof and present the organization of the remaining sections of the paper. Definition of the model {#Definition} ----------------------- *Hilbert space* The Hilbert space of pure state vectors of a system consisting of one non-relativistic electron interacting with the quantized electromagnetic field is given by $$\label{eq-I-1} {{\mathcal{H}}}\; := \; {{\mathcal{H}}}_{{el}}\, \otimes \, {{\mathcal{F}}}\, ,$$ where ${{\mathcal{H}}}_{{el}}= L^2({\mathbb{R}}^3)$ is the Hilbert space for a single Schrödinger electron; for expository convenience, we neglect the spin of the electron. The Hilbert space, ${{\mathcal{F}}}$, used to describe the states of the transverse modes of the quantized electromagnetic field (the *photons*) in the Coulomb gauge is given by the Fock space $$\label{eq-I-2} {{\mathcal{F}}}\ := \ \bigoplus_{N=0}^\infty {{\mathcal{F}}}^{(N)} {\: ,}\hspace{6mm} {{\mathcal{F}}}^{(0)} = {\mathbb{C}}\, {\Omega}{\: ,}$$ where ${\Omega}$ is the vacuum vector (the state of the electromagnetic field without any excited modes), and $$\label{eq-I-3} {{\mathcal{F}}}^{(N)} \ := \ {{\mathcal{S}}}_N \, \bigotimes_{j=1}^N {\mathfrak{h}}{\: ,}\hspace{6mm} N \geq 1 {\: ,}$$ where the Hilbert space ${\mathfrak{h}}$ of state vectors of a single photon is $$\label{eq-I-4} {\mathfrak{h}}\ := \ L^2( {\mathbb{R}}^3 \times {\mathbb{Z}}_2 ) \,.$$ Here, ${\mathbb{R}}^3$ is momentum space, and ${\mathbb{Z}}_2$ accounts for the two independent transverse polarizations (or helicities) of a photon. In (\[eq-I-3\]), ${{\mathcal{S}}}_N$ denotes the orthogonal projection onto the subspace of $\bigotimes_{j=1}^N {\mathfrak{h}}$ of totally symmetric $N$-photon wave functions, which accounts for the fact that photons satisfy Bose-Einstein statistics. Thus, ${{\mathcal{F}}}^{(N)}$ is the subspace of ${{\mathcal{F}}}$ of state vectors corresponding to configurations of exactly $N$ photons.\ *Units* In this paper, we employ units such that Planck’s constant $\hbar$, the speed of light $c$, and the mass of the electron $m$ are equal to 1.\ *Hamiltonian* The dynamics of the system is generated by the Hamiltonian $$\label{eq-I-6} H \; := \; \frac{\big(-i{{\vec{\nabla}}}_{{{\vec{x}}}} \, + \, \alpha^{1/2} {{\vec{A}}}({{\vec{x}}}) \, \big)^2}{2} \, + \, H^{f}\,.$$ The (three-component) multiplication operator ${{\vec{x}}}\in{\mathbb{R}}^3$ represents the position of the electron. The electron momentum operator is given by ${{\vec{p}}}=-i{{\vec{\nabla}}}_{{\vec{x}}}$. Furthermore, $\alpha>0$ is the fine structure constant (which, in this paper, plays the rôle of a small parameter), and ${{\vec{A}}}({{\vec{x}}})$ denotes the vector potential of the transverse modes of the quantized electromagnetic field in the *Coulomb gauge*, $$\label{eq-I-7} {{\vec{\nabla}}}_{{\vec{x}}}\cdot {{\vec{A}}}({{\vec{x}}}) \ = \ 0 \, ,$$ cutoff at high photon frequencies. $H^f$ is the Hamiltonian of the quantized, free electromagnetic field. It is given by $$\label{eq-I-10} H^f \; := \; \sum_{\lambda = \pm} \int d^3k \; |{{\vec{k}}}| \, a^*_{{{\vec{k}}},\lambda} \, a_{{{\vec{k}}}, \lambda} {\: ,}$$ where $a^*_{{{\vec{k}}}, \lambda}$ and $a_{{{\vec{k}}}, \lambda}$ are the usual photon creation- and annihilation operators satisfying the canonical commutation relations $$\begin{aligned} \label{eq-I-12} [a_{{{\vec{k}}}, \lambda} \, , \, a^*_{{{\vec{k}}}', \lambda'}] & = & \delta_{\lambda \lambda'} \, \delta ({{\vec{k}}}- {{\vec{k}}}') {\: ,}\\ \label{eq-I-11} [a^\#_{{{\vec{k}}}, \lambda} \, , \, a^\#_{{{\vec{k}}}', \lambda'}] & = & 0\end{aligned}$$ for ${{\vec{k}}}, {{\vec{k}}}' \in {\mathbb{R}}^3$ and $\lambda,\lambda' \in {\mathbb{Z}}_2\equiv\{\pm\}$, where $a^\#=a$ or $a^*$. The vacuum vector $\Omega\in{{\mathcal{F}}}$ is characterized by the condition $$a_{{{\vec{k}}}, \lambda} \, {\Omega}\; = \; 0 {\: ,}\label{eq-I-13}$$ for all ${{\vec{k}}}\in {\mathbb{R}}^3$ and $\lambda \in {\mathbb{Z}}_2\equiv\{\pm\}$. The quantized electromagnetic vector potential is given by $$\begin{aligned} \label{eq-I-14} {{\vec{A}}}({{\vec{x}}}) \; := \; \sum_{\lambda = \pm} \int_{{{\mathcal{B}}}_{\Lambda}} \frac{d^3k}{\sqrt{ |{{\vec{k}}}| \,}}\, \big\{ {{\vec{\varepsilon}}}_{{{\vec{k}}}, \lambda} e^{-i{{\vec{k}}}\cdot {{\vec{x}}}} a^*_{{{\vec{k}}}, \lambda} \, + \, {{\vec{\varepsilon}}}_{{{\vec{k}}}, \lambda}^{\,\,*} e^{i{{\vec{k}}}\cdot {{\vec{x}}}} a_{{{\vec{k}}}, \lambda} \big\} {\: ,}\end{aligned}$$ where ${{\vec{\varepsilon}}}_{{{\vec{k}}}, -}$, ${{\vec{\varepsilon}}}_{{{\vec{k}}}, +}$ are photon polarization vectors, i.e., two unit vectors in ${\mathbb{R}}^3 \otimes{\mathbb{C}}$ satisfying $$\label{eq-I-15} {{\vec{\varepsilon}}}_{{{\vec{k}}}, \lambda}^{\,*} \cdot {{\vec{\varepsilon}}}_{{{\vec{k}}}, \mu} \; = \; \delta_{\lambda \mu} {\: ,}\hspace{8mm} {{\vec{k}}}\cdot {{\vec{\varepsilon}}}_{{{\vec{k}}}, \lambda} \; = \; 0 {\: ,}$$ for $\lambda, \mu = \pm$. The equation ${{\vec{k}}}\cdot {{\vec{\varepsilon}}}_{{{\vec{k}}}, \lambda} = 0$ expresses the Coulomb gauge condition. Moreover, ${{\mathcal{B}}}_{\Lambda}$ is a ball of radius $\Lambda$ centered at the origin in momentum space. Here, $\Lambda$ represents an *ultraviolet cutoff* that will be kept fixed throughout our analysis. The vector potential defined in (\[eq-I-14\]) is thus cut off in the ultraviolet. Throughout this paper, it will be assumed that $\Lambda\approx 1$ (the rest energy of an electron), and that $\alpha>0$ is sufficiently small. Under these assumptions, the Hamiltonian $H$ is selfadjoint on $D(H^0)$, i.e., on the domain of definition of the operator $$H^0 \; := \; \frac{(-i{{\vec{\nabla}}}_{{{\vec{x}}}})^2}{2} \, + \, H^f \;.$$ The perturbation $H-H^0$ is small in the sense of Kato. The operator representing the total momentum of the system consisting of the electron and the electromagnetic radiation field is given by $${{\vec{P}}}\,:=\,{{\vec{p}}}+{{\vec{P}}}^f \, ,$$ with ${{\vec{p}}}=-i{{\vec{\nabla}}}_{{{\vec{x}}}}$, and where $${{\vec{P}}}^f\,:=\, \sum_{\lambda = \pm} \int d^3k \; {{\vec{k}}}\, a^*_{{{\vec{k}}}, \lambda} \, a_{{{\vec{k}}}, \lambda}$$ is the momentum operator associated with the photon field. The operators $H$ and ${{\vec{P}}}$ are essentially selfadjoint on a common domain, and since the dynamics is invariant under translations, they commute, $[H,{{\vec{P}}}]=\vec 0$. The Hilbert space ${{\mathcal{H}}}$ can be decomposed into a direct integral over the joint spectrum, ${\mathbb{R}}^3$, of the three components of the momentum operator ${{\vec{P}}}$. Their spectral measure is absolutely continuous with respect to Lebesgue measure, and hence we have that $${{\mathcal{H}}}\,:=\,\int^{\oplus}{{\mathcal{H}}}_{{{\vec{P}}}} \, d^3P \, ,$$ where each fiber space ${{\mathcal{H}}}_{{{\vec{P}}}}$ is a copy of Fock space ${{\mathcal{F}}}$.\ [**[Remark]{}**]{} *Throughout this paper, the symbol ${{\vec{P}}}$ stands for both a vector in ${\mathbb{R}}^3$ and the vector operator on ${{\mathcal{H}}}$, representing the total momentum, depending on context. Similarly, a double meaning is given to arbitrary functions, $f({{\vec{P}}})$, of the total momentum operator.*\ We recall that vectors $\Psi\in{{\mathcal{H}}}$ are given by sequences $$\{\Psi^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)\}_{m=0}^{\infty}\,,$$ of functions, $\Psi^{(m)}$, where $\Psi^{(0)}({{\vec{x}}})\in L^2({\mathbb{R}}^3)$, of the electron position ${{\vec{x}}}$ and of $m$ photon momenta ${{\vec{k}}}_1,\dots,{{\vec{k}}}_m $ and helicities $\lambda_1, \dots, \lambda_m$, with the following properties: - $\Psi^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)$ is totally symmetric in its $m$ arguments $({{\vec{k}}}_j,\lambda_j)_{j=1,\dots,m}$. - $\Psi^{(m)}$ is square-integrable, for all $m$. - If $\Psi$ and $\Phi$ are two vectors in ${{\mathcal{H}}}$ then $$\begin{aligned} \lefteqn{(\Psi\,,\,\Phi) }\\ &=&\sum_{m=0}^{\infty}\big(\sum_{\lambda_j=\pm}\,\int\,d^3x\,\prod_{j=1}^{m}\,d^3k_j\,\overline{\Psi^{(m)}({{\vec{x}}};{{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)}\,\Phi^{(m)}({{\vec{x}}};{{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)\big)\,.\nonumber\end{aligned}$$ We identify a square integrable function $g({{\vec{x}}})$ with the sequence $$\{\Psi^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)\}_{m=0}^{\infty}\,,$$ where $\Psi^{(0)}({{\vec{x}}})\equiv g({{\vec{x}}})$, and $\Psi^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)\equiv 0$ for all $m>0$; analogously, a square integrable function $g^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)$, $m\geq 1$, is identified with the sequence $$\{\Psi^{(m')}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_{m'},\lambda_{m'})\}_{m'=0}^{\infty}\,,$$ where $\Psi^{(m)}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_m,\lambda_m)\equiv g^{(m)} $, and $\Psi^{(m')}({{\vec{x}}}; {{\vec{k}}}_1,\lambda_1;\dots; {{\vec{k}}}_{m'},\lambda_{m'})\equiv 0$ for all $m'\neq m$. From now on, a sequence describing a quantum state with a fixed number of photons is identified with its nonzero component wave function; vice versa, a wave function corresponds to a sequence according to the previous identification. The elements of the fiber space ${{\mathcal{H}}}_{{{\vec{P}}}^*}$ are obtained by linear combinations of the (improper) eigenvectors of the total momentum operator ${{\vec{P}}}$ with eigenvalue ${{\vec{P}}}^*$, e.g., the plane wave $e^{i{{\vec{P}}}^*\cdot{{\vec{x}}}}$ is the eigenvector describing a state with an electron and no photon. Given any ${{\vec{P}}}\in{\mathbb{R}}^3$, there is an isomorphism, $I_{{{\vec{P}}}}$, $$I_{{{\vec{P}}}}\,:\,{{\mathcal{H}}}_{{{\vec{P}}}}\,\longrightarrow\,{{\mathcal{F}}}^{b}\,,$$ from the fiber space ${{\mathcal{H}}}_{{{\vec{P}}}}$ to the Fock space ${{\mathcal{F}}}^{b}$, acted upon by the annihilation- and creation operators $b_{{{\vec{k}}},\lambda}$, $b^*_{{{\vec{k}}},\lambda}$, where $b_{{{\vec{k}}},\lambda}$ corresponds to $e^{i{{\vec{k}}}\cdot{{\vec{x}}}} a_{{{\vec{k}}},\lambda}$, and $b_{{{\vec{k}}},\lambda}^*$ to $e^{-i{{\vec{k}}}\cdot{{\vec{x}}}} a_{{{\vec{k}}},\lambda}^* $, and with vacuum $\Omega_{f}:=I_{{{\vec{P}}}}(e^{i{{\vec{P}}}\cdot{{\vec{x}}}})$. To define $I_{{{\vec{P}}}}$ more precisely, we consider a vector $\psi_{(f^{(n)};{{\vec{P}}})}\in {{\mathcal{H}}}_{{{\vec{P}}}}$ with a definite total momentum describing an electron and $n$ photons. Its wave function in the variables $({{\vec{x}}};{{\vec{k}}}_1,\lambda_1;\dots,{{\vec{k}}}_n,\lambda_n)$ is given by $$e^{i({{\vec{P}}}-{{\vec{k}}}_1-\cdots-{{\vec{k}}}_n)\cdot{{\vec{x}}}}f^{(n)}({{\vec{k}}}_1,\lambda_1;\dots;{{\vec{k}}}_n,\lambda_n)$$ where $f^{(n)}$ is totally symmetric in its $n$ arguments. The isomorphism $I_{{{\vec{P}}}}$ acts by way of $$\begin{aligned} \lefteqn{I_{{{\vec{P}}}}\big( e^{i({{\vec{P}}}-{{\vec{k}}}_1-\cdots-{{\vec{k}}}_n)\cdot{{\vec{x}}}}f^{(n)}({{\vec{k}}}_1,\lambda_1;\dots;{{\vec{k}}}_n,\lambda_n)\big)} \\ &= &\frac{1}{\sqrt{n!}}\sum_{\lambda_1,\dots,\lambda_n}\int \, d^3k_1\dots d^3k_n \,f^{(n)}({{\vec{k}}}_1,\lambda_1;\dots;{{\vec{k}}}_n,\lambda_n)\, b_{{{\vec{k}}}_1,\lambda_1}^* \cdots b_{{{\vec{k}}}_n,\lambda_n}^* \, \Omega_f \,.\nonumber\end{aligned}$$ Because the Hamiltonian $H$ commutes with the total momentum, it preserves the fibers ${{\mathcal{H}}}_{{{\vec{P}}}}$ for all ${{\vec{P}}}\in{\mathbb{R}}^3$, i.e., it can be written as $$H\,=\,\int^{\oplus} H_{{{\vec{P}}}}\,d^3P\,,$$ where $$H_{{{\vec{P}}}}\,:\,{{\mathcal{H}}}_{{{\vec{P}}}}\longrightarrow{{\mathcal{H}}}_{{{\vec{P}}}}\,.$$ Written in terms of the operators $b_{{{\vec{k}}},\lambda}$, $b^*_{{{\vec{k}}},\lambda}$, and of the variable ${{\vec{P}}}$, the fiber Hamiltonian $H_{{{\vec{P}}}}$ is given by $$\label{eq-fibHam} H_{{{\vec{P}}}} \; := \; \frac{\big({{\vec{P}}}-{{\vec{P}}}^f +\alpha^{1/2} {{\vec{A}}}\big)^2}{2} \; + \;H^{f}\,,$$ where $$\begin{aligned} {{\vec{P}}}^f &= & \sum_{\lambda}\, \int d^3k\, {{\vec{k}}}\, b^*_{{{\vec{k}}},\lambda} \, b_{{{\vec{k}}},\lambda} \, , \\ H^f & =& \sum_{\lambda}\, \int d^3k \, |{{\vec{k}}}| b^*_{{{\vec{k}}},\lambda} b_{{{\vec{k}}},\lambda} \,,\end{aligned}$$ and $${{\vec{A}}}\, := \,\sum_{\lambda}\, \int_{{{\mathcal{B}}}_{\Lambda}}\, \frac{d^3k}{\sqrt{ |{{\vec{k}}}| \,}} \, \big\{ {{\vec{\varepsilon}}}_{{{\vec{k}}},\lambda}b^*_{{{\vec{k}}},\lambda} \, + \, {{\vec{\varepsilon}}}^{\,\,*}_{{{\vec{k}}},\lambda} b_{{{\vec{k}}},\lambda} \big\} \,.$$ Let $$\mathcal{S}:=\lbrace\, {{\vec{P}}}\in{\mathbb{R}}^3\,:\,|{{\vec{P}}}|<\frac{1}{3}\,\rbrace\,.$$ In order to give a mathematically precise meaning to the constructions presented in the following, we introduce an infrared cut-off at a photon frequency $\sigma>0$ in the vector potential. The calculation of the second derivative of the energy of a dressed electron – in the following called the “ground state energy” – as a function of ${{\vec{P}}}$ in the limit where $\sigma\to0$, and for ${{\vec{P}}}\in{{\mathcal{S}}}$, represents the main problem solved in this paper. Hence we will, in the sequel, study the regularized fiber Hamiltonian $$\label{eq:H-fiber} H_{\vec{P}}^{\sigma} \, := \, \frac{\big(\vec{P}-\vec{P}^{f} +\alpha^{1/2} \vec{A}^{\sigma} \big)^2}{2} \, + \, H^{f}\,,$$ acting on the fiber space $\mathcal{H}_{\vec{P}}$, for $\vec{P}\in \mathcal{S}$, where $$\begin{aligned} \vec{A}^{\sigma} \, := \,\sum_{\lambda}\, \int_{\mathcal{B}_{\Lambda}\setminus \mathcal{B}_{\sigma}}\, \frac{d^3k}{\sqrt{ |{{\vec{k}}}| \,}} \, \big\{ {{\vec{\varepsilon}}}_{{{\vec{k}}},\lambda}b^*_{{{\vec{k}}},\lambda}\, + \, {{\vec{\varepsilon}}}^{\,\,*}_{{{\vec{k}}},\lambda} b_{{{\vec{k}}},\lambda} \big\}\end{aligned}$$ and where $\mathcal{B}_{\sigma}$ is a ball of radius $\sigma$ centered at the origin. In the following, we will consider a sequence of infrared cutoffs $$\sigma_j \, := \, \Lambda\epsilon^{j}$$ with $0<\epsilon<\frac{1}{2}$ and $j\in\mathbb{N}_{0}:={\mathbb{N}}\cup \{0\}$.\ *Notations* 1\) We use the notation $\| A \|_{{{\mathcal{H}}}}=\|A|_{{{\mathcal{H}}}}\|$ for the norm of a bounded operator $A$ acting on a Hilbert space ${{\mathcal{H}}}$. Typically, ${{\mathcal{H}}}$ will be some subspace of ${{\mathcal{F}}}^b$. 2\) Throughout the paper, we follow conventions such that $$\frac{1}{2\pi i}\oint_{\gamma}\frac{1}{z}dz=-1\quad,\quad \oint_{\gamma}\frac{1}{\bar{z}}d\bar{z}=\overline{( \oint_{\gamma}\frac{1}{z}dz)}\,,$$ where $\gamma$ is an integration path in the complex space enclosing the origin. Outline of the proof {#Outline} -------------------- Next, we outline the key ideas used in the proofs of our main results in Eqs. (\[eq:I.6\]), (\[eq:I.6bis\]), and (\[eq:I.7\]). For ${{\vec{P}}}\in{{\mathcal{S}}}$, $\alpha$ small enough, and $\sigma>0$, $E^{\sigma}_{{{\vec{P}}}}$ is an isolated eigenvalue of $H_{{{\vec{P}}}}^{\sigma}|_{{{\mathcal{F}}}_{\sigma}}$; see Section II and Eq. (\[eq:II.4\]). Because of the analyticity of $H_{{{\vec{P}}}}^{\sigma}$ in the variable ${{\vec{P}}}$, it follows that $$\begin{aligned} \label{eq:I.36} & &\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}=\partial_i^2E^{\sigma}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}\\ &=&1-2 \langle \frac{1}{2\pi i}\oint_{\gamma_{\sigma}}\frac{1}{H^{\sigma}_{{{\vec{P}}}}-z}[\partial_iH^{\sigma}_{{{\vec{P}}}}]\frac{1}{H^{\sigma}_{{{\vec{P}}}}-z}dz\,\Psi_{{{\vec{P}}}}^{\sigma},\,[\partial_iH^{\sigma}_{{{\vec{P}}}}]\Psi_{{{\vec{P}}}}^{\sigma}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}} \,, \nonumber\end{aligned}$$ where $\partial_i=\partial/\partial P^{i}$, $\hat{i}$ is the unit vector in the direction $i$, $\Psi_{{{\vec{P}}}}^{\sigma}$ is the normalized ground state eigenvector of $H_{{{\vec{P}}}}^{\sigma}$ constructed in [@ChenFroehlichPizzo2]; $\gamma_{\sigma}$ is a contour path in the complex energy plane enclosing $E^{\sigma}_{{{\vec{P}}}}$ and no other point of the spectrum of $H_{{{\vec{P}}}}^{\sigma}$, and such that the distance of $\gamma_{\sigma}$ from $\text{spec}\,(H_{{{\vec{P}}}}^{\sigma})$ is of order $\sigma$. At first glance, the expression on the R.H. S. of (\[eq:I.36\]) might become singular as $\sigma\to0$, because the spectral gap above $E^{\sigma}_{{{\vec{P}}}}=\inf \text{spec} \, (H^{\sigma}_{{{\vec{P}}}}|_{{{\mathcal{F}}}_{\sigma}})$ is of order $\sigma$. To prove that the limit $\sigma\to0$ is, in fact, well defined, we make use of a $\sigma$-dependent Bogoliubov transformation, $W_{\sigma}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma})$; (see Section II, Eq. (\[eq:II.3\])). This transformation has already been employed in [@ChenFroehlichPizzo2] to analyze mass shell properties. In fact, conjugation of $H^{\sigma}_{{{\vec{P}}}}$ by $W_{\sigma}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma})$ yields an infrared regularized Hamiltonian $$K_{{{\vec{P}}}}^{\sigma}\,:=\,W_{\sigma}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma})H_{{{\vec{P}}}}^{\sigma}W_{\sigma}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma})\,$$ with the property that the corresponding ground state, $\Phi^{\sigma}_{{{\vec{P}}}}$, has a non-zero limit, as $\sigma\to0$. The Hamiltonian $K_{{{\vec{P}}}}^{\sigma}$ has a “canonical form” derived in [@ChenFroehlichPizzo2] (see also [@Pizzo], where a similar operator has been used in the analysis of the Nelson model): $$K_{{{\vec{P}}}}^{\sigma}\,=\,\frac{(\vec{\Gamma}^{\sigma}_{{{\vec{P}}}})^2}{2}+\sum_{\lambda}\int_{{\mathbb{R}}^3}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma}(\hat{k})b^*_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k+{{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma}\,,$$ where $\delta_{{{\vec{P}}}}^{\sigma}(\hat{k})$ is defined in Eq. (\[eq-III.18\]), ${{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma}$ is a c-number defined in Eq. (\[eq-II.42\]), and $\vec{\Gamma}^{\sigma}_{{{\vec{P}}}}$ is a vector operator defined in Eq. (\[eq:II.40\]) starting from Eqs. (\[eq:III.16\]), (\[eq-III.17\]), (\[eq:II.34bis\]). By construction, $$\label{eq:I.39} \langle \Phi^{\sigma}_{{{\vec{P}}}}\,,\,\vec{\Gamma}^{\sigma}_{{{\vec{P}}}}\,\Phi^{\sigma}_{{{\vec{P}}}}\rangle=0\,.$$ This is a crucial property in the proof of existence of a limit of $\Phi^{\sigma}_{{{\vec{P}}}}$ as $\sigma\to0$.\ Eq. (\[eq:I.39\]) is also an important ingredient in the proof of (\[eq:I.6\]), because, by applying the unitary operator $W_{\sigma}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma})$ to each term of the scalar product on the R.H.S. of (\[eq:I.36\]) and using (\[eq:I.39\]) (see Section \[SectIII\]), one finds that $$\begin{aligned} (\ref{eq:I.36})&=&1-2\frac{1}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|^2}\langle \frac{1}{2\pi i}\oint_{\gamma_j}\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}[\partial_iE_{{{\vec{P}}}}^{\sigma_j}-(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}]\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}dz_j\,\Phi_{{{\vec{P}}}}^{\sigma_j}\,,\nonumber\\ & &\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\,,\,[\partial_iE_{{{\vec{P}}}}^{\sigma_j}-(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}]\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}} \label{eq:I.39bis} \\ &=&1-2 \langle \frac{1}{2\pi i}\oint_{\gamma_{\sigma}}\frac{1}{K^{\sigma}_{{{\vec{P}}}}-z}\,(\Gamma^{\sigma}_{{{\vec{P}}}})^i\, \frac{1}{K^{\sigma}_{{{\vec{P}}}}-z}dz\,\frac{\Phi_{{{\vec{P}}}}^{\sigma}}{\|\Phi_{{{\vec{P}}}}^{\sigma}\|}\,,\,(\Gamma^{\sigma}_{{{\vec{P}}}})^i\,\frac{\Phi_{{{\vec{P}}}}^{\sigma}}{\|\Phi_{{{\vec{P}}}}^{\sigma}\|}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}}\,. \quad\quad\quad \label{eq:I.40}\end{aligned}$$ Notice that, if one starts from the two expressions on the R.H.S. of Eq. (\[eq:I.36\]) and on the R.H.S. of Eq. (\[eq:I.39bis\]) respectively, both formally expanded in powers of $\alpha^{1/2}$, the Bogoliubov transformation can be seen as a tool to re-collect an infinite number of terms and to show a nontrivial identity, thanks only to Eq. (\[eq:I.39\]) and to a vanishing contour integration (see Eq. (\[eq:III.48\])). Next, still using Eq. (\[eq:I.39\]), one can show that (\[eq:I.40\]) remains uniformly bounded in $\sigma$.\ To see this we use the inequality $$\label{eq:I.41} \big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma}-z}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma}\big \rangle \big|\leq {{\mathcal{O}}}(\frac{1}{\alpha^{\frac{1}{2}}\, \sigma^{2\delta}})\,,$$ for an arbitrarily small $\delta>0$, with $z\in\gamma_{\sigma}$ and $\alpha$ small enough depending on $\delta$. This inequality will be proven inductively (see Theorem \[Th.induction\]) by introducing sequences of infrared cut-offs $\sigma_j$, where $\sigma_j\to0$ as $j\to \infty$. The proof by induction is combined with an improved (as compared to the result in [@ChenFroehlichPizzo2]) estimate of the rate of convergence of $\{\Phi^{\sigma}_{{{\vec{P}}}}\}$ as $ \sigma\to0$ . By telescoping, one can plug these improved estimates into (\[eq:I.40\]) to end up with the desired uniform bound. The control of the rate of convergence of the R.H.S. in (\[eq:I.40\]), as $\sigma\to0$, combined with the smoothness in ${{\vec{P}}}$, for arbitrary infrared cutoff $\sigma>0$, finally entails the Hölder-continuity in ${{\vec{P}}}$ of the limiting quantity $$\Sigma_{|{{\vec{P}}}|}\,:=\,1-\lim_{\sigma\to0} 2 \langle \frac{1}{2\pi i}\oint_{\gamma_{\sigma}}\frac{1}{K^{\sigma}_{{{\vec{P}}}}-z}\,(\Gamma^{\sigma}_{{{\vec{P}}}})^i\,\frac{1}{K^{\sigma}_{{{\vec{P}}}}-z}dz\,\frac{\Phi_{{{\vec{P}}}}^{\sigma}}{\|\Phi_{{{\vec{P}}}}^{\sigma}\|}\,,\,(\Gamma^{\sigma}_{{{\vec{P}}}})^i\,\frac{\Phi_{{{\vec{P}}}}^{\sigma}}{\|\Phi_{{{\vec{P}}}}^{\sigma}\|}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}} \,.$$ The Hölder-continuity in ${{\vec{P}}}$ of $ \Sigma_{|{{\vec{P}}}|}$ and of $\lim_{\sigma\to0}\frac{\partial E^{\sigma}_{|{{\vec{P}}}|}}{\partial |{{\vec{P}}}|}$, combined with the fundamental theorem of calculus, imply that $E_{{{\vec{P}}}}$ is twice differentiable and $\frac{\partial^2 E_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}\equiv \Sigma_{|{{\vec{P}}}|}$.\ Our paper is organized as follows. In Section \[SectII\], we recall how to construct the ground states of the Hamiltonians $H_{{{\vec{P}}}}^{\sigma}$ and $K_{{{\vec{P}}}}^{\sigma}$ by *iterative analytic perturbation theory*. This section contains an explicit derivation of the formula of the transformed Hamiltonians and of related algebraic identities that will be used later on. In Section \[SectIII\], we first derive inequality (\[eq:I.41\]) and the improved convergence rate of $\{\Phi^{\sigma}_{{{\vec{P}}}}\}$ as $\sigma\to0$, by using some key ingredients described in Section \[SectII\]. Section \[SectIII.1\] is devoted to an analysis of (\[eq:I.36\]) that culminates in the following main results. ***[Theorem]{}*** *For $\alpha$ small enough, $\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}$ converges as $\sigma\to0$. The limiting function $ \Sigma_{|{{\vec{P}}}|}:=\lim_{\sigma\to0}\,\frac{\partial E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}$ is Hölder continuous in ${{\vec{P}}}\in{{\mathcal{S}}}$. The limit $$\label{eq:IV.71} \lim_{\alpha\to0}\Sigma_{|{{\vec{P}}}|}=1\,$$ holds true uniformly in ${{\vec{P}}}\in{{\mathcal{S}}}$.* ***[Corollary]{}*** *For $\alpha$ small enough, the function $E_{{{\vec{P}}}}:=\lim_{\sigma\to0}E_{{{\vec{P}}}}^{\sigma}$, ${{\vec{P}}}\in{{\mathcal{S}}}$, is twice differentiable, and $$\frac{\partial^2E_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}= \Sigma_{|{{\vec{P}}}|}\quad\text{for all}\quad{{\vec{P}}}\in{{\mathcal{S}}}\,.$$*\ [**[Remark]{}**]{} For the complete proof of the construction of the ground states of the Hamiltonians $H_{{{\vec{P}}}}^{\sigma}$ and $K_{{{\vec{P}}}}^{\sigma}$ by *iterative analytic perturbation theory*, the reader is advised to consult ref. [@ChenFroehlichPizzo2]. Sequences of ground state vectors {#SectII} ================================= In this section, we report on results contained in [@ChenFroehlichPizzo2] concerning the ground states of the Hamiltonians $H_{{{\vec{P}}}}^{\sigma_j}$, where ${{\vec{P}}}\in{{\mathcal{S}}}$ and $j\in {\mathbb{N}}_0$, and the existence of a limiting vector for the sequence of ground state vectors of the transformed Hamiltonians, $K_{{{\vec{P}}}}^{\sigma_j}$, where the Bogoliubov transformation used to obtain $K_{{{\vec{P}}}}^{\sigma_j}$ from $H_{{{\vec{P}}}}^{\sigma_j}$ (derived in [@ChFr]) is determined by $$\begin{aligned} b^{\,*}_{{{\vec{k}}},\lambda}&\rightarrow&W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})b^{\,*}_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,=\,b^{\,*}_{{{\vec{k}}},\lambda}-\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\\ b_{{{\vec{k}}},\lambda}&\rightarrow&W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})b_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,=\,b_{{{\vec{k}}},\lambda}-\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\,,\end{aligned}$$ for ${{\vec{k}}}\in {{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}$, with $$\label{eq:II.3} W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,:=\,\exp\big(\alpha^{\frac{1}{2}}\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}d^3k\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\cdot(\vec{\epsilon}_{{{\vec{k}}},\lambda}b^*_{{{\vec{k}}},\lambda}-h.c.)\big)\,,$$ where $\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})$ is defined in (\[eq-III.18\]). Ground states of the Hamiltonians $ H_{\vec{P}}^{\sigma_j}$. {#SectII.0} ------------------------------------------------------------ In [@ChenFroehlichPizzo2], the first step consists in constructing the ground states of the regularized fiber Hamiltonians $ H_{\vec{P}}^{\sigma_j}$. As shown in [@ChenFroehlichPizzo2], $ H_{\vec{P}}^{\sigma_j}$ has a unique ground state, $\Psi_{{{\vec{P}}}}^{\sigma_j}$, that can be constructed by *iterative analytic perturbation theory*, as developed in [@Pizzo]. To recall how this method works some preliminary definitions and results are needed: - We introduce the Fock spaces $$\label{eq:II.4} {{\mathcal{F}}}_{\sigma_j}:={{\mathcal{F}}}^b(L^2(({\mathbb{R}}^3\setminus{{\mathcal{B}}}_{\sigma_j})\times {\mathbb{Z}}_2))\,\quad,\quad\,{{\mathcal{F}}}^{\sigma_j}_{\sigma_{j+1}}:={{\mathcal{F}}}^b(L^2(({{\mathcal{B}}}_{\sigma_j}\setminus{{\mathcal{B}}}_{\sigma_{j+1}})\times {\mathbb{Z}}_2))\,.$$ Note that $${{\mathcal{F}}}_{\sigma_{j+1}}\,=\,{{\mathcal{F}}}_{\sigma_j}\otimes{{\mathcal{F}}}^{\sigma_j}_{\sigma_{j+1}}\,.$$ If not specified otherwise, $\Omega_f$ denotes the vacuum state in anyone of these Fock spaces. Any vector $\phi$ in ${{\mathcal{F}}}_{\sigma_j}$ can be identified with the corresponding vector, $\phi\otimes\Omega_f$, in ${{\mathcal{F}}}$, where $\Omega_f$ is the vacuum in ${{\mathcal{F}}}^{\sigma_j}_{0}$. - Momentum-slice interaction Hamiltonians are defined by $$\Delta H_{{{\vec{P}}}}|^{\sigma_j}_{\sigma_{j+1}}:=\alpha^{\frac{1}{2}}\,{{\vec{\nabla}}}_{{{\vec{P}}}}H_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{A}|^{\sigma_j}_{\sigma_{j+1}}+\frac{\alpha}{2}\,({\vec{A}}|^{\sigma_j}_{\sigma_{j+1}})^2\,,$$ where $$\begin{aligned} \vec{A}|^{\sigma_j}_{\sigma_{j+1}} \, := \,\sum_{\lambda}\, \int_{\mathcal{B}_{\sigma_j}\setminus \mathcal{B}_{\sigma_{j+1}}}\, \frac{d^3k}{\sqrt{ |{{\vec{k}}}| \,}} \, \big\{ {{\vec{\varepsilon}}}_{{{\vec{k}}},\lambda}b^*_{{{\vec{k}}},\lambda}\, + \, {{\vec{\varepsilon}}}^{\,\,*}_{{{\vec{k}}},\lambda} b_{{{\vec{k}}},\lambda} \big\}\,;\end{aligned}$$ - Four real parameters, $\epsilon,\,\rho^+,\,\rho^-$, and $\mu$, will appear in our analysis. They have the properties $$\begin{aligned} & &0<\rho^-<\mu<\rho^+<1-C_{\alpha}<\frac{2}{3}\label{eq:III.8bis}\\ & &0<\epsilon<\frac{\rho^-}{\rho^+}\label{eq:III.9}\end{aligned}$$ where $C_{\alpha}$, with $\frac{1}{3}<C_{\alpha}<1$, for $\alpha$ small enough, is a constant such that the inequality $$\label{eq:II.10} E_{{{\vec{P}}}-{{\vec{k}}}}^{\sigma}>E_{{{\vec{P}}}}^{\sigma}-C_{\alpha}|{{\vec{k}}}|\,$$ holds for all ${{\vec{P}}}\in{{\mathcal{S}}}$ and any ${{\vec{k}}}\neq0$. Here $E_{{{\vec{P}}}-{{\vec{k}}}}^{\sigma}:=\text{inf spec}H_{{{\vec{P}}}-{{\vec{k}}}}^{\sigma}$. We note that $C_{\alpha}\to\frac{1}{3}$, as $\alpha\to 0$; (see Statement ($\mathcal{I}4$) of Theorem 3.1. in [@ChenFroehlichPizzo2]). By iterative analytic perturbation theory (see [@ChenFroehlichPizzo2]), one derives the following results, valid for sufficiently small $\alpha$, depending on our choice of $\Lambda, \epsilon,\rho^-, \mu,$ and $ \rho^+$ (see also Figure 1 below): - $E_{{{\vec{P}}}}^{\sigma_j}$ is an isolated simple eigenvalue of $H_{{{\vec{P}}}}^{\sigma_j}|_{{{\mathcal{F}}}_{\sigma_j}}$ with spectral gap larger or equal to $\rho^-\sigma_j$. Furthermore, $E_{{{\vec{P}}}}^{\sigma_j}$ is also the ground state energy of $H_{{{\vec{P}}}}^{\sigma_j}|_{{{\mathcal{F}}}_{\sigma_{j+1}}}$, and it is an isolated simple eigenvalue of $H_{{{\vec{P}}}}^{\sigma_j}|_{{{\mathcal{F}}}_{\sigma_{j+1}}}$ with spectral gap larger or equal to $\rho^+\sigma_{j+1}$. - The ground-state energies $E_{{{\vec{P}}}}^{\sigma_j}$ and $E_{{{\vec{P}}}}^{\sigma_{j+1}}$ of the Hamiltonians $H_{{{\vec{P}}}}^{\sigma_j}$ and $H_{{{\vec{P}}}}^{\sigma_{j+1}}$, respectively, (acting on the same space ${{\mathcal{F}}}_{\sigma_{j+1}}$) satisfy the inequalities $$0\leq E_{{{\vec{P}}}}^{\sigma_{j+1}} \leq E_{{{\vec{P}}}}^{\sigma_{j}}+c\,\alpha\,\sigma_j^2\,,$$ where $c$ is independent of $j$ and of $\alpha$. - The ground state vectors, $\Psi_{{{\vec{P}}}}^{\sigma_{j+1}}$, of $H_{{{\vec{P}}}}^{\sigma_{j+1}}$ can be recursively constructed starting from $\Psi_{{{\vec{P}}}}^{\sigma_0}\equiv \Omega_f$ with the help of the spectral projection $$\frac{1}{2\pi i}\oint_{\gamma_{j+1}}dz_{j+1}\frac{1}{H_{{{\vec{P}}}}^{\sigma_{j+1}}-z_{j+1}}\,.$$ More precisely, $$\begin{aligned} \Psi_{{{\vec{P}}}}^{\sigma_{j+1}}&:=&\frac{1}{2\pi i}\oint_{\gamma_{j+1}}dz_{j+1}\frac{1}{H_{{{\vec{P}}}}^{\sigma_{j+1}}-z_{j+1}}\Psi_{{{\vec{P}}}}^{\sigma_{j}}\otimes\Omega_f\,\\ &= &\frac{1}{2\pi i}\sum_{n=0}^{\infty}\oint_{\gamma_{j+1}}dz_{j+1}\frac{1}{H_{{{\vec{P}}}}^{\sigma_{j}}-z_{j+1}}[-\Delta H_{{{\vec{P}}}}|^{\sigma_j}_{\sigma_{j+1}}\frac{1}{H_{{{\vec{P}}}}^{\sigma_{j}}-z_{j+1}}]^n\,\Psi_{{{\vec{P}}}}^{\sigma_{j}}\otimes\Omega_f \,,\quad\quad\quad\,\,\label{eq:II.13}\end{aligned}$$ where $\gamma_{j+1}:=\{z_{j+1}\,\in{\mathbb{C}}\,|\,|z_{j+1}-E_{{{\vec{P}}}}^{\sigma_j}|=\mu\sigma_{j+1}\}$, with $\mu$ as in (\[eq:III.8bis\]). $\Psi_{{{\vec{P}}}}^{\sigma_{j+1}}$ is the (unnormalized) ground state vector of $H_{{{\vec{P}}}}^{\sigma_{j+1}}|_{{{\mathcal{F}}}_{\sigma}}$ for any $0\leq\sigma\leq \sigma_{j+1}$. {width="140mm" height="60mm"} Transformed Hamiltonians. ------------------------- In this section, we consider the (Bogoliubov-transformed) Hamiltonians $$K_{{{\vec{P}}}}^{\sigma_j}\,:=\,W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})H_{{{\vec{P}}}}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,$$ with ground state vectors $\Phi_{{{\vec{P}}}}^{\sigma_j}$, $j=0,1,2,3,\dots$. Some algebraic manipulations to express $K_{{{\vec{P}}}}^{\sigma_j}$ in a “canonical form" appear to represent a crucial step before iterative perturbation theory can be applied to the sequence of these transformed Hamiltonians. In addition, some intermediate Hamiltonians, denoted $\hat{K}_{{{\vec{P}}}}^{\sigma_j}$, must be introduced to arrive at the right kind of convergence estimates. The same algebraic relations that are used to obtain the “canonical form" of $K_{{{\vec{P}}}}^{\sigma_j}$ also play an important role in the proof of our main result concerning the limiting behavior, as $\sigma\to 0$, of the second derivative of the ground state energy $E_{{{\vec{P}}}}^{\sigma}$. It is therefore useful to derive the “canonical form" of $K_{{{\vec{P}}}}^{\sigma_j}$ and the relevant algebraic identities in some detail.\ The Feynman-Hellman formula (which holds because $(H_{{{\vec{P}}}}^{\sigma_j})_{{{\vec{P}}}\in {{\mathcal{S}}}}$ is an analytic family of type A, and $E_{{{\vec{P}}}}^{\sigma_j}$ is an isolated eigenvalue) yields the identity $$\label{eq-III.15} {{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\,=\,{{\vec{P}}}-\langle {{\vec{P}}}^f-\alpha^{\frac{1}{2}}{{\vec{A}}}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}\,,$$ where, given an operator $B$ and a vector $\psi$ in the domain of $B$, we use the notation $$\label{eq:III.16} \langle B \rangle_{\psi}\,:=\,\frac{\langle \psi\,,\,B\,\psi \rangle}{\langle \psi\,,\,\psi \rangle}\,.$$ As stated in [@ChenFroehlichPizzo2], for $\alpha$ small enough, $$\sup_{{{\vec{P}}}\in \mathcal{S}}|{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}|<1 \quad \forall j \in {\mathbb{N}}_0.$$ We define $$\begin{aligned} \vec{\beta}^{\sigma_j}&:=&{{\vec{P}}}^f-\alpha^{\frac{1}{2}}{{\vec{A}}}^{\sigma_j}\label{eq-III.17}\\ \delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})&:=&1-\hat{k}\cdot{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}, \quad\hat{k}:=\frac{{{\vec{k}}}}{|{{\vec{k}}}|}\,,\quad |{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}|<1 \label{eq-III.18}\\ c^{\,*}_{{{\vec{k}}},\lambda}&:=&b^{\,*}_{{{\vec{k}}},\lambda}+\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\\ c_{{{\vec{k}}},\lambda}&:=&b_{{{\vec{k}}},\lambda}+\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\,.\end{aligned}$$ We rewrite $H_{{{\vec{P}}}}^{\sigma_j}$ as $$H_{{{\vec{P}}}}^{\sigma_j}\,=\,\frac{({{\vec{P}}}-\vec{\beta}^{\sigma_j})^2}{2}+H^f\,,$$ and, using (\[eq-III.15\]) and (\[eq-III.17\]), $${{\vec{P}}}\,=\,{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}+\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}\,.$$ We then obtain $$\begin{aligned} H_{{{\vec{P}}}}^{\sigma_j}&=&\frac{{{\vec{P}}}^2}{2}-({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}+\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}})\cdot \vec{\beta}^{\sigma_j}+\frac{\vec{\beta}^{{\sigma_j}^2}}{2}+H^f\\ &=&\frac{{{\vec{P}}}^2}{2}+\frac{\vec{\beta}^{{\sigma_j}^2}}{2}-\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}\cdot \vec{\beta}^{\sigma_j}\\ & &+\sum_{\lambda}\int_{{\mathbb{R}}^3\setminus({{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j})}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})b^{\,*}_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k\\ & &+\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})c^{\,*}_{{{\vec{k}}},\lambda}c_{{{\vec{k}}},\lambda}d^3k\\ & &-\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,.\quad\quad\end{aligned}$$ Adding and subtracting $1/2\,\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}^2$, one finds that $$\begin{aligned} \label{eq:II.28bis} H_{{{\vec{P}}}}^{\sigma_j}&=&\frac{{{\vec{P}}}^2}{2}-\frac{\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}^2}{2}+\frac{(\vec{\beta}^{\sigma_j}-\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}})^2}{2}\\ & &+\sum_{\lambda}\int_{{\mathbb{R}}^3\setminus({{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j})}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})b^{\,*}_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k\\ & &+\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})c^{\,*}_{{{\vec{k}}},\lambda}c_{{{\vec{k}}},\lambda}d^3k\\ & &-\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,.\quad\quad\label{eq:II.31bis}\end{aligned}$$ Next, we implement the Bogoliubov transformation $$\begin{aligned} b^{\,*}_{{{\vec{k}}},\lambda}&\rightarrow&W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})b^{\,*}_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,=\,b^{\,*}_{{{\vec{k}}},\lambda}-\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})} \, ,\\ b_{{{\vec{k}}},\lambda}&\rightarrow&W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})b_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,=\,b_{{{\vec{k}}},\lambda}-\alpha^{\frac{1}{2}}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\,,\end{aligned}$$ for ${{\vec{k}}}\in {{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}$, where $W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})$ is defined in (\[eq:II.3\]). It is evident that $W_{\sigma_j}$ acts as the identity on ${{\mathcal{F}}}^b(L^2({{\mathcal{B}}}_{\sigma_j}\times {\mathbb{Z}}_2))$ and on ${{\mathcal{F}}}^b(L^2(({\mathbb{R}}^3\setminus{{\mathcal{B}}}_{\Lambda})\times {\mathbb{Z}}_2))$. We define the vector operators $$\label{eq:II.34bis} \vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}\,:=\,W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\\ -\langle W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}) \rangle_{\Omega_f}\,,$$ noting that, by (\[eq-III.15\]), (\[eq-III.17\]), and (\[eq:II.34bis\]) $$\begin{aligned} \label{eq:III.35bis} \langle \vec{\beta}^{\sigma_j} \rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}&=&{{\vec{P}}}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\\ &= &\frac{\langle \Phi_{{{\vec{P}}}}^{\sigma_j}\,,\,\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle}{\langle \Phi_{{{\vec{P}}}}^{\sigma_j}\,,\,\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle}+\langle W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}) \rangle_{\Omega_f}\,, \label{eq:III.36bis}\end{aligned}$$ where $\Phi_{{{\vec{P}}}}^{\sigma_j}$ is the ground state of the Bogoliubov-transformed Hamiltonian $$K_{{{\vec{P}}}}^{\sigma_j}\,:=\,W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})H_{{{\vec{P}}}}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,.$$ It is easy to see that $$\label{eq:III.38} W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})-\langle \vec{\beta}^{\sigma_j} \rangle_{ \Psi_{{{\vec{P}}}}^{\sigma_j}}=\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}-\langle \vec{\Pi}_{{{\vec{P}}}}^{\sigma_j} \rangle_{\Phi_{{{\vec{P}}}}^{\sigma_j}}\,.$$ After inspecting straightforward operator domain questions (see [@ChenFroehlichPizzo2]), the “canonical form" of $K_{{{\vec{P}}}}^{\sigma_j}$ is given by $$\label{eq:III.39} K_{{{\vec{P}}}}^{\sigma_j}\,=\,\frac{(\vec{\Gamma}^{\sigma_j}_{{{\vec{P}}}})^2}{2}+\sum_{\lambda}\int_{{\mathbb{R}}^3}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})b^*_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k+{{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma_j}\,,$$ where $$\label{eq:II.40} \vec{\Gamma}^{\sigma_j}_{{{\vec{P}}}}\,:=\,\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}-\langle \vec{\Pi}_{{{\vec{P}}}}^{\sigma_j} \rangle_{\Phi_{{{\vec{P}}}}^{\sigma_j}}\,,$$ so that $$\langle \vec{\Gamma}^{\sigma_j}_{{{\vec{P}}}}\rangle_{\Phi_{{{\vec{P}}}}^{\sigma_j}}=0\,,$$ and $$\begin{aligned} {{\mathcal{E}}}^{\sigma_j}_{{{\vec{P}}}}&:=&\frac{{{\vec{P}}}^2}{2}-\frac{({{\vec{P}}}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})^2}{2} \label{eq-II.42}\\ &&-\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})\,\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,.\nonumber\end{aligned}$$ Eqs. (\[eq:III.38\]), (\[eq:III.39\]) follow by using that $$\begin{aligned} W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})c^{\,*}_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})&=&b^{\,*}_{{{\vec{k}}},\lambda}\,,\\ W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})c_{{{\vec{k}}},\lambda}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})&=&b_{{{\vec{k}}},\lambda}\,,\end{aligned}$$ for ${{\vec{k}}}\in{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_j}$.\ An intermediate Hamiltonian, $\hat{K}_{{{\vec{P}}}}^{\sigma_{j+1}}$, is defined by $$\hat{K}_{{{\vec{P}}}}^{\sigma_{j+1}}\,:=\,W_{\sigma_{j+1}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})H^{\sigma_{j+1}}_{{{\vec{P}}}}W_{\sigma_{j+1}}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,,$$ where $$W_{\sigma_{j+1}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\,:=\,\exp\big(\alpha^{\frac{1}{2}}\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}d^3k\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\cdot(\vec{\epsilon}_{{{\vec{k}}},\lambda}b^*_{{{\vec{k}}},\lambda}-h.c.)\big)\,.$$ We decompose $\hat{K}_{{{\vec{P}}}}^{\sigma_{j+1}}$ into several different terms, similarly as $K_{{{\vec{P}}}}^{\sigma_j}$. We recall that $$H_{{{\vec{P}}}}^{\sigma_{j+1}}\,=\,\frac{({{\vec{P}}}-\vec{\beta}^{\sigma_{j+1}})^2}{2}+H^f\,,$$ and, by (\[eq:III.35bis\]), $${{\vec{P}}}\,=\,{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}+\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}\,.$$ It follows that (see also (\[eq:II.28bis\])-(\[eq:II.31bis\])) $$\begin{aligned} H_{{{\vec{P}}}}^{\sigma_{j+1}}&=&\frac{{{\vec{P}}}^2}{2}-({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}+\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}})\cdot \vec{\beta}^{\sigma_{j+1}}+\frac{\vec{\beta}^{{\sigma_{j+1}}^2}}{2}+H^f\\ &=&\frac{{{\vec{P}}}^2}{2}+\frac{\vec{\beta}^{{\sigma_{j+1}}^2}}{2}-\langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}\cdot \vec{\beta}^{\sigma_{j+1}}\\ & &+\sum_{\lambda}\int_{{\mathbb{R}}^3\setminus({{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_{j+1}})}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})b^{\,*}_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k\\ & &+\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})c^{\,*}_{{{\vec{k}}},\lambda}c_{{{\vec{k}}},\lambda}d^3k\\ & &-\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,.\quad\quad\end{aligned}$$ We now add and subtract $1/2 \, \langle \vec{\beta}^{\sigma_j}\rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}^2$ and conjugate the resulting operator with the unitary operator $W_{\sigma_{j+1}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})$. After inspecting straightforward operator domain questions (see [@ChenFroehlichPizzo2]), we find that $$\begin{aligned} \label{eq:II.54bis} \hat{K}_{{{\vec{P}}}}^{\sigma_{j+1}}&=&\frac{(\vec{\Gamma}^{\sigma_j}_{{{\vec{P}}}}+\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j}+\vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j})^2}{2}\\ & &+\sum_{\lambda}\int_{{\mathbb{R}}^3}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})b^{\,*}_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}d^3k+\hat{{{\mathcal{E}}}}_{{{\vec{P}}}}^{\sigma_{j+1}}\,,\end{aligned}$$ where $$\begin{aligned} \vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j}&:=&-\alpha^{\frac{1}{2}}\sum_{\lambda}\int_{{{\mathcal{B}}}_{\sigma_j}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}{{\vec{k}}}\,\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}b_{{{\vec{k}}},\lambda}+h.c.}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,,\label{eq:III.57}\\ & &-\alpha^{\frac{1}{2}}\vec{A}|_{\sigma_{j+1}}^{\sigma_j}\\ \vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j}&:=&\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\sigma_j}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}{{\vec{k}}}\,\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,,\\ & &+\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\sigma_j}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}\,[\vec{\epsilon}_{{{\vec{k}}},\lambda}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}+h.c.]\frac{d^3k}{\sqrt{|{{\vec{k}}}|}}\nonumber\\ \hat{{{\mathcal{E}}}}_{{{\vec{P}}}}^{\sigma_{j+1}}&:= & \frac{{{\vec{P}}}^2}{2}-\frac{({{\vec{P}}}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})^2}{2}\\ &&-\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\Lambda}\setminus{{\mathcal{B}}}_{\sigma_{j+1}}}|{{\vec{k}}}|\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})\,\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\cdot\vec{\epsilon}^{\,\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_j}(\hat{k})}d^3k\,.\nonumber\end{aligned}$$ We also define the operators $$\label{eq:III.61} \hat{\vec{\Pi}}_{{{\vec{P}}}}^{\sigma_j}\,:=\,W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})\,,$$ and $$\label{eq:III.62} \hat{\vec{\Gamma}}^{\sigma_j}_{{{\vec{P}}}}\,:=\,\hat{\vec{\Pi}}_{{{\vec{P}}}}^{\sigma_j}-\langle \hat{ \vec{\Pi}}_{{{\vec{P}}}}^{\sigma_j} \rangle_{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_j}}\,,$$ which are used in the proofs of convergence of the ground state vectors. Here, $\hat{\Phi}_{{{\vec{P}}}}^{\sigma_j}$ denotes the ground state vector of the Hamiltonian $$\hat{K}_{{{\vec{P}}}}^{\sigma_j}:=W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})K_{{{\vec{P}}}}^{\sigma_j}W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})\,.$$ Notice that $$\label{eq:identity} \hat{\vec{\Gamma}}^{\sigma_j}_{{{\vec{P}}}}=W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j}W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})W_{\sigma_{j}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})\,.$$ An important identity used in [@ChenFroehlichPizzo2] and in the sequel of the present paper is ($j\geq 1$) $$\begin{aligned} \label{eq:III.63} \hat{\vec{\Gamma}}^{\sigma_j}_{{{\vec{P}}}}-\vec{\Gamma}^{\sigma_{j-1}}_{{{\vec{P}}}}&=&{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}}+\vec{{{\mathcal{L}}}}^{\sigma_{j-1}}_{\sigma_j}\\ & &+\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\sigma_{j-1}}\setminus{{\mathcal{B}}}_{\sigma_{j}}}{{\vec{k}}}\,\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}}\cdot\vec{\epsilon}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_{j-1}}(\hat{k})}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}}\cdot\vec{\epsilon}^{\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_{j-1}}(\hat{k})}d^3k \nonumber\\ & &+\alpha\sum_{\lambda}\int_{{{\mathcal{B}}}_{\sigma_{j-1}}\setminus{{\mathcal{B}}}_{\sigma_{j}}}\,[\vec{\epsilon}_{{{\vec{k}}},\lambda}\frac{{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}}\cdot\vec{\epsilon}^{\,*}_{{{\vec{k}}},\lambda}}{|{{\vec{k}}}|^{\frac{3}{2}}\delta_{{{\vec{P}}}}^{\sigma_{j-1}}(\hat{k})}+h.c.]\frac{d^3k}{\sqrt{|{{\vec{k}}}|}}\,.\nonumber\end{aligned}$$ Eq. (\[eq:III.63\]) can be derived using (\[eq:II.34bis\]), (\[eq:III.36bis\]), (\[eq:III.38\]), (\[eq:II.40\]), (\[eq:III.61\]), and (\[eq:III.62\]). Convergence of the sequence $\{\Phi_{{{\vec{P}}}}^{\sigma_j}\}_{j=0}^{\infty}$. -------------------------------------------------------------------------------- To pass from momentum scale $j$ to $j+1$, we proceed in two steps: First, we construct an intermediate vector, $\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}$, defined by $$\label{eq:III.64} \hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}\,:=\,\sum_{n=0}^{\infty}\frac{1}{2\pi i}\oint_{\gamma_{j+1}}dz_{j+1}\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}[-\Delta K_{{{\vec{P}}}}|_{\sigma_{j+1}}^{\sigma_j}\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}]^n\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,$$ where $$\begin{aligned} \Delta K_{{{\vec{P}}}}|_{\sigma_{j+1}}^{\sigma_j}&:=& \hat{K}_{{{\vec{P}}}}^{\sigma_{j+1}}-\hat{{{\mathcal{E}}}}_{{{\vec{P}}}}^{\sigma_{j+1}}+{{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma_j}- K_{{{\vec{P}}}}^{\sigma_j}\\ &= &\frac{1}{2}\big(\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j}\cdot(\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j}+ \vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j})+h.c.\big)+\frac{1}{2}(\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j}+ \vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j})^2\,.\label{eq:III.53}\quad\quad\quad\end{aligned}$$ Subsequently, we construct $\Phi_{{{\vec{P}}}}^{\sigma_{j+1}}$ by setting $$\label{eq:II.bis.67} \Phi_{{{\vec{P}}}}^{\sigma_{j+1}}\,:=\,W_{\sigma_{j+1}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j+1}})W_{\sigma_{j+1}}^{*}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}\,.$$ The series in (\[eq:III.64\]) is term-wise well-defined and converges strongly to a non-zero vector, provided $\alpha$ is small enough (*independently* of $j$). The proof of this claim is based on operator-norm estimates of the type used in controlling the Neumann expansion in (\[eq:II.13\]), which requires an estimate of the spectral gap that follows from the unitarity of $W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})$ and Result 1) described after Eq. (\[eq:II.10\]). A key point in our proof of convergence of the sequence $\{\Phi_{{{\vec{P}}}}^{\sigma_j}\}$ is to show that the term $$\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j}\cdot(\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j}+\vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j})+h.c.$$ appearing in (\[eq:III.53\]), which is superficially “marginal" in the infrared, by power counting, is in fact “irrelevant" (using the terminology of renormalization group theory). This is a consequence of the orthogonality condition $$\label{eq:III.55} \langle \Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j} \Phi_{{{\vec{P}}}}^{\sigma_{j}}\rangle=0\,,$$ which, when combined with an inductive argument, implies that $$\|(\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}})^{\frac{1}{2}}\,[\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j}\cdot(\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j\,(+)}+\vec{{{\mathcal{I}}}}_{\sigma_{j+1}}^{\sigma_j})]\,(\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}})^{\frac{1}{2}}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\|$$ (where $\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j\,(+)}$ stands for the part which contains only photon creation operators) is of order ${{\mathcal{O}}}(\epsilon^{\eta j})$, for some $\eta >0$ specified in [@ChenFroehlichPizzo2]. In particular, this suffices to show that $$\label{eq:III.58} \| \hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}-\Phi_{{{\vec{P}}}}^{\sigma_{j}}\|\leq{{\mathcal{O}}}(\epsilon^{\frac{j+1}{2}(1-\delta)})\,$$ for any $0<\delta<1$ provided $\alpha$ is sufficiently small. Finally, in Theorem 3.1 of ref. [@ChenFroehlichPizzo2], it is proven that there is a non-zero vector in the Hilbert space corresponding to $\lim_{j\to\infty}\Phi_{{{\vec{P}}}}^{\sigma_{j}}$, and that the rate of convergence is at least ${{\mathcal{O}}}(\sigma^{\frac{1}{2}(1-\delta)})$ for any $0<\delta<1$ provided $\alpha$ is sufficiently small.\ [**[Remark]{}**]{} In Theorem 3.1 of ref. [@ChenFroehlichPizzo2], for $\lim_{j\to\infty}\Phi_{{{\vec{P}}}}^{\sigma_{j}}$, the range of values of $\alpha$ such that the rate of convergence, ${{\mathcal{O}}}(\sigma^{\frac{1}{2}(1-\delta)})$, holds is not claimed to be uniform in $\delta$. The stronger result obtained in the next section (see (\[eq:III.bis.2\]) and (\[eq:III.34bis\])) implies that this range (corresponding to the rate ${{\mathcal{O}}}(\sigma^{\frac{1}{2}(1-\delta)})$) is actually $\delta$-independent. ### Key ingredients {#SectII.2} To prove convergence of the sequence $\{\Phi_{{{\vec{P}}}}^{\sigma_j}\}$ of ground state vectors of the Hamiltonians $K_{{{\vec{P}}}}^{\sigma_j}$, some further conditions on $\alpha$, $\epsilon$, and $\mu$ (see (\[eq:III.8bis\]), (\[eq:III.9\])) are required, in particular an upper bound on $\mu$ and an upper bound on $\epsilon$ strictly smaller than the ones imposed by inequalities (\[eq:III.8bis\]), (\[eq:III.9\]); (for details, see Lemma A.3 in [@ChenFroehlichPizzo2]). We note that the more restrictive conditions on $\mu$ and $\epsilon$ imply new bounds on $\rho^{-}$ and $\rho^{+}$. Moreover, $\epsilon$ must satisfy a lower bound $\epsilon>C \alpha^{\frac{1}{2}}$, with a multiplicative constant $C>0$ sufficiently large. Some key inequalities needed in our analysis of the convergence properties of $\{\Phi_{{{\vec{P}}}}^{\sigma_j}\}$ are summarized below. They will be marked by the symbol $(\mathcal{B})$. In order to reach some important improvements in our estimates of the convergence rate of $\Phi_{{{\vec{P}}}}^{\sigma_{j}}$, as $j\to\infty$ (discussed in the next section), a refined estimate is needed that is stated in $(\mathcal{B}2)$, and a new inequality, see $(\mathcal{B}5)$, (analogous to $(\mathcal{B}3)$ and $(\mathcal{B}4)$) is required. - *Estimates on the shift of the ground state energy and its gradient*\ There are constants $C_1$, $C_{2}'$ such that the following inequalities hold.\ $(\mathcal{B}1)$ $$|E_{{{\vec{P}}}}^{\sigma_j}-E_{{{\vec{P}}}}^{\sigma_{j+1}}|\leq C_1\,\alpha\,\epsilon^j\,;$$ see [@ChenFroehlichPizzo2].\ - $(\mathcal{B}2)$ $$\label{eq:III.73} |{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j+1}}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}}|\leq C'_2\big(\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}-\Phi_{{{\vec{P}}}}^{\sigma_{j}}\|+\alpha^{\frac{1}{4}}\epsilon^{j+1}\big)\,.$$ This is an improvement over a corresponding estimate in [@ChenFroehlichPizzo2]: It can be proven *after* the results stated in Theorem 3.1 in [@ChenFroehlichPizzo2], in particular the uniform bound from below on $\langle \Phi_{{{\vec{P}}}}^{\sigma_{j}}, \Phi_{{{\vec{P}}}}^{\sigma_{j}}\rangle$, $\langle \Phi_{{{\vec{P}}}}^{\sigma_{j}}, \Phi_{{{\vec{P}}}}^{\sigma_{j}}\rangle >\frac{2}{3}$, and following the steps in the proof of Lemma A.2 in [@ChenFroehlichPizzo2]. - *Bounds relating expectations of operators to expectations of their absolute values*\ There are constants $C_3, C_4, C_5>1$ such that the following inequalities hold.\ $(\mathcal{B}3)$ For $z_{j+1}\in\gamma_{j+1}$, $$\begin{aligned} & & \big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big|(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \\ & &\leq C_3\big|\big\langle(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big\rangle\big|\,,\end{aligned}$$ where $(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}$ is the $i^{th}$ component of $\vec{\Gamma}_{{{\vec{P}}}}^{\sigma_j}$. $(\mathcal{B}4)$ For $z_{j+1}\in\gamma_{j+1},$ $$\begin{aligned} & & \big\langle ({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^l\,(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big|({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^l\,(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \\ & &\leq C_4\big|\big\langle({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^l\,(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^l\,(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big\rangle\big|\,,\quad\quad\quad\quad\end{aligned}$$ where $({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^l$ is the $l^{th}$ component of $\vec{{{\mathcal{L}}}}_{\sigma_{j+1}}^{\sigma_j\,(+)}$. $(\mathcal{B}5)$ For $z_{j+1}\in\gamma_{j+1},$ $$\begin{aligned} \label{eq:II.78} & & \big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big|^2(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \\ & &\leq C_5\big|\big\langle(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,(\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}})^2(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big\rangle\big|\,.\nonumber\end{aligned}$$ To prove $(\mathcal{B}3)$ and $(\mathcal{B}4)$, it suffices to exploit the fact that the spectral support (with respect to $K_{{{\vec{P}}}}^{\sigma_j}$) of the two vectors $(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}$ and $({{\mathcal{L}}}_{\sigma_{j+1}}^{\sigma_j\,(+)})^{l}(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}$ is strictly above the ground state energy of $K^{\sigma_j}_{{{\vec{P}}}}$, since they are both orthogonal to the ground state, $\Phi_{{{\vec{P}}}}^{\sigma_{j}}$, of this operator. In the proof of bound $(\mathcal{B}5)$, it is also required that $\rho^->3\mu\epsilon$, as will be assumed in the following.\ ***[Remarks]{}*** \(1) The constants $C_1,\dots, C_5$ are independent of $\alpha, \epsilon, \mu$, and $j\in{\mathbb{N}}$, provided that $\alpha, \epsilon$, and $ \mu$ are sufficiently small. (2) For the convenience of the reader, we recapitulate the relations between the parameters entering the construction: $$\begin{aligned} & &0<\rho^-<\mu<\rho^+<1-C_{\alpha}<\frac{2}{3}\,,\label{eq:II.79bis}\\ & &0<\epsilon<\frac{\rho^-}{\rho^+}\,,\\ & &\epsilon>C \alpha^{1/2}\,, \label{eq:II.83b}\\ & &\rho^->3\mu\epsilon\,.\label{eq:II.82bisbis}\end{aligned}$$ Moreover, we stress that the final result is a small coupling result, i.e., valid for small values of $\alpha$, and that, for technical reasons, small values of the parameters $\epsilon, \mu$ within the constraints listed above (that imply more restrictive bounds on $ \rho^-, \rho^+$) are required.\ The crucial estimate for the bound on $\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}-\Phi_{{{\vec{P}}}}^{\sigma_{j}}$ obtained in [@ChenFroehlichPizzo2] (see (\[eq:III.58\])) is $$\label{eq:III.65} \big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\leq \frac{R_0}{\alpha \epsilon^{j\delta}}\,,$$ where $R_0$ is independent of $j$, and $\delta$, $0<\delta<1$, can be taken arbitrarily small for $\alpha$ and $ \epsilon$ sufficiently small (depending on $\delta$). This estimate will be improved in the next section. As a consequence, our estimate of the convergence rate of $\{\Phi_{{{\vec{P}}}}^{\sigma_{j}}\}$ will be improved. As a corollary, the second derivative of $E^{\sigma}_{{{\vec{P}}}}$ is proven to converge, as $\sigma\to0$. Improved estimate of the convergence rate of $\{\Phi_{{{\vec{P}}}}^{\sigma}\}$, as $\sigma\to0$, and uniform bound on the second derivative of $E^{\sigma}_{{{\vec{P}}}}$. {#SectIII} ========================================================================================================================================================================== Our arguments in Section III rely on the results previously proven in [@ChenFroehlichPizzo2] and described in Section II, which hold for $\alpha$ small enough. Therefore, in the following, we assume the constraints (\[eq:II.79bis\])-(\[eq:II.82bisbis\]), and we make use of the estimates on the spectral gaps (see $(\mathcal{A}1)$ in Section \[SectII.0\]) and of the bounds $(\mathcal{B}1)$-$(\mathcal{B}5)$ (see Section \[SectII.2\]). We also make use of the lower bounds $$\label{eq:III.3bis} \langle \hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}, \hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j+1}}\rangle\,,\,\langle \Phi_{{{\vec{P}}}}^{\sigma_{j}}, \Phi_{{{\vec{P}}}}^{\sigma_{j}}\rangle >\frac{2}{3}$$ uniformly in $j\in{\mathbb{N}}_0$, which appear in the proof of Theorem 3.1 of ref. [@ChenFroehlichPizzo2]. Assuming these bounds we can simplify the proof by induction in the theorem below.\ \[Th.induction\] For $\alpha, \epsilon$ sufficiently small (depending on $\delta$), the inequality $$\label{eq:IV.1} \big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\leq \frac{{{\mathcal{R}}}_0}{\alpha^{\frac{1}{2}}\, \epsilon^{2j\delta}}\,$$ holds true where $0<\delta<1$ and ${{\mathcal{R}}}_0$ is a constant independent of $j\in{\mathbb{N}}_0:={\mathbb{N}}\cup0$. Furthermore, for ${{\mathcal{R}}}_0$ and $\alpha$ small enough, inequality (\[eq:IV.1\]) implies $$\label{eq:III.bis.2} \|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\leq\alpha^{\frac{1}{4}}\epsilon^{j(1-\delta)}\,.$$ *Proof by induction.* - ** We assume that, at scale $j-1(\geq 0)$, the following estimate holds $$\label{eq:IV.3} \big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big| \leq \frac{{{\mathcal{R}}}_0}{\alpha^{\frac{1}{2}}\, \epsilon^{2(j-1)\delta}}\,.$$ This estimate readily implies that, for ${{\mathcal{R}}}_0$ and $ \alpha$ small enough, but uniformly in $j$, $$\begin{aligned} \label{eq:III.70} & &\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\,\\ & =&\,\|\sum_{n=1}^{\infty}\frac{1}{2\pi i}\oint_{\gamma_{j}}dz_{j}\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}[-\Delta K_{{{\vec{P}}}}|_{\sigma_{j}}^{\sigma_{j-1}}\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}]^n\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\,\nonumber\\ &\leq&\alpha^{\frac{1}{4}}\epsilon^{j(1-\delta)}\,.\label{eq:III.71}\end{aligned}$$ An improved estimate on $\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|$ is based on the following bounds: - $$\|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}\Delta K_{{{\vec{P}}}}|_{\sigma_{j}}^{\sigma_{j-1}}\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\leq {{\mathcal{O}}}({{\mathcal{R}}}_0^{\frac{1}{2}}\alpha^{\frac{1}{4}}\epsilon^{j(1-\delta)})\,,$$ whose proof requires the use of the “pull-through formula" (see, e.g., [@Schweber]), a Neumann expansion of the resolvent, the inequality in Eq. (\[eq:II.78\]), and Eq. (\[eq:IV.3\]); the reader can follow the similar steps used in Lemma A3 of ref. [@ChenFroehlichPizzo2]; - $$\|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}\Delta K_{{{\vec{P}}}}|_{\sigma_{j}}^{\sigma_{j-1}}\|_{{{\mathcal{F}}}_{\sigma_j}}\leq{{\mathcal{O}}}(\alpha^{\frac{1}{2}})\,;$$ this estimate can be derived from standard bounds and using the “pull-through formula". - ** By unitarity of $W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})W_{\sigma_{j}}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})$, we have that $$\begin{aligned} \lefteqn{\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle\big|}\nonumber\\ &= &\big|\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big(\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big)^2(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\,.\end{aligned}$$ For $\alpha$ small enough and $\epsilon>C\,\alpha^{\frac{1}{2}}$, where $C>0$ is large enough, we may use $(\mathcal{B}1)$ to re-expand the resolvent and find that $$\begin{aligned} \lefteqn{\big|\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big(\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big)^2(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|}\\ & &\leq 2\big|\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|^2(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\,.\end{aligned}$$ It follows that $$\begin{aligned} & &2\big|\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|^2(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\\ & &\leq 4\big \|\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|\,((\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\Phi_{{{\vec{P}}}}^{\sigma_{j-1}})\big\|^2 \label{eq:III.79} \\ & & \quad+4\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big| \,.\,\,\,\label{eq:III.80}\end{aligned}$$ Our recursion, combined with (\[eq:II.78\]), relates (\[eq:III.80\]) to the initial expression in (\[eq:IV.1\]), with $j$ replaced by $j-1$, while (\[eq:III.79\]) is a remainder term. Next we note that $$\begin{aligned} & & 4\big \|\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|\,((\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\Phi_{{{\vec{P}}}}^{\sigma_{j-1}})\big\|^2\\ &\leq & 8\big \|\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|\,((\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_{j}})^{i}\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}})\big\|^2\label{eq:III.82}\\ & &+8\big \|\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|\,(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}(\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}})\big\|^2\label{eq:III.83}\\ &\leq &\frac{\mathcal{R}_1}{\epsilon^{2j\delta}}+\frac{\mathcal{R}_2}{\epsilon^{2j\delta}}\,.\label{eq:III.84}\end{aligned}$$ Here $\mathcal{R}_1\leq{{\mathcal{O}}}(\epsilon^{-2})$ and $\mathcal{R}_2\leq{{\mathcal{O}}}(\epsilon^{-2})$ are constants independent of $\alpha$, $\mu$, and $j\in{\mathbb{N}}$, provided that $\alpha$, $\mu$ are sufficiently small, and $\epsilon>C\alpha^{\frac{1}{2}}$. In detail: - Property ($\mathcal{B}4$) and the two norm-bounds $$\|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\|_{{{\mathcal{F}}}_{\sigma_j}}\leq{{\mathcal{O}}}(\epsilon^{-(j+1)})\quad,\quad\quad\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\leq\alpha^{\frac{1}{4}}\epsilon^{j(1-\delta)}$$ (see (\[eq:III.70\])) justify the step from (\[eq:III.83\]) to (\[eq:III.84\]); - concerning the step from (\[eq:III.82\]) to (\[eq:III.84\]), it is enough to consider Eq. (\[eq:III.63\]) and the two bounds $$\label{eq:III.20b} \|({{\mathcal{L}}}^{\sigma_{j-1}}_{\sigma_j})^{i}\hat{\Phi}_{{{\vec{P}}}}^{\sigma_j}\|\leq{{\mathcal{O}}}(\alpha^{\frac{1}{2}}\epsilon^{j-1})\quad,\quad\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\leq\alpha^{\frac{1}{4}}\epsilon^{j(1-\delta)}\,.$$ (hint: for the first inequality in (\[eq:III.20b\]), use the expression in (\[eq:III.64\]).) To bound the term (\[eq:III.80\]), we use $(\mathcal{B}5)$ and the key orthogonality property (\[eq:III.55\]). For $z_{j}\in\gamma_{j}$ and $z_{j+1}\in\gamma_{j+1}$, we find that for $\epsilon/\rho^{-}$ sufficiently small $$\begin{aligned} & &4\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big|^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big|\\ & &\leq 4C_5\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j+1}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big|\quad\quad\quad \label{eq:III.87}\\ & &\leq 8C_5^2\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big|\,.\label{eq:III.88}\end{aligned}$$ In passing from (\[eq:III.87\]) to (\[eq:III.88\]), we again use the constraint on the spectral support (with respect to $K_{{{\vec{P}}}}^{\sigma_{j-1}}$) of the vector $(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}$. Assuming that the parameters $\epsilon$ and $\alpha$ are so small that the previous constraints are fulfilled and that $$\label{eq:III.89} 0<{{\mathcal{R}}}_1+{{\mathcal{R}}}_2\leq\,(1-8C_5^2\epsilon^{2\delta})\frac{{{\mathcal{R}}}_0}{\alpha^{\frac{1}{2}}}\,,$$ we then conclude that $$\begin{aligned} & &\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\,,\,\big(\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-z_{j+1}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j}}\big \rangle \big|\\ &\leq &\frac{\mathcal{R}_1}{\epsilon^{2j\delta}}+\frac{\mathcal{R}_2}{\epsilon^{2j\delta}}\\ & &+8C_5^2\big|\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\,,\,\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}\big)^2(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big \rangle \big|\,\quad\\ &\leq&\frac{{{\mathcal{R}}}_0}{\alpha^{\frac{1}{2}}\, \epsilon^{2j\delta}}\,.\end{aligned}$$ Notice that the bound in (\[eq:III.89\]) induces a $\delta-$dependent constraint on the admissible values of $\epsilon$ and, due to (\[eq:II.83b\]), on $\alpha$. - ** Since $$(\Gamma_{{{\vec{P}}}}^{\sigma_{0}})^{i}\equiv ({{\vec{P}}}^f)^{i}\quad,\quad\Phi_{{{\vec{P}}}}^{\sigma_{0}}\equiv\Omega_f\,,$$ inequality (\[eq:IV.1\]) is trivially fulfilled for $j=0$; thus (\[eq:IV.1\]) holds for all $j\in{\mathbb{N}}_0$ and for ${{\mathcal{R}}}_0$ *arbitrarily* small, provided $\alpha$ is small enough. As we explain below, an improved estimate of the rate of convergence of the sequence $\{\Phi_{\vec{P}}^{\sigma_{j}}\}_{j=0}^{\infty}$ follows from the bound in (\[eq:III.bis.2\]), but we stress that only the estimates in Eqs. (\[eq:IV.1\]), (\[eq:III.bis.2\]) will be used for the uniform bound on the second derivative of $E^{\sigma}_{{{\vec{P}}}}$ in next section. In fact, one can combine the bound in Eq. (\[eq:III.bis.2\]) with the estimate $$\label{eq:III.30bis} \|\Phi_{\vec{P}}^{\sigma_{j}}-\widehat{\Phi}_{\vec{P}}^{\sigma_{j}}\| \, \leq \, C \, \alpha^{\frac{1}{2}} \,|{{\vec{\nabla}}}E_{\vec{P}}^{\sigma_{j-1}}-{{\vec{\nabla}}}E_{\vec{P}}^{\sigma_{j}}|\,|\ln(\epsilon^j)|\,,$$ where $C$ is independent of $\alpha$, $\epsilon$, $\mu$, and $j\in{\mathbb{N}}$, provided that $\alpha$, $\epsilon$, and $\mu$ are sufficiently small. The estimate in (\[eq:III.30bis\]) is obtained starting from the definition in Eq. (\[eq:II.bis.67\]) and using the soft photon bound $$\label{eq-II-9bisbis} \| \, b_{{{\vec{k}}},\lambda}\Psi_{{{\vec{P}}}}^{\sigma_j} \, \| \, \leq C \, \alpha^{1/2} \, \frac{{{\bf 1}}_{\sigma_j,\Lambda}({{\vec{k}}})}{|{{\vec{k}}}|^{3/2}}\,,\quad {{\bf 1}}_{\sigma_j,\Lambda}({{\vec{k}}}):=\{{{\vec{k}}}\,:\,\sigma_j<|{{\vec{k}}}|\leq \Lambda\}\,,$$ that follows from inequality (\[eq:II.10\]) and the identity $$\label{eq-II-9bis} b_{{{\vec{k}}},\lambda}\Psi_{{{\vec{P}}}}^{\sigma_j}\,=\,-\, \alpha^{\frac{1}{2}}\,\frac{{{\bf 1}}_{\sigma_j,\Lambda}({{\vec{k}}})}{|{{\vec{k}}}|^{\frac{1}{2}}} \, \frac{1}{H_{{{\vec{P}}}-{{\vec{k}}}}^{\sigma_j}+|{{\vec{k}}}|-E_{{{\vec{P}}}}^{\sigma_j}} \, {{\vec{\varepsilon}}}_{{{\vec{k}}},\lambda}\cdot{{\vec{\nabla}}}_{{{\vec{P}}}} H_{{{\vec{P}}}}^{\sigma_j}\,\Psi_{{{\vec{P}}}}^{\sigma_j}\,,$$ which is derived in [@ChFr] by using a “pull-through argument". By a standard procedure (see, e.g., [@Pizzo]), one obtains similar results for the ground state vectors of the $\sigma$-dependent Hamiltonians $K_{{{\vec{P}}}}^{\sigma}$, for arbitrary $\sigma>0$. A precise statement concerning the rate of convergence is as follows: The *normalized* ground state vectors (that, with an abuse of notation, we call $\Phi_{{{\vec{P}}}}^{\sigma}$) $$\Phi_{{{\vec{P}}}}^{\sigma}\,:=\frac{\frac{1}{2\pi i}\oint_{\gamma_{\sigma}}dz \frac{1}{K_{{{\vec{P}}}}^{\sigma}-z}\Omega_f }{\|\frac{1}{2\pi i}\oint_{\gamma_{\sigma}}\frac{1}{K_{{{\vec{P}}}}^{\sigma}-z}\Omega_f \|}\,,$$ where $\gamma_{\sigma}:=\{z\in{\mathbb{C}}\,|\,|z-E_{{{\vec{P}}}}^{\sigma}|=\frac{\rho^{-}}{2}\,\sigma\}$, converge strongly to a vector $\Phi_{{{\vec{P}}}}$, as $\sigma\to0$, with $$\label{eq:III.34bis} \|\Phi_{{{\vec{P}}}}^{\sigma}-\Phi_{{{\vec{P}}}}\|\leq {{\mathcal{O}}}(\alpha^{\frac{1}{4}}\,\big(\frac{\sigma}{\Lambda}\big)^{1-\delta})$$ for any $0<\delta(<1)$, provided $\alpha$ is in an interval $(0,\alpha_{\delta})$ where $\alpha_{\delta}>0$, $\alpha_{\delta}\to 0$ as $\delta\to0$. The $\delta$-dependence of the interval $\alpha_{\delta}$ is an indirect consequence of the upper bound on $\epsilon$ that must be imposed through (\[eq:III.89\]) to implement the proof by induction. The relations (\[eq:II.79bis\])-(\[eq:II.82bisbis\]) induce a $\delta$-dependence on the other parameters and in particular on $\alpha$. Another $\delta$-type dependence of the estimated rate of convergence, ${{\mathcal{O}}}(\big(\frac{\sigma}{\Lambda}\big)^{1-\delta})$, comes from the logarithmic term in (\[eq:III.30bis\]). However, this does not spoil the uniformity in $\delta$ of the interval of admissible values of $\alpha$ but only affects the multiplicative constant on the R.H.S. of (\[eq:III.34bis\]). Moreover, with further work, the estimate in Eq. (\[eq:III.30bis\]) can be improved to remove the logarithmic term. Convergence of the second derivative of the ground state energy $E^{\sigma}_{{{\vec{P}}}}$. {#SectIII.1} ------------------------------------------------------------------------------------------- Because of rotational symmetry we have that $ E^{\sigma}_{{{\vec{P}}}}\equiv E^{\sigma}_{|{{\vec{P}}}|}$. Moreover, $(H_{{{\vec{P}}}}^{\sigma})_{{{\vec{P}}}\in{{\mathcal{S}}}}$ is an analytic family of type A in ${{\vec{P}}}\in{{\mathcal{S}}}$, with an isolated eigenvalue $E^{\sigma}_{|{{\vec{P}}}|}$. Thus, the second derivative $\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}$ is well defined and $$\label{eq:IV.33a} \frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}\,=\,\partial_i^2E^{\sigma}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}\,,\quad i=1,2,3\,,$$ where $\partial_i\,:=\,\frac{\partial}{\partial P^i}$. Without loss of generality, the following results are proven for the standard sequence $(\sigma_j)_{j=0}^{\infty}$ of infrared cutoffs. By simple arguments (see [@Pizzo]), limiting behavior as $\sigma\to0$ is shown to be “sequence-independent". By analytic perturbation theory we have that $$\begin{aligned} & &\partial_i^2E^{\sigma_j}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}\\ &=&1-2\langle \frac{1}{2\pi i}\oint_{\gamma_j}\frac{1}{H^{\sigma_j}_{{{\vec{P}}}}-z_j}[P^{i}-(\beta^{\sigma_j})^{i}]\frac{1}{H^{\sigma_j}_{{{\vec{P}}}}-z_j}dz_j\,\Psi_{{{\vec{P}}}}^{\sigma_j},\,[P^{i}-(\beta^{\sigma_j})^{i}]\Psi_{{{\vec{P}}}}^{\sigma_j}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}} \,, \nonumber\end{aligned}$$ here $\Psi_{{{\vec{P}}}}^{\sigma_j}$ is the normalized ground state eigenvector of $H_{{{\vec{P}}}}^{\sigma_j}$. Next, we make use of the Bogoliubov transformation implemented by $W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})$ to show that $$\begin{aligned} & &\langle \frac{1}{2\pi i}\oint_{\gamma_j}\frac{1}{H^{\sigma_j}_{{{\vec{P}}}}-z_j}[P^{i}-(\beta^{\sigma_j})^{i}]\frac{1}{H^{\sigma_j}_{{{\vec{P}}}}-z_j}dz_j\,\Psi_{{{\vec{P}}}}^{\sigma_j},\,[P^{i}-(\beta^{\sigma_j})^{i}]\Psi_{{{\vec{P}}}}^{\sigma_j}\rangle \\ &=&\frac{1}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|^2}\langle \frac{1}{2\pi i}\oint_{\gamma_j}\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}[P^{i}-W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})(\beta^{\sigma_j})^{i}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})]\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}dz_j\,\Phi_{{{\vec{P}}}}^{\sigma_j}\,,\nonumber\\ & &\quad\quad\quad\quad\quad,\,[P^{i}-W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})(\beta^{\sigma_j})^{i}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})]\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle \,,\end{aligned}$$ where $\Phi_{{{\vec{P}}}}^{\sigma_j}$ is the ground state eigenvector of $K_{{{\vec{P}}}}^{\sigma_j}$ (iteratively constructed in Section II). Recalling the definitions $$\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}\,:=\,W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}) -\langle W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}) \rangle_{\Omega_f}\,,$$ $$\vec{\Gamma}^{\sigma_j}_{{{\vec{P}}}}\,:=\,\vec{\Pi}_{{{\vec{P}}}}^{\sigma_j}-\langle \vec{\Pi}_{{{\vec{P}}}}^{\sigma_j} \rangle_{\Phi_{{{\vec{P}}}}^{\sigma_j}}\,,$$ and because of the identity (Feynman-Hellman, see (\[eq:III.36bis\])) $$\begin{aligned} \langle \vec{\beta}^{\sigma_j} \rangle_{\psi_{{{\vec{P}}}}^{\sigma_j}}&=&{{\vec{P}}}-{{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}\\ &= &\langle \vec{\Pi}_{{{\vec{P}}}}^{\sigma_j} \rangle_{\Phi_{{{\vec{P}}}}^{\sigma_j}}+\langle W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})\vec{\beta}^{\sigma_j}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j}) \rangle_{\Omega_f}\,,\end{aligned}$$ we find that $$P^{i}-W_{\sigma_j}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})(\beta^{\sigma_j})^{i}W_{\sigma_j}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_j})=-(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}+\partial_iE_{{{\vec{P}}}}^{\sigma_j}\,;$$ hence, $$\begin{aligned} & &\partial_i^2E^{\sigma_j}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}\\ &=&1-2\frac{1}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|^2}\langle \frac{1}{2\pi i}\oint_{\gamma_j}\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}[\partial_iE_{{{\vec{P}}}}^{\sigma_j}-(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}]\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}dz_j\,\Phi_{{{\vec{P}}}}^{\sigma_j}\,,\nonumber\\ & &\quad\quad\quad\quad\quad\quad\quad\quad\quad\quad\,,\,[\partial_iE_{{{\vec{P}}}}^{\sigma_j}-(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}]\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}} \,.\end{aligned}$$ Using the eigenvalue equation $$K^{\sigma_j}_{{{\vec{P}}}}\Phi_{{{\vec{P}}}}^{\sigma_j}=E_{{{\vec{P}}}}^{\sigma_j}\Phi_{{{\vec{P}}}}^{\sigma_j},$$ the terms proportional to $(\partial_iE_{{{\vec{P}}}}^{\sigma_j})^2$ and to the mixed terms – i.e., proportional to the product of $\partial_iE_{{{\vec{P}}}}^{\sigma_j}$ and $(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}$ – are seen to be identically $0$, because the contour integral vanishes for each $i=1,2,3$; e.g., $$\begin{aligned} \label{eq:III.48} \lefteqn{\oint_{\gamma_j}\langle \frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}[\partial_iE_{{{\vec{P}}}}^{\sigma_j}]\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}\,\Phi_{{{\vec{P}}}}^{\sigma_j},\,[\partial_iE_{{{\vec{P}}}}^{\sigma_j}]\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle d\bar{z}_j\quad\quad}\quad\\ &=&\oint_{\gamma_j}\langle\Phi_{{{\vec{P}}}}^{\sigma_j}\,,\,\Phi_{{{\vec{P}}}}^{\sigma_j}\rangle \big(\frac{\partial_iE_{{{\vec{P}}}}^{\sigma_j}}{E^{\sigma_j}_{{{\vec{P}}}}-\bar{z_j}}\big)^2d\bar{z}_j=0\,.\nonumber\end{aligned}$$ It follows that $$\begin{aligned} & &\partial_i^2E^{\sigma_j}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}\\ &=&1+\frac{1}{\pi i}\oint_{\gamma_j}d\bar{z}_j\langle \frac{1}{K^{\sigma}_{{{\vec{P}}}}-z_j}\,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{1}{K^{\sigma}_{{{\vec{P}}}}-z_j}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|},\,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}}\,\quad\\ &=&1+\frac{1}{\pi i}\oint_{\gamma_j}d\bar{z}_j\frac{1}{E^{\sigma_j}_{{{\vec{P}}}}-\bar{z_j}}\langle \,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-z_j}\,\,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\,,\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\rangle |_{{{\vec{P}}}=P^{i}\hat{i}}\,.\quad\quad\quad\quad\label{eq:IV.48}\end{aligned}$$\ We are now ready for the key estimate. \[LemmaIII.1\] The estimate below holds true ($j\in {\mathbb{N}}$): $$\begin{aligned} \label{eq:IV.19} \lefteqn{\big|\oint_{\gamma_{j-1}}\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j-1}}(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\big \rangle \frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j-1}} d\bar{z}_{j-1}}\,\nonumber\\ & &\quad-\oint_{\gamma_{j}}\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-\bar{z}_{j}}(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\big \rangle \frac{1}{E^{\sigma_j}_{{{\vec{P}}}}-\bar{z_j}} d\bar{z}_{j}\big|\leq \epsilon^{j(1-2\delta)}\,,\quad\quad\quad\label{eq:IV.49}\end{aligned}$$ for any $0<\delta(<1/2)$, and for $\alpha$ and $\epsilon$ small enough depending on $\delta$. *Proof.* By unitarity of $W_{\sigma_{j}}({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j-1}})W_{\sigma_{j}}^*({{\vec{\nabla}}}E_{{{\vec{P}}}}^{\sigma_{j}})$, $$\begin{aligned} \lefteqn{\oint_{\gamma_{j}}\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_j}-\bar{z}_{j}}(\Gamma_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\big \rangle \frac{1}{E^{\sigma_j}_{{{\vec{P}}}}-\bar{z}_j} d\bar{z}_{j}}\\ &=&\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-\bar{z}_{j}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle \frac{1}{E^{\sigma_{j}}_{{{\vec{P}}}}-\bar{z}_{j}}d\bar{z}_{j}\,.\end{aligned}$$ By assumption, $\alpha$ is so small that the Neumann series expansions of the resolvents below converge in ${{\mathcal{F}}}^{b}_{\sigma_j}$: $$\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-\bar{z}_{j}}=\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}+\Sigma_{1}^{\infty}(K_{{{\vec{P}}}}^{\sigma_{j-1}},\bar{z}_{j})\,,$$ $$\frac{1}{E^{\sigma_{j}}_{{{\vec{P}}}}-\bar{z}_{j}}=\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z_{j}}}+\Delta(E^{\sigma_{j-1}}_{{{\vec{P}}}},\bar{z_{j}})\,,$$ where: $$\begin{aligned} & &\Sigma_{1}^{\infty}(K_{{{\vec{P}}}}^{\sigma_{j-1}},\bar{z}_{j})\\ &:=&\sum_{l=1}^{\infty}\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}[-(\Delta K_{{{\vec{P}}}}|^{\sigma_{j-1}}_{\sigma_j}+\hat{{{\mathcal{E}}}}_{{{\vec{P}}}}^{\sigma_j}-{{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma_{j-1}})\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}]^l\,,\nonumber\end{aligned}$$ and $\Delta K_{{{\vec{P}}}}|^{\sigma_{j-1}}_{\sigma_j}$ is defined in Eq. (\[eq:III.53\]); $$\Delta(E^{\sigma_{j-1}}_{{{\vec{P}}}},\bar{z}_{j}):=\frac{1}{E^{\sigma_{j}}_{{{\vec{P}}}}-\bar{z}_{j}}(E^{\sigma_{j-1}}_{{{\vec{P}}}}-E^{\sigma_{j}}_{{{\vec{P}}}})\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}}\,.$$ We proceed by using the obvious identity: $$\begin{aligned} & &\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_j}-\bar{z}_{j}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\frac{1}{E^{\sigma_{j}}_{{{\vec{P}}}}- \bar{z}_{j}} d\bar{z}_{j}\\ &=&\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}} d\bar{z}_{j}\,\label{eq:IV.57}\\ & &+\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\Sigma_{1}^{\infty}(K_{{{\vec{P}}}}^{\sigma_{j-1}},\bar{z}_{j})(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle \frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}} d\bar{z}_{j}\,\,\,\,\quad\quad\label{eq:IV.58}\\ & & +\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-\bar{z}_{j}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\Delta(E^{\sigma_{j-1}}_{{{\vec{P}}}},\bar{z}_{j})d\bar{z}_{j}\,.\label{eq:IV.59} $$ Each of the expressions (\[eq:IV.57\]) and (\[eq:IV.58\]) can be rewritten by adding and subtracting $(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}$. For (\[eq:IV.57\]) we get $$\begin{aligned} & &(\ref{eq:IV.57})\nonumber\\ & =& \oint_{\gamma_{j}}\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\,,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\big \rangle \frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}}d\bar{z}_{j}\quad\quad\quad\quad\label{eq:IV.61}\\ & & +\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}} \|}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\,,\,\label{eq:IV.62}\\ & &\quad\quad\quad, \,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}[(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}} \|}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}]\big \rangle\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}} d\bar{z}_{j}\nonumber\\ & & +\oint_{\gamma_{j}}\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}} \|}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\,,\label{eq:IV.63}\\ & &\quad\quad\quad,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\big \rangle\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}} d\bar{z}_{j}\,\nonumber\\ & & +\oint_{\gamma_{j}}\big\langle (\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}\,,\label{eq:IV.64}\\ & &\quad\quad\quad,\,\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}[(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}]\big \rangle\frac{1}{E^{\sigma_{j-1}}_{{{\vec{P}}}}-\bar{z}_{j}} d\bar{z}_{j}\,\nonumber\,.\end{aligned}$$ The difference in Eq. (\[eq:IV.49\]) corresponds to the sum of the terms (\[eq:IV.58\])-(\[eq:IV.59\]) and of the terms (\[eq:IV.62\])-(\[eq:IV.64\]). In fact, (\[eq:IV.61\]) corresponds to the first term in (\[eq:IV.49\]) after a contour deformation from $\gamma_{j-1}$ to $\gamma_j$. The sum of the remainder terms (\[eq:IV.58\]), (\[eq:IV.59\]), and (\[eq:IV.62\])-(\[eq:IV.64\]) can be bounded by $\epsilon^{j(1-2\delta)}$, for ${{\mathcal{R}}}_0$ and $\alpha$ small enough but independent of $j$, for any ${{\vec{P}}}\in{{\mathcal{S}}}$. (We recall that ${{\mathcal{R}}}_0$ can be taken arbitrarily small, provided $\alpha$ is small enough). The details are as follows. - For (\[eq:IV.62\])-(\[eq:IV.64\]) use the following inequalities $$\begin{aligned} & &\big\|\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}\big)(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big\| \leq {{\mathcal{O}}}(\frac{{{\mathcal{R}}}_0^{\frac{1}{2}}}{\alpha^{\frac{1}{4}}\, \epsilon^{(j-1)\delta}})\,,\label{eq:IV.65}\\ & &\big\|[(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_{j}})^{i}-(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\big]\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big\| \leq {{\mathcal{O}}}(\alpha^{\frac{1}{4}} \epsilon^{j(1-\delta)})\,,\label{eq:IV.66}\\ & &\big\|\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}} \big\|_{{{\mathcal{F}}}_{\sigma_j}} \leq {{\mathcal{O}}}(\frac{1}{\epsilon^{j}})\,,\\ & &\big\|\big(\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-\bar{z}_{j}}\big)(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\,(\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}})\big\| \leq {{\mathcal{O}}}(\frac{\alpha^{\frac{1}{4}} \epsilon^{j(1-\delta)}}{\epsilon^{j}})\,,\\ & &\big\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}-\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\big\| \leq \alpha^{\frac{1}{4}} \epsilon^{j(1-\delta)}\,.\label{eq:IV.68}\end{aligned}$$ In order to derive the inequality in Eq. (\[eq:IV.66\]), one uses Eqs. (\[eq:III.63\]), (\[eq:III.73\]), and (\[eq:III.57\])-(II.58). - For (\[eq:IV.58\]), after adding and subtracting $(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\frac{\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}}{\|\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|}$, one also has to use that $$\|[-(\Delta K_{{{\vec{P}}}}|^{\sigma_{j-1}}_{\sigma_j}+\hat{{{\mathcal{E}}}}_{{{\vec{P}}}}^{\sigma_j}-{{\mathcal{E}}}_{{{\vec{P}}}}^{\sigma_{j-1}})]\frac{1}{K_{{{\vec{P}}}}^{\sigma_{j-1}}-z_{j}}(\Gamma_{{{\vec{P}}}}^{\sigma_{j-1}})^{i}\Phi_{{{\vec{P}}}}^{\sigma_{j-1}}\|\leq {{\mathcal{O}}}(\alpha^{\frac{1}{2}}\epsilon^{j-1}\,\frac{{{\mathcal{R}}}_0^{\frac{1}{4}}}{\alpha^{\frac{1}{4}}\, \epsilon^{(j-1)\delta}})\,;$$ - To bound (\[eq:IV.59\]), note that $$\begin{aligned} (\ref{eq:IV.59})&=&-2\pi i\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-E_{{{\vec{P}}}}^{\sigma_j}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\\ & &+2\pi i\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-E_{{{\vec{P}}}}^{\sigma_{j-1}}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\,\quad\quad\quad\\ &= & 2\pi i\big\langle (\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\,,\,\frac{(E_{{{\vec{P}}}}^{\sigma_{j-1}}-E_{{{\vec{P}}}}^{\sigma_j})}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-E_{{{\vec{P}}}}^{\sigma_j}}\,\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-E_{{{\vec{P}}}}^{\sigma_{j-1}}}(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_j})^{i}\,\frac{\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}}{\|\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\|}\big \rangle\quad\quad\quad\quad\quad\quad\end{aligned}$$ where $|E_{{{\vec{P}}}}^{\sigma_{j-1}}-E_{{{\vec{P}}}}^{\sigma_{j}}|\leq {{\mathcal{O}}}(\alpha\,\epsilon^{j-1})$. Then use that $\epsilon\geq C\alpha^{\frac{1}{2}}$ and the following inequality analogous to (\[eq:IV.65\]) $$\big\|\big(\frac{1}{\hat{K}_{{{\vec{P}}}}^{\sigma_{j}}-E_{{{\vec{P}}}}^{\sigma_j}}\big)(\hat{\Gamma}_{{{\vec{P}}}}^{\sigma_{j}})^{i}\,\hat{\Phi}_{{{\vec{P}}}}^{\sigma_{j}}\big\| \leq {{\mathcal{O}}}(\frac{{{\mathcal{R}}}_0^{\frac{1}{2}}}{\alpha^{\frac{1}{4}}\, \epsilon^{j\delta}})\,.$$ For $\alpha$ small enough, $\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}$ converges, as $\sigma\to0$. The limiting function, $ \Sigma_{|{{\vec{P}}}|}:=\lim_{\sigma\to0}\,\frac{\partial^2E^{\sigma}_{|{{\vec{P}}}|}}{(\partial|{{\vec{P}}}|)^2}$, is Hölder-continuous in ${{\vec{P}}}\in{{\mathcal{S}}}$ (for an exponent $\eta>0$). The limit $$\label{eq:IV.71} \lim_{\alpha \to0}\,\Sigma_{|{{\vec{P}}}|}=1\,$$ holds true uniformly in ${{\vec{P}}}\in{{\mathcal{S}}}$. *Proof* It is enough to prove the result for a fixed choice of a sequence $\{\sigma_j\}_{j=0}^{\infty}$. The estimate in Lemma \[LemmaIII.1\] implies the existence of $\lim_{j\to\infty}\,\partial_i^2E^{\sigma_j}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^i\hat{i}}$. We now observe that $\partial_i^2E^{\sigma_0}_{|{{\vec{P}}}|}|_{{{\vec{P}}}=P^{i}\hat{i}}=1$ (see Eq. (\[eq:IV.48\])), because $$\label{eq:IV.72} (\Gamma_{{{\vec{P}}}}^{\sigma_{0}})^{i}\equiv ({{\vec{P}}}^f)^{i}\quad,\quad\Phi_{{{\vec{P}}}}^{\sigma_{0}}\equiv\Omega_f\,.$$ According to the constraint in Eq. (\[eq:II.83b\]), we can take $\epsilon={{\mathcal{O}}}(\alpha^{\frac{1}{2}(1-\delta)})$ so that, for $\alpha$ small enough, Lemma \[LemmaIII.1\] and (\[eq:IV.72\]) yield $$\label{eq:IV.73} \big|\frac{1}{\pi i}\oint_{\gamma_j}d\bar{z}_j\frac{1}{E^{\sigma_j}_{{{\vec{P}}}}-\bar{z}_j} \langle \,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\,,\,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{1}{K^{\sigma_j}_{{{\vec{P}}}}-\bar{z}_j}\,\,(\Gamma^{\sigma_j}_{{{\vec{P}}}})^{i}\,\frac{\Phi_{{{\vec{P}}}}^{\sigma_j}}{\|\Phi_{{{\vec{P}}}}^{\sigma_j}\|}\rangle|_{{{\vec{P}}}=P^{i}\hat{i}}\big|<{{\mathcal{O}}}(\alpha^{\frac{1}{2}(1-\delta)(1-2\delta)})\,,$$ uniformly in $j\in{\mathbb{N}}$. Hence the limit (\[eq:IV.71\]) follows. The Hölder-continuity in ${{\vec{P}}}$ of $\Sigma_{|{{\vec{P}}}|}$ is a trivial consequence of the analyticity in ${{\vec{P}}}\in{{\mathcal{S}}}$ of $E^{\sigma}_{{{\vec{P}}}}$, for any $\sigma>0$, and of Lemma \[LemmaIII.1\]; see [@Pizzo] for similar results. For $\alpha$ small enough, the function $E_{{{\vec{P}}}}:=\lim_{\sigma\to0}E_{{{\vec{P}}}}^{\sigma}$, ${{\vec{P}}}\in{{\mathcal{S}}}$, is twice differentiable, and $$\frac{\partial^2E_{|{{\vec{P}}}|}}{(\partial |{{\vec{P}}}|)^2}=\Sigma_{|{{\vec{P}}}|}\,.$$ *Proof* The result follows from the Hölder-continuity of $\Sigma_{|{{\vec{P}}}|}$, of $\lim_{\sigma\to0} \frac{\partial E_{|{{\vec{P}}}|}^{\sigma}}{\partial|{{\vec{P}}}|}$, and from the fundamental theorem of calculus applied to the functions $E_{{{\vec{P}}}}$ and $\lim_{\sigma\to0} \frac{\partial E_{|{{\vec{P}}}|}^{\sigma}}{\partial|{{\vec{P}}}|}$, because - $$\frac{\partial E_{|{{\vec{P}}}|}^{\sigma}}{\partial|{{\vec{P}}}|} \quad \text{and}\quad \frac{\partial^2 E_{|{{\vec{P}}}|}^{\sigma}}{(\partial|{{\vec{P}}}|)^2}$$ converge pointwise, for ${{\vec{P}}}\in {{\mathcal{S}}}$, as $\sigma\to0$, - $$\big|\frac{\partial E_{|{{\vec{P}}}|}^{\sigma}}{\partial|{{\vec{P}}}|}\big|\quad \text{and}\quad\big|\frac{\partial^2 E_{|{{\vec{P}}}|}^{\sigma}}{(\partial|{{\vec{P}}}|)^2}\big|$$ are uniformly bounded in $\sigma$, for all ${{\vec{P}}}\in{{\mathcal{S}}}$. ### Acknowledgements {#acknowledgements .unnumbered} The authors thank Thomas Chen for very useful discussions. [10]{} V.  Bach, J.  Froehlich, A.  Pizzo. Infrared-finite algorithms in QED II. The expansion of the groundstate of an atom interacting with the quantized radiation field. V. Bach,T. Chen, J. Fr[ö]{}hlich, and I. M. Sigal. The Renormalized Electron Mass in Non-Relativistic Quantum Electrodynamics. http://arxiv.org/abs/math-ph/0507043 T. Chen. Infrared Renormalization in Nonrelativistic QED and Scaling Criticality. http://arxiv.org/abs/math-ph/0601010 T. Chen, J. Fr[ö]{}hlich. Coherent Infrared Representations in Nonrelativistic QED, [*To appear in Proc. Symp. Pure Math. (B. Simon 60-th Birthday Volume)*]{}. http://arxiv.org/abs/math-ph/0601009 T. Chen, J. Fr[ö]{}hlich, and A. Pizzo. Infraparticle Scattering States in Non-Relativistic QED: I. The Bloch-Nordsieck paradigm http://arxiv.org/abs/math-ph/07092493 T. Chen, J. Fr[ö]{}hlich, and A. Pizzo. Infraparticle Scattering States in Non-Relativistic QED: II. Mass Shell Properties. M. Fierz, W. Pauli. F. Hiroshima, H. Spohn. C. Hainzl, R. Seiringer. E.T. Lieb, M. Loss. E.T. Lieb, M. Loss. A. Pizzo. One-particle (improper) states in [N]{}elson’s massless model. , [**4**]{} (3), 439–486, 2003. A. Pizzo. Scattering of an *Infraparticle*: The One-particle (improper) Sector in Nelson’s massless model. , [**4**]{} (3), 439–486, 2003. S.S. Schweber. An introduction to Relativistic Quantum Field Theory. . Reed Simon. Methods of modern mathematical physics. Vol. I-II-III-IV . [^1]: also at IHES, Bures-sur-Yvette

--- abstract: 'The main goal of this paper is to give some construction results of BiHom-post-Lie algebras which are a generalization of both post-Lie-algebras and Hom-post-Lie algebras. They are the algebraic structures behind the $\mathcal{O}$-operator of BiHom-Lie algebras. They can be also regarded as the splitting into three parts of the structure of a BiHom-Lie-algebra. Moreover we develop the representation theory of BiHom-post-Lie algebras on a vector space $V$. We show that there is naturally an induced representation of its sub-adjacent Lie algebra.' author: - '**H. Adimi, T. Chtioui, S. Mabrouk, S. Massoud**' - | H. Adimi$^{1}$ [^1], T. Chtioui$^{2}$ [^2] , S. Mabrouk$^{3}$ [^3] , S. Massoud$^{4}$ [^4]\ [1. Université Mohamed El Bachir El Ibrahimi de Bordj Bou Arréridj El-Anasser, 34030 - Algérie, ]{}\ [2. Faculty of Sciences, University of Sfax, BP 1171, 3000 Sfax, Tunisia]{}\ [3. University of Gafsa, Faculty of Sciences Gafsa, 2112 Gafsa, Tunisia]{}\ [4. Faculty of Sciences, University of Sfax, BP 1171, 3000 Sfax, Tunisia]{} title: 'Constructions and representation theory of BiHom-post-Lie algebras' --- **Key words**: BiHom-post-Lie algebras, BiHom-Lie algebras, representations, $\mathcal{O}$-operator. **Mathematics Subject Classification**: 17D05,17D10,17A15. Introduction ============ The notion of post-algebras goes back to Rosenbloom (1942). An equivalent formulation of the class of post-algebras was given by Rousseau (1969, 1970) which became a starting point for deep research. After then authors inspired many applications in computer science of various generalizations of post-algebras.Post-Lie algebras have been introduced by Vallette in 2007 [@B.; @Vallette] in connection with the homology of partition posets and the study of Koszul operads. However, J. L. Loday studied pre-Lie algebras and post-Lie algebras within the context of algebraic operad triples, see for more details [Dialg99,Trialg01]{}. In the last decade, many works [@Burde16; @Fard; @sl2] intrested in post-Lie algebra structures, motivated by the importance of pre-Lie algebras in geometry and in connection with generalized Lie algebra derivations.Recently,Post-Lie algebras which are non-associative algebras played an important role in different areas of pure and applied mathematics. They consist of a vector space $A$ equipped with a Lie bracket $[\cdot ,\cdot ]$ and a binary operation $\triangleright $ satisfying the following axioms$$\begin{aligned} x\triangleright \lbrack y,z]& =[x\triangleright y,z]+[y,x\triangleright z], \\ \lbrack x,y]\triangleright z& =as_{\triangleright }(x,y,z)-as_{\triangleright }(y,x,z).\end{aligned}$$If the bracket $[\cdot ,\cdot ]$ is zero, we have exactly a pre-Lie structure. The varieties of pre- and post-Lie algebras play a crucial role in the definition of any pre and post-algebra through black Manin operads product, see details in [@BBGN2012; @GubKol2014]. Whereas pre-Lie algebras are intimately associated with euclidean geometry, post-Lie algebras occur naturally in the differential geometry of homogeneous spaces, and are also closely related to Cartan’s method of moving frames .Fard et al.[@Fard] studied universal enveloping algebras of post-Lie algebras and the free post-Lie algebra. In [@bakayoko] Bakayoko studied hom-post-Lie algebras, as a twisted generalization post Lie algebra. He studied modules over Hom-post-Lie algebras and gave some constructions he showed also that modules over Hom-post-Lie algebras are close by twisting either by Hom-post-Lie algebra endomorphisms or module structure maps. We can obtain Hom post-Lie modules from only a given multiplicative Hom-post-Lie algebra in a non-trivial sense.Motivated by a categorical study of Hom-algebra and new type of categories, the authors of [@bihomass] introduced a generalized algebraic structure dealing with two commuting multiplicative linear maps, called BiHom-algebras including BiHom-associative algebras and BiHom-Lie algebras. When the two linear maps are the same, then BiHom-algebras become Hom-algebras in some cases. Many varieties of algebras are generalized to the BiHom-version, see details in . The paper is organized as follows. In section 2, we recall some basic definitions and properties of post-Lie algebras and BiHom-Lie algebras. In section 3, We construct BiHom-Lie algebras structure associated to any BiHom-post-Lie algebra. The bracket is given by$$\left\{ x,y\right\} =x\triangleright y-\alpha ^{-1}\beta \left( y\right) \triangleright \alpha \beta ^{-1}\left( x\right) +\left[ x,y\right] .$$In addition, we investigate the notion of an $\mathcal{O}$-operator of weight $\lambda $ to construct a BiHom-post-Lie algebra structure on the $A$-module $\mathbb{K}$-algebra of a BiHom-Lie algebra $(A,[\cdot ,\cdot ],\alpha ,\beta )$. The Last section is devoted to introduce the representation theory of a BiHom-post-Lie algera and its sub-adjacent BiHom-Lie algebra. Throughout this paper, all vector spaces are over a field $\mathbb{K}$ of characteristic zero. Preliminaries ============= In what follows, we recall some concepts and facts used in this paper \[post-lie alg\] A (left) post-Lie algebra $(A,[\c,\c],\rhd)$ consists of a Lie algebra $(A,[\c,\c])$ and a binary product $\rhd:A \times A \to A$ such that, for all elements $x, y, z \in A$ the following relations hold $$\begin{aligned} & x \rhd [y,z]=[x\rhd y,z]+[y,x \rhd z], \label{postlie1} \\ &[x,y]\rhd z= as_{\rhd}(x,y,z)-as_{\rhd}(y,x,z), \label{postlie2}\end{aligned}$$ where $as_{\rhd}(x,y,z)=x\rhd (y \rhd z)-(x\rhd y)\rhd z$. It is clear that a post-Lie algebra with an abelian Lie algebra structure reduces to a pre-Lie algebra. If we define $L_{\rhd}: A \to gl(A)$ by $L_{\rhd}(x)y=x\rhd y$, then by Eq. , $L_{\rhd}(x)$ is a derivation on $(A,[\c,\c])$. A morphism $f:(A,[\c,\c],\rhd)\to (A',[\c,\c]',\rhd')$ of post-Lie algebras is a linear map satisfying $$f([x,y])=[f(x),f(y)], \ \ f(x\rhd y)=f(x)\rhd f(y), \ \ \forall x,y\in A.$$ Let $(A,[\c,\c],\rhd)$ be a post-Lie algebra. Then the bracket $$\begin{aligned} \label{post-lie =>lie} \{x,y\}=x\rhd y-y\rhd x+[x,y]\end{aligned}$$ defines a Lie algebra structure on $A$. We denote this algebra by $A^C$ and it is called the sub-adjacent Lie-algebra of $(A,[\c,\c],\rhd)$. \[RepPostLie\] A representation of a post-Lie algebra $(A,[\c,\c],\rhd)$ on a vector space $V$ is a tuple $(V,\rho,\mu,\nu)$, such that $(V,\rho)$ is a representation of $(A,[\c,\c])$ and $\mu,\nu: A \to gl(V)$ are linear maps satisfying [$$\begin{aligned} & \nu([x,y]) =\rho(x) \nu(y) -\rho(y) \nu(x) , \label{rep pos lie 1}\\ & \rho(x\rhd y) =\mu(x) \rho(y)-\rho(y) \mu(x) , \label{rep pos lie 2}\\ &\mu([x,y]) =\mu(x)\mu(y)-\mu(x\rhd y) -\mu(y)\mu(x)+\mu(y\rhd x) , \label{rep pos lie 3}\\ & \nu(y)\rho(x) =\mu(x)\nu(y)-\nu(y)\mu(x)-\nu(x\rhd y) +\nu(y)\nu(x) , \label{rep pos lie 4}\end{aligned}$$]{} for any $x,y \in A$. Let $(A,[\c,\c],\rhd)$ be a post-Lie algebra and $(V,\rho,\mu,\nu)$ a representation of $(A,[\c,\c],\rhd)$. By identity , we deduce that $(V,\mu)$ is a representation of the sub-adjacent Lie algebra $(A,\{\c,\c\})$ of $(A,[\c,\c],\rhd)$. Further, it is obvious that $(A,ad,L_{\rhd},R_{\rhd})$ is a representation of $(A,[\c,\c],\rhd)$ which is called the adjoint representation. Let $(V,\rho,\mu,\nu)$ be a representation of a post-Lie algebra $(A,[\c,\c],\rhd)$. Then $(V,\r+\mu-\nu)$ is a representation of the sub-adjacent Lie algebra $(A,\{\c,\c\})$ of $(A,[\c,\c],\rhd)$. A BiHom-Lie algebra is a tuple $(A,[\cdot,\cdot],\a,\b)$ consisting of a linear space $A$, a linear map $[\cdot,\cdot]:\otimes^2 A\longrightarrow A$ and two linear map $\a,\b:A\longrightarrow A $, satisfying $$\begin{aligned} &\a \circ \b =\b \circ \a, \nonumber\\ &\a( [x,y]) =[\a(x),\a(y)],\ \b( [x,y]) = [\b(x),\b(y)], \nonumber\\ & [\b(x),\a(y)]= -[\b(y),\a(x)],\nonumber\\ & \circlearrowleft_{x,y,z \in A} [\b^2(x),[\b(y),\a(z)]]= 0\label{Bihom-jaco},\end{aligned}$$ for any $x,y,z \in A$. A BiHom-Lie algebra $(A,[\cdot,\cdot],\a,\b)$ is called regular if $\a$ and $\b$ are bijective. \[TwistLieAlg\] Let $(A,[\cdot,\cdot])$ be a Lie algebra and $\a,\b$ two commuting morphisms on $A$. Then $(A,[\cdot,\cdot]_{\a,\b}=[\cdot,\cdot]\circ (\a\otimes\b),\alpha,\beta)$ is a BiHom-Lie algebra. \[defi:bihom-lie representation\] A representation of a BiHom-Lie algebra $(A,[\cdot,\cdot],\alpha,\beta)$ on a vector space $V$ with respect to commuting linear maps $\phi,\psi:V\rightarrow V$ is a linear map $\rho:A\longrightarrow gl(V)$, such that for all $x,y\in A$, the following equalities are satisfied $$\begin{aligned} \label{bihom-lie-rep-1}\rho(\alpha(x))\circ \f&=&\f\circ \rho(x),\\ \label{bihom-lie-rep-2} \rho(\b(x))\circ \p&=&\p \circ \rho(x),\\ \label{bihom-lie-rep-3}\rho([\b(x),y])\circ \p &=&\rho(\alpha \b(x))\circ\rho(y)-\rho(\b(y))\circ\rho(\a(x)).\end{aligned}$$ We denote such a representation by $(V,\rho,\f,\p)$. For all $x\in A$, we define $ad_{x}:A \to A$ by $$\begin{aligned} ad_{x}(y)=[x,y],\quad\forall y \in A.\end{aligned}$$ Then $ad:A \longrightarrow gl(A)$ is a representation of the BiHom-Lie algebra $(A,[\cdot,\cdot],\a,\b)$ on $A$ with respect to $\alpha$ and $\b$, which is called the adjoint representation. Let $(A,[\cdot,\cdot])$ be a Lie algebra, $\a,\b$ two commuting morphisms on $A$. Consider a representation $(V,\rho)$ of $A$ and two commuting linear maps $\phi,\psi:V\rightarrow V$ satisfying $$\rho(\alpha(x))\circ \f=\f\circ \rho(x),\ \ \rho(\b(x))\circ \p=\p \circ \rho(x).$$ Define a linear map $\widetilde{\rho}:A\longrightarrow gl(V)$ by $\widetilde{\rho}(x)(v)=\rho(\alpha(x))( \p(v))$. \[twistmodule\] With the above notation $(V,\widetilde{\rho},\f,\p)$ is a representation of $(A,[\cdot,\cdot]_{\a,\b},\alpha,\beta)$. \[lem:semidirectp\] Let $(A,[\cdot,\cdot],\alpha,\b)$ be a regular BiHom-Lie algebra, $(V,\f,\p)$ a vector space with two commuting bijective linear transformations and $\rho: A \rightarrow gl(V)$ a linear map. Then $(V,\r,\f,\p)$ is a representation of $(A,[\cdot,\cdot],\a,\b)$ if and only if $(A\oplus V,[\cdot,\cdot]_\rho,\alpha+\f,\b+\p)$ is a BiHom-Lie algebra, where $[\cdot,\cdot]_\rho$, $\alpha+\f$ and $\b+\p$ are defined by $$\begin{aligned} \label{eq:sum} [x+u,y+v]_{\rho}&=&[x,y]+\rho(x)v-\rho(\a^{-1}\b(y))\f\p^{-1}u,\\ (\alpha+\f)(x+u)&=&\alpha(x)+\f(u),\\ (\b+\p)(x+u)&=&\b(x)+\p(u),\end{aligned}$$ for all $x,y\in A,~u,v\in V$. Now we recall the definition of an $\mathcal{O}$-operator on a BiHom-Lie algebra associated to a given representation, which generalize the Rota-Baxter operator of weight $0$ introduced in and defined as a linear operator $R$ on a BiHom-Lie algebra $(A,[\cdot,\cdot],\alpha,\beta)$ such that $R\circ \a=\a \circ R$, $R \circ \b=\b \circ R$ and $$[R(x),R(y)]=R([R(x),y]+[x,R(y)]),\ \forall x,y \in A.$$ Let $(A,[\cdot,\cdot],\alpha,\beta)$ be a BiHom-Lie algebra and $(V,\rho,\f,\p)$ be a representation. A linear map $T: V \to A$ is called an $\mathcal{O}$-**operator** associated to $\rho$, if it satisfies $$\begin{aligned} \label{O-operator} [T(u),T(v)]=T(\rho(T(u))v-\rho(T(\f^{-1}\p(v)))\f\p^{-1}(u)),\ \ \forall u,v \in V.\end{aligned}$$ BiHom-post-Lie algebras ======================= In this section, we recall the notion of BiHom-post-Lie algebras ( see ). We will provide some construction results. Let $(A,[\c,\c],\a,\b)$ be BiHom-Lie algebra. Let $(V,\{\c,\c\},\f,\p)$ be a BiHom-Lie algebra and $\rho:A \to gl(V)$ be a linear map. We say that $(V,\{\c,\c\},\rho,\f,\p,)$ is an $A$-module $\mathbb{K}$-algebra if $(V,\rho,\f,\p,)$ is a representation of $(A,[\c,\c],\a,\b)$ such that the following compatibility condition holds $$\begin{aligned} \label{repKalgebras} &\r(\a\b(x))\{u,v\}=\{\r(\b(x))u,\p(v)\}+\{\p(u),\r(\a(x))v\},\end{aligned}$$ for all $x \in A,\ u,v \in V$. It is known that $(A,ad,\a,\b)$ is a representation of $A$ called the adjoint representation. Then $(A,[\cdot,\cdot],ad,\a,\b)$ is an $A$-module $\mathbb{K}$-algebra. Let $(A,[\c,\c],\a,\b)$ and $(V,\{\c,\c\},\f,\p)$ be two regular BiHom-Lie algebras and $\r: A \to gl(V)$ be a linear map. Then $(V,\{\c,\c\},\r,\f,\p)$ be an $A$-module $\mathbb{K}$-algebra if and only if the direct sum $A\oplus V$ of vector spaces is turned into a BiHom-Lie algebra (the semi-direct sum) by defining a bracket on $A\oplus V$ by $$\begin{aligned} [x+u,y+v]_{\rho}=&[x,y]+\rho(x)v-\rho(\a^{-1}\b(y))\f\p^{-1}u+\{u,v\},\\ (\alpha+\f)(x+u)=&\alpha(x)+\f(u),\\ (\b+\p)(x+u)=&\b(x)+\p(u),\end{aligned}$$ for all $x,y\in A,~u,v\in V$. We denote this algebra by $A\ltimes_{\r,\f,\p}^{\a,\b} V$ or simply $A\ltimes V$. Let $x,y,z\in A$ and $a,b,c\in V$ $$\begin{aligned} &\circlearrowleft_{x+a,y+b,z+c}[(\beta+\psi)^2(x+a),[(\beta+\psi)(y+b),(\alpha+\phi)(z+c)]_{\rho}]_{\rho}\\ =&\circlearrowleft_{x+a,y+b,z+c}([\beta^2(x),[\beta(y),\alpha(z)]]+\rho(\beta^2(x))\rho(\beta(y))\phi(c)-\rho(\beta^2(x))\rho(\beta(z))\phi(b)\\ &+\rho(\beta^2(x))[\psi(b),\phi(c)]_{V}-\rho([\alpha^{-1}\beta^2(y),\beta(z)])\phi\psi(a)+[\psi^2(a),\rho(\beta(y))\phi(c)]_{V}\\ &-[\psi^2(a),\rho(\beta(z))\phi(b)]_{V}+[\psi^2(a),[\psi(b),\phi(c)]_{V}]_{V})=0,\end{aligned}$$ if and only if $(V,\rho,\phi,\psi)$ is a representation on $A$ and satisfies Eq. . Let $(A,[\c,\c],\a,\b)$ be a regular BiHom-Lie algebra and $(V,\{\c,\c\},\r,\f,\p)$ be a regular $A$-module $\mathbb{K}$-algebra. A linear map $T:V\to A$ is called an $\mathcal{O}$-operator of weight $\lambda\in \mathbb{K}$ associated to $(V,\{\c,\c\},\r,\f,\p)$ if its satisfies $$\begin{aligned} & T\f=\a T,\ \ T\p=\b T, \\ & \label{O-operatorLamda} [T(u),T(v)]=T(\rho(T(u))v-\rho(T(\f^{-1}\p(v)))\f\p^{-1}(u)+\lambda\{u,v\}), \end{aligned}$$ for all $u,v\in V$. An $\mathcal{O}$-operator of weight $\lambda$ associated to $(A,[\c,\c],ad,\a,\b)$ is just a Rota-Baxter on $A$ of the same weight. A (left) BiHom-post-Lie algebra is a tuple $(A,[\c,\c],\rhd,\a,\b)$ where $(A,[\c,\c],\a,\b)$ is a BiHom-Lie algebra such that for any $x,y,z \in A$ $$\begin{aligned} & \a(x \rhd y)=\a(x) \rhd \a(y),\ \b(x \rhd y)= \b(x) \rhd \b(y), \\ \label{bihom-post-lie condition1} &\a\b(x)\rhd[y,z]=[\b(x)\rhd y,\b(z)]+[\b(y),\a(x)\rhd z] ,\\ \label{bihom-post-lie condition2}& [\b(x),\a(y)]\rhd \b(z)=as_{\a,\b}(\b(x),\a(y),z)- as_{\a,\b}(\b(y),\a(x),z) ,\end{aligned}$$ where $as_{\a,\b}(x,y,z)=\a(x)\rhd (y\rhd z)-(x\rhd y) \rhd \b(z)$. We call $[\c,\c]$ the *torsion* and $\rhd$ the *connection* of the (left) BiHom-post-Lie algebra. if $\a=\b=id$, then we recover a (left) post-Lie algebra. In addition, if $[\c,\c]=0$, then $A$ reduces to a BiHom-pre-Lie algebra. From now on, we will write BiHom-post-Lie algebra instead of (left) BiHom-post-Lie algebra. A BiHom-post-Lie algebra $(A,[\c,\c],\rhd,\a,\b)$ is said to be regular if $\a$ and $\b$ are bijective. \[TwistBiHomPostLie\] Let $(A,[\c,\c],\rhd)$ be a post-Lie algebra and $\a, \b$ be two commuting morphisms on $A$. Then $(A,[\c,\c]_{\a,\b},\rhd_{\a,\b})$ is a BiHom-post-Lie algebra, Where, for $x,y\in A$, $$\label{TwistOperations} [x,y]_{\a,\b}=[\a(x),\b(y)],\ \ x\rhd_{\a,\b}y=\a(x)\rhd\b(y).$$ Let $(A,[\c,\c],\a,\b)$ be a BiHom-Lie algebra. Then $(A,[\c,\c],\rhd,\a,\b)$ is a BiHom-post-Lie algebra, where $$x\rhd y=[y,x],\ \ \forall x,y\in A.$$ \[subadj lie alg\] Let $(A,[\c,\c],\rhd,\a,\b)$ be a regular BiHom-post-Lie algebra. Then the bracket $$\label{liebracket} \{x,y\}=x\rhd y-\a^{-1}\b(y)\rhd \a\b^{-1}(x)+[x,y]$$ defines a BiHom-Lie algebra structure on $A$. We denote this algebra by $A^C$ and we call it the sub-adjacent BiHom-Lie algebra of $A$. The BiHom-skew symmetry is obvious. We will just check the BiHom-Jacobi identity. For all $x,y,z \in A,$ we have $$\begin{aligned} \{\beta^2(x),\{\beta(y),\alpha(z)\}\}&=\beta^2(x)\rhd(\beta(y)\rhd\alpha(z))-(\alpha^{-1}\beta^2(y))\rhd\alpha\beta(z))\rhd\alpha\beta(x)\\ &+[\beta^2(x),\beta(y)\rhd\alpha(z)]-\beta^2(x)\rhd(\beta(z)\rhd\alpha(y))\\ &+(\alpha^{-1}\beta^2(z)\rhd\beta(y))\rhd\alpha\beta(x)-[\beta^2(x),\beta(z)\rhd\alpha(y)]\\ &+\beta^2(x)\rhd[\beta(y),\alpha(z)]+[\alpha^{-1}\beta^2(y),\beta(z)]\rhd\alpha\beta(x)\\ &+[\beta^2(x),[\beta(y),\alpha(z)]],\end{aligned}$$ $$\begin{aligned} \{\beta^2(y),\{\beta(z),\alpha(x)\}\}&=\beta^2(y)\rhd(\beta(z)\rhd\alpha(x))-(\alpha^{-1}\beta^2(z))\rhd\beta(x))\rhd\alpha\beta(y)\\ &+[\beta^2(y),\beta(z)\rhd\alpha(x)]-\beta^2(y)\rhd(\beta(x)\rhd\alpha(z))\\ &+(\alpha^{-1}\beta^2(x)\rhd\beta(z))\rhd\alpha\beta(y)-[\beta^2(y),\beta(x)\rhd\alpha(z)]\\ &+\beta^2(y)\rhd[\beta(z),\alpha(x)]+[\alpha^{-1}\beta^2(z),\beta(x)]\rhd\alpha\beta(y)\\ &+[\beta^2(y),[\beta(z),\alpha(x)]].\end{aligned}$$ Similarly, $$\begin{aligned} \{\beta^2(z),\{\beta(x),\alpha(y)\}\}&=\beta^2(z)\rhd(\beta(x)\rhd\alpha(y))-(\alpha^{-1}\beta^2(x))\rhd\beta(y))\rhd\alpha\beta(z)\\ &+[\beta^2(z),\beta(zx)\rhd\alpha(y)]-\beta^2(z)\rhd(\beta(y)\rhd\alpha(x))\\ &+(\alpha^{-1}\beta^2(y)\rhd\beta(x))\rhd\alpha\beta(z)-[\beta^2(z),\beta(y)\rhd\alpha(x)]\\ &+\beta^2(z)\rhd[\beta(z),\alpha(y)]+[\alpha^{-1}\beta^2(x),\beta(y)]\rhd\alpha\beta(z)\\ &+[\beta^2(z),[\beta(x),\alpha(y)]].\end{aligned}$$ By the BiHom-Jacobi identity of BiHom-Lie algebras and and we have $$\{\beta^2(x),\{\beta(y),\alpha(z)\}\}+\{\beta^2(y),\{\beta(z),\alpha(x)\}\}+\{\beta^2(z),\{\beta(x),\alpha(y)\}\}=0.$$ Given a BiHom-post-Lie algebra $(A,[\c,\c],\rhd,\a,\b)$. Suppose that $\rhd$ is commutative in the BiHom-sense, that is $\b(x) \rhd \a(y)=\b(y) \rhd \a(x)$. Then the two Lie brackets $[\c,\c]$ and $\{\c,\c\}$ coincide. Now, recall that a regular BiHom-algebra $(A,\c,\a,\b)$ is called admissible BiHom-Lie algebra if $(A,[\c,\c],\a,\b)$ is a BiHom-Lie algebra, where $[x,y]=x \c y- \a^{-1}\b(y) \c \a \b^{-1}(x).$ Let $(A,[\c,\c],\rhd,\a,\b)$ be a regular BiHom-post-Lie algebra. Define the product $\circ$ as $$\label{AdmissibileBiHomLie} x\circ y=x\vartriangleright y+\frac{1}{2}[x,y],\ \forall\ x,y\in A.$$ Then $(A,\circ,\a,\b)$ is an admissible BiHom-Lie algebra. Let $(A,[\c,\c],\rhd,\a,\b)$ be a regular BiHom-post-Lie algebra. Then $(A,L_\rhd,\a,\b)$ is a representation of $(A,\{\c,\c\},\a,\b)$, where the bracket $\{\c,\c\}$ is defined by the identity . It is obvious to check that $L_\rhd(\a(x)) \circ \a= \a \circ L_\rhd(x) $ and $L_\rhd(\b(x)) \circ \b= \b \circ L_\rhd(x)$. To prove , we compute as follows. Let $x,y,z \in A$. Then $$\begin{aligned} &L_\rhd\Big( \{\b(x),y\}\Big)\b(z)= \{\b(x),y\} \rhd \b(z)\\ =& (\b(x)\rhd y)\rhd \b(z)-(\a^{-1}\b(y)\rhd \a\b^{-1}(x))\rhd \b(z) + [\b(x),y] \rhd \b(z) \\ =& (\b(x)\rhd y)\rhd \b(z)-(\a^{-1}\b(y)\rhd \a(x))\rhd \b(z) + \a\b(x)\rhd (y \rhd z) \\ &-(\b(x) \rhd y) \rhd \b(z) -\b(y) \rhd (\a(x) \rhd z) + (\a^{-1}\b(y) \rhd \a(x)) \rhd \b(z) \\ =& \a\b(x)\rhd (y \rhd z) -\b(y) \rhd (\a(x) \rhd z) \\ =& L_\rhd(\a\b(x)) \circ L_\rhd(y) (z)- L_\rhd(\b(y)) \circ L_\rhd(\a(x)) (z)\end{aligned}$$ and the proof is finished. Let $(A,[\c,\c],\rhd,\a,\b)$ be a BiHom-post-Lie algebra. Then $(A,-[\c,\c],\blacktriangleright,\a,\b)$, where $$\label{BlackBiHomPosteLie} x\blacktriangleright y=x\vartriangleright y+[x,y],$$ is also a BiHom-post-Lie algebra. In addition $(A,[\c,\c],\rhd,\a,\b)$ and $(A,-[\c,\c],\blacktriangleright,\a,\b)$ have the same sub-adjacent BiHom-Lie algebra $A^C$. For all $x,y,z\in A,$ we check that $(A,-[\cdot,\cdot],\blacktriangleright,\alpha,\beta)$ is BiHom-post-Lie algebras. in fact we have $$\begin{aligned} -\alpha\beta(x)\blacktriangleright[y,z]&=-(\alpha\beta(x)\rhd[y,z]+[\alpha\beta(x),[y,z]])\\ &=-\alpha\beta(x)\rhd[y,z]-[\beta^2(\alpha\beta^{-1}(x)),[\beta(\beta^{-1}(y)),\alpha(\alpha^{-1}(z))]],\end{aligned}$$ $$\begin{aligned} -[\beta(x)\blacktriangleright y,\beta(z)]=&-[\beta(x)\rhd y+[\beta(x),y],\beta(z)]\\ =&-[\beta(x)\rhd y,\beta(z)]+[\alpha^{-1}\beta^2(\beta^{-1}(z)),[\alpha(x),\alpha\beta^{-1}(y))]]\end{aligned}$$ and $$\begin{aligned} -[\beta(y),\alpha(x)\blacktriangleright z]=&-[\beta(y),\alpha(x)\rhd z+[\alpha(x),z]]\\ &=-[\beta(y),\alpha(x)\rhd z]+[\beta^2(\beta^{-1}(y)),[\beta\alpha^{-1}(z),\alpha(\alpha\beta^{-1}(x))]].\end{aligned}$$ Then we obtain $$\begin{aligned} &-\alpha\beta(x)\blacktriangleright[y,z]+[\beta(x)\blacktriangleright y,\beta(z)]+[\beta(y),\alpha(x)\blacktriangleright z]\\ &=-\alpha\beta(x)\rhd[y,z]-[\beta(x)\rhd y,\beta(z)]-[\beta(y),\alpha(x)\rhd z]\\ &-[\beta^2(\alpha\beta^{-1}(x)),[\beta(\beta^{-1}(y)),\alpha(\alpha^{-1}(z))]]\\ &+[\alpha^{-1}\beta^2(\beta^{-1}(z)),[\alpha(x),\alpha\beta^{-1}(y))]]\\ &+[\beta^2(\beta^{-1}(y)),[\beta\alpha^{-1}(z),\alpha(\alpha\beta^{-1}(x))]]=0.\end{aligned}$$ Hence, the first condition hold using and BiHom-Jacobi-Identity. To check the second condition, we compute as follow $$\begin{aligned} &-[\beta(x),\alpha(y)]\blacktriangleright\beta(z)-as_{\alpha,\beta}^{\blacktriangleright}(\beta(x),\alpha(y),z)+as_{\alpha,\beta}^{\blacktriangleright}(\beta(y),\alpha(x),z)\\ &=-[\beta(x),\alpha(y)]\blacktriangleright\beta(z)+[[\beta(x),\alpha(y)],\beta(z)]+\alpha\beta(x)\rhd(\alpha(y)\rhd z)\\ &-(\beta(x)\rhd\alpha(y))\rhd \beta(z)+\alpha\beta(x)\rhd[\alpha(y),z]+[\alpha\beta(x),\alpha(y)\rhd z]\\ &+[\alpha\beta(x),[\alpha(y),z]]-[\beta(x),\alpha(y)]\rhd\beta(z)-[\beta(x)\rhd\alpha(y),\beta(z)]\\ &-[[\beta(x),\alpha(y)],\beta(z)]-\alpha\beta(y)\rhd(\alpha(x)\rhd z)+(\beta(y)\rhd \alpha(x))\rhd\beta(z)\\ &-\alpha\beta(y)\rhd[\alpha(x),z]-[\alpha\beta(y),\alpha(x)\rhd z]-[\alpha\beta(y),[\alpha(x),z]]\\ &+[\beta(y),\alpha(x)]\rhd\beta(z)+[\beta(y)\rhd \alpha(x),\beta(z)]+[[\beta(y),\alpha(x)],\beta(z)]\\ &=0,\end{aligned}$$ where $as_{\alpha,\beta}^{\blacktriangleright}(x,y,z)=x\blacktriangleright (y\blacktriangleright z)-(x\blacktriangleright y)\blacktriangleright z$. Let $(A,[\c,\c],\rhd,\a,\b)$ be a BiHom-post-Lie algebra. Define the double bracket $\llbracket \c,\c \rrbracket$ on $A\times A$ by $$\begin{aligned} \llbracket (a,x),(b,y) \rrbracket&=(a\rhd b-\a^{-1}\b(b)\rhd\a\b^{-1}(a)+[a,b],\nonumber\\& \label{DoubleBarcket} \qquad\qquad\qquad\qquad a\rhd y-\a^{-1}\b(b)\rhd\a\b^{-1}(x)+[x,y]) \end{aligned}$$ for all $a,b,x,y\in A.$ Then $(A\times A,\llbracket \c,\c \rrbracket,\a^{\times2},\b^{\times2})$ is a BiHom-Lie algebra. Let $x,y,z,a,b,c\in A$, It’s obvious that $$\llbracket (\b(a),\b(x)),(\a(b),\a(y)) \rrbracket=-\llbracket (\b(b),\b(y)),(\a(a),\a(x)) \rrbracket.$$ On the other hand, we have $$\begin{aligned} &\circlearrowleft_{(a,x),(b,y),(c,z)} \llbracket (\b^2(a),\b^2(x)),\llbracket (\b(b),\b(y)),(\a(c),\a(z)) \rrbracket \rrbracket \\=&\big(\circlearrowleft_{a,b,c} \{\b^2(a),\{\b(b),\a(c)\}\},\\&\circlearrowleft_{(a,x),(b,y),(c,z)}\b^2(a)\rhd (\b(b)\rhd \a(z))-\b^2(a)\rhd( \b(c)\rhd\a(y))+\b^2(a)\rhd [\b(y),\a(z)] \\&\qquad\qquad\qquad+(\b(b)\rhd \a(c))\rhd \a\b(x)-(\b(c)\rhd\a(b))\rhd \a\b(x) +[\b(b),\a(c)]\rhd \a\b(x) \\&\qquad\qquad\qquad+[\b^2(x),\b(b)\rhd \a(z)]-[\b^2(x),\b(c)\rhd\a(y)]+[\b^2(x),[\b(y),\a(z)]]\big) \\ =&(0,0),\end{aligned}$$ where the bracket $\{\c,\c\}$ is defined in . Therefor $(A\times A,\llbracket \c,\c \rrbracket,\a^{\times2},\b^{\times2})$ is a BiHom-Lie algebra. It is well known that BiHom-assciative, BiHom-pre-Lie and BiHom-Novikov algebras are BiHom-Lie admissible algebras. Another class of BiHom-Lie admissible algebras is the variety of BiHom-LR algebras. A BiHom-LR algebra is a tuple $(A, \cdot , \alpha , \beta )$ satisfying the following conditins $$\begin{aligned} \label{LRCondition1} &(x\c y)\c\a(z)=(x\c z)\c\a(y) ,\\ \label{LRCondition2} & \b(x)\c(y\c z)=\b(y)\c(x\c z),\ \forall x,y,z\in A.\end{aligned}$$ \[LR to Post\] Let $(A, \cdot , \alpha , \beta )$ be a regular BiHom-LR algebra. Then $(A, [\c,\c] ,\rhd , \alpha , \beta )$ is a BiHom-post-Lie algebra, where $$\label{CommutatorLR} x\rhd y=-x\c y\ \ \textrm{and}\ \ [x,y]= x\c y- \a^{-1}\b(y)\c \a\b^{-1}(x),\ \forall x,y\in A.$$ The BiHom-skew symmetry is obvious. We will check the BiHom-Jacobi identity. For all $x,y,z \in A,$ we have $$\begin{aligned} \circlearrowleft_{x,y,z}\{\beta^2(x),\{\beta(y),\alpha(z)\}\}&=\beta^2(x)\c(\beta(y)\c\alpha(z))-(\alpha^{-1}\beta^2(y))\c\alpha\beta(z))\c\alpha\beta(x)\\ &-\beta^2(x)\c(\beta(z)\c\alpha(y))+(\alpha^{-1}\beta^2(z)\c\beta(y))\c\alpha\beta(x)\\ &+\beta^2(y)\c(\beta(z)\c\alpha(x))-(\alpha^{-1}\beta^2(z))\c\beta(x))\c\alpha\beta(y)\\ &-\beta^2(y)\c(\beta(x)\c\alpha(z))+(\alpha^{-1}\beta^2(x)\c\beta(z))\c\alpha\beta(y)\\ &+\beta^2(z)\c(\beta(x)\c\alpha(y))-(\alpha^{-1}\beta^2(x))\c\beta(y))\c\alpha\beta(z)\\ &-\beta^2(z)\c(\beta(y)\c\alpha(x))+(\alpha^{-1}\beta^2(y)\c\beta(x))\c\alpha\beta(z).\end{aligned}$$ By the identities and of BiHom-LR algebra, we have $$\circlearrowleft_{x,y,z}\{\beta^2(x),\{\beta(y),\alpha(z)\}\}=0.$$ similarly, we have $$\begin{aligned} & \alpha \beta \left( x\right) \rhd \left[ y,z\right] -\left[ \beta \left( y\right) , \alpha \left( x\right) \rhd z \right] -\left[ \beta \left( x\right) \rhd y,\beta \left( z\right) \right] \\ =& -(\alpha \beta \left( x\right) \c \left[ y,z\right] -\left[ \beta \left( y\right) , \alpha \left( x\right) \c z \right] -\left[ \beta \left( x\right) \c y,\beta \left( z\right) \right] )\\ =& -(\alpha \beta \left( x\right) \c \left( y\c z\right)-\alpha \beta \left( x\right) \c \left(\a^{-1}\b(z)\c\a\b^{-1}(y)\right) -\beta \left( y\right) \c( \alpha \left( x\right) \c z )\\&+( \b \left( x\right) \c \a^{-1}\b(z) )\c\a \left( y\right) - (\beta \left( x\right) \c y)\c\beta \left( z\right)+\a^{-1}\beta^2 \left( z\right)\c(\a \left( x\right) \c\a\b^{-1}( y)) )\\ =&0\end{aligned}$$ \[DefBiHomTriDend\] A BiHom-tri-dendriform algebra is a 6-tuple $(A, \prec , \succ , \cdot , \alpha , \beta )$, where $A$ is a linear space and $\prec , \succ , \cdot :A\otimes A\rightarrow A$ and $\alpha , \beta :A\rightarrow A$ are linear maps satisfying $$\begin{aligned} &&\alpha \circ \beta =\beta \circ \alpha , \label{BiHomtridend1} \\ &&\alpha (x\prec y)=\alpha (x)\prec \alpha (y), ~ \alpha (x\succ y)=\alpha (x)\succ \alpha (y), ~ \alpha (x\cdot y)=\alpha (x)\cdot \alpha (y), \label{BiHomtridend4} \\ &&\beta (x\prec y)=\beta (x)\prec \beta (y), ~ \beta (x\succ y)=\beta (x)\succ \beta (y), ~ \beta (x\cdot y)=\beta (x)\cdot \beta (y), \label{BiHomtridend7} \\ &&(x\prec y)\prec \beta (z)=\alpha (x)\prec (y\prec z+y\succ z+y\cdot z), \label{BiHomtridend8} \\ &&(x \succ y)\prec \beta (z)=\alpha (x)\succ (y\prec z), \label{BiHomtridend9} \\ &&\alpha (x)\succ (y\succ z)=(x\prec y+x\succ y+x\cdot y)\succ \beta (z), \label{BiHomtridend10} \\ &&\alpha (x)\cdot (y\succ z)=(x\prec y)\cdot \beta (z), \label{BiHomtridend11} \\ &&\alpha (x)\succ (y\cdot z)=(x\succ y)\cdot \beta (z), \label{BiHomtridend12} \\ &&\alpha (x)\cdot (y\prec z)=(x\cdot y)\prec \beta (z), \label{BiHomtridend13} \\ &&\alpha (x)\cdot (y\cdot z)=(x\cdot y)\cdot \beta (z), \label{BiHomtridend14}\end{aligned}$$ for all $x, y, z\in A$. We call $\alpha $ and $\beta $ (in this order) the structure maps of $A$. A morphism $f:(A, \prec , \succ , \cdot , \alpha , \beta )\rightarrow (A', \prec ', \succ ', \cdot ', \alpha ', \beta ')$ of BiHom-tri-dendriform algebras is a linear map $f:A\rightarrow A'$ satisfying $f(x\prec y)=f(x)\prec ' f(y)$, $f(x\succ y)=f(x)\succ ' f(y)$ and $f(x\cdot y)=f(x)\cdot ' f(y)$, for all $x, y\in A$, as well as $f\circ \alpha =\alpha '\circ f$ and $f\circ \beta =\beta '\circ f$. Let $A =< e_1, e_2 > $ be a tow dimensional vector space. Given the multiplications $$\label{ExTriDen} e_2 \prec e_2 = e_2 \succ e_2 = ae_1, \ e_2 \c e_2 = -ae_1,$$ and the linear maps $ \alpha,\beta: A\to A$ defined by $$\begin{aligned} \label{morphismExTriDen1} &\alpha(e_1) = e_1, \alpha(e_2) = e_1 +e_2, \\ &\beta(e_1) = e_1, \beta(e_2) = 2e_1 +e_2.\end{aligned}$$ Then $(A, \prec , \succ , \cdot , \alpha , \beta )$ is a regular BiHom-tri-dendriform algebra. Let $(A,\prec,\succ,\c,\a,\b)$ be a regular BiHom-tri-dendriform algebra. Then $(A,[\c,\c],\rhd,\a,\b)$ is a BiHom-post-Lie algebra, where $$\begin{aligned} \label{tri to post 1}[x,y]=& x\c y- \a^{-1}\b(y)\c \a\b^{-1}(x),\\ \label{tri to post 2} x \rhd y=& x \succ y -\a^{-1}\b(y)\prec \a\b^{-1}(x),\end{aligned}$$ for any $x,y \in A$. For $x,y,z\in A$ we prove that$$\alpha \beta \left( x\right) \rhd \left[ y,z\right] =\left[ \beta \left( y\right) ,\left( \alpha \left( x\right) \rhd z\right) \right] +\left[ \beta \left( x\right) \rhd y,\beta \left( z\right) \right]$$ We compute:$$\begin{aligned} \alpha \beta \left( x\right) \rhd \left[ y,z\right] &=&\alpha \beta \left( x\right) \rhd \left( y\cdot z\right) -\alpha \beta \left( x\right) \rhd \left( \alpha ^{-1}\beta \left( z\right) \cdot \alpha \beta ^{-1}\left( y\right) \right) \\ &=&\alpha \beta \left( x\right) \succ \left( y\cdot z\right) -\left( \alpha ^{-1}\beta \left( y\right) \cdot \alpha ^{-1}\beta \left( z\right) \right) \prec \alpha ^{2}\left( x\right) \\ \text{ \ \ }-\alpha \beta \left( x\right) &\succ &\left( \alpha ^{-1}\beta \left( z\right) \cdot \alpha \beta ^{-1}\left( y\right) \right) +\left( \alpha ^{-2}\beta \left( z\right) \cdot \left( y\right) \right) \prec \alpha ^{2}\left( x\right) .\end{aligned}$$ Then$$\begin{aligned} \left[ \beta \left( y\right) ,\left( \alpha \left( x\right) \rhd z\right) \right] &=&\left[ \beta \left( y\right) ,\alpha \left( x\right) \succ z\right] -\left[ \beta \left( y\right) ,\alpha ^{-1}\beta \left( z\right) \prec \cdot \alpha ^{2}\beta ^{-1}\left( x\right) \right] \\ &=&\beta \left( y\right) \cdot \left( \alpha \left( x\right) \succ z\right) -\beta \left( x\right) \succ \alpha ^{-1}\beta \left( z\right) \cdot \alpha \left( y\right) \\ &&-\beta \left( y\right) \cdot \left( \alpha ^{-1}\beta \left( z\right) \prec \cdot \alpha ^{2}\beta ^{-1}\left( x\right) \right) +\left( \alpha ^{-2}\beta ^{2}\left( z\right) \prec \alpha \left( x\right) \right) \cdot \alpha \left( y\right)\end{aligned}$$ Similarly, we have $$\begin{aligned} \left[ \beta \left( x\right) \rhd y,\beta \left( z\right) \right] &=&\left[ \beta \left( x\right) \succ y,\beta \left( z\right) \right] -\left[ \alpha ^{-1}\beta \left( y\right) \prec \alpha \left( x\right) ,\beta \left( z\right) \right] \\ &=&\beta \left( x\right) \succ y\cdot \beta \left( z\right) -\alpha ^{-1}\beta ^{2}\left( z\right) \cdot \left( \alpha \left( x\right) \succ \alpha ^{-1}\beta \left( x\right) \right) \\ &&-\left( \alpha ^{-1}\beta \left( y\right) \prec \alpha \left( x\right) \right) \cdot \beta \left( z\right) +\alpha ^{-1}\beta ^{2}\left( z\right) \cdot \left( y\prec \alpha ^{2}\beta ^{-1}\left( x\right) \right)\end{aligned}$$ and the equality hold applying the relations and from Definition \[DefBiHomTriDend\]. Similar, we can proof the second assertion by using -. It is easy to see that Eq. , Eq. and Eq. fit into the commutative diagram $$\xymatrix{ \text{BiHom-tri-dendriform alg.} \ar[rr]^{x\prec y + x\succ y + x\cdot y} \ar[dd]^{x\circ y =x\succ y - \a^{-1}\b(y)\prec \a\b^{-1}(x) }_{[x,y]=x\cdot y -\a^{-1}\b(y)\cdot \a\b^{-1}(x)} && \mbox{BiHom-associative alg.} \ar[dd]^{x\star y - \a^{-1}\b(y)\star \a\b^{-1}(x)} \\ &&\\ \mbox{BiHom-post-Lie alg.} \ar[rr]^{x\circ y-\a^{-1}\b(y)\circ \a\b^{-1}(x)+[x,y]} && \text{BiHom-Lie alg.} } $$ When the operation $\cdot$ of the BiHom-tri-dendriform algebra and the bracket $[\c,\c,]$ of the BiHom-post-Lie algebra are both trivial, we obtain the following commutative diagram (See for more details). $$\xymatrix{ \mbox{BiHom-dendriform alg.} \ar[rr]^{x\prec y + x\succ y} \ar[dd]^{x\succ y - \a^{-1}\b(y)\prec \a\b^{-1}(x)} && \mbox{BiHom-ssociative alg.} \ar[dd]^{x\star y - \a^{-1}\b(y)\star \a\b^{-1}(x)}\\ && \\ \mbox{BiHom-pre-Lie alg.} \ar[rr]^{x\circ y - \a^{-1}\b(y)\circ \a\b^{-1}(x)} && \mbox{BiHom-Lie alg.} }$$ \[BiHomLieToBiHomPostLie\] Let $(A, [\c,\c], \a, \b)$ be a BiHom-Lie algebra and $(V, [\c,\c]_V, \r, \f, \p)$ an $A$-module $\mathbb{K}$-algebra. The linear map $T: V\rightarrow A$ is an $\mathcal{O}$-operator of weight $\lambda\in \mathbb{K}$ associated to $(V, [\c,\c]_V, \r, \f, \p)$. Define two new bilinear operations $\{\c,\c\}, \vartriangleright: V\times V\rightarrow V$ as follows $$\begin{aligned} \{u,v\}=\lambda[u,v]_V,~~u\vartriangleright v=\r(T(u))v,\end{aligned}$$ for any $u,v\in V$. Then $(V, \{\c,\c\}, \vartriangleright, \f, \p)$ is a BiHom-post-Lie algebra and $T$ is a morphism of BiHom-Lie algebras from the associated BiHom-Lie algebra of $(V, \{\c,\c\}, \vartriangleright, \f, \p)$ to $(A, [\c,\c], \a, \b)$. Furthermore, $T(V)$ is a BiHom-Lie subalgebra of $(A, [\c,\c], \a, \b)$ and there is an induced BiHom-post-Lie algebra structure on $T(V)$ given by $$\label{InducedBiHomPostLie} [T(u),T(v)]_{T(V)}=T(\{u,v\})\ \ , T(u)\vartriangleright_{T(V)}T(v)=T(u\vartriangleright v),\ \ \forall\ u,v\in V.$$ We use the last condition of representation of BiHom-Lie algebras on $\mathbb{K}$-algebra. $$\begin{aligned} &\phi\psi(a)\rhd\{b,c\}-{\psi(a)\rhd b,\psi(c)}-\{\psi(b),\phi(a)\rhd c\}\\ &=\phi\psi(a)\rhd(\lambda[b,c]_{V})-\{\rho(T(\psi(a)))b,\psi(c)\}-\{\psi(b),\rho(T(\phi(a)))c\}\\ &=\rho(T(\phi\psi(a)))(\lambda[b,c]_{V})-\lambda[\rho(T(\psi(a)))b,\psi(c)]_{V}-\lambda[\psi(b),\rho(T(\phi(a)))c]_{V}\\ &=\lambda(\rho(\alpha\beta(T(a)))[b,c]_{V}-[\rho(\beta(T(a)))b,\psi(c)]-[\psi(b),\rho(\alpha(T(a)))c]_{V}\\ &=0.\end{aligned}$$ Using the condition of Definition \[defi:bihom-lie representation\], we check $$\begin{aligned} &\{\psi(a),\phi(b)\}\rhd\psi(c)-\phi\psi(a)\rhd(\phi(b)c)\\ &-(\psi(a)\rhd\phi(b))\rhd\psi(c)+\phi\psi(b)\rhd(\phi(a)\rhd c)+(\psi(b)\rhd\phi(a))\rhd\psi(c)\\ &=\rho(T(\rho(T(a)\phi(b))T(c)-\rho(T(\rho(T(\phi^{-1}\psi\phi(b)))\phi(a))\psi(a)\\ &+\rho(T(\{\psi(a),\phi(b)\}))\psi(c)-\phi\psi(a)(b \rhd c)+\psi\phi(b)\rhd(\phi(a) \rhd c)\\ &=\rho([T(\psi(a)),T(\phi(b)))])\psi(c)-\rho(T(\phi\psi(a)))\rho(T(a))c+\rho(T(\psi\phi(b)))\rho(T(\phi(a)))c\\ &=\rho(\beta(T(a)),T(\phi(b))])\psi(c)-\rho(\alpha\beta(T(a)))\rho(T(\phi(b)))c+\rho(\beta(T(\phi(b))))\rho(\alpha(T(a)))c\\ &=\rho([\beta(T(a)),T(\phi(b))])-\rho(\alpha\beta(T(a)))\rho(T(\phi(b)))c+\rho(\beta(T(\phi(b))))\rho(\alpha(T(a)))c\\ &=0.\end{aligned}$$ An obvious consequence of Theorem \[BiHomLieToBiHomPostLie\] is the following construction of a BiHom-post-Lie algebra in terms of Rota-Baxter operator of weight $\lambda$ on a BiHom-Lie algebra. Let $(A, [\c,\c], \a, \b)$ be a BiHom-Lie algebra. Then there exists a compatible BiHom-post-Lie algebra structure on $A$ if and only if there exists an invertible $\mathcal{O}$-operator of weight $\lambda\in \mathbb{K}$ on $A$. Let $(A,[\c,\c],\rhd,\a,\b)$ be a BiHom-post-Lie algebra and $(A,[\c,\c],\a,\b)$ be the associated BiHom-Lie algebra. Then the identity map $id: A \to A$ is an invertible $\mathcal{O}$-operator of weight $1$ on $(A,[\c,\c],\a,\b)$ associated to $(A,[\c,\c],ad,\a,\b)$ . Conversely, suppose that there exists an invertible $\mathcal{O}$-operator $T$ (of weight $\lambda$) of $(A,[\c,\c],\a,\b)$ associated to an $A$-module $\mathbb{K}$-algebra $(V, [\c,\c]_V, \r, \f, \p)$ . Then, using Proposition \[BiHomLieToBiHomPostLie\], there is a BiHom-post-Lie algebra structure on $T(V)=A$ given by $$\{T(u),T(v)\}=\lambda T([u,v]_V)\ \ , T(u)\vartriangleright T(v)=T(\rho(T(u))v),\ \ \forall\ u,v\in V.$$ If we st $x=T(u)$ and $y=T(v)$, then we get $$\{x,y\}=\lambda T([T^{-1}(x),T^{-1}(y)]_V)\ \ , x\rhd y=T(\rho(x)T^{-1}(y)).$$ This is compatible BiHom-post-Lie algebra structure on $(A,[\c,\c],\a,\b)$. Indeed, $$\begin{aligned} & x\rhd y-\a^{-1}\b(y)\rhd \a\b^{-1}(x)+ \{x,y\} \\ &= T(\rho(x)T^{-1}(y)-\rho(\a^{-1}\b(y))T^{-1}\a\b^{-1}(x)+\lambda [T^{-1}(x),T^{-1}(y)]_V)\\ &=[TT^{-1}(x),TT^{-1}(y)]=[x,y].\end{aligned}$$ The proof is finished. \[PostLieByRotaBaxter\] Let $(A, [\c,\c], \a, \b)$ be a BiHom-Lie algebra and the linear map $R: A\rightarrow A$ is a Rota-Baxter operator of weight $\lambda\in \mathbb{K}$. Then there exists a BiHom-post-Lie structure on $A$ given by $$\begin{aligned} \{x,y\}=\lambda[x,y],~~x\vartriangleright y=[R(x),y],\ \forall \ x,y\in A.\end{aligned}$$ If in addition, $R$ is invertible, then there is a compatible BiHom-Post-Lie algebra structure on $A$ given by $$[x,y]'=\lambda R([R^{-1}(x),R^{-1}(y)])\ \ , x\rhd y=R([x,R^{-1}(y)]),\ \forall x,y \in A.$$ Let $L=sl(2,\K)$ whose standard basis consists of $$X=\left( \begin{array}{cc} 0 & 1 \\ 0 & 0 \\ \end{array} \right), Y=\left( \begin{array}{cc} 0 & 0 \\ 1 & 0 \\ \end{array} \right), H=\left( \begin{array}{cc} 1 & 0 \\ 0 & -1 \\ \end{array} \right).$$ Then $[H,X]=2X,\quad [H,Y]=-2Y,\quad [X,Y]=H.$ Define two linear maps $\alpha, \beta: L\rightarrow L$ by $$\alpha(X)=\lambda^2 X, \quad \alpha(Y)=\frac{1}{\lambda^2}Y,\quad \alpha(H)=H,$$ $$\beta(X)=\gamma^2X, \quad \beta(Y)=\frac{1}{\gamma^2}Y,\quad \beta(H)=H,$$ where $\lambda,\gamma$ are parameters in $\K$. Obviously we check that $\alpha$ and $\beta$ are two morphisms of the Lie algebra $(L,[\c,\c])$. Consider the linear map $ [\c,\c]_{\a,\b}\colon L\otimes L\rightarrow L$ $$\begin{gathered} [a,b]_{\a,\b} = [ \alpha (a) ,\beta (b) ] ,\qquad \text{for all} \ \ a,b\in L\end{gathered}$$ defined in the basis $X,\ Y, \ H$ by $$[H,X]_{\a,\b}=2\gamma^2 X,\ \ [H,Y]_{\a,\b} =-\frac{2}{\gamma^2}Y,\ \ [X,Y]_{\a,\b}=\frac{\lambda^2}{\gamma^2}H.$$ Then $L_{( \alpha ,\beta ) }:=(L, [\c,\c]_{\a,\b}, \alpha , \beta )$ is a BiHom-Lie algebra. Now, define the linear map $R:L\to L$ by $$\label{RotaBaxterSl2} R(X)=0,\ \ R(Y)=4Y\ \ \textrm{and}\ \ R(H)=2H.$$ Then $R$ is a Rota-Baxter operator of weight $-4$ on $L_{( \alpha ,\beta ) }$. Using Corollary \[PostLieByRotaBaxter\], we can construct a BiHom-Post-Lie algebra on $L_{( \alpha ,\beta ) }$ given by $$\{H,X\}=-8\gamma^2 X,\ \ \{H,Y\} =\frac{8}{\gamma^2}Y,\ \ \{X,Y\}=-4\frac{\lambda^2}{\gamma^2}H$$ and $$X\rhd Y=X\rhd H=0,\ \ Y\rhd X=-4\frac{\gamma^2}{\lambda^2}H,\ \ Y\rhd H=-\frac{8}{\lambda^2}Y,\ \ H\rhd X=4\gamma^2X,\ \ H\rhd Y=-\frac{4}{\gamma^2}Y.$$ Let $(A, [\c,\c], \a, \b)$ be a BiHom-Lie algebra such that $A=A_1\oplus A_2$, where $A_1$ and $ A_2$ are two BiHom-Lie subalgebras, and the linear map $R: A\rightarrow A$ given by $$R(x_1+x_2)=-\lambda x_2, \forall\ x_1\in A_1, \ x_2\in A_2.$$ It is easy to check that $R$ is a Rota-Baxter operator of weight $\lambda\in \mathbb{K}$ on $A$. Then $(A,\vartriangleright, \{\c,\c\}, \a, \b)$ is a BiHom-post-Lie algebra, where $$\begin{aligned} \{x,y\}=\lambda[x,y],~~x\vartriangleright y=[R(x),y],\ \forall \ x,y\in A.\end{aligned}$$ Representation theory ===================== In this section, we introduce the notion of representations of a BiHom-post-Lie algebra $(A,[\c,\c],\rhd,\a,\b)$ on a vector space $V$. \[rep bihom postlie\] A representation of a BiHom-post-Lie algebra $(A,[\c,\c],\rhd,\a,\b)$ on a vector space $V$ is a tuple $(V,\rho,\mu,\nu,\f,\p)$, such that $(V,\rho,\f,\p)$ is a representation of $(A,[\c,\c],\a,\b)$ and $\mu,\nu: A \to gl(V)$ are linear maps satisfying $$\begin{aligned} &\mu(\alpha(x)) \f=\f \mu(x),\ \ \mu(\b(x)) \p=\p \mu(x), \\ &\nu(\alpha(x) \f=\f \nu(x),\ \ \nu(\b(x)) \p=\p \nu(x), \\ & \nu([x,y]) \f \p=\rho(\b(x)) \nu(y) \f-\rho(\b(y)) \nu(x) \p , \label{rep bihom pos lie 1}\\ & \rho(\b(x)\rhd y) \p=\mu(\a\b(x)) \rho(y)-\rho(\b(y)) \mu(\a(x)) , \label{rep bihom pos lie 2}\\ &\mu([\b(x),\a(y)]) \p=\mu(\a\b(x))\mu(\a(y))-\mu(\b(x)\rhd \a(y)) \p \nonumber\\ & \hspace{0.5 cm}-\mu(\a\b(y))\mu(\a(x))+\mu(\b(y)\rhd \a(x)) \p , \label{rep bihom pos lie 3}\\ & \nu(\b(y))\rho(\b(x)) \f=\mu(\a\b(x))\nu(y) \f-\nu(\b(y))\mu(\b(x))\f \nonumber\\ & \hspace{0.5 cm}-\nu(\a(x)\rhd y)\f\p +\nu(\b(y))\nu(\a(x)) \p , \label{rep bihom pos lie 4}\end{aligned}$$ for any $x,y \in A$. Let $(A,[\c,\c],\rhd,\a,\b)$ be a BiHom-post-Lie algebra and $(V,\rho,\mu,\nu,\f,\p)$ a representation of $(A,[\c,\c],\rhd,\a,\b)$. By identity , we deduce that $(V,\mu,\f,\p)$ is a representation of the sub-adjacent BiHom-Lie algebra $(A,\{\c,\c\},\a,\b)$ of $(A,[\c,\c],\rhd,\a,\b)$. In addition, it is obvious that $(A,ad,L_{\rhd},R_{\rhd},\a,\b)$ is a representation of $(A,[\c,\c],\rhd,\a,\b)$ which is called the adjoint representation. \[semi-directproduct post-lie alg\] A tuple $(V,\rho,\mu,\nu,\f,\p)$ is a representation of a BiHom-post-Lie algebra\ $(A,[\c,\c],\rhd,\a,\b)$ if and only if $(A\oplus V,[\c,\c]_{\rho},\rhd_{\mu,\nu},\a+\f,\b+\p)$ is a BiHom-post-Lie algebra, where for any $x,y \in A$ and $u,v \in V$ $$\begin{aligned} &(\alpha+\f)(x+u)=\alpha(x)+\f(u),\\ &(\b+\p)(x+u)=\b(x)+\p(u), \\ &[x+u,y+v]_{\rho}=[x,y]+\rho(x)v-\rho(\a^{-1}\b(y))\f\p^{-1}(u),\\ &(x+u)\rhd_{\mu,\nu}(y+v)=x\rhd y+\mu(x)v+\nu(y)u.\end{aligned}$$ By the conditions and in Definition \[rep bihom postlie\] and the identity , we have $$\begin{aligned} &[(\beta+\psi)(x+a),(\alpha+\phi)(y+b)]_{\rho}\rhd_{\mu,\nu}(\beta+\psi)(z+c)\\ &-as_{\alpha+\phi,\beta+\psi}((\beta+\psi)(x+a),(\alpha+\phi)(y+b),z+c)\\ &+as_{\alpha+\phi,\beta+\psi}((\beta+\psi)(y+b),(\alpha+\phi)(x+a),z+c)\\ &=[\beta(x),\alpha(y)]\rhd\beta(z)+\mu([\beta(x),\alpha(y)])\psi(c)+\nu(\beta(z))(\rho(\beta(x))\phi(b)-\rho(\beta(y))\phi(a))\\ &-\alpha\beta(x)\rhd(\alpha(y)\rhd z)-\mu(\alpha\beta(x))\mu(\alpha(y))c-\mu(\alpha\beta(x))\nu(z)\phi(b)-\nu(\alpha(y)\rhd z)\phi\psi(a)\\ &+(\beta(x)\rhd\alpha(y))\rhd\beta(z)+\mu(\beta(x)\rhd\alpha(y))\psi(c)+\nu(\beta(z))\mu(\beta(x))\phi(b)+\nu(\beta(z))\nu(\alpha(y))\psi(a)\\ &+\alpha\beta(y)\rhd(\alpha(x)\rhd z)+\mu(\alpha\beta(y))\mu(\alpha(x))c+\mu(\alpha\beta(y))\nu(z)\phi(a)+\nu(\alpha(x)\rhd z)\phi\psi(b)\\ &-(\beta(y)\rhd\alpha(x))\rhd\beta(z)-\mu(\beta(y)\rhd\alpha(x))\psi(c)-\nu(\beta(z))\mu(\beta(y))\phi(a)-\nu(\beta(z))\nu(\alpha(x))\psi(b)\\ &=0.\end{aligned}$$ Similarly we can check the identity for the algebra $A\bigoplus V$ using the axioms – and . Let $(A,[\c,\c],\rhd)$ be a post-Lie algebra and $\a,\b$ be two commuting post-Lie algebra morphisms on $A$. Consider a representation $(V,\rho,\mu,\nu)$ $(A,[\c,\c],\rhd)$ and $\f,\p: V \to V$ be two commuting linear maps obeying two the following conditions $$\begin{aligned} & \rho(\alpha(x))\circ \f=\f\circ \rho(x),\ \ \rho(\b(x))\circ \p=\p \circ \rho(x), \\ & \mu(\alpha(x))\circ \f=\f\circ \mu(x),\ \ \mu(\b(x))\circ \p=\p \circ \mu(x), \\ & \nu(\alpha(x))\circ \f=\f\circ \nu(x),\ \ \nu(\b(x))\circ \p=\p \circ \nu(x).\\\end{aligned}$$ Define the three linear maps $\widetilde{\rho}, \widetilde{\mu}, \widetilde{\nu}: A\longrightarrow gl(V)$ by\ $\widetilde{\rho}(x)(v)=\rho(\alpha(x))( \p(v))$, $\widetilde{\mu}(x)(v)=\mu(\alpha(x))( \p(v))$ and $\widetilde{\nu}(x)(v)=\nu(\alpha(x))( \p(v))$. \[twistmodule\] The tuple$(V,\widetilde{\rho},\widetilde{\mu},\widetilde{\nu},\f,\p)$ is a representation of $(A,[\cdot,\cdot]_{\a,\b},\rhd_{\a,\b},\alpha,\beta)$. straightforward. Let $(V,\rho,\mu,\nu,\f,\p)$ be a representation of a BiHom-post-Lie algebra $(A,[\c,\c],\rhd,\a,\b)$. Define the linear map $\pi:A \to gl(V)$ by $$\begin{aligned} \pi(x):=\rho(x)+\mu(u)-\nu(\a\b^{-1}(x))\f^{-1}\p,\ \forall x \in A.\end{aligned}$$ With the above notations, $(V,\pi,\f,\p)$ is a representation of the sub-adjacent BiHom-Lie algebra $(A,\{\c,\c\},\a,\b)$ of $(A,[\c,\c],\rhd,\a,\b)$. By Theorem \[semi-directproduct post-lie alg\], we have the semi-direct product BiHom-post-Lie algebra $(A\oplus V,[\c,\c]_{\rho},\rhd_{\mu,\nu},\a+\f,\b+\p)$. Consider its sub-adjacent Lie algebra structure $[\c,\c]^C$, we have [$$\begin{aligned} & [x+u,y+v]^C=(x+u)\rhd_{\mu,\nu} (y+v)- (\a^{-1}\b(y)+\f^{-1}\p(v)) \rhd_{\mu,\nu} (\a\b^{-1}(x)+\f\p^{-1}(u)) +[x+u,y+v]_{\rho} \\ &= x\rhd y+\mu(x)v+\nu(y)u- \a^{-1}\b(y) \rhd \a\b^{-1}(x)-\mu(\a^{-1}\b(y))\f\p^{-1}(u)-\nu(\a\b^{-1}(x))\f^{-1}\p(v) \\ & + [x,y]+\rho(x)v-\rho(\a^{-1}\b(y))\f\p^{-1}(u) \\ &=\{x,y\} +\pi(x)v-\pi(\a^{-1}\b(y))\f\p^{-1}(u).\end{aligned}$$ ]{} Then $(V,\pi,\f,\p)$ is a representation of the sub-adjacent BiHom-Lie algebra $(A,\{\c,\c\},\a,\b)$. [999]{} C. Bai, O.Bellier, L. Guo, X. Ni. Splitting of operations, Manin products, and Rota—Baxter operators. Int. Math. Res. Notes. 2013. Vol. 3. P. 485–524. I. Bakayoko. Hom-post-Lie modules, $\mathcal{O}$-operators and some functors on Hom-algebras. arXiv preprint arXiv:1610.02845 (2016). D. Burde, K. Dekimpe. Post-Lie algebra structures on pairs of Lie algebras. J. Algebra. 2016. Vol. 464. P. 226–245. C. Charles, K. E-Fard and H. M-Kaas. What is a post-Lie algebra and why is it useful in geometric integration. European Conference on Numerical Mathematics and Advanced Applications. Springer, Cham, 2017. T. Chtioui, S. Mabrouk, A. Makhlouf, BiHom-alternative, BiHom-Malcev and BiHom-Jordan algebras, arXiv:1811.10394v1 (2018). T. Chtioui, S. Mabrouk, A. Makhlouf, BiHom-pre-alternative algebras and BiHom-alternative quadri-algebras." arXiv preprint arXiv:1903.03994 (2019). Y. Cheng, and H. Qi. Representations of BiHom-Lie algebras. arXiv preprint arXiv:1610.04302 (2016). K. E-Fard, A. Lundervold, H. M-Kaas. On the Lie enveloping algebra of a post-Lie algebra. J. Lie Theory. 2015. Vol. 25, no. 4. P. 1139–1165. Floystad, Gunnar, and H. M-Kaas. Pre-and Post-Lie Algebras: The Algebro-Geometric View. The Abel Symposium. Springer, Cham, 2016. G. Graziani, A. Makhlouf, C. Menini and F. Panaite, BiHom-Associative Algebras, BiHom-Lie Algebras and BiHom-Bialgebras. Symmetry, Integrability and Geometry. Methods and Applications 11 (2015). Gubarev, V. Yu. Universal Enveloping Lie Rota-Baxter Algebras of Pre-Lie and Post-Lie Algebras. Algebra and Logic: (2017) 1-14. Gubarev, V.Yu., P.S. Kolesnikov. Operads of decorated trees and their duals Comment. Math. Univ. Carolin. 2014. Vol. 55, No 4. P. 421–445. H. M-Kaas and A. Lundervold., On post-Lie algebras, Lie-Butcher series and moving frames. Foundations of Computational Mathematics 13.4 (2013): 583-613. J.L. Loday Dialgebras, Dialgebras and related operads. Berlin: Springer-Verl., 2001. P. 1–61. (Lectures Notes in Math. Vol. 1763) J.-L. Loday, M. Ronco. Trialgebras and families of polytopes Comtep. Math. 2004. Vol. 346. P. 369–398. L. Liu, A. Makhlouf, C. Menini, F. Panaite, (2017). BiHom-pre-Lie algebras, BiHom-Leibniz algebras and Rota-Baxter operators on BiHom-Lie algebras. arXiv preprint arXiv:1706.00474. Liu, Ling, et al. Rota-Baxter operators on BiHom-associative algebras and related structures. arXiv preprint arXiv:1703.07275 (2017). Tang, Rong, et al. Homotopy Rota-Baxter operators, homotopy $\mathcal {O} $-operators and homotopy post-Lie algebras. arXiv preprint arXiv:1907.13504 (2019). B. Vallette. Homology of generalized partition posets. J. Pure Appl. Algebra 208 (2007), no. 2, 699-725. P. Yu, Q. Liu, C. Bai and L. Guo. PostLie algebra structures on the Lie algebra $\mathrm{sl}(2,\mathbb{C})$ . Electron. J. Linear Algebra. 2012. Vol. 23. P. 180–197. [^1]: Corresponding author, E-mail: h.adimi@univ-bba.dz [^2]: Corresponding author, E-mail: chtioui.taoufik@yahoo.fr [^3]: Corresponding author, E-mail: Mabrouksami00@yahoo.fr [^4]: Corresponding author, E-mail: sonia.massoud2015@gmail.com

--- abstract: 'We present a deep generative model for Zero-Shot Learning (ZSL). Unlike most existing methods for this problem, that represent each class as a *point* (via a semantic embedding), we represent each seen/unseen class using a class-specific *latent-space distribution*, conditioned on class attributes. We use these latent-space distributions as a prior for a *supervised* variational autoencoder (VAE), which also facilitates learning highly discriminative feature representations for the inputs. The entire framework is learned end-to-end using only the seen-class training data. At test time, the label for an unseen-class test input is the class that maximizes the VAE lower bound. We further extend the model to a ([ ]{}) semi-supervised/transductive setting by leveraging unlabeled unseen-class data via an *unsupervised* learning module, and ([ ]{}) few-shot learning where we also have a small number of labeled inputs from the unseen classes. We compare our model with several state-of-the-art methods through a comprehensive set of experiments on a variety of benchmark data sets.' author: - | Wenlin Wang$^{1}$[[^1]]{}, Yunchen Pu$^{1}$, Vinay Kumar Verma$^{3}$, Kai Fan$^{2}$, Yizhe Zhang$^{2}$\ [**[Changyou Chen$^{4}$, Piyush Rai$^{3 *}$, Lawrence Carin$^{1}$]{}**]{}\ $^{1}$Department of Electrical and Computer Engineering, Duke University\ $^{2}$Compuational Biology and Bioinformatics, Duke University\ $^{3}$Department of Computer Science and Engineering, IIT Kanpur, India\ $^{4}$Department of Computer Science and Engineering, SUNY at Buffalo\ [, [[{vkverma, piyush}@cse.iitk.ac.in]{}]{}, [[cchangyou@gmail.com]{}]{} ]{} title: 'Zero-Shot Learning via Class-Conditioned Deep Generative Models' --- Introduction ============ A goal of autonomous learning systems is the ability to learn new concepts even when the amount of supervision for such concepts is scarce or non-existent. This is a task that humans are able to perform effortlessly. Endowing machines with similar capability, however, has been challenging. Although machine learning and deep learning algorithms can learn reliable classification rules when supplied with abundant labeled training examples per class, their generalization ability remains poor for classes that are not well-represented (or not present) in the training data. This limitation has led to significant recent interest in zero-shot learning (ZSL) and one-shot/few-shot learning [@socher2013zero; @lampert2014attribute; @fei2006one; @lake2015human; @vinyals2016matching; @ravi2017optimization]. We provide a more detailed overview of existing work on these methods in the Related Work section. In order to generalize to previously unseen classes with no labeled training data, a common assumption is the availability of side information about the classes. The side information is usually provided in the form of class attributes (human-provided or learned from external sources such as Wikipedia) representing semantic information about the classes, or in the form of the similarities of the unseen classes with each of the seen classes. The side information can then be leveraged to design learning algorithms [@socher2013zero] that try to transfer knowledge from the seen classes to unseen classes (by linking corresponding attributes). Although this approach has shown promise, it has several limitations. For example, most of the existing ZSL methods assume that each class is represented as a fixed point (e.g., an embedding) in some semantic space, which does not adequately account for intra-class variability [@akata2015evaluation; @mukherjee2016gaussian]. Another limitation of most existing methods is that they usually lack a proper generative model [@kingma2014auto; @rezende2014stochastic; @kingma2014semi] of the data. Having a generative model has several advantages [@kingma2014auto; @rezende2014stochastic; @kingma2014semi], such as unraveling the complex structure in the data by learning expressive feature representations and the ability to seamlessly integrate unlabeled data, leading to a transductive/semi-supervised estimation procedure. This, in the context of ZSL, may be especially useful when the amount of labeled data for the seen classes is small, but otherwise there may be plenty of unlabeled data from the seen/unseen classes. {width="\textwidth"} Motivated by these desiderata, we design a deep generative model for the ZSL problem. Our model (summarized in Figure \[fig:idea\]) learns a set of *attribute-specific* latent space distributions (modeled by Gaussians), whose parameters are outputs of a trainable deep neural network (defined by $p_\psi$ in Figure \[fig:idea\]). The attribute vector is denoted as $\av$, and is assumed given for each training image, and it is inferred for test images. The class label is linked to the attributes, and therefore by inferring attributes of a test image, there is an opportunity to recognize classes at test time that were not seen when training. These latent-space distributions serve as a prior for a variational autoencoder (VAE) [@kingma2014auto] model (defined by a decoder $p_\theta$ and an encoder $q_\phi$ in Figure \[fig:idea\]). This combination further helps the VAE to learn discriminative feature representations for the inputs. Moreover, the generative aspect also facilitates extending our model to semi-supervised/transductive settings (omitted in Figure \[fig:idea\] for brevity, but discussed in detail in the Transductive ZSL section) using a deep *unsupervised* learning module. All the parameters defining the model, including the deep neural-network parameters $\psi$ and the VAE decoder and encoder parameters $\theta, \phi$, are learned end-to-end, using only the seen-class labeled data (and, optionally, the available unlabeled data when using the semi-supervised/transductive setting). Once the model has been trained, it can be used in the ZSL setting as follows. Assume that there are classes we wish to identify at test time that have not been seen when training. While we have not seen images before from such classes, it is assumed that we know the attributes of these previously unseen classes. The latent space distributions $p_\psi(\zv|\av)$ for all the unseen classes (Figure \[fig:idea\], best seen in color, shows this distribution for one such unseen class using a red-dotted ellipse) are inferred by conditioning on the respective class attribute vectors $\av$ (including attribute vectors for classes not seen when training). Given a test input $\xv_*$ from some unseen class, the associated class attributes $\av_*$ are predicted by first mapping $\xv_*$ to the latent space via the VAE recognition model $q_\phi(\zv_*|\xv_*)$, and then finding $\av_*$ that maximizes the VAE lower bound. The test image is assigned a class label $y_*$ linked with $\av_*$. This is equivalent to finding the class latent distribution $p_\psi$ that has the smallest KL divergence w.r.t. the variational distribution $q_\phi(\zv_*|\xv_*)$. Variational Autoencoder ======================= The variational autoencoder (VAE) is a deep generative model [@kingma2014auto; @rezende2014stochastic], capable of learning complex density models for data via latent variables. Given a nonlinear generative model $p_{\theta}(\xv|\zv)$ with input $\xv \in \mathbb{R}^D$ and associated latent variable $\zv \in \mathbb{R}^L$ drawn from a prior distribution $p_0(\zv)$, the goal of the VAE is to use a recognition model $q_{\phi}(\zv|\xv)$ (also called an inference network) to approximate the posterior distribution of the latent variables, i.e., $p_{\theta}(\zv|\xv)$, by maximizing the following variational lower bound $$\begin{aligned} \Lcal_{\theta,\phi}^{\mbox{v}}(\xv) = \E_{q_\phi(\zv|\xv)}[\log p_\theta(\xv|\zv)] - \mbox{KL}(q_\phi(\zv|\xv) || p_0(\zv))~.\end{aligned}$$ Typically, $q_{\phi}(\zv|\xv)$ is defined as an isotropic normal distribution with its mean and standard deviation the output of a deep neural network, which takes $\xv$ as input. After learning the VAE, a probabilistic “encoding” $\zv$ for the input $\xv$ can be generated efficiently from the recognition model $q_{\phi}(\zv|\xv)$. We leverage the flexibility of the VAE to design a *structured*, supervised VAE that allows us to incorporate class-specific information (given in the form of class attribute vectors $\av$). This enables one to learn a deep generative model that can be used to predict the labels for examples from classes that were not seen at training time (by linking inferred attributes to associated labels, even labels not seen when training). Deep Generative Model for ZSL ============================= We consider two settings for ZSL learning: inductive and transductive. In the standard inductive setting, during training, we only assume access to labeled data from the seen classes. In the transductive setting [@kodirov2015unsupervised], we also assume access to the *unlabeled* test inputs from the unseen classes. In what follows, under the *Inductive ZSL* section, we first describe our deep generative model for the inductive setting. Then, in the *Transductive ZSL* section, we extend this model for the transductive setting, in which we incorporate an *unsupervised* deep embedding module to help leverage the *unlabeled* inputs from the unseen classes. Both of our models are built on top of a variational autoencoder [@kingma2014auto; @rezende2014stochastic]. However, unlike the standard VAE [@kingma2014auto; @rezende2014stochastic], our framework leverages attribute-specific latent space distributions which act as the prior (Figure \[fig:idea\]) on the latent codes of the inputs. This enables us to adapt the VAE framework for the problem of ZSL. #### Notation In the ZSL setting, we assume there are $S$ seen classes and $U$ unseen classes. For each seen/unseen class, we are given side information, in the form of $M$-dimensional class-attribute vectors [@socher2013zero]. The side information is leveraged for ZSL. We collectively denote the attribute vectors of all the classes using a matrix $\Amat \in \R^{M\times (S+U)}$. During training, images are available only for the seen classes, and the labeled data are denoted $\Dcal_s = \{(\xv_n, \av_n)\}_{n=1}^{N}$, where $\xv_n \in \mathbb{R}^D$ and $\av_n = \Amat_{y_n}$, $\Amat_{y_n}\in \mathbb{R}^M$ denotes the $y_n^{th}$ column of $\Amat$ and $y_n \in \{1,\dots, S\}$ is the corresponding label for $\xv_n$. The remaining classes, indexed as $\{S+1, \dots,S+U\}$, represent the unseen classes (while we know the $U$ associated attribute vectors, at training we have no corresponding images available). Note that each class has a unique associated attribute vector, and we infer unseen classes/labels by inferring the attributes at test, and linking them to a label. Inductive ZSL {#sec:ind} ------------- We model the data $\{\xv_n\}_{n=1}^N$ using a VAE-based deep generative model, defined by a decoder $p_\theta(\xv_n|\zv_n)$ and an encoder $q_\phi(\zv_n|\xv_n)$. As in the standard VAE, the decoder $p_{\theta}(\xv_n|\zv_n)$ represents the generative model for the inputs $\xv_n$, and $\theta$ represents the parameters of the deep neural network that define the decoder. Likewise, the encoder $q_\phi(\zv_n|\xv_n)$ is the VAE *recognition model*, and $\phi$ represents the parameters of the deep neural network that define the encoder. However, in contrast to the standard VAE prior that assumes each latent embedding $\zv_n$ to be drawn from the same latent Gaussian (e.g., $p_{\psi}(\zv_n) = \Ncal(0,\Imat)$), we assume each $\zv_n$ to be drawn from a *attribute-specific* latent Gaussian, $p_{\psi}(\zv_n|\av_n) = \Ncal(\muv(\av_n), \Sigmamat(\av_n))$, where (\_n) = f\_(\_n),   (\_n) = () \[eq:meancov\] where we assume $f_\mu(\cdot)$ and $f_\sigma(\cdot)$ to be linear functions, [*i.e.*]{}, $f_\mu(\av_n) = \Wmat_\mu \av_n$ and $f_\sigma(\av_n) = \Wmat_\sigma \av_n$; $\Wmat_\mu$ and $\Wmat_\sigma$ are learned parameters. One may also consider $f_\mu(\cdot)$ and $f_\sigma(\cdot)$ to be a deep neural network; this added complexity was not found necessary for the experiments considered. Note that once $\Wmat_\mu$ and $\Wmat_\sigma$ are learned, the parameters $\{\muv(\av), \Sigmamat(\av)\}$ of the latent Gaussians of unseen classes $c=S+1,\ldots,S+U$ can be obtained by plugging in their associated class attribute vectors $\{\Amat_c\}_{c=S+1}^{S+U}$, and inferring which provides a better fit to the data. Given the class-specific priors $p_{\psi}(\zv_n|\av_n)$ on the latent code $\zv_n$ of each input, we can define the following variational lower bound for our VAE based model (we omit the subscript $n$ for simplicity) $$\begin{aligned} \vspace{-1em} \Lcal_{\theta,\phi,\psi}(\xv, \av) = \E_{q_\phi(\zv|\xv)}[\log p_\theta(\xv|\zv)] - \mbox{KL}(q_\phi(\zv|\xv) || p_\psi(\zv|\av))\label{eq:elbo} \vspace{-0.5em} \end{aligned}$$ **Margin Regularizer** The objective in (\[eq:elbo\]) naturally encourages the inferred variational distribution $q_\phi(\zv|\xv)$ to be close to the class-specific latent space distribution $p_\psi(\zv|\av)$. However, since our goal is classification, we augment this objective with a *maximum-margin* criterion that promotes $q_\phi(\zv|\xv)$ to be as far away as possible from all other class-specific latent space distributions $p_{\psi}(\zv|\Amat_c)$, $\Amat_c \neq \av$. To this end, we replace the $- \mbox{KL}(q_\phi(\zv|\xv) || p_\psi(\zv|\av))$ term in our original VAE objective (\[eq:elbo\]) by $-[\mbox{KL}(q_\phi(\zv|\xv) || p_\psi(\zv|\av)) - R^*]$ where “margin regularizer” term $R^*$ is defined as the minimum of the KL divergence between $q_\phi(\zv|\xv)$ and all other class-specific latent space distributions: [$$\begin{aligned} R^* &= \min_{c: c\in \{1.., y-1, y+1,..,S\}} \{\mbox{KL}(q_\phi(\zv|\xv)||p_\psi(\zv|\Amat_c))\} ~\nonumber\\ &= -\max_{c: c\in \{1.., y-1, y+1,..,S\}} \{-\mbox{KL}(q_\phi(\zv|\xv)||p_\psi(\zv|\Amat_c))\}\end{aligned}$$]{} Intuitively, the regularizer $-[\mbox{KL}(q_\phi(\zv|\xv) || p_\psi(\zv|\av)) - R^*]$ encourages the true class and the *next best* class to be separated maximally. However, since $R^*$ is non-differentiable, making the objective difficult to optimize in practice, we approximate $R^*$ by the following surrogate: $$\begin{aligned} R = -\log\sum_{c=1}^S \exp(-\mbox{KL}(q_\phi(\zv|\xv)||p_\psi(\zv|\Amat_c)))\end{aligned}$$ It can be easily shown that $$\begin{aligned} R^* \leq R \leq R^* + \log S\end{aligned}$$ Therefore when we maximize $R$, it is equivalent to maximizing a lower bound on $R^*$. Finally, we optimize the variational lower bound together with the margin regularizer as [$$\begin{aligned} \hat{\Lcal}_{\theta,\phi,\psi}(\xv, \av) &= \E_{q_\phi(\zv|\xv)}[\log p_\theta(\xv|\zv)] - \mbox{KL}(q_\phi(\zv|\xv) || p_\psi(\zv|\av)) \nonumber \\ &\underbrace{-\lambda\log\sum_{c=1}^S \exp(-\mbox{KL}(q_\phi(\zv|\xv)||p_\psi(\zv|\Amat_c)))}_{R}\label{eq:loss}\end{aligned}$$]{} where $\lambda$ is a hyper-parameter controlling the extent of regularization. We train the model using the *seen-class* labeled examples $\Dcal_s = \{(\xv_n, \av_n)\}_{n=1}^{N}$ and learn the parameters $(\theta, \phi, \psi)$ by maximizing the objective in (\[eq:loss\]). Once the model parameters have been learned, the label for a new input $\hat{\xv}$ from an *unseen* class can be predicted by first predicting its latent embedding $\hat{\zv}$ using the VAE recognition model, and then finding the “best” label by solving $$\begin{aligned} \hat{y} &= \arg \max_{y \in \Ycal_u} \Lcal_{\theta,\phi,\psi}(\hat{\xv}, \Amat_y) \nonumber \\ &= \arg \min_{y \in \Ycal_u} \mbox{KL}(q_\phi(\hat{\zv}|\hat{\xv}) || p_\psi(\hat{\zv}|\Amat_y)) \end{aligned}$$ where $\Ycal_u = \{S+1,\ldots,S+U\}$ denotes the set of unseen classes. Intuitively, the prediction rule assigns $\hat{\xv}$ to that unseen class whose class-specific latent space distribution $p_\psi(\hat{\zv}|\av)$ is most similar to the VAE posterior distribution $q_\phi(\hat{\zv}|\hat{\xv})$ of the latent embeddings. Unlike the prediction rule of most ZSL algorithms that are based on simple Euclidean distance calculations of a point embedding to a set of “class prototypes” [@socher2013zero], our prediction rule naturally takes into account the possible *multi-modal* nature of the class distributions and therefore is expected to result in better prediction, especially when there is a considerable amount of intra-class variability in the data. Transductive ZSL {#sec:trans} ---------------- We now present an extension of the model for the *transductive* ZSL setting [@kodirov2015unsupervised], which assumes that the test inputs $\{\hat{\xv}_i\}_{i=1}^{N^\prime}$ from the unseen classes are also available while training the model. Note that, for the inductive ZSL setting (using the objective in (\[eq:loss\]), the $\mbox{KL}$ term between an unseen class test input $\hat{\xv}_i$ and its class based prior is given by $-\mbox{KL}(q_\phiv(\zv|\hat{\xv}_i)||p_\psiv(\zv|\av)))$. If we had access to the true labels of these inputs, we could add those directly to the original optimization problem ((\[eq:loss\])). However, since we do not know these labels, we propose an unsupervised method that can still use these unlabeled inputs to *refine* the inductive model presented in the previous section. A naïve approach for directly leveraging the unlabeled inputs in (\[eq:loss\]) without their labels would be to add the following reconstruction error term to the objective $$\begin{aligned} \label{eq:transRecons} \tilde{\Lcal}_{\theta,\phi,\psi}(\hat{\xv}, \av) &= \E_{q_\phi(\zv|\xv)}[\log p_\theta(\hat{\xv}|\zv)] \end{aligned}$$ However, since this objective completely ignores the label information of $\hat{\xv}$, it is not expected to work well in practice and only leads to marginal improvements over the purely inductive case (as corroborated in our experiments). To better leverage the unseen class test inputs in the transductive setting, we augment the inductive ZSL objective (\[eq:loss\]) with an additional unlabeled data based regularizer that uses only the unseen class test inputs. This regularizer is motivated by the fact that the inductive model is able to make reasonably confident predictions (as measured by the predicted class distributions for these inputs) for unseen class test inputs, and these confident predicted class distributions can be emphasized in this regularizer to guide those ambiguous test inputs. To elaborate the regularizer, we first define the inductive model’s predicted *probability* of assigning an unseen class test input $\hat{\xv}_i$ to class $c \in \{S+1,\ldots,S+U\}$ to be $$\begin{aligned} \label{eq:probxc} q(\hat{\xv}_i,c)=\frac{\exp(-\mbox{KL}(q_\phiv(\zv|\hat{\xv}_i)||p_\psiv(\zv|\Amat_c)))}{\sum_{c} \exp(-\mbox{KL}(q_\phiv(\zv|\hat{\xv}_i)||p_\psiv(\zv|\Amat_c)))}\end{aligned}$$ Our proposed regularizer (defined below in (\[eq:transKL\])) promotes these class probability estimates $q(\hat{\xv}_i,c)$ to be sharper, i.e., the most likely class should dominate the predicted class distribution $q(\hat{\xv}_i,c)$) for the unseen class test input $\hat{\xv}_i$. Specifically, we define a sharper version of the predicted class probabilities $q(\hat{\xv}_i,c)$ as $p(\hat{\xv}_i,c)=\frac{q(\hat{\xv}_i,c)^2 / g(c)}{ \sum_{c'}q(\hat{\xv}_i,c')^2 / g(c')}$, where $g(c) = \sum_{i=1}^{N^\prime} q(\hat{\xv}_i,c)$ is the marginal probability of unseen class $c$. Note that normalizing the probabilities by $g(c)$ prevents large classes from distorting the latent space. We then introduce our $\mbox{KL}$ based regularizer that encourages $q(\hat{\xv}_i,c)$ to be close to $p(\hat{\xv}_i,c)$. This can be formalized by defining the sum of the KL divergences between $q(\hat{\xv}_i,c)$ and $p(\hat{\xv}_i,c)$ for all the unseen class test inputs, i.e, \[eq:transKL\] (P() || Q()) \_[i=1]{}\^[N’]{}\_[c=S+1]{}\^[S+U]{} p(\_i,c) A similar approach of *sharpening* was recently utilized in the context of learning deep embeddings for clustering problems [@xie2016unsupervised] and data summarization [@wang2016deep], and is reminiscent of self-training algorithms used in semi-supervised learning [@nigam2000analyzing]. Intuitively, unseen class test inputs with *sharp* probability estimates will have a more significant impact on the gradient norm of (\[eq:transKL\]), which in turn leads to improved predictions on the ambiguous test examples (our experimental results corroborate this). Combining (\[eq:transRecons\]) and (\[eq:transKL\]), we have the following objective (which we seek to *maximize*) defined exclusively over the unseen class unlabeled inputs \[eq:trans\] U() = \_[i=1]{}\^[N’]{}\_[q\_(|\_i)]{}\[p\_(\_i|)\] -(P() || Q()) We finally combine this objective with the original objective ((\[eq:loss\])) for the inductive setting, which leads to the overall objective $\sum_{n=1}^{N} \hat{\Lcal}_{\theta,\phi,\psi}(\xv_n,\av_n) + U(\hat{\Xmat})$, defined over the seen class labeled training inputs $\{(\xv_n, \av_n)\}_{n=1}^{N}$ and the unseen class unlabeled test inputs $\{\hat{\xv}_i\}_{i=1}^{N^\prime}$. Under our proposed framework, it is also straightforward to perform few-shot learning [@lake2015human; @vinyals2016matching; @ravi2017optimization] which refers to the setting when a small number of labeled inputs may also be available for classes $c=S+1,\ldots,S+U$. For these inputs, we can directly optimize (\[eq:loss\]) on classes $c=S+1,\ldots,S+U$. Related Work {#sec:relwork} ============ Several prior methods for zero-shot learning (ZSL) are based on embedding the inputs into a semantic vector space, where nearest-neighbor methods can be applied to find the most likely class, which is represented as a point in the same semantic space [@socher2013zero; @norouzi2013zero]. Such approaches can largely be categorized into three types: ($i$) methods that learn the projection from the input space to the semantic space using either a linear regression or a ranking model [@akata2015evaluation; @lampert2014attribute], or using a deep neural network[@socher2013zero]; ($ii$) methods that perform a “reverse” projection from the semantic space to the input space[@zhang2016learning], which helps to reduce the *hubness problem* encountered when doing nearest neighbor search at test time [@radovanovic2010hubs]; and ($iii$) methods that learn a shared embedding space for the inputs and the class attributes [@zhang2016zero; @changpinyo2016synthesized]. Another popular approach to ZSL is based on modeling each unseen class as a linear/convex combination of seen classes [@norouzi2013zero], or of a set of shared “abstract” or “basis” classes [@romera2015embarrassingly; @changpinyo2016synthesized]. Our framework can be seen as a flexible generalization to the latter type of models since the parameters $\Wmat_\mu$ and $\Wmat_\sigma$ defining the latent space distributions are shared by the seen and unseen classes. One general issue in ZSL is the *domain shift* problem – when the seen and unseen classes come from very different domains. Standard ZSL models perform poorly under these situations. However, utilizing some additional unlabeled data from those unseen domains can somewhat alleviates the problem. To this end, [@kodirov2015unsupervised] presented a transductive ZSL model which uses a dictionary-learning-based approach for learning unseen-class classifiers. In their approach, the dictionary is adapted to the unseen-class domain using the unlabeled test inputs from unseen classes. Other methods that can leverage unlabeled data include [@fu2015transductive; @rohrbach2013transfer; @li2015semi; @zhao2016zero]. Our model is robust to the *domain shift* problem due to its ability to incorporate unlabeled data from unseen classes. Somewhat similar to our VAE based approach, recently [@kodirov2017semantic] proposed a semantic autoencoder for ZSL. However, their method does not have a proper generative model. Moreover, it assumes each class to be represented as a fixed point and cannot extend to the transductive setting. Deep encoder-decoder based models have recently gained much attention for a variety of problems, ranging from image generation [@rezende2016one] and text matching [@shen2017deconvolutional]. A few recent works exploited the idea of applying sematic regularization to the latent embedding spaced shared between encoder and decoder to make it suitable for ZSL tasks [@kodirov2017semantic; @tsai2017learning]. However, these methods lack a proper generative model; moreover ($i$) these methods assume each class to be represented as a fixed point, and ($ii$) these methods cannot extend to the transductive setting. Variational autoencoder (VAE) [@kingma2014auto] offers an elegant probabilistic framework to generate continues samples from a latent gaussian distribution and its supervised extensions [@kingma2014semi] can be used in supervised and semi-supervised tasks. However, supervised/semi-supervised VAE [@kingma2014semi] assumes all classes to be seen at the training time and the label space $p(y)$ to be discrete, which makes it unsuitable for the ZSL setting. In contrast to these methods, our approach is based on a deep generative framework using a supervised variant of VAE, treating each class as a distribution in a latent space. This naturally allows us to handle the intra-class variability. Moreover, the supervised VAE model helps learning highly discriminative representations of the inputs. Some other recent works have explored the idea of generative models for zero-shot learning [@li2017zero; @verma2017simple]. However, these are primarily based on linear generative models, unlike our model which can learn discriminative and highly nonlinear embeddings of the inputs. In our experiments, we have found this to lead to significant improvements over linear models [@li2017zero; @verma2017simple]. Deep generative models have also been proposed recently for tasks involving learning from limited supervision, such as one-shot learning [@rezende2016one]. These models are primarily based on feedback and attention mechanisms. However, while the goal of our work is to develop methods to help recognize previously unseen classes, the focus of methods such as  [@rezende2016one] is on tasks such as generation, or learning from a very small number of labeled examples. It will be interesting to combine the expressiveness of such models within the context of ZSL. Experiments =========== We evaluate our framework for ZSL on several benchmark datasets and compare it with a number of state-of-the-art baselines. Specifically, we conduct our experiments on the following datasets: ($i$) Animal with Attributes (AwA) [@lampert2014attribute]; ($ii$) Caltech-UCSD Birds-200-2011 (CUB-200) [@wah2011caltech]; and ($iii$) SUN attribute (SUN) [@patterson2012sun]. For the large-scale dataset (ImageNet), we follow [@fu2016semi], for which 1000 classes from ILSVRC2012 [@russakovsky2015imagenet] are used as seen classes, while 360 non-overlapped classes of ILSVRC2010 [@deng2009imagenet] are used as unseen classes. The statistics of these datasets are listed in Table \[Table:datasets\]. In all our experiments, we consider VGG-19 fc7 features [@simonyan2014very] as our raw input representation, which is a 4096-dimensional feature vector. For the semantic space, we adopt the default class attribute features provided for each of these datasets. The only exception is ImageNet, for which the semantic word vector representation is obtained from word2vec embeddings [@mikolov2013distributed] trained on a skip-gram text model on 4.6 million Wikipedia documents. For the reported experiments, we use the standard train/test split for each dataset, as done in the prior work. For hyper-parameter selection, we divide the training set into training and validation set; the validation set is used for hyper-parameter tuning, while setting $\lambda=1$ across all our experiments. For the VAE model, a multi-layer perceptron (MLP) is used for both encoder $q_\phiv(\zv | \xv)$ and decoder $p_\thetav(\xv | \zv)$. The encoder and decoder are defined by an MLP with two hidden layers, with $1000$ nodes in each layer. ReLU is used as the nonlinear activation function on each hidden layer and dropout with constant rate $0.8$ is used to avoid overfitting. The latent space $\zv$ was set to be $100$ for small datasets and $500$ for ImageNet. Our results with variance are reported by repeating with 10 runs. Our model is written in Tensorflow and trained on NVIDIA GTX TITAN X with 3072 cores and 11GB global memory. We compare our method (referred to as VZSL) with a variety of state-of-the-art baselines using VGG-19 fc7 features and specifically we conduct our experiments on the following tasks: - **Inductive ZSL:** This is the standard ZSL setting where the unseen class latent space distributions are learned using only seen class data. - **Transductive ZSL:** In this setting, we also use the unlabeled test data while learning the unseen class latent space distributions. Note that, while this setting has access to more information about the unseen class, it is only through unlabeled data. - **Few-Shot Learning:** In this setting [@lake2015human; @vinyals2016matching; @ravi2017optimization], we also use a small number of labeled examples from each unseen class. In addition, through a visualization experiment (using t-SNE [@maaten2008visualizing]), we also illustrate our model’s behavior in terms its ability to separate the different classes in the latent space. Inductive ZSL {#inductive-zsl} ------------- Table \[table:inductive\_zsl\] shows our results for the inductive ZSL setting. The results of the various baselines are taken from the corresponding papers or reproduced using the publicly available implementations. From Table \[table:inductive\_zsl\], we can see that: ($i$) our model performs better than all the baselines, by a reasonable margin on the small-scale datasets; ($ii$) On large-scale datasets, the margin of improvement is even more significant and we outperform the best-performing state-of-the art baseline by a margin of $37.4\%$; ($iii$) Our model is superior when including the reconstruction term, which shows the effectiveness of the generative model; ($iv$) Even without the reconstruction term, our model is comparable with most of the other baselines. The effectiveness of our model can be attributed to the following aspects. First, as compared to the methods that embed the test inputs in the semantic space and then find the most similar class by doing a Euclidean distance based nearest neighbor search, or methods that are based on constructing unseen class classified using a weighted combination of seen class classifiers [@zhang2015zero], our model finds the “most probable class” by computing the distance of each test input from *class distributions*. This naturally takes into account the shape (possibly multi-modal) and spread of the class distribution. Second, the reconstruction term in the VAE formulation further strengthens the model. It helps leverage the intrinsic structure of the inputs while projecting them to the latent space. This aspect has been shown to also help other methods such as [@kodirov2017semantic] (which we use as one of the baseline), but the approach in [@kodirov2017semantic] lacks a generative model. This explains the favorable performance of our model as compared to such methods. Transductive ZSL {#transductive-zsl} ---------------- Our next set of experiments consider the transductive setting. Table \[table:transductive\_zsl\] reports our results for the transductive setting, where we compare with various state-of-the-art baselines that are designed to work in the transductive setting. As Table \[table:transductive\_zsl\] shows, our model again outperforms the other state-of-the-art methods by a significant margin. We observe that the generative framework is able to effectively leverage unlabeled data and significantly improve upon the results of inductive setting. On average, we obtain about $8\%$ better accuracies as compared to the inductive setting. Also note that in some cases, such as CUB-200, the classification accuracies drop significantly once we remove the VAE reconstruction term. A possible explanation to this behavior is that the CUB-200 is a relative difficult dataset with many classes are very similar to each other, and the inductive setting may not achieve very confident predictions on the unseen class examples during the inductive pre-training process. However, adding the reconstruction term back into the model significantly improves the accuracies. Further, compare our entire model with the one having only (\[eq:transRecons\]) for the unlabeled, there is a margin for about $5\%$ on AwA and CUB-200, which indicates the necessity of introduced $\mbox{KL}$ term on unlabeled data. Few-Shot Learning (FSL) ----------------------- In this section, we report results on the task of FSL [@salakhutdinov2013learning; @mensink2014costa] and transductive FSL [@frome2013devise] [@socher2013zero]. In contrast to standard ZSL, FSL allows leveraging a few labeled inputs from the unseen classes, while the transductive FSL additionally also allows leveraging unseen class unlabeled test inputs. To see the effect of knowledge transfer from the seen classes, we use a multiclass SVM as a baseline that is provided the same number of labeled examples from each unseen class. In this setting, we vary the number of labeled examples from 2 to 20 (for SUN, we only use 2, 5 and 10 due to the small number of labeled examples). In Figure \[fig:fsl\], we also compared with standard inductive ZSL which does not have access to the labeled examples from the unseen classes. Our results are shown in Figure \[fig:fsl\]. ![[]{data-label="fig:fsl"}](figures/FSL.pdf) As can be seen, even with as few as 2 or 5 additional labeled examples per class, the FSL significantly improves over ZSL. We also observe that the FSL outperform a multiclass SVM which demonstrates the advantage of the knowledge transfer from the seen class data. Table \[table:transductive\_fsl\] reports our results for the transductive FSL setting where we compare with other state-of-the-art baselines. In this setting too, our approach outperforms the baselines. {width="\textwidth"} t-SNE Visualization ------------------- To show the model’s ability to learn highly discriminative representations in the latent embedding space, we perform a visualization experiment. Figure \[fig:tSNE\_exp4\] shows the t-SNE [@maaten2008visualizing] visualization for the raw inputs, the learn latent embeddings, and the *reconstructed* inputs on AwA dataset, for both inductive ZSL and transductive ZSL setting. As can be seen, both the reconstructions and the latent embeddings lead to reasonably separated classes, which indicates that our generative model is able to learn a highly discriminative latent representations. We also observe that the inherent correlation between classes might change after we learn the latent embeddings of the inputs. For example, “giant+panda” is close to “persian+cat” in the original CNN features space but far away from each other in our learned latent space under transductive setting. A possible explanation could be that the sematic features and image features express information from different views and our model learns a representation that is sort of a compromise of these two representations. Conclusion ========== We have presented a deep generative framework for learning to predict unseen classes, focusing on inductive and transductive zero-shot learning (ZSL). In contrast to most of the existing methods for ZSL, our framework models each seen/unseen class using a class-specific latent-space distribution and also models each input using a VAE-based decoder model. Prediction for the label of a test input from any unseen class is done by matching the VAE posterior distribution for the latent representation of this input with the latent-space distributions of each of the unseen class. This distribution matching method in the latent space provides more robustness as compared to other existing ZSL methods that simply use a point-based Euclidean distance metric. Our VAE based framework leverages the intrinsic structure of the input space through the generative model. Moreover, we naturally extend our model to the transductive setting by introducing an additional regularizer for the unlabeled inputs from unseen classes. We demonstrate through extensive experiments that our generative framework yields superior classification accuracies as compared to existing ZSL methods, on both inductive ZSL as well as transductive ZSL tasks. The proposed framework can scale up to large datasets and can be trained using any existing stochastic gradient based method. Finally, although we use isotropic Gaussian to model each model each seen/unseen class, it is possible to model with more general Gaussian or any other distribution depending on the data type. We leave this possibility as a direction for future work.\ **Acknowledgements**: This research was supported in part by grants from DARPA, DOE, NSF and ONR. [^1]: Corresponding authors

--- abstract: 'Understanding the features of and mechanisms behind language learning can provide insights into the general process of knowledge acquisition. Recent methods from network science applied to language learning have advanced the field, particularly by noting associations between densely connected words and acquisition. However, the importance of sparse areas of the network, or knowledge gaps, remains unexplored. Here we create a semantic feature network in which words correspond to nodes and in which connections correspond to semantic similarity. We develop a new analytical approach built on principles of applied topology to query the prevalence of knowledge gaps, which we propose manifest as cavities within the network. We detect topological cavities of multiple dimensions in the growing semantic feature network of children ages 16 to 30 months. The pattern of cavity appearance matches that of a constrained null model, created by predefining the affinity of each node for connections. Furthermore, when word acquisition time is computed from children of mothers with differing levels of education, we find that despite variation at the word level, the global organization as measured by persistent homology remains comparable. We show that topological properties of a node correlate with filling in cavities better than simple lexical properties such as the length and frequency of the corresponding word. Finally, we show that the large-scale architecture of the semantic feature network is topologically accommodating to many node orders. We discuss the importance of topology in language learning, and we speculate that the formation and filling of knowledge gaps may be a robust feature of knowledge acquisition.' author: - 'Ann E. Sizemore' - 'Elisabeth A. Karuza' - Chad Giusti - 'Danielle S. Bassett' bibliography: - 'bibfile.bib' title: Knowledge gaps in the early growth of semantic networks --- Introduction {#introduction .unnumbered} ============ Formal analysis of the mechanisms driving knowledge acquisition remains a foundational area of research in cognitive science. In the domain of word learning, behavioral evidence suggests that this process is mediated in part by various properties of words at an individual level, such as the frequency of a given word or the extent to which it evokes a mental image [@duff2012role; @ambridge2015ubiquity]. Recently, tools from network science have offered a means of examining how lexical acquisition might also be mediated by higher-order relationships between many words, or the network topology underlying input that is available to the learner [@karuza2016local]. Under this approach, words are typically represented by the nodes of the network, while shared semantic or sound-based associations can be used to construct edges between them [@hills2009longitudinal; @goldstein2014influence; @steyvers2005large]. Broadly, evidence indicates that learners are particularly sensitive to how densely connected a given word is relative to others words in a network. Because these previous studies focus on the areas of the network that have been learned by children, they have left open the question, “How do those words not yet known affect learning?” More precisely, as children produce new words in the semantic network, does their attained semantic network contain any knowledge gaps, or voids where a word is missing? Since edges correspond to meaning, a gap in the network suggests a unifying concept that is not yet understood. In the network science formalism, such knowledge gaps in a growing network correspond to topological cavities that are born, and then later filled in with the addition of new nodes and edges. We propose that characterizing the evolution of these cavities in a semantic network offers unique insight into lexical organization in children, and we investigate whether knowledge gaps might serve as a useful proxy for the difficulty associated with acquiring feature-based concepts. To answer these questions, we employ concepts and tools from applied topology that allow us to detect topological cavities within a growing semantic network. The specific network that we study is a semantic network in which words are given a weight (and ordering) derived from the month that the word was produced by toddlers aged 16-30 months. Though such node-weighted networks are commonly observed in biology, they are challenging to analyze because common tools from network science, such as traditional graph metrics, can account for weighted edges, but not for weighted nodes. To address this challenge, we develop a formalism and construction that transforms any node-weighted system into a sequence of binary graphs analyzable by both topological data analysis and standard tools of network science including graph metrics. Specifically, we transform the growing semantic feature network into a sequence of binary graphs called a *filtration*, with one new node (corresponding to one new word) added at each step. We call this sequence of graphs built from a node-ordered network a *node-filtered order complex* (hereafter referred to as the n-order complex for brevity), inspired by the order complex defined for edge-weighted networks [@giusti2015clique]. We can then compute *persistent homology* [@carlsson2009topology; @zomorodian2005computing], which tracks the formation and possible filling in of topological cavities of different dimensions throughout a filtration. By encoding the growing semantic feature network of children ages 16-30 months as a n-order complex, we use persistent homology to ask if topological cavities – corresponding to knowledge gaps – form and then fill in throughout the learning process. In contrast to expectations derived from a growing network null model, we find a collection of long-persisting cavities of varying dimensions. Interestingly, the pattern of cavity formation suggests that the semantic network is organized under heavy constraints. We adjudicate between conflicting hypotheses that topological cavities might either be robust to or vary systematically with the nature of input available to the learner, in this case indexed by the mother’s level of education. We observe at most minor differences in topological cavity existence despite random variation in node order; on average, any ordering of words produces a similar topological signature. Our results suggest that these topological cavities might be a conserved feature of the learning process, and that semantic network growth is a robust phenomenon that can accommodate many local changes without abrupt restructuring of its large-scale organization. Materials and Methods {#materials-and-methods .unnumbered} ===================== Growing semantic network construction {#growing-semantic-network-construction .unnumbered} ------------------------------------- We constructed a 120-node semantic feature network with node ordering following the procedure outlined in [@hills2009longitudinal]. Specifically, we used the *Wordbankr* package [@frank2016wordbank] which contains data from the MacArthur-Bates Communicative Development Inventory (MB-CDI) [@dale1996lexical]. This database contains which of 541 English words 5511 toddlers ages 16-30 months could produce, as recorded via parental report. For each word we calculate the month at which $\geq 50\%$ of children could produce the word [@hills2009longitudinal]. Within one month, words are ordered according to the percentage of children producing each word, resulting in a total ordering of words. To form a node-ordered network we represent words as nodes and connect two nodes (words) if they share a semantic feature within the McRae feature list [@mcrae2005semantic]. These semantic features were derived from adult norming data from 725 adults and are organized into categories based on feature type. When an individual generates features for a given concept, these features are interpretations of the abstract concept that are constructed for the sole purpose of description [@barsalou2003abstraction]. Then feature norms offer a unique understanding of representation, which varies across and within individuals, and which often results in a collection of distinguishing features (as opposed to general features of many words) [@mcrae2005semantic], making them useful for modeling and theory testing [@hampton1979polymorphous; @wu2009perceptual; @devlin1998category; @moss2002emergence]. We use all feature categories excluding encyclopedic and taxonomic, which are unlikely to be accessible to toddlers [@hills2009longitudinal]. All words included in both the McRae and Wordbankr databases were used in our final semantic network. Persistent homology {#persistent-homology .unnumbered} ------------------- Below we include a brief description of persistent homology. We refer the interested reader to [@carlsson2009topology; @zomorodian2005computing; @ghrist2008barcodes] and the Appendix for more details. Given a graph $G = (V,E)$, we begin by assigning a $k$-simplex to each set of $(k+1)$ all-to-all connected nodes, called a $(k+1)$-clique. Recall a $k$-simplex is the convex hull of $k+1$ affinely positioned points, and the clique complex $X(G)$ is the collection of all simplices defined by appropriately sized cliques of $G$. Within a clique complex, a $k$-cycle is a closed path of $k$-simplices. Note that a $k$-cycle either encloses a collection of higher dimensional simplices (called a $k$-boundary) or a topological cavity. We call two $k$-cycles equivalent if their set difference is a $k$-boundary. This creates an equivalence relation, and the number of non-trivial equivalence classes indicates the number of topological cavities of dimension $k$. As is common in the literature, we refer to an equivalence class of $k$-cycles as a $k$-cycle. We construct a graph filtration, or sequence of graphs each included in the next, from the growing binary network by letting $G_n$ be the graph after the addition of node $n$, with $G_n = G_{n-1} \cup \{n\}$ and all of its connections between node $n$ and nodes $1,\dots, n-1$. This process produces a graph filtration indexed by nodes, and consequentially a node-indexed filtration of clique complexes. We call this filtration of clique complexes induced by the binary graph and node-ordering the *node-filtered order complex*. This filtration of clique complexes allows us to map the $k$-cycles in $X(G_n)$ to $k$-cycles in $X(G_{n+1})$. We call these cycles – which are tracked throughout the filtration – *persistent cycles*. We compute the persistent homology [@carlsson2009topology; @zomorodian2005computing] in dimensions 1–3 using the Eirene software [@henselmanghrist16]. Models of n-order complexes {#models-of-n-order-complexes .unnumbered} --------------------------- For each model n-order complex, we generate 1000 instances and provide MATLAB code and detailed descriptions at the Filtered Network Model Reference (filterednetworkmodelref.weebly.com). Measures for correlation calculations {#measures-for-correlation-calculations .unnumbered} ------------------------------------- For correlations with the number of persistent cycles killed at each node, we compute the following graph statistics using [@rubinov2010complex] on the binary semantic feature network: node degree, clustering coefficient, and betweenness centrality. The degree of a node is the number of edges incident to the node. The clustering coefficient measures the connectivity of a node’s neighbors, calculated by the ratio of existing triangles to the number of triangles possible. Precisely, $$C_n = \frac{2t_n}{k_n(k_n -1)}$$ where $t_n$ is the number of triangles formed by node $n$ and its neighbors [@watts1998collective]. Additionally, we inquire whether the centrality or number of shortest paths passing through a node might correlate with the number of cycles killed. In particular, we define the betweenness centrality [@kintali2008betweenness] of a node as $$BC_n = \sum_{s,t,n, s\neq t\neq n} \frac{\lambda_{n}(s,t)}{\lambda(s,t)}$$ with $\lambda(s,t)$ being the number of shortest paths between nodes $s$, $t$, and $\lambda_n(s,t)$ being the number of such paths passing through node $n$. Results {#results .unnumbered} ======= To begin, we construct an ordering on 120 nouns derived from the first month at which $\geq$50% of children between the ages of 16 and 30 months can produce each word (Fig. \[fig:1\]a, left) [@frank2016wordbank]. Multiple words could be first produced within one month, so we create a total ordering by sorting words learned within one month by ascending percentage of children producing each word. We next form a binary semantic feature network with 120 nouns as nodes, and with edges connecting words that share a property or function (Fig. \[fig:1\]a, right) [@mcrae2005semantic]. Together, the word ordering and binary network pair assemble into the growing semantic network [@hills2009longitudinal], with the node added at step $n$ connecting to all of its neighbors added at steps $1,\dots, n-1$ (Fig. \[fig:1\]a, middle). We are interested in the presence of knowledge gaps, which we hypothesize manifest as voids within the growing semantic network that exist only for a finite number of months. For example, in Fig. \[fig:1\]b (top), the words ‘balloon’, ‘bear’, ‘cheese’, and ‘banana’ connect in a pattern that leaves a hole within the network. If the word ‘bus’ is learned later, this word connects to each of ‘balloon’, ‘bear’, ‘cheese’, and ‘banana’, so that there is no longer a void within the network. When a void in the network is extinguished, we say the knowledge gap is *filled in* (Fig. \[fig:1\]b, bottom). The features of interest in such a network are then the nodes responsible for filling in the cavity, which correspond to the temporarily missing words. Detecting cavities in node-ordered networks {#detecting-cavities-in-node-ordered-networks .unnumbered} ------------------------------------------- To identify topological cavities within the growing semantic network, we use a method from applied topology called persistent homology, which returns the (1) number, (2) dimension, and most importantly (3) longevity of topological cavities within a growing network, all of which we will more rigorously define in this section. Before we discuss growing graphs, we outline the process of detecting cavities in a single binary network. Given a binary graph $G$, we first translate our graph into a combinatorial object on which we can perform the later computations. Instead of a simple graph described by nodes and edges, we allow all groups of completely connected nodes to define entities. Formally, we create the *clique complex* $X(G)$, a collection of all the *cliques*, or all-to-all connected subgraphs, in the network. In Fig. \[fig:2\]a, we depict this process as ‘coloring in’ the graph $G$ to build the clique complex $X(G)$. For example, we color in 1-cliques (nodes), 2-cliques (edges), 3-cliques (triangles), and so on, giving us higher dimensional information about the structure[^1]. Now with our graph encoded as a clique complex, we can use *homology* to detect cavity-surrounding motifs of edges, triangles, tetrahedra, and higher dimensional analogs (Fig. \[fig:2\]b). Loops of edges form 1-cycles, loops of triangles form 2-cycles, and loops of tetrahedra form 3-cycles. For example, Fig. \[fig:2\]b shows cavity-surrounding cycles of each dimension on the top row, while those on the bottom are tessellated by higher-dimensional cliques formed with the purple node. Homology distinguishes between cavity-surrounding loops and those tessellated by higher cliques, thereby returning detailed information about the mesoscale architecture of the complex. In particular, homology detects *equivalence classes* of $k$-cycles, with two $k$-cycles being in the same equivalence class if their symmetric difference is a collection of higher dimensional cliques (see Appendix for details). By abuse it is common to refer to an equivalence class of $k$-cycles as a $k$-cycle, and we will adopt this abbreviated description throughout the remainder of the paper. To summarize: homology counts the number of cavities in each dimension of a clique complex constructed from a binary graph. While this approach is hypothetically useful, our data describes a *growing* network instead of a single binary graph, so we cannot simply compute its homology as above described. However, notice that we get a binary graph after the addition of each new node, and that the binary graph $G_n$ created after the addition of node $n$ is a subgraph of $G_{n+1}$ for all $n$. This sequence of objects (here, graphs) with $G_n \subseteq G_{n+1}$ is called a *filtration* (Fig. \[fig:sfig\_filts\], top and Fig. \[fig:sfig\_ph2\]b, top). If we construct a filtration of binary graphs $G_n$, we immediately gain a filtration of clique complexes $X(G_n)$ with $X(G_n)\subseteq X(G_{n+1})$ necessarily true since $G_n \subseteq G_{n+1}$ (Fig. \[fig:sfig\_ph2\]b, middle). For example, using the ordering and clique complex in Fig. \[fig:2\]c, Fig. \[fig:2\]d illustrates the described filtration of clique complexes for steps 9-13 (addition of nodes 9-13), with new nodes (outlined in white) connecting to any neighbor already in the complex. We call the filtration of clique complexes created from a growing network the *node-filtered order complex* (see Appendix for further details), inspired by the order complex in [@giusti2015clique]. The order complex creates a filtration of clique complexes from an edge-weighted network using the edge weights to induce an edge ordering. Here, the node filtered order complex (which we shorten to n-order complex for brevity) can be completely defined by the pair $(G,s)$ with $G$ a binary graph and $s$ the ordering of vertices, possibly induced by a weighting on the nodes. Finally, at each node addition we can map the clique complex $X(G_n)$ into the next $X(G_{n+1})$, so we can follow cycles, and consequentially cavities, throughout the filtration. For example, the addition of node 10 creates a cavity surrounded by a 1-cycle, which persists in the complexes $X(G_{10}), X(G_{11}), X(G_{12})$ until it is tessellated with the addition of node 13. The barcode plot in the top of Fig. \[fig:2\]e records this persistent cavity as a horizontal line running throughout the duration of this persistent cavity, or its *lifetime*. We call the node at which the cavity begins the *birth* node, and we call the node at which the cavity is tessellated the *death* node. Thus, the lifetime is formally $death-birth$. The number of cavities of dimension $k$ at each step (node addition) in the filtration is recorded in the Betti curves $\beta_k(n)$ and shown in the bottom of Fig. \[fig:2\]e for the n-order complex of Fig. \[fig:2\]d. Tracking these persistent cycles throughout a filtration is called *persistent homology* [@carlsson2009topology; @zomorodian2005computing], which – in summary – allows us to extract the number and dimension along with the longevity of topological cavities throughout the growth process. Topology of generative growing network models {#topology-of-generative-growing-network-models .unnumbered} --------------------------------------------- Before approaching knowledge gaps in the semantic network, we first pause to ask if and how persistent homology can distinguish randomness from structure within artificial models of growing graphs (equivalently n-order complexes), which will additionally help us to gain an intuition for processes creating particular cavity existence patterns. Specifically, we test four growing graph models with varying degrees of predefined structure. Each model constructs a binary graph by assigning a probability to the existence of each edge as a function of one or both parent nodes. We call these *generative* models because they construct *de novo* both the binary graph and the node order. For the following models we let the node ordering $s$ simply be $1:N$ with $N$ the total number of nodes. The most basic (and random) model that we tested assigns each edge entering the graph with the addition of node $n$ a probability $p(n) = c \in [0,1]$ of existing. We call this model the *constant probability* model, and we show its persistent homology in Fig. \[fig:3\]a with $p(n) = 0.3$. The next model, the *proportional probability* model, attaches edges from node $n$ to all nodes $1, \dots, n-1$ with probability $p(n) = n/N$ (Fig. \[fig:3\]b). Next, we generate a modular network composed of four equal-sized communities, and we refer to it as the *modular growth* model (Fig. \[fig:3\]c). We randomly assign nodes to communities; edges added with node $n$ exist with probability $p_{in}$ between $n$ and nodes within its community, and with probability $p_{out}$ between $n$ and members of different communities. In our final generative model, each node $n$ is assigned an affinity $a_n$ for edges such that when node $n$ is added, edges between node $n$ and $m = 1, \dots, n-1$ exist with probability $p(n,m) = \frac{a_m}{\max(\vec{a})}$ with $\vec{a} = (a_1, a_2, \dots, a_N)$ being the vector of affinities. We call this model the *edge affinity* model, and we show results for this model with affinities given by a random permutation of $(1:120)^2$ in Fig. \[fig:3\]d. For each of these models, we choose parameters so as to produce graphs whose edge density closely matches the edge density of the empirically measured semantic network, $\sim 0.3$. The constant probability model generates growing graphs with the least amount of imposed structure, producing hundreds of persistent 2-cycles that never die. The $\beta_2$ curve on average dominates the Betti curves and we observe few if any persistent 1-cycles or 3-cycles. The proportional probability model instead shows an increasing trend of Betti curve peaks with increasing dimension. Additionally, all persistent cycles of dimensions 1-3 die by around node 100, as later nodes are likely to tessellate cavities. Interestingly, the modular growth model produces Betti curves dominated by dimension 2, similar to the constant probability model. Though this model produces networks with high modularity throughout the node addition process, we see qualitatively similar properties between this and the constant probability model. Still, the density of within-community connections restricts the maximum height of the $\beta_2$ peak and drives the creation of persistent 3-cycles. In contrast to these null models, we expect the growing semantic structure to be organized according to external constraints: there exist (external) properties of nodes that do not fluctuate based on the current state of the network. If a node (word) has some aptitude for connections (similar to many other words), such an aptitude should not change as the network grows. Such a constancy is unlike that observed in, for example, a preferential attachment process but is explicitly accounted for in our edge affinity model. Interestingly, we observe far fewer persistent cycles of each dimension in the edge affinity model than in the previous models. Furthermore, we observe a pattern of increasing peaks of persistent cycles as we move to higher dimensions. These results demonstrate that a growing process constrained by the external constraint of constant edge affinity will yield fewer topological cavities than that of a more random growth process. Gaps in the growing semantic feature network {#gaps-in-the-growing-semantic-feature-network .unnumbered} -------------------------------------------- Now that we have developed some intuition for the structure detected by persistent homology, we ask if topological cavities exist within the growing semantic network and if so, what information such cavities might provide about the learning process. We observe multiple persistent cavities of dimensions 1-3, most of which die before the 30 month mark (Fig. \[fig:4\]a). The Betti curves show increasing peaks with larger dimensions as more nodes are added. Next, we ask which nodes (words) enter the growing graph when persistent cycles are born or killed. We list these words next to the corresponding bar, and we show persistent cycle birth and death nodes visually in the inset of Fig. \[fig:4\]a as a persistent cycle network, with words ordered alphabetically and an edge for each persistent cavity emanating from the birth node and terminating at the death node. The edge thickness is proportional to the corresponding cavity lifetime (index of death minus index of birth) and the edge color indicates cavity dimension. We observe nodes generally begin or kill one or no persistent cycles, with a few exceptions such as ‘bench’, ‘peas’, and ‘couch’. Persistent cavities that never die (have a death time of $\inf$) are not shown. The fact that very few cavities exist in the final semantic network suggests that knowledge gaps not only form, but must also evolve and are extinguished during the learning process. Next, we ask if there are simple rules by which cavities form and evolve in the growing semantic network. We notice that nodes added late in the growth process have higher chances of having a high degree at the time of their addition than nodes added early in the growth process. Thus, one might hypothesize that the empirically observed pattern of Betti curves follows simply from a pattern of higher-connectivity nodes added throughout the filtration. Contrary to this simplistic expectation, we observe instead that the degree of nodes varies considerably across time with no salient trend of either a decreasing or increasing node degree (Fig. \[fig:4\]b). Indeed, when the final node is added there exists great variability in node degrees when plotted in the order of node addition. This complexity motivates a more thorough effort to model the growth process to infer underlying mechanisms of cavity formation and evolution, which we turn to next. The persistent homology of a growing network is classically used to infer global organizational properties, and we can therefore use this tool to understand the growing semantic network from a global perspective. We begin by comparing the persistent homology of the growing semantic network to the generative models of Fig. \[fig:3\]. We observe that the edge affinity model generates n-order complexes with the most similar persistent homology to that of the growing semantic network. Upon closer inspection, we observe that the largest difference between the Betti curves of the affinity model and growing semantic network stems from the likelihood that cavities die. All but five persistent cavities in the growing semantic network die by node 120, while in the edge affinity model the barcodes show the majority of persistent cycles never die. Moreover the comparisons with models shown in Fig. \[fig:3\] strongly imply that the evolving architecture of the growing semantic network is highly non-random, as the Betti curve peaks are not even of the same magnitude between the growing semantic and random n-order complex models. Taken together, these results suggest that the growing semantic network topology during learning is non-random and that node constraints such as a fixed affinity for connections might play a role in the evolving architecture. To further probe mechanisms guiding the evolution of the growing semantic structure, we construct *derived* n-order complex models that begin with the semantic network and alter either the node ordering or edge placement to determine which (if either) explains the observed evolving architecture. Beginning with the influence of node order, we compute the persistent homology of the *randomized nodes* n-order complex model, which retains the binary graph of semantic feature connections but randomly permutes the node order. We observe strikingly similar persistent homology between this model (Fig. \[fig:4\]c) and the growing semantic network, though note that any n-order complex built with the same binary graph $G$ will necessarily have the same homology at $G_N$, which limits the variability in persistent homology. Next we keep the binary network of semantic connections but now order the nodes by decreasing degree (Fig. \[fig:4\]d) or by their topological distance from the first node (Fig. \[fig:4\]e), which we call the *decreasing degree* and *distance from $v_0$* models, respectively (any ties are randomly permuted). Both of these more engineered models exhibit persistent homology that is less similar to the growing semantic network than the randomized nodes model, suggesting that neither learning the most connected words first nor learning those with the shortest semantic distance to the first word can account for the evolution of the growing semantic network. Finally, we keep the node ordering constant but now randomly rewire edges while preserving node degree; we call this the *randomized edges* model and note that it is similar in spirit to the configuration model [@bender1978asymptotic; @maslov2002specificity]. We observe a highly random persistent homology signature as described by the Betti curves and barcodes (Fig. \[fig:4\]f), suggesting that the pattern of semantic feature connections between words is more important in explaining knowledge gap formation than the order in which those words are produced by children. Influence of or robustness to maternal education level {#influence-of-or-robustness-to-maternal-education-level .unnumbered} ------------------------------------------------------ In the previous sections, we demonstrated that knowledge gaps not only exist but are created and filled in throughout early semantic learning. Knowledge gaps might correspond to learning relatively difficult concepts, and one could hypothesize that gaps would occur more frequently in the growing semantic networks of children with mothers having achieved a higher level of education, which – along with socioeconomic status [@hoff2005socioeconomic; @schwab2016language] – can significantly impact child-directed speech and learning [@dollaghan1999maternal]. Yet, a contrary hypothesis is that the laws of evolution predispose the growing semantic networks in children to be relatively robust to variations in the environment. To adjudicate between these two conflicting hypotheses, we create three distinct growing semantic networks from children with mothers whose highest education was some or all of secondary school, some or all of college, and some or all of graduate school. Then we have the same binary network for each of the three networks, but the ordering of nodes has now changed. We label these networks the *secondary*, *college*, and *graduate* growing networks, respectively. We compute the persistent homology and find no trend of increasing topological cavity number or lifetimes (Fig. \[fig:5\]a-c), despite the differences in word ordering. Since at a global level the persistent homology of the three growing networks varies little, we next ask if the same words correspond to nodes killing persistent cavities in each growing network. Figure \[fig:5\] (also see Fig. \[fig:sfig\_education1\]) shows persistent cycle networks for each of the three growing networks with nodes ordered and placed alphabetically as in Fig. \[fig:4\]a. This visualization allows for comparison of nodes that begin and kill persistent cavities across the three growing networks. For example, a persistent 1-cycle is seen beginning at ‘doll’ and ending at ‘pony’ in each of the *secondary*, *college*, and *graduate* growing networks (indicated by the red arrow). We observe that while a few node pairs begin and end persistent cycles in each of the *secondary*, *college*, and *graduate* growing networks, generally node pairs do not begin and end persistent cycles, or at least persistent cycles of the same dimension, across each of the education levels. Characterizing the manner in which knowledge gaps are extinguished {#characterizing-the-manner-in-which-knowledge-gaps-are-extinguished .unnumbered} ------------------------------------------------------------------ In the previous sections, we have shown that knowledge gaps are created and later filled in with similar rates despite differences in maternal education. These observations motivate our final effort to determine if particular properties of the nodes or their corresponding words increase the likelihood of a node tessellating cavities. For each of the *secondary*, *college*, *graduate*, and original all-inclusive growing semantic network barcodes, we count the number of persistent cavities killed by each node. Since these cavity-killing nodes correspond to temporarily missing words, one might hypothesize that these corresponding words are more difficult to learn. We use a simple proxy for word difficulty: word length. However we find no significant correlation between the number of cycles killed and word length (Fig. \[fig:6\]a, $p>0.1$ for all growing networks, Pearson correlation coefficient). Additionally, we ask if the frequency with which caregivers use words when speaking to children could play a role in cavity filling, expecting lower-frequency words to be more difficult for children to learn. However, again we observe no significant correlation between frequency and number of persistent cycles killed (Fig. \[fig:app\_corr\]). These results suggest that simple word descriptors such as length and frequency do not predict a word’s tendency to fill in knowledge gaps. Next we test if topological characteristics of nodes – rather than their non-topological statistics such as length and frequency – might better explain the number of persistent cycles killed. To address this question, we study the number of persistent cycles killed against node degree, centrality, and clustering coefficient (Fig. \[fig:6\]b-d). While node degree and betweenness centrality are positively correlated with the number of persistent cycles killed ($p<0.01$) as expected, the clustering coefficient shows a negative correlation ($p<0.01$). Initially, this result might appear counterintuitive because to cone (fill in) a cycle, a node must by definition create many triangles. Yet, if we combine this result with the positive correlations of persistent cycle killing to node degree and betweenness, we can construct a toy example of a possible cavity-killing node neighborhood. Shown in Fig. \[fig:6\]e, the central node outlined in white tessellates two cycles when added (cycles and coning triangles highlighted), but also has a low clustering coefficient. Taken together, we suggest that the connectivity pattern of words within the semantic network better predicts the tendency of that word to fill a knowledge gap than simple lexical features of the words themselves. Discussion {#discussion .unnumbered} ========== In this study, we query the existence of knowledge gaps manifesting as topological cavities within the growing semantic feature network of toddlers. Using persistent homology and the formalism of node-filtered order complexes, we find that such knowledge gaps both form and are often later filled in throughout the learning process. We observe that the global architecture of the growing semantic network is similar to that of a constrained generative model. Furthermore we report similar persistent homology across growing semantic networks of children from mothers with differing education, and we find that this pattern of topological cavity existence remains present after node order randomization, but not after edge rewiring. Together these results suggest that knowledge gaps are robust features of word production order and that the global topology of the semantic feature network is resilient to local alterations induced by node reordering. The accommodating topology of the growing semantic network {#the-accommodating-topology-of-the-growing-semantic-network .unnumbered} ---------------------------------------------------------- Understanding the growing semantic network through the lens of persistent homology offers a novel perspective on the nature of the learning process. Previous research has provided evidence supporting the influence of network topology on many types of language networks including those constructed from phonological [@arbesman2010structure; @siew2013community] and syntactic relations [@corominas2009ontogeny; @vcech2009word]. Here we observe that the persistent homology of the growing semantic network follows a regular pattern throughout the majority of the learning process, indicating an organized and potentially predictable growth pattern. Yet when we consider the node (word) level, the fine-scale topology (node degree) varies considerably and does not suggest a predictable addition pattern. Furthermore if we permute the order of the nodes uniformly at random, we recover similar global topology to that observed in the unpermuted network. We describe the semantic network topology as accommodating, since its large-scale architecture changes little despite variations in small-scale inputs (new nodes with differing degrees). Previous studies show that the order of word learning in children depends on multiple variables including word frequency [@brent2001role; @huttenlocher1991early], parental interaction [@hart1995meaningful], and communication quality [@hirsh2015contribution]. We speculate that part of the learning mechanism might include a global accommodating topology that develops similarly in children despite variations in input (order of words learned) due to differing environments. Though many possibilities for word production order yield similar persistent homology, we observe that the global structure disintegrates if we randomly rewire the network edges while preserving the degree of each node (Fig. \[fig:4\]). This finding – together with the results described in the previous paragraph – suggests that the higher order connectivity patterns between words, instead of individual word properties such as time of production, have a greater impact on the evolving global structure of the semantic network. This observation raises the important question of whether this resilience to node reordering and emphasis on fixing relations between words is restricted to the English language, or whether these phenomena are general properties of language networks. Previous research points to the similarity of word networks across languages [@youn2016universal] and supports the hypothesis that similar global patterns would be observed in languages other than English. Yet, the differences in node (word) connectivity patterns within this structure could offer insight into subtle distinguishing features between different languages [@goddard2008cross]. Hints toward novel learning mechanisms {#hints-toward-novel-learning-mechanisms .unnumbered} -------------------------------------- The presence of persistent topological cavities of multiple dimensions in the growing semantic network offers novel insights into the learning process. One might expect that when one grows one’s vocabulary, one tends to learn words that are similar to words already known. Such a process would leave few if any knowledge gaps, corresponding to topological cavities, within the network, and should be well-modeled by a topological-distance-from-initial-node rule. By contrast, we observe the salient presence of topological cavities in a growing semantic network that is best modeled by an edge-affinity rule. Yet, it is also important to acknowledge that the recovered gaps are not simply gaps in the final semantic network itself. Instead by the age of 30 months all but five gaps have been filled in by other words. Since we observed cavities in the growing semantic network irrespective of the mother’s education, we speculate that knowledge gaps that form *and* die may themselves be a feature of the semantic learning process. If indeed these knowledge gaps represent learning a more difficult word or concept, and filling in the created gap with intermediate concepts as they are later added, then these cavities may be a natural part of the learning process. As an extension to more explicit learning in a classroom, one could ask if cavities exist as students learn other subjects as well, in particular math and science where reaching for an understanding of distant or difficult concepts may create higher numbers of cavities or longer-lived persistent cavities. Many models exist for semantic networks in which edges are defined by word association such as can be estimated from a free association task or other metrics. Yet, models for the growing semantic *feature* networks as we study here are scarce if present, and efforts to use preferential attachment or close variations have not been successful in capturing the feature network’s development [@hills2009longitudinal]. The difference between the previously defined preferential attachment model and the affinity model that we introduce here is that the likelihood of new connections for each node evolves based on how the network has already grown in the former, while the latter only relies on predefined properties of nodes. As the semantic network a child is able to produce grows, the child is likely already sensitive to semantic relationships in their external world even before they acquire the label attached to a previously unnamed object. This sensitivity might explain why the affinity model better captures the topological properties of the growing semantic network. One possible way that the above concept can manifest in our encoding is that labels for features are allowed to exist in the network before a child is able to produce this label. For example, if a child produces ‘cheese’ and ‘bus’ but not ‘yellow’ at a given time, the lack of the child’s ability to produce ‘yellow’ does not mean that ‘cheese’ and ‘bus’ are not still both yellow and thus are connected. Similarly, since each word exists and connects to other words in the external world, a word will always have the same affinity for others. For example, any new animal with legs will always be ready to connect to all other objects that have legs, regardless of which of these words is known to the child. Furthermore, we speculate that words corresponding to nodes with high affinity may generally be *polysemous* words, or those with multiple meanings, thus increasing the likelihood of connections to other words within the semantic network [@sole2015ambiguity]. Overall our results are consistent with an externally constrained topology of the semantic network, as suggested in [@hills2009longitudinal; @steyvers2005large]. Topology in growing processes {#topology-in-growing-processes .unnumbered} ----------------------------- Growing networks are implemented in multiple systems including contagion propagation [@taylor2015topological], distribution networks in biological systems [@papadopoulos2016embedding], and social networks [@jin2001structure]. Consequentially, numerous methods exist for their analysis [@holme2012temporal]. For example, representing a growing process as a dynamic network or directed graph (or directed dynamic graph) would allow for analyses with those corresponding sets of tools [@chowdhury2016persistent; @sizemore2017dynamic]. Though persistent homology for node-weighted systems is not a novel concept [@taylor2015topological; @hofer2017deep; @courtney2017weighted], we suggest the formalism presented here as a practical mode of encoding such systems, in which both graph metrics and topological data analysis can be simultaneously applied. Additionally, these models hearken to previous theoretical studies including node-exchangeable graphs [@aldous1985exchangeability; @hoover1979relations], growing simplicial complexes [@courtney2017weighted; @bianconi2016emergent], and random clique complexes [@kahle2013limit]. Furthermore, we propose that the persistent homology of a n-order complex at the level of persistent features (as opposed to the more commonly studied global structure) may be more easily interpretable than that of an edge-weighted network. In particular, here individual words initiate and terminate persistent cycles. In many biological contexts, nodes are objects with attached empirical observations and metadata. The analyses of such systems might include analyzing the number of persistent cycles a node begins or kills in relation to this metadata[^2]. Additionally the n-order complex is invariant under any monotonic (rank-preserving) transformation of the node weights which is a noted benefit for applications to noisy experimental data [@giusti2015clique; @petri2013topological]. Topics suitable for the node-filtered complex encoding and subsequent analyses include tracking information dissemination through brain networks [@mivsic2015cooperative], signaling cascades in protein interaction networks [@vinayagam2011directed], sound propagation on force chains [@bassett2012influence], contagion spreading on social networks, and information transfer throughout enzyme architecture after allosteric effector binding [@cockrell2013new]. Broadly, the formality presented in this study may be useful for “filling in” open questions from multiple areas of science. Conclusions {#conclusions .unnumbered} ----------- In conclusion, we offer a novel perspective on the growing semantic feature network of toddlers that highlights the persistence of knowledge gaps in contrast to the formation of densely connected clusters. Using the node-filtered order complex formalism and persistent homology, we reveal the existence of such knowledge gaps and their persistence as children age. Furthermore we provide evidence supporting the notion that the gaps in the network will exist despite differences in word production times, and we speculate that these gaps are an important and general component of the learning process. Acknowledgments {#acknowledgments .unnumbered} =============== The authors thank Leonardo Torres, Dr. Tina Eliassi-Rad, and Brennan Klein for helpful discussions. This work was supported by the National Science Foundation CAREER PHY-1554488 to DSB. The authors also acknowledge support from the John D. and Catherine T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Paul G. Allen Foundation, the Army Research Laboratory through contract number W911NF-10-2-0022, the Army Research Office through contract numbers W911NF-14-1-0679 and W911NF-16-1-0474, the National Institute of Health (2-R01-DC-009209-11, 1R01HD086888-01, R01-MH107235, R01-MH107703, R01MH109520, 1R01NS099348 and R21-M MH-106799), the Office of Naval Research, and the National Science Foundation (BCS-1441502, CAREER PHY-1554488, BCS-1631550, and CNS-1626008). The content is solely the responsibility of the authors and does not necessarily represent the official views of any of the funding agencies. Supplementary Information {#supplementary-information .unnumbered} ========================= Details of topological methods {#details-of-topological-methods .unnumbered} ------------------------------ This section outlines the details of encoding the growing semantic network as a filtration and computing the persistent homology. We devote significant real estate to the encoding and briefly describe persistent homology since it is more thoroughly discussed elsewhere [@carlsson2009topology; @zomorodian2005computing; @ghrist2008barcodes; @hatcher2002algebraic]. ### Node-weighted networks and induced filtrations {#node-weighted-networks-and-induced-filtrations .unnumbered} Our motivation comes from data described most naturally as a network with weights on the nodes. Such systems can arise from protein-protein interaction networks with protein expression as node weights, structural brain networks with region activity as node weights, or a social network with time of contamination as nodes weights. Though generally classic graph statistics do not extend easily to networks with node weights, we present a construction that allows the simultaneous computation of both graph statistics and persistent homology on node-weighted networks. It is important to note that we derive our inspiration from an object generated from edge-weighted networks called the *order complex* [@giusti2015clique]. Given a graph with edge weights, we obtain an ordering on the edges by decreasing edge weights. Then we create a sequence of graphs, $G_0 \subset G_1 \subset \dots G_{|E|}$ with each $G_i$ the graph containing the $i$ highest ranked edges. This sequence of graphs is called the *order complex* of the weighted network (or corresponding symmetric weighted matrix). Now returning to node-weighted networks, we get a node ordering from the decreasing node vales. Then from this node ordering $s$ and graph $G$ with $N$ nodes we can similarly construct a sequence of graphs $G_0 \subset G_1 \subset \dots \subset G_N$ with $G_n$ containing the first $n$ nodes in $s$ and any connections between these nodes which exist in $G$. We call this sequence of binary graphs the *node-filtered order complex* of a node-weighted network, denoted nord$(G,w)$ with $G$ the binary graph and $w:N \rightarrow \mathbb{R}$ the function assigning node weights. For brevity we often refer to this object as the n-order complex. Note then the n-order complex is completely determined by the pair $(G,w)$ or $(G,s)$ with $s$ the node ordering. The n-order complex retains intrinsic developmental aspects of the node-weighted system, and furthermore allows for computation of both common binary graph metrics and persistent homology on these objects. Most graph metrics are not generalizable to include orders (weights) on the nodes, but if we instead construct the n-order complex, we now can compute such metrics on each $G_n$ in the filtration.\ *Construction of n-order complex into order complex* In practice, the software that we use to compute the persistent homology expects oder complexes. Then to compute the persistent homology of a growing network, we encode the associated n-order complex as an order complex. We will show by construction that any n-order complex $\text{nord}((G,s))$ can be translated into a weighted network $M$ with each $G_i$ in the filtration of $\text{nord}((G,s))$ equal to $G'_i$ in the filtration of $\text{ord}(M)$. Given a node-ordered network $(G,s)$, we construct the node-filtered order complex $\text{nord}((G,s))$. Then each $G_i$ is a binary graph and can be written as an $N \times N$ binary symmetric matrix $M_i$ (with $N-n$ extra padding rows and columns). Let $M = \sum_{i=0}^N M_i$. Now we have created a real-valued symmetric matrix $M$ that encodes a weighted network with the highest edge weights corresponding to the earliest added edges. If we then create $\text{ord}(M)$ with binary graphs $G_i'$, then by our construction it must be that $G'_i = G_i$ for all $i = 0, \dots, N$. Thus any n-order complex can be written as a weighted matrix $M$ with $\text{ord}(M) = \text{nord}((G,s))$. It is important to note that to get $G_i = G_i'$, we must assume that either all nodes always exist and we use the “growing graph" concept to describe how edges are added, or for both we do not include nodes in the graph until they have a neighbor. Finally, if we denote by $\mathcal{N}$ the set of filtrations achievable by node-weighted networks and $\mathcal{O}$ the set of filtrations achievable by edge-weighted networks, by the above discussion we must have $\mathcal{N} \subseteq \mathcal{O}$. ![Filtration of a node-ordered network (top) and equivalent edge-weighted network yielding the same graph filtration (bottom).[]{data-label="fig:sfig_filts"}](filtration_app.png){width="5in"} To examine the reverse relation, take the $3$-clique with nodes $a$, $b$, and $c$ with weights $e_{a,b} = 3$, $e_{b,c} = 2$, $e_{a,c} = 1$. The resulting filtration could not be created from a node-weighted network, since the final edge $e_{a,c}$ would be added to two nodes that already exist. In other words, it could not be added as the result of adding a new node. Therefore $\mathcal{N} \subsetneq \mathcal{O}$. ### Persistent homology {#persistent-homology-1 .unnumbered} Next we formally describe persistent homology. We begin with the task of detecting topological cavities in an unweighted graph. Given $G = (V,E)$ we translate this binary, unweighted graph into a combinatorial object called the clique complex by “coloring in” all cliques (all-to-all connected subgraphs) of $G$. Formally every $(k+1)$-clique, a completely connected subgraph of $G$ containing $k+1$ nodes (Fig. \[fig:sfig\_ph1\]a, top), is replaced with a $k$-simplex (Fig. \[fig:sfig\_ph1\]a, bottom). A $k$-simplex $\sigma = \{v_0, v_1, \dots, v_k\}$ is the convex hull of $k+1$ affinely positioned nodes. The collection of simplices created from the cliques of $G$ is called the *clique complex* $X(G)$. The clique complex of $G$ is an abstract simplicial complex, meaning that we have a vertex set $V$ (the original vertex set of $G$) along with a collection $K$ of subsets of $V$ that is closed under taking subsets. So elements of $K$ are simplices, and using geometric intuition we see clearly that any subset of a simplex must also be a simplex, called a *face*. We can write $X(G) = \{X_0(G), X_1(G), \dots, X_M(G)\}$ with each $X_k(G)$ being the collection of $k$-simplices of $G$ called the $k$-skeleton. To summarize thus far we have taken our graph $G$ and translated it into the combinatorial object called the clique complex $X(G)$. To locate topological cavities, we will need to perform algebra with elements of the clique complex. We create the *chain group* $C_k(X(G))$, a vector space with basis elements $\sigma$ corresponding to $k$-simplices of $X(G)$. Then elements of $C_k(X(G))$, called $k$-chains, are linear combinations of these basis elements. Though we can certainly choose coefficients from any group (for example, $\mathbb{Z}$), for computational purposes we work in the field $\mathbb{Z}_2$. To streamline notation, we will write $C_i(G)$ to mean $C_i(X(G))$ for simplicity.\ *Boundary operator.* To locate the topological cavities, we need to first comprehend the makeup and arrangements of simplices in our simplicial complex. For example, if we only have edges, we cannot tell which closed loops are true cavities without information about the positions of higher dimensional simplices within the complex. In particular, when searching for $k$-dimensional cavities we need to know the $k$-dimensional footprints of $(k+1)$-dimensional simplices. These footprints are the boundaries of $(k+1)$-simplices that can be computed using the boundary operator $\partial_{k+1}$. The boundary operator $\partial_{k+1}: C_{k+1} \rightarrow C_k$ is defined $$\partial_{k+1}(\sigma_{v_0, v_1, \dots, v_k}) = \sum_i (-1)^i\sigma_{v_0, v_1, \dots, v_{i-1}, v_{\hat{i}}, v_{i+1}, \dots, v_k}$$ with $v_{\hat{i}}$ omitted. Note the $(-1)^i$ records the directionality of chains, but since we work in $\mathbb{Z}_2$ we can drop this term. The boundary operator allows us to detect cavities due to a few particularly useful properties. First, the boundary operator extends linearly, so for $a,b \in C_k$, $\partial_k(a+b) = \partial_k(a) + \partial_k(b)$. Geometrically this means that the boundary of a collection of simplices is what we would intuit: the $(k-1)$-simplices that form a “shell” around the $k$-chain $a+b$ (for example, see Fig. \[fig:sfig\_ph1\]b). Next, let us examine what happens when we take the boundary of a cycle, a closed path of simplices. Again following geometric intuition, the boundary of a cycle is the end minus the beginning, which are the same in a closed path, and therefore the boundary should be $0$. Indeed, cycles of dimension $k$ are precisely the elements in $C_k$ sent to $0$ by $\partial_k$, or $\ker(\partial_k)$. Now note that one way we could construct a cycle is to take a $(k+1)$-simplex $\sigma$ and remove the interior – equivalently send $\sigma$ to its $k$-boundary. Then $\partial_{k+1}(\sigma)$ is a cycle, and thus $\partial_k(\partial_{k+1}(\sigma)) = 0$. If the boundary of any simplex is a cycle, then by linearity we get that the boundary of any $(k+1)$-chain is a $k$-cycle. Thus $\text{im}(\partial_{k+1}) \subseteq \ker(\partial_k)$. We call elements of $\text{im}(\partial_{k+1})$ $k$-boundaries. To summarize, we have $\ker(\partial_k)$ the $k$-cycles, $\text{im}(\partial_{k+1})$ the $k$-boundaries, and $\partial_k \circ \partial_{k+1} = 0$ (so $\text{im}(\partial_{k+1}) \subseteq \ker(\partial_k)$).\ *Equivalent cycles.* We have just seen how all $k$-boundaries are necessarily $k$-cycles. But what if $\text{im}(\partial_{k+1}) \subsetneq \ker(\partial_k)$? Then there exist $k$-cycles in $\ker(\partial_k) - \text{im}(\partial_{k+1})$ that do not surround a collection of higher dimensional simplices. Thus, they must instead enclose a void of dimension $k+1$ called a $k$-cavity. Since the cavities themselves are the features of interest, we do not want to simply enumerate $\ker(\partial_k) - \text{im}(\partial_{k+1})$, but instead we need to have all cycles surrounding the same cavity count as one (to avoid grossly overcounting). If two cycles surround the same cavity (we will assume each only surrounds one cavity for the sake of this example) then their difference must be some collection of higher dimensional simplices. More precisely, if we let $\ell_1$ and $\ell_2$ denote these two $k$-cycles, then $\ell_1 - \ell_2 \in \text{im}(\partial_{k+1})$. We call these two cycles *equivalent*. In fact, we say that any two cycles $a,b \in \ker(\partial_k)$ are equivalent if $a-b \in \text{im}(\partial_{k+1})$. For example, we see in Fig. \[fig:sfig\_ph2\]a the two cycles $a_1$ and $a_2$ are equivalent because $a_2 - a_1$ is the boundary of a $2$-simplex. However, $a_1 \not\sim b$, since their difference is not a boundary of a collection of 2-simplices. We could also take the $1$-cycle $c$ which surrounds both cavities, though note that it is not equivalent to any of $a_1$, $a_2$, or $b$ but is instead the sum $a_2 + b \sim a_1+b$. The defined equivalence relation partitions $\ker(\partial_k)$ into equivalence classes $[\ell_0] = \{\ell \in Z_k | \ell_0 - \ell \in \text{im}(\partial_{k+1})\}$. Then each (non-trivial) equivalence class corresponds to a topological cavity within the simplicial complex. By abuse it is common to refer to an equivalence class of $k$-cycles as a $k$-cycle.\ *Homology.* At this point in the exposition, we have detailed the intuitions and definitions necessary to concretely define the homology groups of simplicial complexes. The homology group is simply the group formed by the equivalence classes as we defined above. Formally stated $H_k(X(G)) := \ker(\partial_k)/\text{im}(\partial_{k+1})$. Each non-trivial equivalence class corresponds to a topological cavity, so $\dim(H_k(X(G)))$ is the number of $k$-cavities within the simplicial complex $X$. The dimension of $H_k(X(G))$ is called the $k^{th}$ Betti number $\beta_k$ and the list $\{\beta_0, \beta_1, \dots, \beta_m\}$ are the *Betti numbers* of $X(G)$.\ *Revisiting filtrations.* Earlier we introduced filtrations as a way to encode node-weighted networks. Consider one unit of the filtration, the map $i_k:G_k \hookrightarrow G_{k+1}$. Since every node and edge in $G_k$ maps to itself in $G_{k+1}$, we see that the map $i$ extends to clique complexes, with every simplex in $X(G_k)$ mapping to itself in $X(G_{k+1})$. This gives us the map $i'_k: X(G_k) \hookrightarrow X(G_{k+1})$. Then, the filtration of graphs induces a filtration of clique complexes (Fig. \[fig:sfig\_ph2\]b). Furthermore, the inclusion $X(G_k) \hookrightarrow X(G_{k+1})$ also means that we can easily map elements of the chain groups $C_*(X(G_k)) \hookrightarrow C_*(X(G_{k+1}))$, since we can take the above inclusion $i'_k$ as mapping the basis elements of $C_*(X(G_k))$ to those in $C_*(X(G_{k+1}))$. We depict these concepts in Fig. \[fig:sfig\_ph2\]b. Finally since we have these nice maps from one chain complex into the next, we can map cycles to cycles and consequentially the homology groups $H_*(X_k) \rightarrow H_*(X_{k+1})$ (Fig. \[fig:sfig\_ph2\]b, bottom). This means we can not only find $k$-cavities at each filtration index, but we can *follow* each $k$-cavity from the first point it exists in the filtration (called the birth), as it evolves throughout the filtration, and is killed (called the death) by simplices tessellating the cavity. Some cavities never die, so we assign them a death time of $\inf$. We call the *lifetime* of a persistent cycle the $death - birth$. The birth and death can be given in terms of the edge density [@giusti2015clique; @sizemore2016classification], filtration index, or for this study the number of nodes added. For example, the persistent 1-cycle in Fig. \[fig:sfig\_ph2\]b is born with the addition of node 6 and dies when node 9 is added, resulting in a lifetime $=3$. Models of node-filtered order complexes {#models-of-node-filtered-order-complexes .unnumbered} --------------------------------------- While this exposition is motivated by early semantic learning, creating simple models with controlled properties will help us gain an intuition for possible behaviors of n-order complexes. Though we describe the following models in the main text, for conceptual organization we revisit the definitions and group by model type. We describe two main categories of n-order complex models: generative and derived (Fig. \[fig:sfig\_models\]). Recall a n-order complex can be completely defined by the pair $(G,s)$ with $G$ a binary graph and $s$ an ordering of the nodes. A generative model creates the complex according to a set of rules: for n-order complexes we can generate a pair $(G,s)$ by first assuming $s = 1, 2, \dots, N$ and then constructing $G$. Derived models instead begin with either $G$ or $s$ and use this to construct the model. We can construct simple n-order complex models using a function $p:\{1,2, \dots, N\} \rightarrow [0,1]$ such that when node $n$ is added, each edge between node $n$ and all previous nodes exist with probability $p(n)$. In the main text we include results for $$p(n) = c\\,$$ $$p(n) = (n/N)^d\\,$$ called the constant probability model and proportional probability model, respectively. We could also enforce some global architecture such as community structure on the graph. If we let $c_n$ be the community of node $n$, then we can iteratively build a binary graph $G$ at step $n$ $$p(n,m) = \Big\{ \begin{array}{ll} p_{in} & c_n = c_m; \\ p_{out} & c_n \neq c_m \\ \end{array} \Big\}$$ with $p_{in} > p_{out}$ defining within or between module edge probabilities and $m < n$. We assign the community affiliation vector randomly with the desired number of communities and call this the modular n-order complex model. Instead of constructing a graph with a particular global structure, it may be the case that we may have some local information such as a predetermined affinity of each node for connections. The node affinity does not change as the network grows, making this inherently different than for example the preferential attachment model [@barabasi1999emergence]. Then given an affinity vector $a = (a_1, a_2, \dots, a_N)$ with $a_m$ the affinity of node $m$, we can construct a n-order complex using the following rule: when the $n$th node is added, the probability of an edge forming between node $n$, $m$ is $$p(n,m) = c\frac{a_m}{\max(a)} \mathrm{~.}$$ We call this the edge affinity model. If a node with normalized affinity $=1$ is not ideal, one can multiply $\frac{a_m}{\max(a)}$ by a constant $c$ to adjust the maximum probability that any node will acquire edges after it is added. The second class of models that we consider we call derived models, which, in contrast to generative models, alter features of an existing network and therefore require some prior knowledge of the system. We further group these into two basic types based on whether the edges of the binary graph $G$ or the node ordering $s$ changes. In the first, we maintain the original node ordering $s$ but, as an example, could randomly rewire the edges of $G$ while preserving degree distribution (similar to the configuration model [@bender1978asymptotic; @maslov2002specificity]) which we call the randomized edges model. In the second, we maintain the original graph $G$ and reorder the nodes, either randomly (randomized node model) or perhaps based on a node property of $G$ such as degree (decreasing degree model) or topological distance from a given node (distance from $v_0$ model). Additional parameters {#additional-parameters .unnumbered} --------------------- Parameters for the four presented generative n-order complex models were chosen to match the edge density of the semantic feature network ($\sim 0.3$). One might ask how these parameters affect the persistent homology of the growing graphs. In Fig \[fig:sfig\_er\] we show the persistent homology of the constant probability model with $p = 0.2$ and $p = 0.4$, the proportional probability model with $d = 0.5$ or $d =2$, the modular model with $p_{in} = 0.8$, $p_{out} = 0.2$ and $p_{in} = 0.6$, $p_{out} = 0.4$, and the edge affinity model with affinity vectors as random permutations of $(1:120)^3$ (left) and $(1:120)$ (right). We observe that the persistent homology varies considerably between outputs of the constant probability model constructed with differing parameters, suggesting that the edge density of the network plays a large role in the persistent homology of this model. Further information for maternal education levels {#further-information-for-maternal-education-levels .unnumbered} ------------------------------------------------- To supplement Fig. \[fig:4\], we include in Fig. \[fig:sfig\_education1\] the barcodes of the *secondary*, *college*, and *graduate* growing semantic networks with the starting and ending words of each persistent cavity. Additional persistent cycle death correlates {#additional-persistent-cycle-death-correlates .unnumbered} -------------------------------------------- Frequency of a word in child-directed speech is known to correlate with age of acquisition [@goodman2008does]. We asked if caregiver output frequency correlates with the number of persistent cycles each node kills. We extracted frequency counts of child-directed speech from [@macwhinney2009childes; @li2000acquisition; @goodman2008does]. Of our original 120 words, 87 were found in this database so we restrict the following calculations to those 87 words. As described in the main text, we calculated the Pearson correlation coefficient between the persistent cycle death count and the node degree, clustering coefficient, betweenness centrality, word length, and word frequency for the semantic network using all children, and separately when broken down by education level (Fig. \[fig:app\_corr\]). We observe a slightly decreasing trend of word frequency as corresponding nodes kill more persistent cycles, but this is not significant ($p > 0.3$ for each of the education levels). Trends for number of persistent cavities killed with node degree, clustering, and betweenness are similar to those seen with all nodes (Fig. \[fig:6\]) and remain significant ($p<0.01$) in all cases except for the clustering coefficient of the original all-included growing semantic network ($p = 0.022$). To summarize, we observe the connectivity patterns of words better determine the tendency of a word to fill in a knowledge gap than do simple lexical features. [^1]: Precisely we assign a $k$-simplex to each $(k+1)$-clique within $G$ to create the clique complex. See Appendix for definitions and details. [^2]: though assigning full responsibility of persistent cycles to individual cliques of any size should always be done with care [@bendich2015stabilizing]

--- address: | for the OPAL collaboration\ Albert-Ludwigs-Universität Freiburg\ Hermann-Herder-Str.3,\ D-79104 Freiburg, Germany\ E-mail: soldner@ruhpb.physik.uni-freiburg.de author: - 'STEFAN SÖLDNER-REMBOLD' title: 'TESTING QCD IN PHOTON-PHOTON INTERACTIONS' --- =cmr8 1.5pt \#1\#2\#3\#4[[\#1]{} [**\#2**]{}, \#3 (\#4)]{} Electron-photon scattering ========================== If one of the scattered electrons in $\ee$ collisions is detected (tagged), the process $\ee \rightarrow \ee + \mbox{hadrons}$ (Fig. \[fig-egfig\]) can be regarded as deep-inelastic scattering of an electron[^1] on a quasi-real photon which has been radiated by the other electron beam. The cross-section is written as $$\frac{{\rm d}^2\sigma_{\rm e\gamma\rightarrow {\rm e+hadrons}}}{{\rm d} x{\rm d}Q^2} =\frac{2\pi\alpha^2}{x\,Q^{4}} \left[ \left( 1+(1-y)^2\right) F_2^{\gamma}(x,Q^2) - y^{2} F_{\rm L}^{\gamma}(x,Q^2)\right], \label{eq-eq1}$$ where $\alpha$ is the fine structure constant and $$Q^2=-q^2=-(k-k')^2$$ is the negative four-momentum squared of the virtual photon $\gamma^*$ and $$x=\frac{Q^2}{2p\cdot q}=\frac{Q^2}{Q^2+W^2+P^2} \;\;\;\mbox{;}\;\;\; y=\frac{p\cdot q}{p\cdot k}$$ are the usual dimensionless variables of deep-inelastic scattering. $W^2=(q+p)^2$ is the squared invariant mass of the hadronic final state. The negative four-momentum squared, $P^2=-p^2$, of the quasi-real target photon is approximately zero. In leading order (LO) the photon structure function $F_2^{\gamma}(x,Q^2)$ is related to the sum over the quark densities of the photon weighted by the quark charge $e_{\rm q}$ $$F_2^{\gamma}(x,Q^2)=2x\sum_{\rm q} e^2_{\rm q} f_{\rm q/\gamma}(x,Q^2)$$ with $f_{\rm q/\gamma}(x,Q^2)$ being the probability to find a quark flavour q with the momentum fraction $x$ (sometimes denoted by $x_{\gamma}$) in the photon. For measuring $F_2^{\gamma}(x,Q^2)$ the values of $Q^2$ and $y$ can be reconstructed from the energy, [l]{}[2.5in]{} $E_{\rm tag}$, and the angle, $\theta_{\rm tag}$, of the tagged electron and the beam energy $E_{\rm beam}$: $$Q^2\approx 2E_{\rm beam} E_{\rm tag}(1-\cos\theta_{\rm tag})$$ $$y\approx 1-\frac{E_{\rm tag}}{E_{\rm beam}}\cos^2\frac{\theta_{\rm tag}}{2}.$$ In order to identify an electron in the detector, $E_{\rm tag}$ has to be large, i.e. $y^2\ll 1$. The contribution of the term proportional to the longitudinal structure function $F_{\rm L}^{\gamma}$ is therefore negligible (Eq. \[eq-eq1\]). The reconstruction of $x$, however, relies heavily on the measurement of the invariant mass $W$ from the energies $E_{\rm h}$ and momenta $\vec{p}_{\rm h}$ of the final state hadrons h: $$W^2=\left( \sum_{\rm h} E_{\rm h} \right)^2-\left( \sum_{\rm h} \vec{p}_{\rm h}\right )^2.$$ Unfolding of the $x$ dependence of $F_2^{\gamma}$ requires that the hadronic final state is well measured and well simulated by the Monte Carlo (MC) models. The photon structure function $F_2^{\gamma}$ {#sec-egamma} -------------------------------------------- Even though the concept of the photon structure function $F_2^{\gamma}$ has been developed in analogy to the formalism of the nucleon structure functions $F_2^{\rm N}$, there are important differences: $F_2^{\gamma}(x,Q^2)$ increases with $Q^2$ for all $x$ and this positive scaling violation is expected already within the parton model. Furthermore, $F_2^{\gamma}$ is large for high $x$, whereas $F_2^{\rm N}$ decreases at large $x$. These differences are due to the additional perturbative $\gamma\rightarrow\qqbar$ splitting which does not exist for the nucleon. For large $x$ and asymptotically large $Q^2$ the value of $F_2^{\gamma}$ can therefore be calculated from perturbative QCD [@bib-witten]. The next-to-leading order (NLO) result [@bib-buras] can be written as $$\frac{F_2^{\gamma}}{\alpha}=\frac{a(x)}{\alpha_{\rm s}(Q^2)}+b(x),$$ where $a(x)$ and $b(x)$ are calculable functions which diverge for $x\rightarrow 0$ and $\alpha_{\rm s}$ is the strong coupling constant. The first term corresponds to the LO result by Witten [@bib-witten]. The measurement of $F_2^{\gamma}$ could be a direct measurement of $\Lambda_{\rm QCD}$ if it were not for the large non-perturbative contributions due to hadronic states. (15.0,6.2) (0,3.25) -- -- -- -- (5.0,1.2)[(a)]{} (11.1,1.2)[(b)]{} $F_2^{\gamma}(x,Q^2)$ can be measured at LEP in the range $x>10^{-3}$ and $1<Q^2<10^3$ GeV$^2$ (Fig. \[fig-cov\]a). The $Q^2$ evolution of $F_2^{\gamma}$ is shown in Fig. \[fig-cov\]b using the currently available $F_2^{\gamma}$ measurements for 4 active flavours. The data are compared to the LO GRV [@bib-grv] and the SaS-1D [@bib-sas] parametrisations, and to a higher order (HO) prediction based on the NLO GRV parametrisation for light quarks and on the NLO charm contribution calculated in Ref. [@bib-laenen]. The data are measured in different $x$ ranges. The comparison of the LO GRV curves for these $x$ ranges shows that for $Q^2>100$ GeV$^2$ significant differences are expected. An augmented asymptotic prediction for $F_2^{\gamma}$ is also shown. The contribution to $F_2^{\gamma}$ from the three light flavours is approximated by Witten’s LO asymptotic form [@bib-witten]. This has been augmented by adding a charm contribution from the Bethe-Heitler formula [@WIT-7601], and an estimate of the hadronic part of $F_2^{\gamma}$ based on the hadronic part of the LO GRV parametrisation. In the region of medium $x$ values studied here, this asymptotic prediction in general lies higher than the GRV and SaS predictions but it is still in agreement with the data. The importance of the hadronic contribution to $F_2^{\gamma}$ decreases with increasing $x$ and $Q^2$, and it accounts for only 15 % of $F_2^{\gamma}$ at $Q^2= 59$ GeV$^2$ and $x = 0.5$. As predicted by QCD the evolution of $F_2^{\gamma}$ leads to a logarithmic rise with $Q^2$, but theoretical and experimental uncertainties are currently too large for a precision test of perturbative QCD. All currently available $F_2^{\gamma}$ measurements [@bib-f2g] are compared to the NLO GRV [@bib-grv] and the LO SaS-1D [@bib-sas] parametrisation in Fig. \[fig-sf\]. If the photon is purely hadron-like at low $x$, a rise of the photon structure function is expected at low $x$ for not too small $Q^2$, similar to the rise of the proton structure function. Only with the complete LEP2 data will it be possible to access regions in $x$ and $Q^2$ where the rise of $F_2^{\gamma}$ could really be observed. An interesting low $x$ measurement of $F_2^{\gamma}$ by OPAL lies in the ranges $2.5\times10^{-3}<x<0.2$ and $1.1<Q^2<6.6$ GeV$^2$. L3 has recently presented their first $F_2^{\gamma}$ measurement for $Q^2=1.9$ and $5.0$ GeV$^2$. These measurements are consistent with a possible rise within large errors. Jet production and NLO calculations {#sec-jet} ==================================== [r]{}[0.47]{} If the virtualities $Q^2$ and $P^2$ are approximately zero, i.e. both photons are quasi-real, LEP2 is a $\gg$ collider with $\gg$ centre-of-mass energies in the approximate range $10<W<120$ GeV. In LO different event classes can be defined in $\gg$ interactions. The photons can either interact as bare photons (“direct”) or as hadronic fluctuation (“resolved”). Direct and resolved events can be separated by measuring the fraction $\xgpm$ of the photon’s momentum participating in the hard interaction for the two photons. Ideally, the direct events are expected to have $\xgpm=1$, whereas for double-resolved events both values $\xgp$ and $\xgm$ are expected to be much smaller. For a given parton centre-of-mass energy the cross-sections vary only with the parton scattering angle $\theta^{*}$. The leading order direct process $\gg \rightarrow \qqbar$ is mediated by $t$-channel spin-$\frac{1}{2}$ quark exchange which leads to an angular dependence $\propto (1-\left|\cos\theta^{*2}\right|)^{-1}$. In double-resolved processes all matrix elements involving quarks and gluons have to be taken into account, with a large contribution from spin-$0$ gluon exchange. After adding up all relevant processes, perturbative QCD predicts an angular dependence of approximately $\propto (1-\left|\cost\right|)^{-2}$. Figure \[fig-theta\] shows the corrected $|\cost|$ distribution of dijet events with $\xgpm>0.8$ and with $\xgpm < 0.8$ compared to the calculation for various LO matrix elements which qualitatively reproduce the data. NLO jet cross-sections for $\gg$ interactions have been calculated by many authors [@bib-kleinwort; @bib-aurenche] using the cone jet finding algorithm [@bib-cone]. The transverse momentum $p_{\rm T}$ of the final-state partons (or the jet) defines the hard scale. The jet cross-section is written as a convolution of the parton density of the photon with the matrix elements for the scattering of two partons. In the kinematic range covered by LEP the $F_2^{\gamma}$ measurements are mainly probing the quark content of the photon, whereas jet production can be used to constrain the relatively unknown gluon distribution in the photon. Inclusive one-jet and dijet cross-sections have been measured in $\gg$ scattering at an $\ee$ centre-of-mass energy of $\sqee=58$ GeV at TRISTAN [@bib-amy; @bib-topaz] and at $\sqee=130-172$ GeV by OPAL [@bib-opalgg; @bib-opalgg2]. The $\ETJET$ distribution for dijet events in the range $|\etajet|<2$ measured by OPAL [@bib-opalgg2] at $\sqee=161-172$ GeV is shown in Fig. \[fig-ettwojet\]a. The measurements are compared to a NLO calculation [@bib-kleinwort] which uses the NLO GRV parametrisation [@bib-grv]. The direct, single- and double-resolved parts and their sum are shown separately. The data points are in good agreement with the calculation except in the first bin where theoretical and experimental uncertainties are large. (15.0,6.0) (0,3.0) -- -- -- -- (1.5,3.4)[(a)]{} (9.0,3.4)[(b)]{} To study the sensitivity to the choice of parametrisation for the parton distributions of the photon, OPAL has also measured the inclusive dijet cross-section as a function of $|\etajet|$ for events with a large double-resolved contribution obtained by requiring $\xgpm<0.8$ (Fig. \[fig-ettwojet\]b). The inclusive dijet cross-section predicted by the two LO QCD models PYTHIA [@bib-pythia] and PHOJET [@bib-phojet] differ significantly even if the same parton distribution functions (here LO GRV) are used, reducing the sensitivity to the parametrisation. Different parametrisations were used as input to PYTHIA. LO GRV [@bib-grv] and SaS-1D [@bib-sas] describe the data equally well, but LAC1 [@bib-LAC1], which increases the cross-section for gluon-initiated processes, overestimates the inclusive dijet cross-section significantly. A correct treatment of multiple parton interactions (MI) is also important. PYTHIA plus LAC1 with and without MI differs by more than a factor of two. The influence of MI can be constrained by studying energy flows outside the jets. Inclusive charged hadron production =================================== Measurements of inclusive charged hadron production complement similar studies of jet production. OPAL has measured the differential cross-sections $\dspt$ as a function of the transverse momentum $\pt$ of charged hadrons at $\sqee=161-172$ GeV. Until now, $\pt$ distributions of charged hadrons have only been measured for single-tagged events by TASSO [@bib-tasso] and MARK II [@bib-mark2] at an average $\langle Q^2 \rangle$ of 0.35 GeV$^2$ and 0.5 GeV$^2$, respectively. The $\pt$ distributions in $\gg$ interactions are expected to be harder than in $\gamma p$ or hadron-p interactions due to the direct component. This is demonstrated in Fig. \[fig-wa69\]a by comparing the $p_{\rm T}$ distribution for $\gg$ interactions to $p_{\rm T}$ distributions measured in $\gamma$p and hp (h$=\pi,$K) interactions by WA69 [@bib-wa69]. The WA69 data is normalised to the $\gg$ data in the low $p_{\rm T}$ region at $\pt\approx 200$ MeV/$c$ using the same factor for the hp and the $\gamma$p data. The $p_{\rm T}$ distribution of WA69 has been measured in the Feynman-$x$ range $0.0<x_{\rm F}<1.0$. The hadronic invariant mass of the hp and $\gamma$p data is about $W=16$ GeV which is of similar size as the the average $\langle W \rangle$ of the $\gg$ data in the range $10<W<30$ GeV. Whereas only a small increase is observed in the $\gamma$p data compared to the h$\pi$ data at large $\pt$, there is a significant increase of the relative rate in the range $\pt>2$ GeV/$c$ for $\gg$ interactions due to the direct process. (15.0,6.0) (0,3.0) -- -- -- -- (1.5,3.4)[(a)]{} (9.0,3.4)[(b)]{} The cross-sections $\dspt$ are also compared to NLO calculations [@bib-binnewies] which are calculated using the QCD partonic cross-sections to NLO for direct, single- and double-resolved processes. The hadronic cross-section is a convolution of the Weizsäcker-Williams effective photon distribution, the parton distribution functions and the fragmentation functions of Ref. [@bib-bkk] which are obtained from a fit to $\ee$ data from TPC and ALEPH. The NLO GRV parametrisation is used [@bib-grv] with $\Lambda^{(5)}_{\overline{\rm MS}}=131$ MeV and $m_{\rm c}=1.5$ GeV$/c^2$. The renormalization and factorization scales are set equal to $\xi\pt$ with $\xi=1$. The change in slope around $\pt=3$ GeV/$c$ in the NLO calculation is due to the charm threshold, below which the charm distribution in the resolved photon and the charm fragmentation functions are set to zero. A minimum $p_{\rm T}$ of 1 GeV/$c$ is required to ensure the validity of the perturbative QCD calculation. The NLO calculation is shown separately for double-resolved, single-resolved and direct interactions. At large $\pt$ the direct interactions dominate. It should be noted that these classifications are scale dependent in NLO. The scale dependence of the NLO calculation was studied by setting $\xi=0.5$ and 2. This leads to a variation of the cross-section of about $30\%$ at $\pt=1$ GeV/$c$ and about $10\%$ for $\pt>5$ GeV/$c$. The NLO calculations of $\dspt$ lie about $25\%$ below the data for $10<W<125$ GeV. Total cross sections {#sec-total} ==================== The total cross-sections $\sigma$ for hadron-hadron and $\gamma$p collisions are well described by the parametrisation $\sigma=X s^{\epsilon}+Y s^{-\eta}$, where $\sqrt{s}$ is the centre-of-mass energy of the interaction. Assuming factorisation for the Pomeron term $X$, the total hadronic $\gg$ cross-section $\sigma_{\gg}$ can be related to the pp (or $\ppbar$) and $\gamma$p total cross-sections at high $W=\sqrt{s}_{\gg}$, where the Pomeron trajectory should dominate: $$\sigma_{\gg} =\frac{\sigma_{\gamma{\rm p}}^2}{\sigma_{\rm pp }}. \label{eq-tot2}$$ A slow rise of the total cross-sections with energy is predicted, corresponding to $\epsilon\approx0.08$. Before LEP $\sigma_{\gg}(W)$ has been measured by PLUTO [@bib-pluto], TPC/2$\gamma$ [@bib-tpc] and MD1 [@bib-md1] in the region $W<10$ GeV, before the onset of the high energy rise of $\sigma_{\gg}$. Using LEP data taken at $\sqee=130-161$ GeV L3 [@bib-l3tot] has demonstrated that $\sigma_{\gg}(W)$ is consistent with the universal Regge behaviour of total cross-sections in the range $5\le W \le 75$ GeV. The L3 measurement is shown [r]{}[0.45]{} in Fig. \[fig-stot\] together with an OPAL measurement in the range $10<W<110$ GeV using data taken at $\sqee=161-172$ GeV. The observed energy dependence of the cross-section is similar, but the values for $\sigmagg$ are about $20\%$ higher. The errors are strongly correlated between the $W$ bins in both experiments. About $5~\%$ discrepancy is due to the different MC generators used for unfolding. The origin of the remaining discrepancy is not yet understood. Based on the DL-model [@bib-DL], the assumption of a universal high energy behaviour of $\gg$, $\gamma$p and pp cross-sections is tested. The parameters $X$ and $Y$ are fitted to the total $\gg$, $\gamma$p and pp cross-sections in order to predict $\sigmagg$ via Eq. \[eq-tot2\] using $\sqrt{s}_{\gg}=\sqrt{s}_{\rm \gamma p}=\sqrt{s}_{\rm pp }$. The process dependent fit values for $X$ and $Y$ and the universal parameters $\epsilon = 0.0790 \pm 0.0011$ and $\eta = 0.4678 \pm 0.0059$ are taken from Ref. [@bib-pdg]. This simple ansatz gives a reasonable description of the total $\gg$ cross-section $\sigmagg$. The models of Schuler and Sjöstrand [@bib-GSTSZP73] and the model of Engel and Ranft [@bib-phojet] are also shown. Conclusions =========== In general, $\gg$ interactions are similar to hadron-hadron interactions. At centre-of-mass energies $\sqrt{s}_{\gg}>10$ GeV the energy dependence of the total $\gg$ cross-section is consistent with the rise observed in hadronic interactions. QCD is tested in $\gg$ interactions at LEP by comparing to LO and NLO QCD calculations. In case of the photon, the perturbative splitting $\gamma\to\qqbar$ must also be taken into account which modifies the QCD predictions. This is observed in the scaling violations of the photon structure function $F_2^{\gamma}$. Information about the gluon content of the photon can be extracted from measurements of jet production. NLO calculations are in reasonable agreement with the data. Comparing transverse momentum distribution of $\gg$ interactions with hadron-proton or $\gamma$-proton data shows the relative increase of hard interactions in $\gg$ processes due to the direct component. References {#references .unnumbered} ========== [99]{} E. Witten, [*Nucl. Phys.*]{} [**B120**]{} (1977) 189. W.A. Bardeen, A.J. Buras, [*Phys. Rev.*]{} [**D20**]{} (1979) 166; [*Phys. Rev.*]{} [**D21**]{} (1980) 2041. M. Glück et al., [*Phys. Rev.*]{} [**D46**]{} (1992) 1973; Phys. Rev. [**D45**]{} (1992) 3986. G. A. Schuler, T. Sjöstrand, [*Z. Phys*]{}. [**C68**]{} (1995) 607. E. Laenen et al, [*Phys. Rev.*]{} [**D49**]{} (1994) 5753. E. Witten, [*Nucl. Phys.*]{} [**B104**]{} (1976) 445. A compilation can be found in D. Morgan et al, [*J. Phys. G*]{}, [*Nucl. Part. Phys.*]{} [**20**]{} (1994) A1. More recent references are given in S. Söldner-Rembold, hep-ex/9711005, proc. of LP97, Hamburg. The L3 values are taken from M. Wadhwa, talk given at Moriond (1998). T. Kleinwort, G. Kramer, [*Z. Phys.*]{} [**C75**]{} (1997) 489; [*Nucl. Phys.*]{} [**B477**]{} (1996) 3; [*Phys. Lett.*]{} [**B370**]{} (1996) 141. P. Aurenche et al, [*Progr. Theor. Phys.*]{} [**92**]{} (1994) 175. CDF Coll., F. Abe et al, [*Phys. Rev.*]{} [*D45*]{} (1992) 1448; J. Huth et al, [*Proc. of the 1990 DPF Summer Study on High Energy Physics*]{}, Snowmass, Colarado (1992) 134. AMY Coll., B. J. Kim et al., [*Phys. Lett.*]{} [**B325**]{} (1994) 248. TOPAZ Coll., H. Hayashii et al., [*Phys. Lett.*]{} [**B314**]{} (1993) 149. OPAL Coll., K. Ackerstaff et al., [*Z. Phys.*]{} [**C73**]{} (1997) 433. OPAL Coll., paper LP-201 submitted to Lepton-Photon 1997. T. Sjöstrand, [*Comp. Phys. Commun.*]{} [**82**]{} (1994) 74 and LU-TP-95-20. R. Engel, J. Ranft, [*Phys. Rev.*]{} [**D54**]{} (1996) 4244; R. Engel, [*Z. Phys*]{}. [**C66**]{} (1995) 203. H. Abramowicz, K. Charchula, A. Levy, [*Phys. Lett.*]{} [**B269**]{} (1991) 458. TASSO Coll., R. Brandelik et al., Phys. Lett. B107 (1981) 290. MARK II Coll., D. Cords et al., Phys. Lett. B302 (1993) 341. WA69 Coll., R.J. Apsimon et al., Z. Phys. C43 (1989) 63. J. Binnewies, B.A. Kniehl and G. Kramer, Phys. Rev. D53 (1996) 6110. J. Binnewies, B.A. Kniehl and G. Kramer, Phys. Rev. D52 (1995) 497. PLUTO Coll., Ch. Berger et al., [*Phys. Lett.*]{} [**B149**]{} (1984) 421. TPC/2$\gamma$ Coll., H. Aihara et al., [*Phys. Rev.*]{} [**D41**]{} (1990) 2667. S. E. Baru et al., [*Z. Phys*]{}. [**C53**]{} (1992) 219. L3 Coll., M. Acciarri et al., [*Phys. Lett.*]{} [**B408**]{} 1997 450. A. Donnachie, P. V. Landshoff, [*Phys. Lett.*]{} [**B296**]{} (1992) 227. R. M. Barnett et al., [*Phys. Rev.*]{} [**D54**]{} (1996) 1. G. A. Schuler, T. Sjöstrand, [*Z. Phys*]{}. [**C73**]{} (1997) 677. [^1]: In this paper positrons are also referred to as electrons

--- author: - 'M. Gerin and H. Liszt' date: received title: ALMA hints at the existence of an unseen reservoir of diffuse molecular gas in the Galactic bulge --- [We aim to understand the unexpected presence of mm-wave molecular absorption at -200 $< {\rm v} < -140$  in a direction that is well away from regions of the Galactic bulge where CO emission at such velocities is prominent.]{} [We compared 89 GHz Cycle 2 ALMA absorption spectra of [[[$\mathrm{HCO^+}$]{}]{}]{}, HCN, and HNC toward the extragalactic continuum source B1741-312 at l=-2.14, b=-1.00 with existing CO, H I, and dust emission and absorption measurements. We placed the atomic and molecular gas in the bulge and disk using circular and non-circular galactic kinematics, deriving N(H I) from a combination of 21cm emission and absorption and we derive N([[[$\mathrm{H_2}$]{}]{}]{}) from scaling of the [[[$\mathrm{HCO^+}$]{}]{}]{} absorption. We then inverted the variation of near-IR reddening E(J-K) with distance modulus and scale E(J-K) to a total gas column density N(H) that may be compared to N(H I) and N([[[$\mathrm{H_2}$]{}]{}]{}). ]{} [At galactocentric radii  $>$ 1.5 kpc, conventional measures such as the standard CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor and locally observed N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) ratio separately imply that H I and [[[$\mathrm{H_2}$]{}]{}]{} contribute about equally to N(H), and the gas-derived N(H) values are in broad agreement with those derived from E(J-K). Within the Galactic bulge at $<$ 1.5 kpc, H I contributes less than 10% of the material inferred from E(J-K), so that the molecular absorption detected here is needed to understand the extinction.]{} Introduction ============ The distribution of molecular gas in the Galactic nucleus has been heavily studied in the so-called central molecular zone (CMZ) [@MorSer96] that extends $\pm$ 1.2 in Galactic longitude or $\pm$ 180 pc projected distance at the IAU standard distance R$_0$ = 8.5 kpc. The molecular component outside the CMZ, but within the bulge, is far less well understood owing to limited latitude coverage and sparse sampling of the CO survey [@BitAlv+97] and to uncertainty in the conversion from CO brightness to [[[$\mathrm{H_2}$]{}]{}]{} column density [@BolWol+13]. In the Galactic bulge, the atomic gas has been fully if somewhat coarsely mapped [@BurLis83; @HI4PI], and the mass of H I, $3\times 10^7$ [[$M_\odot$]{}]{} is comparable to that of the molecular gas within the much more compact CMZ. The atomic gas shows a broad range of phenomena associated with large-scale gas flow in the Galactic bar outside the CMZ [@Fux99] but only a small portion of that is seen in CO and other molecules [@RodCom+06]. Molecular emission is present in a few very broad-lined features associated with the shredding and intake of molecular gas at the standing dust lane shock in the Galactic bar, and CO is detected in the terminal-velocity feature associated with inward gas flow along the dust lane [@Lis06Shred; @Lis08Multi]. Even CO emission is largely absent from most of the velocity range observed in H I, however, presumably because of limited sensitivity to weak, broad emission. This led to the suggestion [@LisBur78] that a large reservoir of molecular gas had yet to be observed in the bulge. Some of this gas may have been serendipitously observed in absorption in an ALMA experiment studying the chemistry of CF and HOC in local diffuse molecular gas, and this paper explains why that might be the case. The structure of this work is as follows. In Sect. 2 we describe the new and existing observational material that is discussed. In Sect. 3 we present our observational results, from which we derive column densities of diffuse atomic and molecular gas in the Galactic disk and bulge, and compare these with gas column densities implied by the variation of near-IR dust reddening E(J-K) with distance modulus. Section 4 is a summary and discussion. Observations and data reduction =============================== ![SGPS H I, Southern Mini CO, and ALMA [[[$\mathrm{HCO^+}$]{}]{}]{}, HCN, and HNC profiles. The strongest CO emission feature is the counterpart of a strong and very widespread H I self-absorption feature.](B1741-NewFig1.eps){height="9.6cm"} ![Latitude-velocity diagram of CO emission at $l =-2.125$  with the ALMA [[[$\mathrm{HCO^+}$]{}]{}]{} absorption profile superposed such that its 0-level is at the latitude of B1741-312.](B1741-NewFig2.eps){height="9.6cm"} ALMA absorption measurements ---------------------------- We observed [[[$\mathrm{HCO^+}$]{}]{}]{}, HCN, and HNC in absorption near 89 GHz in ALMA Cycle 2 project 2013.1.01194.S, whose scientific goal was a comparative study of the chemistries of CF and HOC with respect to HF and [[[$\mathrm{H_2}$]{}]{}]{}O in diffuse molecular gas. The blazar background source B1741-312 at $l=-2.1366$, $b=-0.9968$ was observed in nine spectral line windows, three of which covered the J=1-0 transitions of HCN, [[[$\mathrm{HCO^+}$]{}]{}]{}, and HNC at 88.631 GHz, 89.188 GHz, and 90.663 GHz, respectively, as discussed here. The appearance of the HCN absorption is complicated by the inherent hyperfine splitting into three components with local thermal equilibrium (LTE) optical depths in the ratio 1:5:3 in velocity order. Observations of HOC, CF, C$_3$H, and c-C$_3$H in the nearby gas will be discussed in a forthcoming paper. Although the intended subject of study toward B1741-312 was relatively local gas near 0-velocity, the spectral setup was held in common with those used for Sgr A and Sgr B in the same science goal, therefore extending almost to -200 : the lowest-frequency transition, that of HCN, extends the farthest to negative velocity, detecting (with [[[$\mathrm{HCO^+}$]{}]{}]{}) a feature at -190  that was narrowly missed in the highest frequency spectral window, that for HNC. The channel spacing of the ALMA data presented here is 244 kHz, or approximately 0.8 , and the resolution is twice as high. The continuum flux was 0.59 Jy and the source was observed for 720s, leading to channel-channel rms optical depth fluctuations of 0.0062 at zero optical depth. The spectra were reduced using the standard ALMA pipeline in January and February 2016 after the data had been delivered. Spectra were extracted from the final data cubes at the pixel of peak continuum flux. The velocities of these and all other data discussed here are measured with respect to the kinematic definition of the Local Standard of Rest (LSR). Other data: H I and CO emission ------------------------------- We present CO J=1-0 emission profiles from the Southern Mini-Telescope CO survey of [@BitAlv+97] at 8 spatial resolution on a 1/8 grid and H I emission from the Southern Galactic Plane Survey (SGPS) survey data of [@McCDic+05] at 145 spatial resolution, which we used in conjunction with the H I absorption line profile of [@DicKul+83] to derive an optical-depth corrected N(H I) as $$N(H~I) = 1.823\times 10^{18}\pcc \int \frac{\tau(v)T_B(v)}{1-\exp{(-\tau(v))}} dv, \eqno(1)$$ where $\tau$ and T$_B$ are the optical depth and brightness temperature and the units of velocity are . Extinction ---------- We also use data at 6 resolution from the 3D IR extinction map of the bulge of [@SchChe+14], extracting the variation of E(J-K) along the line of sight at (l,b) = (-2.1,-1.0) and using the relationships A$_{\rm K}$/ = 0.11, A$_{\rm K}$/E(J-K) = 0.689, or  = 6.26 E(J-K) and N(H) $= 1.171\times 10^{22}\pcc$ E(J-K), where N(H) is the total column density of H-nuclei in atomic and molecular form, N(H) = N(H I)+2N([[[$\mathrm{H_2}$]{}]{}]{}). Observational results ===================== Atomic and molecular absorption and emission spectra toward and around B1741-312 are shown in Fig. 1. Strong molecular absorption lines are seen from local gas around 0-velocity, and the molecular feature at v = 7  corresponds to a strong H I self-absorption feature originally discovered by [@Hee55] that provided the first evidence of inhomogeneity in the temperature structure of the interstellar H I gas. The high column density of local gas seen around 0-velocity dominates the total absorption and emission in this direction, but molecular absorption also appears around -25, -150, and -190 . The spectrum of HNC, lying at the highest rest frequency, did not quite extend out to the narrow absorption near -190 . Molecular absorption at -190  from the Galactic bulge is narrow in [[[$\mathrm{HCO^+}$]{}]{}]{} and HCN, as perhaps arising from a single isolated cloud. The absorption around -150  is wider and more complex, with two kinematic sub-components separated by 6  in HNC and more continuous absorption over 20  in [[[$\mathrm{HCO^+}$]{}]{}]{}. The absorption depth ratios of the three species and the somewhat greater ubiquity of [[[$\mathrm{HCO^+}$]{}]{}]{} in the high negative-velocity bulge features are typical of molecular absorption from diffuse molecular gas in the interstellar medium (ISM) near the Sun [@LisLuc01], or the Galactic disk [@GodFal+10]. H I emission is apparent around -190  toward B1741-312 in Fig. 1, but is seen only much more weakly at -150 . CO emission is detected at -22 and with low significance at -150 , but not at $-190$ . CO emission at high negative velocities is, however, present in the vicinity of B1741-312. In Fig. 2 we superpose the [[[$\mathrm{HCO^+}$]{}]{}]{} absorption profile on a latitude-velocity cut at l = -2.125 in CO emission, illustrating the peculiar geometry and non-circular kinematics of gas in the bulge. CO emission from the bulge, at $|$v$| > 30$ , appears preferentially above the Galactic equator owing to the tip and tilt of the inner galaxy gas [@BurLis78; @LisBur78] with the high positive-velocity gas on the far side of the gas distribution appearing at higher latitude [^1]. CO emission around -190  is present in Fig. 2 only at $ b > -0.5$ at a minimum projected distance of 75 pc. The distribution of CO emission is shown more comprehensively in Fig. 3, which locates B1741-312 on the sky as seen in slices 20 wide of the CO emission data cube. CO emission is detected at -190  within some 10-15 of the position of B1741-312, a projected separation of 25 - 35 pc. The molecular absorption at -150  is seen just at the periphery of a nearby CO emission feature, as also suggested in Fig. 2. Placement of kinematic features ------------------------------- Gas kinematics in the Galactic disk are dominated by circular rotation while motions in the Galactic bar and bulge are characterized by strongly non-circular velocities [@Fux99]. Figure 4 shows the variation in LSR velocity with distance and galactocentric radius for pure circular motion with a flat rotation curve with R$_0$ = 8.5 kpc, $\Theta = 220$ , and for a non-circular velocity field inside  = 1.5 kpc using approximately equal circular and radial motions 170-180  to explain the geometry and gas motions in the inner galaxy, including the stronger CO emission at higher latitude in Fig. 2 [@BurLis78]. In pure rotation, velocities v $<$ -140  arise exclusively within 300 - 500 pc of the center: with non-circular motion the velocity field is bifurcated, and velocities v $<$ -140  arise at both larger and smaller radii in the region interior to  = 1.5 kpc (the outermost radius of the underlying model). The line-of-sight velocity gradient for disk gas in pure rotation is quite shallow within 5-6 kpc of the Sun, and the LSR velocity at the edge of the bulge gas distribution at  = 1.5 kpc in Fig. 4 is -40 . A separation into bulge gas at v $\la -140$  and disk gas at v $\ga$ -40  works quite well for the molecules seen toward B1741-312 and for most of the H I because there are no features or strong emission in the intermediate velocity range. H I emission from the 3 kpc arm is seen in the intermediate velocity range, and the 3 kpc arm is usually placed in the transition region between the bulge and disk gas distributions [@Fux99]. Derived results --------------- To provide a quantitative assessment of the importance of the molecular absorption, Table 1 summarizes the observed line properties for a set of velocity ranges framing the features that were detected in [[[$\mathrm{HCO^+}$]{}]{}]{}. The table entries from the top down are as follows: - [: galactocentric radius. For the two most negative velocity ranges, the gas is ascribed to the bulge at  = 320 - 1500 pc without greater specificity given the bifurcation in the velocity shown in Fig. 4 and discussed in Sect. 3.1: 320 pc is the pericenter distance. The other ranges of galactocentric radius were assigned on the basis of the flat rotation curve with R$_0$ = 8.5 kpc, $\Theta$ = 220 . The identical radius ranges given for the two highest velocity intervals both correspond to v $\ge -15$  in pure rotation. The molecular gas seen at 7   must be near the Sun, given the strong H I self-absorption, even though positive velocities are not allowed by rotational kinematics within the solar circle toward B1741-312. H I emission at v $\ga 20$  probably arises in the inner galaxy at an uncertain galactocentric radius, and its contribution to the column density, 7.3% of the total, has been ignored in this table.]{} - [W$_{\rm X}$: the integrated observed optical depth for HCN, [[[$\mathrm{HCO^+}$]{}]{}]{}, and HNC (units of ) and integrated brightness temperature for H I and CO observed in emission, in units of K-.]{} - [N(X): column densities (units of $\pcc$) calculated using Eq. 1 for H I,  = 2.729 K for [[[$\mathrm{HCO^+}$]{}]{}]{}, and with N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) $= 3 \times 10^{-9}$, the local value for solar metallicity as a benchmark for [[[$\mathrm{H_2}$]{}]{}]{}. A fixed conversion taken from the values determined near the Sun was also used by [@SchChe+14], whose 3D E(J-K) extinction maps we used to derive the variation of N(H) as inferred from dust measures.]{} - [X$_{\rm CO}$: N([[[$\mathrm{H_2}$]{}]{}]{})/[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{}, the CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor. We note that the assumed [[[$\mathrm{HCO^+}$]{}]{}]{} abundance yields a canonical value $2\times 10^{20}$ [[[$\mathrm{H_2}$]{}]{}]{}$\pcc$ (K-)$^{-1}$ for the CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor for the two velocity intervals at v $\ge -15$ , independent of any assumption about the rotation curve or placement of the gas in the Galaxy. [@SchChe+14] derived a similar CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor in the vicinity of B1741-312 when N(H) was derived from the total E(J-K) and compared with profile-integrated total CO measurements from the NANTEN2 telescope. This occurs because the gas column is dominated by the local material seen around zero-velocity.]{} {height="14cm"} The tabulated values of N(H I) and N([[[$\mathrm{H_2}$]{}]{}]{}) show a marked change from an approximately equal mixture of atomic and molecular gas in the disk at  $>$ 3 kpc to a predominantly molecular medium in the bulge at  $<$ 1.5 kpc. Deriving the [[[$\mathrm{H_2}$]{}]{}]{} column density from CO emission using a standard CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor N([[[$\mathrm{H_2}$]{}]{}]{})/[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{} = $2\times 10^{20}\pcc$ (K-)$^{-1}$ [@BolWol+13] also yields approximately equal proportions locally, but CO emission underrepresents the molecular gas at  $<$ 1.5 kpc. This is in strong contrast to the low values for the CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor that apply in the CMZ where CO is overluminous [@BolWol+13; @SchChe+14]. The transition region at   = 1.5 - 3 kpc represented by the velocity range -41 $\le$ v $\le$ -15 is predominantly atomic when N([[[$\mathrm{H_2}$]{}]{}]{}) is derived from [[[$\mathrm{HCO^+}$]{}]{}]{}, but with relatively bright CO, leading to a small implied CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor. The proportion of molecular gas would be much larger and about equal to that of H I if the standard local CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion were applied. Importance of the molecular absorption -------------------------------------- In Fig. 5 we show the variation of cumulative total column density N(H) working outward to the solar circle, as derived under various assumptions. The solid curve results from interpolating between the binned entries in the 3D extinction data cube of [@SchChe+14] at $l=-2.1$, $b=-1$ using a constant scaling N(H) $= 1.171\times 10^{22}\pcc$ E(J-K), as discussed in Sect. 2.3. The extinction data do not extend beyond a heliocentric distance of 10.5 kpc, but this should include the entire bulge contribution, while only a very small contribution is expected from the far-side disk gas at the latitude of B1741-312. Even the radio spectra do not sample much material outside the solar circle, which is seen at z-heights exceeding 300 pc on the far side of the Galaxy. We also show in Fig. 5 values for N(H I) and N(H) where the latter is derived either using a constant abundance X([[[$\mathrm{HCO^+}$]{}]{}]{}) $ = 3\times 10^{-9}$ or a canonical CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor N([[[$\mathrm{H_2}$]{}]{}]{})/[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{} = $2\times 10^{20} {{{\ensuremath{\mathrm{H_2}}}}}\pcc$ (K-)$^{-1}$. The value plotted is integrated over the full corresponding velocity range. We do not pretend to know where within the range the matter is distributed. The least uncertain values are those for N(H I), so that if the radii are properly assigned, it is apparent that the extinction requires an additional contribution whose magnitude is like that which we have seen in molecular absorption. In the outer regions of the disk, the existence of the molecular gas can be inferred from CO emission using a standard CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor, but this is not the case in the bulge at  $<$ 1.5 kpc. High [[[$\mathrm{H_2}$]{}]{}]{}-fractions in the bulge gas ---------------------------------------------------------- Features associated with the bulge have a remarkably high fraction of hydrogen in molecular form, $\sim 0.8$ and $\sim 0.95$ for the -190 and -150  velocity components, respectively. Using the analytical framework for the H I/H$_2$ transition developed by [@sternberg:14] and [@bialy:16], we derived the parameter $\frac{I_{\rm UV}}{n_{100}}$ where $I_{\rm UV}$ is the intensity of the FUV radiation field in units of the Draine field and $n_{100}=\frac{n(H)}{100 \rm cm^{-3}}$ is the gas density in units of 100 H-nuclei $\pccc$. We find $\frac{I_{\rm UV}}{n_{100}} \sim 0.5$ and $\sim 0.2$ for these two clouds, indicating that they are exposed to a low far-ultraviolet (FUV) radiation field. We note that the density must be modest, n$_{100} \leq 1$, since no significant CO emission is detected. The theory indicates that [[[$\mathrm{H_2}$]{}]{}]{} is mostly protected by self-shielding in such an environment, with extinction contributed by dust grains providing a minor contribution to the protection of H$_2$ against photodissociation. It is known from analysis of the spatial distribution of star-forming regions in the inner galaxy [e.g., @churchwell:09] that star formation is weak in the Galactic bulge, possibly as a consequence of the non-circular motions induced by the bar. The old stellar population of the bulge is expected to produce a radiation field dominated by visible and near-IR radiation, which can heat the dust [@robitaille:12] but is less effective at dissociating molecular hydrogen. The diffuse gas in the bulge therefore follows the general trend of matter in the disk, which has an increasing fraction of hydrogen in molecular form at smaller , more than 80% at $\approx$ 2kpc [@koda:16]. Influence of metallicity variation ---------------------------------- Of course the tabulated values are affected by uncertainty in the metallicity in the bulge gas and disk gas, and an increased metallicity, if reflected in a higher [[[$\mathrm{HCO^+}$]{}]{}]{}  relative abundance and dust/gas ratio, would lead to smaller N([[[$\mathrm{H_2}$]{}]{}]{}) inferred from [[[$\mathrm{HCO^+}$]{}]{}]{} and N(H) inferred from extinction. A factor three increase in N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) would yield nearly equal atomic and molecular gas columns at v = -190 , but no reasonable increase in metallicity would equalize the [[[$\mathrm{H_2}$]{}]{}]{}/H I ratio at v = -150 , which is 37. We note that [@SchChe+14], from whose work we drew the variation of E(J-K), also used a constant gas-reddening ratio to derive N(H) and [[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{}. Another estimate of the total atomic gas deficit ------------------------------------------------ The more recent value of the IRAS dust-inferred optical reddening  in the direction of B1741 at 6 resolution is  = 6.3 mag [@SchFin11], which is somewhat beyond the stated range of validity of this estimator. Converting this into total column density with a recent value of the N(H)/ ratio [@Lis14EBVI; @Lis14EBVII; @HenDra17] yields a high total column density N(H) $= 6.3 \times 8 \times 10^{21} = 5.0 \times 10^{22}$ H-nuclei $\pcc$. The implied gas deficit N(H) - N(H I) $\approx 3.5 \times 10^{22}$ H-nuclei $\pcc$ is much higher than can be explained by either the [[[$\mathrm{HCO^+}$]{}]{}]{} absorption or the CO emission. ![Variation of LSR velocity with heliocentric distance (left) and galactocentric radius (right) for circular and non-circular velocity fields in the direction of B1741-312 at $l=-2.14$, $b=-1.0$. The curves labeled “Galactic rotation” are for a flat rotation curve with R$_0$ = 8.5 kpc, $\Theta = 220 $ , and the non-circular velocity field interior to  = 1.5 kpc is taken from [@BurLis78]. The diagrams were constructed by taking equal steps in heliocentric distance, and the density of points represents the line-of-sight geometry and velocity gradients.](B1741-NewFig4.eps){height="7.5cm"} ![Variation in cumulative hydrogen column densities N(H I) and N(H) = N(H I) + 2N([[[$\mathrm{H_2}$]{}]{}]{}) vs. galactocentric radius. The solid blue curve for the total column density N(H) was derived by inverting the variation of E(J-K) with distance modulus in the 3D extinction data cube of [@SchChe+14], as discussed in Section 2.3. The binned estimates for N(H) from H I and [[[$\mathrm{HCO^+}$]{}]{}]{} or H I and CO assume N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) $= 3 \times 10^{-9}$ or N([[[$\mathrm{H_2}$]{}]{}]{}) $= 2 \times 10^{20}\pcc$[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{}. The bins correspond to the ranges of galactocentric radius listed in Table 1.](B1741-NewFig5.eps){height="7.5cm"} ------------------------------------------------------------ ------------------------------ ---------------------------- --------------------------- --------------------------- --------------------------- Velocity Range ( ) Quantity -196,-180 -164,-135 -41,-15 -15,-1 -1, 19  \[pc\] 320 - 1500 320 - 1500 1500-3035 3035-8500 3035-8500 W$_{\rm H I}$ \[K \] 79.9(3.1) 32.8(3.8) 1062(6.1) 1524(7.6) 2622(10.4) W$_{{{{\ensuremath{\mathrm{HCO^+}}}}}}$ \[ \] 0.802(0.025) 3.381(0.036) 1.108 (0.031) 9.10(0.03) 12.22(0.32) W$_{\rm HCN}$ \[ \] 0.640(0.027) 1.310(0.036) 0.444(0.032) 8.62(0.03) 9.65(0.03) W$_{\rm HNC}$ \[ \] 0.333(0.034) 0.062(0.031) 1.42(0.03) 1.72 (0.03) [[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{}  \[K \] 0 (0.44) 1.40(0.69) 5.55(0.50) 14.53(0.42) 22.37(0.49) N(H I) \[$\pcc$\] $1.51(0.06)\times10^{20}$ $6.13(0.71)\times10^{19}$ $2.77(0.02)\times10^{21}$ $4.05(0.02)\times10^{21}$ $6.16(0.02)\times10^{21}$ N([[[$\mathrm{HCO^+}$]{}]{}]{}) \[$\pcc$\] $8.58(0.26)\times10^{11}$ $ 3.62(0.03)\times10^{12}$ $1.18(0.03)\times10^{12}$ $9.74(0.28)\times10^{12}$ $1.31(0.03)\times10^{13}$ N([[[$\mathrm{H_2}$]{}]{}]{}) \[$\pcc$\] $2.86\times10^{20}$ $1.21\times10^{21}$ $3.92 \times10^{20}$ $ 3.25\times10^{21} $ $ 4.36 \times10^{21}$ X$_{\rm CO}$\[$\pcc$/K- \] $>$ 2.17 $\times10^{20}$$^b$ $8.64\times10^{20}$ $7.06\times10^{19}$ $2.24\times10^{20}$ $ 1.95\times10^{20}$ ------------------------------------------------------------ ------------------------------ ---------------------------- --------------------------- --------------------------- --------------------------- \ $^a$ See Sect. 2.3 for a description of table entries\ $^b$ Using the $3\sigma$ upper limit for [[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{} \ Summary ======= We began by noting the accidental observation of mm-wave [[[$\mathrm{HCO^+}$]{}]{}]{}, HCN and HNC absorption at -150  and -190  toward the extragalactic mm-wave continuum background source B1741-312 seen at l = -2.1366, b = -0.9968 (Fig. 1). There is weak CO emission in the vicinity of the position of B1741 (Figs. 2 and 3), but prominent CO emission at such velocities is confined to smaller longitudes and latitudes nearer the Galactic equator. Emission and absorption at velocities v $\la -150$  cannot be precisely located along the line of sight owing to imprecise knowledge of the bar-like inner-galaxy velocity field, but must arise in gas located within the Galactic bulge (Fig. 4). Similarly, strong disk absorption and emission at -40  $\le$ v $\le$ 0  can be understood in terms of a flat disk rotation velocity curve with R$_0$ = 8.5 kpc, $\Theta$ = 220 , but the bright emitting gas at 0  $\le$ v $\le$ 20  is not permitted by rotational kinematics and cannot be precisely located either, although we know that it must be nearby because it is self-absorbed in H I. The atomic and molecular absorption profiles were interpreted on this basis in Table 1, where we assigned broad ranges in galactocentric radius to the kinematic intervals containing the observed molecular features. We derived column densities N(H I), N([[[$\mathrm{HCO^+}$]{}]{}]{}), and N([[[$\mathrm{H_2}$]{}]{}]{}) assuming as a baseline the locally-determined value N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) $ = 3\times 10^{-9}$. This [[[$\mathrm{HCO^+}$]{}]{}]{}-based estimate of N([[[$\mathrm{H_2}$]{}]{}]{}) yields canonical values of the CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor N([[[$\mathrm{H_2}$]{}]{}]{})/[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{} $ = 2.0-2.2 \times 10^{20} {{{\ensuremath{\mathrm{H_2}}}}}\ \pcc$ (K-)$^{-1}$ for either of the velocity intervals around 0  in Table 4. Broadly speaking, the balance between the atomic and molecular gas constituents shifts from approximate equality in the disk to a mostly molecular medium within the bulge. Figure 5 shows the galactocentric radial distribution of total column density N(H) derived by inverting the run of IR reddening E(J-K) with distance modulus in the 3D extinction data cube of [@SchChe+14] using a fixed relationship N(H) $= 1.171\times 10^{22}\pcc$ E(J-K) as described in Sect. 2.3. H I can account for only a small portion of the required gas column inside the bulge at  $< 1.5$ kpc. The gas shortfall can be explained by the observed [[[$\mathrm{HCO^+}$]{}]{}]{} absorption using the locally determined fractional abundance N([[[$\mathrm{HCO^+}$]{}]{}]{})/N([[[$\mathrm{H_2}$]{}]{}]{}) $= 3\times 10^{-9}$ , but CO emission is weak or absent and would require a CO-[[[$\mathrm{H_2}$]{}]{}]{} conversion factor $\approx 10^{21}$ [[[$\mathrm{H_2}$]{}]{}]{} $\pcc$ (K-)$^{-1}$. This stands in marked contrast to the overluminous CO emission from the central molecular zone where the N([[[$\mathrm{H_2}$]{}]{}]{})/[[[${\rm W}_{{\ensuremath{\mathrm{CO}}}}$]{}]{}]{} ratio is believed to be 3-10 times lower than the value near the Sun. The physical properties and total mass of the molecular gas in the bulge remain to be determined by observation. We suggested that the atomic-molecular transition between the disk and bulge results from a weak UV radiation field inside the bulge with its older stellar population, encouraging [[[$\mathrm{H_2}$]{}]{}]{} formation at relatively low number and column densities in otherwise atomic gas. Suppression of CO emission at brightness levels of 1-2 K (typical of existing survey detection limits) would require densities that are characteristic of mostly atomic H I clouds in the disk, n(H) $\la 30-50 \pccc$. Information on the in situ number density in the absorbing molecular gas can be derived by observing [[[$\mathrm{HCO^+}$]{}]{}]{}, HCN, or HNC in emission around the position of B1741-312, given the column densities measured in absorption and the fact that the emission brightness at the limit of detectability varies only as the n(H)-N(X) product [@LisPet16]. ALMA Cycle 4 observations of a small handful of other suitably strong mm-wave continuum sources are underway, but none are as fortuitously positioned at such small longtitude and latitude as B1741-312. We note that even if the sort of detailed discussion presented here cannot currently be extended to other sightlines, existing 3D measurements of the reddening can be compared to H I emission at bulge velocities to determine whether the atomic gas deficit seen here is a general phenomenon, as we would predict. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2013.1.01194.S . ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan), together with NRC (Canada), NSC and ASIAA (Taiwan), and KASI (Republic of Korea), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO and NAOJ. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This work was supported by the French program “Physique et Chimie du Milieu Interstellaire” (PCMI) funded by the Conseil National de la Recherche Scientifique (CNRS) and Centre National d’Etudes Spatiales (CNES). We thank Edwige Chapillon and Philippe Salomé for help with the ALMA data processing and the referee for helpful comments. HSL is grateful to the ENS and FREI for their hospitality during the completion of this manuscript. [28]{} natexlab\#1[\#1]{} , S. & [Sternberg]{}, A. 2016, Astroph. J., 822, 83 , M., [Alvarez]{}, H., [Bronfman]{}, L., [May]{}, J., & [Thaddeus]{}, P. 1997, Astrophys. J., Suppl. Ser., 125, 99 , A. D., [Wolfire]{}, M., & [Leroy]{}, A. K. 2013, Ann. Rev. Astrophys. Astron., 51, 207 , W. B. & [Liszt]{}, H. S. 1978, Astroph. J., 225, 815 —. 1983, A&A Supp., 52, 63 , E., [Babler]{}, B. L., [Meade]{}, M. R., [Whitney]{}, B. A., [Benjamin]{}, R., [Indebetouw]{}, R., [Cyganowski]{}, C., [Robitaille]{}, T. P., [Povich]{}, M., [Watson]{}, C., & [Bracker]{}, S. 2009, Publ. Astron. Soc. Pac., 121, 213 , T. M. & [Thaddeus]{}, P. 2008, Astroph. J., 683, L143 , J. M., [Kulkarni]{}, S. R., [Heiles]{}, C. E., & [Van Gorkom]{}, J. H. 1983, Astrophys. J., Suppl. Ser., 53, 591 , R. 1999, A&A, 345, 787 , B., [Falgarone]{}, E., [Gerin]{}, M., [Hily-Blant]{}, P., & [de Luca]{}, M. 2010, A&A, 520, A20 , D. S. 1955, Astroph. J., 121, 569 , B. S. & [Draine]{}, B. T. 2016, ArXiv e-prints , [Ben Bekhti]{}, N., [Fl[ö]{}er]{}, L., [Keller]{}, R., [Kerp]{}, J., [Lenz]{}, D., [Winkel]{}, B., [Bailin]{}, J., [Calabretta]{}, M. R., [Dedes]{}, L., [Ford]{}, H. A., [Gibson]{}, B. K., [Haud]{}, U., [Janowiecki]{}, S., [Kalberla]{}, P. M. W., [Lockman]{}, F. J., [McClure-Griffiths]{}, N. M., [Murphy]{}, T., [Nakanishi]{}, H., [Pisano]{}, D. J., & [Staveley-Smith]{}, L. 2016, A&A, 594, A116 , J., [Scoville]{}, N., & [Heyer]{}, M. 2016, Astroph. J., 823, 76 , H. 2014, Astroph. J., 783, 17 —. 2014, Astroph. J., 780, 10 , H. & [Lucas]{}, R. 2001, A&A, 370, 576 , H. S. 2006, A&A, 447, 533 —. 2008, A&A, 486, 467 , H. S. & [Burton]{}, W. B. 1978, Astroph. J., 226, 790 , H. S. & [Pety]{}, J. 2016, Astroph. J., 823, 124 , N. M., [Dickey]{}, J. M., [Gaensler]{}, B. M., [Green]{}, A. J., [Haverkorn]{}, M., & [Strasser]{}, S. 2005, Astrophys. J., Suppl. Ser., 158, 178 , M. & [Serabyn]{}, E. 1996, Ann. Rev. Astrophys. Astron., 34, 645 , T. P., [Churchwell]{}, E., [Benjamin]{}, R. A., [Whitney]{}, B. A., [Wood]{}, K., [Babler]{}, B. L., & [Meade]{}, M. R. 2012, A&A, 545, A39 , N. J., [Combes]{}, F., [Martin-Pintado]{}, J., [Wilson]{}, T. L., & [Apponi]{}, A. 2006, A&A, 455, 963 , E. F. & [Finkbeiner]{}, D. P. 2011, Astroph. J., 737, 103 , M., [Chen]{}, B. Q., [Jiang]{}, B. W., [Gonzalez]{}, O. A., [Enokiya]{}, R., [Fukui]{}, Y., [Torii]{}, K., [Rejkuba]{}, M., & [Minniti]{}, D. 2014, A&A, 566, A120 , A., [Le Petit]{}, F., [Roueff]{}, E., & [Le Bourlot]{}, J. 2014, Astroph. J., 790, 10 [^1]: CO emission at -54  that is centered on the Galactic equator arises in the so-called near 3 kpc arm [@DamTha08] and does not share the tilted geometry.

--- abstract: 'Theory of expectation values is presented as an alternative to S-matrix theory for quantum fields. This change of emphasis is conditioned by a transition from the accelerator physics to astrophysics and cosmology. The issues discussed are the time-loop formalism, the Schwinger–Keldysh diagrams, the effective action, the vacuum currents, and the effect of particle creation.' author: - 'G.A. Vilkovisky' title: | Expectation Values and Vacuum Currents\ of Quantum Fields[^1] --- Introduction {#introduction .unnumbered} ============ High-energy physics will probably have to undergo major changes. The accelerators will cease being its experimental base, and it will become a part of astrophysics. Simultaneously, the S-matrix will cease being the central object of high-energy theory because the emphasis on this object is entirely owing to the accelerator setting of the problem. If there is a background radiation that originates from some initial state in the past, then where is the S-matrix here? Astrophysics and cosmology offer the evolution problems rather than the scattering problems. The gravitational collapse is a typical initial-value problem. It is such by its physical setting irrespective of whether the state of the system is classical or quantum. The nature of measurement also changes. No final state is prepared. One measures observables like temperatures or mechanical deflections and subjects these measurements to a statistical treatment to obtain the value of the observable. This means that one measures expectation values in the given initial state. S-matrix theory should give way to expectation-value theory. There is a proof that accelerator physics is dead: Gabriele Veneziano is leaving CERN for Collège de France. At this historic moment, my mission is to convert him into a new faith. The present preaching consists of 4 lectures: 1. Formal aspects of expectation-value theory. 2. The in-vacuum state and Schwinger–Keldysh diagrams. 3. The effective action. 4. Vacuum currents and the effect of particle creation. Literature to Lectures 1 and 2 is in [@1]–[@16]. Additional literature to Lecture 3 is in [@17]–[@41] and to Lecture 4 in [@42]–[@56]. lecture1.tex lecture2.tex lecture3.tex lecture4.tex [99.]{} Bryce DeWitt: *The Global Approach to Quantum Field Theory*, vols 1,2 (Oxford University Press, Oxford New York 2003) B.S. DeWitt: Dynamical theory of groups and fields. In: *Relativity, Groups and Topology. 1963 Les Houches Lectures*, ed by C. DeWitt, B.S. DeWitt (Gordon and Breach, New York 1964) pp 587–820 G. Jona-Lasinio: Nuovo Cimento **34**, 1790 (1964) B.S. DeWitt: Phys. Rep. **19**, 295 (1975) E.S. Fradkin, G.A. Vilkovisky: Lett. Nuovo Cimento **19**, 47 (1977) J. Schwinger: Field theory methods in non-field theory contexts. In: *Proc. 1960 Brandeis Summer School* (Brandeis University Press, Brandeis 1960) pp 282–285 J. Schwinger: J. Math. Phys. **2**, 407 (1961) L.V. Keldysh: Zh. Eksp. Teor. Fiz. **47**, 1515 (1964) Yu.A. Golfand: Yad. Fiz. **8**, 600 (1968) P. Hajicek: Time-loop formalism in quantum field theory. In: *Proc. 2nd Marcel Grossmann Meeting on General Relativity (Trieste, 1979)*, ed by R. Ruffini (North Holland, Amsterdam 1982) pp 483–491 E.S. Fradkin, D.M. Gitman: Fortschr. der Phys. **29**, 381 (1981) J.L. Buchbinder, E.S. Fradkin, D.M.Gitman: Fortschr. der Phys. **29**, 187 (1981) R.D. Jordan: Phys. Rev. D **33**, 44 (1986) E. Calzetta, B.L. Hu: Phys. Rev. D **35**, 495 (1987) A.O. Barvinsky, G.A. Vilkovisky: Nucl. Phys. B **282**, 163 (1987) R.C. Hwa, V.L. Teplitz: *Homology and Feynman Integrals* (Benjamin, New York Amsterdam 1966) G.A. Vilkovisky: Class. Quantum Grav. **9**, 895 (1992) J.S. Schwinger: Phys. Rev. **82**, 664 (1951) J.L. Synge: *Relativity: The General Theory* (North Holland, Amsterdam 1960) G. ’t Hooft, M. Veltman: Ann. Inst. Henri Poincare **XX**, 69 (1974) P.B. Gilkey: J. Diff. Geom. **10**, 601 (1975) L.S. Brown: Phys. Rev. D **15**, 1469 (1977) L.S. Brown, J.P. Cassidy: Phys. Rev. D **15**, 2810 (1977) A.O. Barvinsky, G.A. Vilkovisky: Phys. Rep. **119**, 1 (1985) G.A. Vilkovisky: Heat kernel: rencontre entre physiciens et mathématiciens. In: *R.C.P. 25*, vol 43 (Publication de l’Institut de Recherche Mathématique Avancée, Strasbourg 1992) pp 203–224 A.M. Polyakov: Phys. Lett. B **103**, 207 (1981) G.A. Vilkovisky: The Gospel according to DeWitt. In: *Quantum Theory of Gravity*, ed by S.M. Christensen (Hilger, Bristol 1984) pp 169–209 A.A. Ostrovsky, G.A. Vilkovisky: J. Math. Phys. **29**, 702 (1988) I.G. Avramidi: Yad. Fiz. **49**, 1185 (1989) A.O. Barvinsky, G.A. Vilkovisky: Nucl. Phys. B **333**, 471 (1990) A.O. Barvinsky, G.A. Vilkovisky: Nucl. Phys. B **333**, 512 (1990) A.O. Barvinsky, Yu.V. Gusev, G.A. Vilkovisky, V.V. Zhytnikov: J. Math. Phys. **35**, 3525 (1994) A.O. Barvinsky, Yu.V. Gusev, G.A. Vilkovisky, V.V. Zhytnikov: J. Math. Phys. **35**, 3543 (1994) A.O. Barvinsky, Yu.V. Gusev, G.A. Vilkovisky, V.V. Zhytnikov: Nucl. Phys. B **439**, 561 (1995) A.O. Barvinsky, Yu.V. Gusev, V.V. Zhytnikov, G.A. Vilkovisky: Class. Quantum Grav. **12**, 2157 (1995) A.O. Barvinsky, Yu.V. Gusev, G.A. Vilkovisky, V.V. Zhytnikov: Covariant perturbation theory (IV). Third order in the curvature. Report, University of Manitoba, Winnipeg (1993) pp 1–192 A.G. Mirzabekian, G.A. Vilkovisky, V.V. Zhytnikov: Phys. Lett. B **369**, 215 (1996) Y. Nambu: Phys. Rev. **100**, 394 (1955) N. Nakanishi: Prog. Theor. Phys. **24**, 1275 (1960) N. Nakanishi: *Graph Theory and Feynman Integrals* (Gordon and Breach, New York 1970) J. Schwinger: *Particles, Sources, and Fields*, vol 2 (Addison-Wesley, Reading 1973) A.A. Grib, S.G. Mamayev, V.M. Mostepanenko: *Quantum Effects in Intense External Fields* (Atomizdat, Moscow 1980) N.D. Birrell, P.C.W. Davies: *Quantum Fields in Curved Space* (Cambridge University Press, Cambridge 1982) N.M.J. Woodhouse: Phys. Rev. Lett. **36**, 999 (1976) A.G. Mirzabekian, G.A. Vilkovisky: Phys. Lett. B **317**, 517 (1993) A.G. Mirzabekian: Zh. Eksp. Teor. Fiz. **106**, 5 (1994) \[Engl. trans.: JETP **79**, 1 (1994)\] A.G. Mirzabekian, G.A. Vilkovisky: Phys. Rev. Lett. **75**, 3974 (1995) A.G. Mirzabekian, G.A. Vilkovisky: Class. Quantum Grav. **12**, 2173 (1995) A.G. Mirzabekian, G.A. Vilkovisky: Phys. Lett. B **414**, 123 (1997) A.G. Mirzabekian, G.A. Vilkovisky: Ann. Phys. **270**, 391 (1998) G.A. Vilkovisky: Phys. Rev. D **60**, 065012 (1999) G.A. Vilkovisky: Phys. Rev. Lett. **83**, 2297 (1999) R. Pettorino, G.A. Vilkovisky: Ann. Phys. **292**, 107 (2001) G.A. Vilkovisky: Ann. Phys. **321**, 2717 (2006) G.A. Vilkovisky: Phys. Lett. B **634**, 456 (2006) G.A. Vilkovisky: Phys. Lett. B **638**, 523 (2006) [^1]: The course of 4 lectures given at Collège de France in May 2006.

==1 [fullgraphs]{} `HU-MATH-2010-23`, `HU-EP-10/88`, `arxiv:1012.3984`\ overview article: `arxiv:1012.3982` {width="5cm"} #### Abstract: We review the constructions and tests of the dilatation operator and of the spectrum of composite operators in the flavour $SU(2)$ subsector of $\mathcal{N}=4$ SYM in the planar limit by explicit Feynman graph calculations with emphasis on analyses beyond one loop. From four loops on, the dilatation operator determines the spectrum only in the asymptotic regime, i.e. to a loop order which is strictly smaller than the number of elementary fields of the composite operators. We review also the calculations which take a first step beyond this limitation by including the leading wrapping corrections. #### Mathematics Subject Classification (2010): 81T18; 81T13; 81T60 #### Keywords: Feynman diagrams; Supersymmetric Yang-Mills theory; Superspace; Integrability; Introduction {#sec:introduction} ============ In the context of the ${\text{AdS}}/{\text{CFT}}$ correspondence [@Maldacena:1997re; @Gubser:1998bc; @Witten:1998qj], the discovery of integrability is a key ingredient towards finding the exact spectrum of strings in ${\text{AdS}}_5\times\text{S}^5$ and of composite operators in $\mathcal{N}=4$ SYM theory with gauge group $SU(N)$ in the planar limit, i.e. for $N\to\infty$. As reviewed in chapters [@chapSpinning] and [@chapQstring], on the string side of the duality the spectrum is accessible order by order as a strong coupling expansion in terms of the ’t Hooft coupling by a (semi)classical analysis of string states with large quantized charges. It is also described in terms of respective string Bethe ansätze which are reviewed in chapter [@chapABA]. In the $\mathcal{N}=4$ SYM theory, the weak coupling expansion of the planar spectrum, i.e. the conformal dimensions of composite operators, can be obtained by direct perturbative calculations of various correlation functions. The appearance of UV divergences requires renormalization, which then leads to a mixing among operators with the same bare conformal dimension. The eigenvalues of the new eigenstates under conformal rescalings are given as the sum of the bare scaling dimension and an individual anomalous dimension. The operator mixing can be extracted, e.g. from the correlation functions involving two composite operators. Alternatively, one can directly calculate the diagrams which contribute to the renormalization of these operators. This directly allows one to obtain an expression for the dilatation operator, whose eigenvalues are the anomalous dimensions. Perturbative calculations become very cumbersome at high loop orders and can be avoided, if the observed integrability at one loop, which is reviewed in chapter [@chapChain], also persists to higher loop orders. The dilatation operator can then be determined, using some very general structural information from the underlying Feynman graphs only and some data from the gauge Bethe ansätze. The details of this approach are reviewed in chapter [@chapLR]. Direct Feynman graph calculations of the dilatation operator in the flavour $SU(2)$ subsector to three loops and of some of its eigenvalues and of parts of the Bethe ansätze also to higher loops provide important checks for the assumed integrability. Even if integrability holds to all loop orders, the respective Bethe ansätze and planar dilatation operator allow us to compute the anomalous dimensions only in the asymptotic regime. In this regime, the loop order of the result is constrained to be strictly smaller than the length (the number of elementary fields) of the shortest composite operator involved. At loop orders which are equal to or exceed this number, the so-called wrapping interactions [@Serban:2004jf; @Beisert:2004hm] have to be considered. They are corrections due to the finite size of the composite operators and have their origin in the neglected higher genus contributions to the dilatation operator [@Sieg:2005kd]. In the dual string theory the counterparts of the wrapping interactions are corrections due to the finite circumference of the closed string worldsheet cylinder [@Ambjorn:2005wa]. Their analyses are reviewed in chapters [@chapLuescher] and [@chapTBA]. In this chapter we review the explicit Feynman graph calculations in $\mathcal{N}=4$ SYM theory in the planar limit beyond one loop. It is organized as follows: In Section \[sec:opreno\] we give a short summary of how composite operators are renormalized, and how the dilatation operator is defined in terms of the renormalization constants. In Section \[sec:Dop\] we then review the explicit calculations and tests of the dilatation operator with particular focus on calculations beyond the first order in perturbation theory.[^1] Only the flavour $SU(2)$ subsector will be considered, since most higher loop calculations are performed within this subsector. As examples we recalculate in detail the respective one- and two-loop dilatation operator in $\mathcal{N}=1$ superfield formalism. This approach is much more efficient than the originally used formalism without manifest supersymmetries, and it yields more direct relations between the dilatation operator and the underlying Feynman graphs. We then display the result of a three-loop calculatoin and also summarize the existing checks of the magnon dispersion relation, of the structure of the dilatation operator and of some of its eigenvalues in the asymptotic regime at three and higher loops. In Section \[sec:wrapint\], we review the perturbative calculations which consider the first wrapping corrections and hence yield results beyond the asymptotic regime. The general strategy of these calculations will be explained. In this way, the four-loop anomalous dimension for the length four Konishi descendant in the flavour $SU(2)$ subsector could be determined. Further results for different operators and for the terms of highest transcendentality are then summarized briefly. In Section \[sec:concl\] we give a concluding summary, and in two appendices we present the explicit $\operatorname{D}$-algebra manipulations for the one- and two-loop calculation and the expressions for the relevant integrals. Renormalization of composite operators {#sec:opreno} ====================================== The dilatation operator and anomalous dimensions can be obtained from a perturbative calculation of the correlation functions which involve the composite operators $\mathcal{O}_a$, where $a$ labels the different operators. The encountered UV divergences require a renormalization of the composite operators as $$\label{opren} \mathcal{O}_{a,\text{ren}}(\phi_{i,\text{ren}}) =\mathcal{Z}_{a}{}^b(\lambda,\varepsilon)\mathcal{O}_{b,\text{bare}}(\phi_{i,\text{bare}}) {~,}\qquad \phi_{i,\text{ren}}=\mathcal{Z}_{i}^{1/2}\phi_{i,\text{bare}} {~,}$$ where in an appropriate basis $\mathcal{Z}={\mathds{1}}+\delta\mathcal{Z}$, and the matrix $\delta\mathcal{Z}$ is of order $\mathcal{O}(\lambda)$ in the renormalized coupling constant $\lambda$. It also depends on the regulator $\varepsilon$ and is in general non-diagonal and thus leads to mixing between the different composite operators. The matrix element $\delta\mathcal{Z}_{a}{}^b$ is given by the negative of the sum of the overall UV divergences of the Feynman diagrams in which the vertices of the theory lead to interactions between the elementary fields of operator $\mathcal{O}_b$, such that the resulting external field flavour and ordering coincide with the ones of the operator $\mathcal{O}_a$. One also has to consider contributions from wave function renormalization of the elementary fields $\phi_i$ the operators are composed of. Respective factors $\mathcal{Z}_{i}^{1/2}$ are included within $\mathcal{Z}$. $\mathcal{N}=4$ SYM theory can be regularized by supersymmetric dimensional reduction [@Siegel:1979wq] in $D=4-2\varepsilon$ dimensions. The coupling constant $g_{\text{YM}}$ is then accompanied by the ’t Hooft mass $\mu$ in the combination $g_{\text{YM}}\mu^{\varepsilon}$ to restore the mass dimension of the loop integrals. Thereby, $g_{\text{YM}}$ is not renormalized and hence itself does not depend on $\mu$, such that superconformal invariance is preserved. This was explicitly found to three loops by computing the vanishing of the $\beta$-function in an $\mathcal{N}=1$ superfield formulation [@Grisaru:1980jc; @Grisaru:1980nk; @Caswell:1980yi; @Caswell:1980ru]. The finiteness of $\mathcal{N}=4$ SYM theory was then later shown to all orders [@Mandelstam:1982cb; @Brink:1982wv; @Howe:1982tm; @Howe:1983sr]. A first argument was given in [@Ferrara:1974pu]. In particular, the self-energy of the superfields is finite, i.e.  $\mathcal{Z}_{i}^{1/2}$ is trivial.[^2] In the planar limit, where the coupling constant is $\lambda=g_{\text{YM}}^2N$, the dilatation operator is then extracted from the renormalization constant of the composite operators in as $$\label{DinZ} \mathcal{D} =\mu\frac{{\operatorname{d}\!}}{{\operatorname{d}\!}\mu}\ln\mathcal{Z}(\lambda\mu^{2\varepsilon},\varepsilon) =\lim_{\varepsilon\rightarrow0}\left[2\varepsilon\lambda \frac{{\operatorname{d}\!}}{{\operatorname{d}\!}\lambda}\ln\mathcal{Z}(\lambda,\varepsilon)\right] {~.}$$ The logarithm of $\mathcal{Z}={\mathds{1}}+\delta\mathcal{Z}$ has to be understood as a formal series in powers of $\delta\mathcal{Z}$. All poles of higher order in $\varepsilon$ must cancel in $\ln\mathcal{Z}$, such that it only contains simple $\frac{1}{\varepsilon}$ poles. In effect, the above description extracts the coefficient of the $\frac{1}{\varepsilon}$ pole of $\mathcal{Z}$, and at a given loop order $K$ multiplies it by a factor $2K$. This then yields the dilatation operator as a power series $$\label{Dex} \mathcal{D}=\sum_{k\ge1}g^{2k}\mathcal{D}_k {~,}\qquad g=\frac{\sqrt{\lambda}}{4\pi} {~,}$$ where for later convenience we have absorbed powers of $4\pi$ into the definition of a new coupling constant $g$. Dilatation operator in the SU(2) subsector {#sec:Dop} ========================================== $\mathcal{N}=4$ SYM theory contains six real scalar fields, four complex Weyl fermions and a gauge field that all transform in the adjoint representation of the gauge group $SU(N)$. In the following we denote these fields as components fields, since in a superspace formalism they appear as components of superfields. In order to build the $\mathcal{N}=1$ superfields, the real scalar component fields are complexified and combined together each with one fermion or with its complex conjugate into three chiral superfields $\phi_i$, $i=1,2,3$ or respectively anti-chiral ones $\bar\phi_i$. The three field flavours are transformed into each other by an $SU(3)$ subgroup of the $SU(4)$ R-symmetry group. The remaining gauge field and fermions are combined together into an $\mathcal{N}=1$ vector superfield $V$. An explicit expression of the $\mathcal{N}=4$ SYM action in terms of $\mathcal{N}=1$ superfields and the respective Feynman rules in which the Wick rotation is included can be found, e.g. in [@Sieg:2010tz]. The superspace conventions are as in [@Gates:1983nr], where also an introduction to the $\operatorname{D}$-algebra is given. The latter is required to reduce the supergraphs, i.e. the Feynman diagrams in superspace, to ordinary spacetime objects that are located at a single point in the fermionic coordinates of superspace. Operator mixing in the SU(2) subsector {#subsec:fsubsec} -------------------------------------- In the following, we denote the three chiral field flavours of $\mathcal{N}=4$ SYM theory by $\phi_i=(\phi,\psi,Z)$. The flavour $SU(2)$ subsector contains operators which are composed of only two different types of these fields, e.g. $\phi$ and $Z$. Their color indices are all contracted with each other to yield a gauge invariant object. In general, the gauge contractions form several cycles, and one obtains a multi-trace operator. Such an operator is a normal-ordered product of single-trace operators, i.e. of operators each of which only contains a single cycle of gauge contractions. Mixing only occurs between those operators that have the same numbers of both types of fields $\phi$ and $Z$. Then, it suffices to consider operators which contain a number of fields $\phi$ that does not exceed the number of fields $Z$, since the results for the remaining operators follow immediately by an exchange of the role of the two fields. Usually, the fields $\phi$ are denoted as impurities which appear between fields of type $Z$ within the traces over the gauge group. Furthermore, in the planar limit that we exclusively consider from now on,[^3] the Feynman diagrams that alter the gauge trace structure of the composite operators are suppressed. The renormalization of multi-trace operators then follows immediately from the one of their single-trace constituents. We can therefore restrict the analysis to single-trace operators. In this case, the planar Feynman diagrams can only affect the ordering of the two different types of fields inside the single trace, but they cannot alter their multiplicities and in particular the length $L$ of the composite operators that is defined as the total number of constituent fields. Flavour contractions cannot appear, since the composite operators of the $SU(2)$ subsector do not contain the complex conjugate fields $(\bar\phi,\bar\psi,\bar Z)$. The $SU(2)$ subsector is closed under renormalization, at least perturbatively [@Minahan:2005jq]. The operators $$\label{groneimpstate} \operatorname{tr}\big(Z^L\big){~,}\qquad\operatorname{tr}\big(\phi Z^{L-1}\big)$$ which are the ground state and a state with a single impurity are protected and do not acquire anomalous dimensions. Operators which contain more than a single impurity $\phi$ undergo non-trivial mixing. Since the aforementioned operator mixing only occurs within subsets of single-trace operators that only differ by permutations of their field content, the renormalization constant $\mathcal{Z}$ and hence also the dilatation operator $\mathcal{D}$ can be expressed in terms of flavour permutations that act on the constituent fields of these composite operators. The flavour permutations themselves can be written as products of permutations acting on nearest neighbour sites. For composite operators of fixed length $L$ they are given by [@Beisert:2003tq] $$\label{permstrucdef} {{}\{a_1,\dots,a_n\}{}}=\sum_{r=0}^{L-1}\operatorname{P}_{a_1+r\;a_1+r+1}\cdots \operatorname{P}_{a_n+r\;a_n+r+1}$$ and by the identity ${{}\{\}{}}$ in flavour space that measures the length $L$ of the composite operator it is applied to. The structures consider the insertion of the Feynman subdiagrams in which elementary fields interact at all possible positions within the single trace of the composite operator by the summation. Periodicity with period $L$ is thereby understood. No other insertions have to be considered here, since in the planar limit the interactions have to occur between adjacent fields. The permutation structures admit a definition of the range of the interaction in flavour space obtained from their lists of arguments as $$\label{nneighbourint} \kappa=\max_{a_1,\dots, a_n}-\min_{a_1,\dots, a_n}+2{~.}$$ The range $\kappa$ and hence also the possible arguments $a_1,\dots,a_n$ of the permutation structures are subject to constraints from the underlying Feynman diagrams. In order to find the restrictions for those structures that can appear in the expression of the dilatation operator, we focus on Feynman diagrams in which the elementary interactions occur in a single region that is simply connected also when the composite operator is removed from the diagram. These diagrams may have overall UV divergences that contribute with simple $\frac{1}{\varepsilon}$ poles to the renormalization constant $\mathcal{Z}$ and hence according to also to the dilatation operator. The remaining diagrams, in which the elementary interactions occur in several non simply-connected regions after the removal of the composite operator, cannot contribute with simple $\frac{1}{\varepsilon}$ poles. Their calculation is only required if one wants to determine $\mathcal{Z}$ itself completely, for example in order to check explicitly that in $\ln\mathcal{Z}$ all higher order poles in $\varepsilon$ cancel. Here, we will not consider them further and only focus on the diagrams that can contribute to the dilatation operator. The interaction range $R$ of a diagram of the latter type is defined as the number of adjacent elementary fields of the composite operator that enter the single simply connected interaction region. It can only yield contributions with permutation structures that obey the following conditions: $$\label{permbounds} n\le K{~,}\qquad \kappa\le R{~,}\qquad R\le K+1 {~,}$$ where $K$ denotes the number of loops inside the diagram. The first inequality considers that each nearest-neighbour permutation is associated with at least one loop. The second condition ensures that the range of the interaction in flavour space does not exceed the interaction range $R$ of the Feynman diagram. In a third inequality $R$ itself is bounded from above by the loop order, since each interaction between nearest neighbour fields of the composite operator generates at least one loop. We denote the diagrams that saturate this bound, i.e. the ones with interaction range $R=K+1$ as maximum range diagrams. Since the summation in runs over all insertion points with periodicity $L$, the smallest integer entry can always be fixed, e.g. to $1$ by shifting all $a_i$ by a common integer. According to the biggest integer can then be at most $K$. Further relations between the structures can be found in [@Beisert:2005wv]. The independent permutation structures which obey then form a basis in which the $K$-loop dilatation operator can be written down. The basis with elements is not the best choice in order to express the result of an explicit Feynman diagram calculation, since the different flavour arrangements within a single Feynman diagram generate linear combinations of several permutation structures with fixed relative coefficients. If, instead, the generated combinations themselves are used as basis elements, each Feynman diagram is associated with only one of them [@Fiamberti:2007rj; @Fiamberti:2008sh]. The basis elements obtained from supergraphs are called chiral functions and are defined as $$\label{chifuncdef} \begin{aligned} \chi(a_1,\dots,a_n)={{}\{a_1,\dots,a_n\}{}}\big|_{\operatorname{P}\to\operatorname{P}-{\mathds{1}}} {~,}\end{aligned}$$ where $\operatorname{P}\to\operatorname{P}-{\mathds{1}}$ denotes a replacement of all permutations in by the fixed combination of permutation and identity. The expansion of the resulting products yields $\chi$ in terms of linear combinations of permutation structures. For each $\chi$ we define the range of the interaction in flavour space by applying the definition to its list of arguments. The chiral functions capture the structure of the chiral and anti-chiral superfield lines of the underlying supergraphs. Hence, all supergraphs which only differ by the arrangement of the flavour-neutral vector fields generate contributions with the same chiral function. In particular, at loop order $K$ the chiral functions $\chi(a_1,\dots,a_n)$ with $n=K$ are associated each with a single Feynman graph since they do not contain any vector fields. We denote the respective graphs as chiral graphs. Except of the identity $\chi()={{}\{\}{}}$, all chiral functions yield zero when they are applied to one of the protected states in . The expression of the dilatation operator in terms of chiral functions should hence not explicitly depend on $\chi()$. We will come back to this statement at the end of Section \[subsubsec:magnonscat\]. One-loop dilatation operator {#subsec:oneloop} ---------------------------- The one-loop calculation in the $SU(2)$ subsector was addressed by Berenstein, Maldacena and Nastase in [@Berenstein:2002jq]. They used component fields to compute the term involving the permutation structure ${{}\{1\}{}}$, which permutes the flavour of two neighbouring fields. It is the maximum shuffling term at one loop, since it shifts the position of the impurity by the maximum number of one site at this loop order. Its generalization to higher loops will be discussed in Section \[subsubsec:magnondisp\]. The remaining Feynman diagrams all contribute to the identity operation ${{}\{\}{}}$ in flavour space and were not computed explicitly. Instead, their contribution was reconstructed from the fact that the eigenvalue for the ground state in should be zero. Furthermore, the contributions in which two neighbouring impurities interact with each other were neglected. Using $\mathcal{N}=1$ superfields instead of component fields for the one-loop calculation, only a single Feynman diagram contains a UV divergence and hence contributes to the renormalization constant in . It is evaluated as $$\label{oneloopdiag} \begin{aligned} \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-11\unitlength}\raisebox{\eqoff}{\fmfframe(0,1)(-10,1){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\end{fmfchar*}}} =+\lambda I_1\chi(1){~,}\end{aligned}$$ where the bold horizontal line represents the composite operator of arbitrary length $L\ge 2$, thereby omitting its $L-2$ elementary field lines that do not participate in the local interaction. The $\operatorname{D}$-algebra manipulations are trivial in this case as explicitly displayed in Appendix \[app:Dalg\]. The resulting loop integral is given in Appendix \[app:integrals\]. The further one-loop diagram of gluon exchange is finite, and the one-loop wave function renormalization vanishes. This is different from their behaviour in component formalism, where they have to be considered. According to the description , the one-loop dilatation operator follows from as $$\label{D1} \mathcal{D}_1=-2\chi(1) {~.}$$ Including also the contributions to the trace operator in flavour space, which extends the result to the flavour $SO(6)$ subsector,[^4] the full one-loop calculation in component fields was performed in [@Minahan:2002ve], and the result was recognized as the Hamiltonian of a respective integrable Heisenberg spin chain. Two-loop dilatation operator {#subsec:twoloops} ---------------------------- A two-loop renormalization of composite operators in the $SU(2)$ subsector was performed in [@Gross:2002mh] in component formalism. As in the one-loop case [@Berenstein:2002jq] only the diagrams which contribute to genuine flavour permutations were explicitly calculated, and the coefficient of the identity operation was determined by the condition of a vanishing eigenvalue of the ground state . Furthermore, the contributions in which impurities interact with each other were neglected. $R=1$ $R=2$ $R=3$ ------------- ------- ------- ------- $\chi()$ [ ]{} $\chi(1)$ $-$ [ ]{} [ ]{} $\chi(1,2)$ $-$ $-$ [ ]{} : Diagrams in $\mathcal{N}=1$ superfields (apart from eventual reflections) which can in principle contribute to the two-loop dilatation operator. Graphs which contain the vanishing one-loop self-energies are not drawn. It turns out that all diagrams depicted in gray are also irrelevant. The two-loop chiral self-energy is finite, and the remaining range $R\ge2$ diagrams are irrelevant due to generalized finiteness conditions [@Sieg:2010tz].[]{data-label="tab:2loopdiagrams"} The relevant diagrams for the complete two-loop calculation of the dilatation operator in terms of $\mathcal{N}=1$ superfields are given in Table \[tab:2loopdiagrams\]. The chiral self-energy is identically zero at one loop and finite at higher loops. According to the generalized finiteness conditions derived in [@Sieg:2010tz], all range $R\ge2$ diagrams, in which all vertices appear in loops are also finite. This concerns all remaining diagrams in the first line and in the second line the respective first diagram in the second and third columns. The pole parts of the last two diagrams in this line in the third column cancel against each other [@Fiamberti:2007rj; @Fiamberti:2008sh]. This cancellation is based on the fact that, in order to obtain contributions with overall UV divergences, a sufficient number of spinor derivatives $\operatorname{D}_\alpha$ and $\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}_{\dot \alpha}$ has to remain inside the loops in order to be transformed into spacetime derivatives. This yields constraints on the $\operatorname{D}$-algebra manipulations that amount to the formulation of generalized finiteness conditions in [@Sieg:2010tz]. All diagrams that are irrelevant due to these conditions are depicted in gray. We only have to compute the remaining diagrams and consider also their reflections where necessary. The substructures in the relevant range $R=2$ diagrams with chiral function $\chi(1)$ combine into the one-loop chiral vertex correction that is explicitly given in . We then find $$\label{twoloopdiags} \begin{aligned} \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-12\unitlength}\raisebox{\eqoff}{\fmfframe(-0.5,2)(-10.5,2){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfcmd{fill fullcircle scaled 8 shifted vloc(__vc2) withcolor black ;} \fmfiv{plain,label=$\scriptstyle\textcolor{white}{1}$,l.dist=0}{vloc(__vc2)} \end{fmfchar*}}}= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-12\unitlength}\raisebox{\eqoff}{\fmfframe(0,2)(-10,2){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfi{wiggly}{vm3{dir 180}..{dir -45}vm4} \end{fmfchar*}}}+ \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-12\unitlength}\raisebox{\eqoff}{\fmfframe(0,2)(-10,2){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfi{wiggly}{vm3{dir 0}..{dir -135}vm5} \end{fmfchar*}}}+ \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-12\unitlength}\raisebox{\eqoff}{\fmfframe(-0.5,2)(-10.5,2){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfi{wiggly}{vm4--vm5} \end{fmfchar*}}}= -2\lambda^2I_2\chi(1) {~,}\qquad \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-11\unitlength}\raisebox{\eqoff}{\fmfframe(0,1)(-5,1){\begin{fmfchar*}(20,20) {\fmftop{v1} \fmfbottom{v4} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v4} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v2,v3} \fmffixed{(0.25w,0)}{v4,v5} \fmffixed{(0.25w,0)}{v5,v6} \fmffixed{(0,whatever)}{vc1,vc3} \fmffixed{(0,whatever)}{vc2,vc4} \fmf{plain,tension=0.5,right=0.25}{v1,vc1} \fmf{plain,tension=0.5,left=0.25}{v2,vc1} \fmf{phantom,tension=0.5,right=0.25}{v2,vc2} \fmf{plain,tension=0.5,left=0.25}{v3,vc2} \fmf{plain,tension=0.5,left=0.25}{v4,vc3} \fmf{phantom,tension=0.5,right=0.25}{v5,vc3} \fmf{plain,tension=0.5,left=0.25}{v5,vc4} \fmf{plain,tension=0.5,right=0.25}{v6,vc4} \fmf{plain,tension=1.25,left=0}{vc1,vc3} \fmf{plain,tension=1.25,left=0}{vc2,vc4} \fmffreeze \fmf{plain,tension=1,left=0}{vc2,vc3} \fmf{plain,tension=0.5,right=0,width=1mm}{v4,v6} \fmffreeze \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p6}{vpath(__v3,__vc2)} \fmfiset{p4}{vpath(__v4,__vc3)} \fmfiset{p8}{vpath(__v5,__vc4)} \fmfiset{p9}{vpath(__v6,__vc4)} \fmfiset{p3}{vpath(__vc1,__vc3)} \fmfiset{p7}{vpath(__vc2,__vc4)} \fmfiset{p5}{vpath(__vc2,__vc3)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vm6}{point 0.5*length(p6) of p6} } {\fmfiequ{vm7}{point 0.5*length(p7) of p7} } {\fmfiequ{vm8}{point 0.5*length(p8) of p8} } {\fmfiequ{vm9}{point 0.5*length(p9) of p9} } }\end{fmfchar*}}} =+\lambda^2I_2\chi(1,2) {~,}\end{aligned}$$ where we have to consider also the reflection of the last diagram which contributes with chiral function $\chi(2,1)$. According to the description , the two-loop dilatation operator is then obtained by extracting the $\frac{1}{\varepsilon}$ pole of the sum of these diagrams and multiplying it by $-4$. With the pole part of the respective integral $I_2$ given in this then yields $$\label{D2} \mathcal{D}_2=4\chi(1)-2[\chi(1,2)+\chi(2,1)] {~.}$$ An explicit demonstration of the cancellation of the double poles in $\ln\mathcal{Z}$ as mentioned after can be found in [@Sieg:2010tz], where the one- and two-loop calculations were presented as a demonstration for the efficiency of the used approach. Three-loop dilatation operator {#subsec:threeloops} ------------------------------ At three-loop order a calculation of the dilatation operator directly from Feynman graphs of $\mathcal{N}=1$ superfields was recently performed in [@Sieg:2010tz]. The result reads $$\begin{aligned}\label{D3} \mathcal{D}_3 &=-4(\chi(1,2,3)+\chi(3,2,1))+2(\chi(2,1,3)-\chi(1,3,2))-4\chi(1,3)\\ &\phantom{{}={}} +16(\chi(1,2)+\chi(2,1))-16\chi(1)-4(\chi(1,2,1)+\chi(2,1,2)) {~.}\end{aligned}$$ It determines the planar spectrum in the $SU(2)$ subsector to three loops and hence goes beyond an earlier test of two eigenvalues [@Eden:2004ua], which employs Anselmi’s trick [@Anselmi:1998ms] to reduce the calculation to two loops. The three-loop results confirm the prediction from integrability in [@Beisert:2003tq]. Earlier checks of some of the three-loop eigenvalues are summarized in Section \[subsubsec:checks\]. Partial tests at higher loops {#subsec:higherloops} ----------------------------- To three-loop order and also beyond, certain parts of the respective Bethe ansatz and dilatation operator have been checked by direct Feynman diagram calculations. This concerns the so-called maximum shuffling terms, which contribute to the dispersion relation of the Bethe ansatz. Further terms in the higher loop expressions of the dilatation operator have also been tested explicitly. ### Tests of the magnon dispersion relation {#subsubsec:magnondisp} Even if with the assumed integrability the $SU(2)$ dilatation operator itself has been determined only to the first few loop orders (see chapter [@chapLR] for a review), the magnon dispersion relation of the Bethe ansatz is an all-order expression and directly related to certain Feynman diagrams. For a single magnon of momentum $p$ it is given by [@Beisert:2004hm] $$\label{magnondisp} E(p)=\sqrt{1+16g^2\sin^2\tfrac{p}{2}}-1 {~,}$$ and it is fixed by the underlying symmetry algebra up to an unknown function of the coupling constant [@Beisert:2005tm], which in the $\mathcal{N}=4$ SYM case essentially appears to be given by $g^2$ itself and has already been substituted accordingly.[^5] At a fixed loop order $K$ in the expansion of the above relation, the momentum dependence can be expressed as linear combination of the elements $\cos(k-1)p\,\sin^2\tfrac{p}{2}$ with $1\le k\le K$. In particular, the term with $k=K$ is generated by the so-called maximum shuffling diagrams, which include shifts of the position of a single impurity (which is a magnon in the spin chain notation) by the maximum number of $K$ neighbouring sites. The relevant diagrams are given by $$\label{maxshuffling} \begin{aligned} \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-10.5\unitlength}\raisebox{\eqoff}{\fmfframe(1,1)(1,1){\begin{fmfchar*}(25,20) \fmftop{v1} \fmfbottom{v7} \fmfforce{(0w,h)}{v1} \fmfforce{(0w,0)}{v7} \fmffixed{(0.2w,0)}{v1,v2} \fmffixed{(0.2w,0)}{v2,v3} \fmffixed{(0.2w,0)}{v3,v4} \fmffixed{(0.2w,0)}{v4,v5} \fmffixed{(0.2w,0)}{v5,v6} \fmffixed{(0.2w,0)}{v7,v8} \fmffixed{(0.2w,0)}{v8,v9} \fmffixed{(0.2w,0)}{v9,v10} \fmffixed{(0.2w,0)}{v10,v11} \fmffixed{(0.2w,0)}{v11,v12} \fmffixed{(0,whatever)}{vc1,vc5} \fmffixed{(0,whatever)}{vc2,vc6} \fmffixed{(0,whatever)}{vc3,vc7} \fmffixed{(0,whatever)}{vc4,vc8} \fmf{plain,tension=0.5,right=0.25}{v1,vc1} \fmf{plain,tension=0.5,left=0.25}{v2,vc1} \fmf{phantom,tension=0.5,right=0.25}{v2,vc2} \fmf{plain,tension=0.5,left=0.25}{v3,vc2} \fmf{phantom,tension=0.5,right=0.25}{v4,vc3} \fmf{dots,tension=0.5,left=0.25}{v5,vc3} \fmf{phantom,tension=0.5,right=0.25}{v5,vc4} \fmf{plain,tension=0.5,left=0.25}{v6,vc4} \fmf{plain,tension=0.5,left=0.25}{v7,vc5} \fmf{phantom,tension=0.5,right=0.25}{v8,vc5} \fmf{plain,tension=0.5,left=0.25}{v8,vc6} \fmf{phantom,tension=0.5,right=0.25}{v9,vc6} \fmf{dots,tension=0.5,left=0.25}{v10,vc7} \fmf{phantom,tension=0.5,right=0.25}{v11,vc7} \fmf{plain,tension=0.5,left=0.25}{v11,vc8} \fmf{plain,tension=0.5,right=0.25}{v12,vc8} \fmf{plain,tension=1.25,left=0}{vc1,vc5} \fmf{plain,tension=1.25,left=0}{vc2,vc6} \fmf{dots,tension=1.25,left=0}{vc3,vc7} \fmf{plain,tension=1.25,left=0}{vc4,vc8} \fmffreeze \fmf{plain,tension=1,left=0}{vc5,vc2} \fmf{dots,tension=1,left=0}{vc6,vc3} \fmf{plain,tension=1,left=0}{vc7,vc4} \fmf{plain,tension=0.5,right=0,width=1mm}{v7,v12} \fmffreeze \fmfposition \end{fmfchar*}}} \to \lambda^{K}I_K\chi(1,2,\dots,K-1,K) \end{aligned}$$ and by its reflection. When the sum of these two diagrams is applied to the eigenstate of a single magnon with momentum $p$, it yields the eigenvalue $$\label{chiphaseshifts} \lambda^{K}I_K\big[\chi(1,2,\dots,K)+\chi(K,\dots,2,1)\big]\to -8\lambda^{K}I_K\cos(K-1)p\,\sin^2\tfrac{p}{2} {~.}$$ According to the description , the $\frac{1}{\varepsilon}$ pole of this expression has to be multiplied by $-2K$ to obtain its contribution to the magnon dispersion relation. A comparison with the respective term in the expansion of , thereby taking into account the relation between the couplings, then makes a prediction for the $\frac{1}{\varepsilon}$ pole of the integral $I_K$ as $$\operatorname{Res}_0(\operatorname{K}\operatorname{R}(I_K)) =\frac{1}{(4\pi)^{2K}}\frac{(2K-2)!}{(K-1)!K!}\frac{1}{K} {~.}$$ The explicit expressions for the poles of $I_K$ for some $K$ are listed in . They are consistent with this result. In [@Gross:2002su] it was shown that at generic loop order the pole structure of the maximum shuffling diagrams in component fields is in accord with the BMN square root formula [@Berenstein:2002jq]. The latter was proposed as an all-order expression for the anomalous dimensions in the so-called BMN limit, where the length $L$ of the operators and the coupling $g$ become infinite $L,g\to\infty$, thereby keeping fixed the numbers of impurities inside the operators and also the effective coupling constant $g'=\frac{g}{L}$. For magnon momenta $p_j=\frac{2\pi n_j}{L}\ll 1$ the dispersion relation yields the individual contributions of each magnon $j$ with mode number $n_j$ to the BMN square root formula. Since the scattering of magnons is neglected, their momenta $p_j$ assume a simple form and are solutions of the originally proposed Bethe equations [@Beisert:2004hm] with a magnon S-matrix that becomes trivial in the BMN limit. However, these Bethe equations do not yield the anomalous dimensions of $\mathcal{N}=4$ SYM theory since the S-matrix is incomplete. One has to consider the so-called dressing phase [@Beisert:2006ez] that first appeared at strong coupling [@Arutyunov:2004vx; @Hernandez:2006tk] but is important also at weak coupling [@Beisert:2006ez; @Bern:2006ew], where it alters the magnon momenta at order $\mathcal{O}(g^6)$.[^6] Due to the dressing phase, the S-matrix violates perturbative BMN scaling, i.e. its perturbative expansion diverges if after the replacement $g\to g'L$ the limit $L\to\infty$ is taken, thereby keeping $g'$ fixed and small. The Bethe equations involving this S-matrix then yield anomalous dimensions that violate perturbative BMN scaling from four loops on. However, the BMN square root formula obeys this scaling, and hence it cannot describe the anomalous dimensions of operators with two or more impurities beyond three loops.[^7] Since the dressing phase only affects the scattering of magnons, all tests and derivations of the BMN square root formula that rely on the calculation of phase shifts of a single magnon are insensitive to this failure and succeed. This concerns the previously mentioned all-order test of the maximum shuffling terms [@Gross:2002su] and also an all order derivation employing the $\mathcal{N}=1$ superfield formalism [@Santambrogio:2002sb]. It would be more appropriate to say that in these calculations the magnon dispersion relation in the BMN limit is obtained. The magnon dispersion relation describes the free propagation of one magnon. It it thus built up from all Feynman diagrams with chiral functions that do not yield a vanishing result when applied to the single magnon momentum eigenstate. The number of impurities of the composite operator sets an upper bound on the number of bubbles formed by two neighbouring lines of the composite operator inside the Feynman diagrams. Such a bubble appears for example in the lower right corner of the graph in , and it vanishes unless the two involved field flavours are different. The diagrams contributing to the magnon dispersion relation hence must not contain more than one of these bubbles. This restricts their chiral functions to $\chi(1,\dots,k)$ and $\chi(k,\dots,1)$ after the identities for the permutation structures found in [@Beisert:2005wv] have been used to simplify the chiral functions, e.g. as $\chi(1,2,1)=\chi(2,1,2)=\chi(1)$ in the three loop result . All-order expressions for the coefficients of these terms in the dilatation operator then follow directly from the magnon dispersion relation and can be found in [@Sieg:2010tz]. It should be stressed that the aforementioned contributions also yield non-vanishing results when additional magnons are present outside of the $k+1$ interacting legs. They therefore also contribute to the magnon S-matrix. ### Tests of magnon scattering {#subsubsec:magnonscat} The Feynman diagrams that vanish for a single magnon state, but are non-vanishing if two or more magnons are present within their respective interaction ranges, should exclusively be associated with the magnon S-matrix. Their contributions appear together with the ones of the aforementioned maximum and non-maximum shuffling terms in the dilatation operator. In the $SU(2)$ subsector they first show up at three-loops as the contribution with chiral function $\chi(1,3)$ in .[^8] The further chiral functions $\chi(2,1,3)$, $\chi(1,3,2)$ are also associated with magnon scattering, but they only appear in a combination that is associated with a similarity transformation, i.e. a change in the basis of operators [@Beisert:2003tq; @Beisert:2005wv], that does not affect the eigenvalues. As a more complicated example, we consider the four-loop dilatation operator. It can be determined from the underlying integrability as reviewed in chapter [@chapLR]. In the basis of the chiral functions it reads $$\begin{aligned}\label{D4} \mathcal{D}_4&={}+{}200\chi(1)-150[\chi(1,2)+\chi(2,1)]+8(10+\epsilon_{3a})\chi(1,3)-4\chi(1,4)\\ &\phantom{{}={}} +60[\chi(1,2,3)+\chi(3,2,1)]\\ &\phantom{{}={}} +(8+2\beta+4\epsilon_{3a}-4i\epsilon_{3b}+2i\epsilon_{3c}-4i\epsilon_{3d}) \chi(1,3,2)\\ &\phantom{{}={}} +(8+2\beta+4\epsilon_{3a}+4i\epsilon_{3b}-2i\epsilon_{3c}+4i\epsilon_{3d}) \chi(2,1,3)\\ &\phantom{{}={}} -(4+4i\epsilon_{3b}+2i\epsilon_{3c})[\chi(1,2,4)+\chi(1,4,3)]\\ &\phantom{{}={}} -(4-4i\epsilon_{3b}-2i\epsilon_{3c})[\chi(1,3,4)+\chi(2,1,4)]\\ &\phantom{{}={}}-(12+2\beta+4\epsilon_{3a})\chi(2,1,3,2)\\ &\phantom{{}={}} +(18+4\epsilon_{3a})[\chi(1,3,2,4)+\chi(2,1,4,3)]\\ &\phantom{{}={}} -(8+2\epsilon_{3a}+2i\epsilon_{3b})[\chi(1,2,4,3)+\chi(1,4,3,2)]\\ &\phantom{{}={}} -(8+2\epsilon_{3a}-2i\epsilon_{3b})[\chi(2,1,3,4)+\chi(3,2,1,4)]\\ &\phantom{{}={}} -10[\chi(1,2,3,4)+\chi(4,3,2,1)] {~.}\end{aligned}$$ The coefficients $\epsilon_i$, $i=3a,3b,3c,3d$ in the above result are not fixed by the construction and parameterize the previously mentioned similarity transformations. The coefficient $\beta$ is the leading term of the previously mentioned dressing phase. The magnon dispersion relation is encoded in the first two terms in the first line, the second line and the last line. The further contributions should be associated with magnon scattering. As the contributions from the maximum shuffling diagrams in the last line, also the other terms in the last four lines have chiral functions that saturate all the bounds in . Hence, the underlying Feynman diagrams are chiral and of maximum range and their contributions can be calculated as easily as the one of the maximum shuffling terms . The term in with chiral function $\chi(2,1,3,2)$ only satisfies the first bound in , i.e. the underlying Feynman diagram is chiral but it is not of maximum range. It involves the leading coefficient $\beta$ of the dressing phase, which can be determined from an evaluation of the respective diagram $$\begin{aligned} \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-11\unitlength}\raisebox{\eqoff}{\fmfframe(-1,1)(-1,1){\begin{fmfchar*}(20,20) \fmftop{v1} \fmfbottom{v5} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v5} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v2,v3} \fmffixed{(0.25w,0)}{v3,v4} \fmffixed{(0.25w,0)}{v5,v6} \fmffixed{(0.25w,0)}{v6,v7} \fmffixed{(0.25w,0)}{v7,v8} \fmffixed{(0,whatever)}{vc1,vc5} \fmffixed{(0,whatever)}{vc2,vc3} \fmffixed{(0,whatever)}{vc3,vc6} \fmffixed{(0,whatever)}{vc6,vc7} \fmffixed{(0,whatever)}{vc4,vc8} \fmffixed{(0.5w,0)}{vc1,vc4} \fmffixed{(0.5w,0)}{vc5,vc8} \fmf{plain,tension=1,right=0.125}{v1,vc1} \fmf{plain,tension=0.25,right=0.25}{v2,vc2} \fmf{plain,tension=0.25,left=0.25}{v3,vc2} \fmf{plain,tension=1,left=0.125}{v4,vc4} \fmf{plain,tension=1,left=0.125}{v5,vc5} \fmf{plain,tension=0.25,left=0.25}{v6,vc6} \fmf{plain,tension=0.25,right=0.25}{v7,vc6} \fmf{plain,tension=1,right=0.125}{v8,vc8} \fmf{plain,tension=0.5}{vc1,vc3} \fmf{plain,tension=0.5}{vc2,vc3} \fmf{plain,tension=0.5}{vc3,vc4} \fmf{plain,tension=0.5}{vc5,vc7} \fmf{plain,tension=0.5}{vc6,vc7} \fmf{plain,tension=0.5}{vc7,vc8} \fmf{plain,tension=2}{vc1,vc5} \fmf{plain,tension=2}{vc4,vc8} \fmf{phantom,tension=2}{vc5,vc4} \fmffreeze \fmfposition \fmf{plain,tension=1,left=0,width=1mm}{v5,v8} \fmffreeze \end{fmfchar*}}} \to\lambda^4I_\beta\chi(2,1,3,2) \end{aligned}$$ if the coefficient $\epsilon_{3a}$ of the similarity transformations is known. One finds $\epsilon_{3a}=-4$ for example by computing the diagram which generates $\chi(1,3,2,4)$ or $\chi(2,1,4,3)$. With the pole part of the integral $I_\beta$ given in , the leading coefficient of the dressing phase is then determined as $\beta=4\zeta(3)$. The result was obtained in [@Beisert:2007hz], using component formalism. It agrees with one of the proposals in [@Beisert:2006ez] and with the result extracted from a four-loop calculation of a four-point amplitude in [@Bern:2006ew]. It is also relatively easy to compute the terms with chiral functions which only saturate the second and third bound in , i.e. all terms in with chiral functions that contain $1$ and $4$ in their lists of arguments and hence only stem from Feynman diagrams of maximum range $R=5$. This calculation was performed in [@Fiamberti:2007rj; @Fiamberti:2008sh] in $\mathcal{N}=1$ superfield formalism in the context of calculating the first wrapping correction to be discussed below. The results yield an overdetermined system of equations that uniquely fixes the coefficients $\epsilon_i$ and provides non-trivial checks of the remaining coefficients that are fixed by the underlying integrability. The analogous calculation of the $R=6$ diagrams at five loops can be found in [@Fiamberti:2009jw]. The expressions , , and do not depend on the identity $\chi()$. This guarantees that the anomalous dimension of the BPS operators are zero. The generalized finiteness conditions in [@Sieg:2010tz] predict this to all orders and relate it to the finiteness of the chiral self-energy, i.e. to the preservation of conformal invariance. ### Checks of eigenvalues {#subsubsec:checks} To three loops the results , and for the dilatation operator have been obtained by direct Feynman diagram calculations. At higher loops, only the terms that saturate at least one of the bounds in have been tested as described above. Further checks concern the eigenvalues of the dilatation operator for some composite operators. They should match with the anomalous dimensions obtained in direct Feynman diagram calculations. Of particular interest is thereby the Konishi supermultiplet. As superconformal primary it contains the $\mathcal{N}=1$ Konishi operator [@Konishi:1983hf] that has bare scaling dimension $\Delta_0=2$ and reads $$\label{Konishiop} \mathcal{K}=\operatorname{tr}\big({\operatorname{e}}^{-g_{\text{YM}}V}\bar\phi_i{\operatorname{e}}^{g_{\text{YM}}V}\phi^i\big) {~.}$$ This operator is not chiral, and hence all its superfield components lie beyond the $SU(2)$ subsector. However, the Konishi supermultiplet also contains an operator of this subsector. In order to find it, one has to select the level four descendant of bare dimension $\Delta_0=4$ that is chiral and pick out the relevant $SU(4)$ R-symmetry component given by $$\label{Kdesc} \operatorname{tr}\big({\left[\phi\smash[b]{\mathbin{,}}Z\right]}{\left[\phi\smash[b]{\mathbin{,}}Z\right]}\big) {~.}$$ It contains as lowest superfield component the respective operator built out of the two scalar fields of the flavour $SU(2)$ subsector. All members of a superconformal multiplet acquire the same anomalous dimension. For the Konishi multiplet it is given to four loops in . The one- and two-loop contributions were obtained by explicit Feynman diagram calculations in [@Anselmi:1996mq; @Anselmi:1996dd] and [@Bianchi:1999ge; @Bianchi:2000hn], and then also by an OPE analysis in [@Arutyunov:2001mh], see also [@Bianchi:2001cm]. These results are also found for a twist-two operator with conformal spin $S=2$ that appears within another level four descendant of the Konishi multiplet. It belongs to the closed $SL(2)$ subsector that contains certain operators with general twist and conformal spin $S$. For twist-two operators with generic $S$, the result to two loops has been obtained from Feynman diagrams in [@Kotikov:2003fb]. At three loops it could be extracted [@Kotikov:2004er] as the terms with highest transcendentality, i.e. with highest degrees of the harmonic sums, from the NNLO QCD result for the non-singlet splitting functions of QCD [@Moch:2004pa]. The truncation of the QCD result is based on the observation [@Kotikov:2002ab] that due to special properties of the DGLAP and BFKL equations in $\mathcal{N}=4$ SYM theory a mixing between functions of different transcendentality degrees does not occur. Specializing to $S=2$, the extracted result agrees with the three-loop contribution in . When the dilatation operator given in , and is applied to the state , it also correctly yields the result in .[^9] In fact, the three-loop term was first predicted in [@Beisert:2003tq], where the dilatation operator was constructed from integrability. Later, an explicit Feynman diagram calculation [@Eden:2004ua], which employs Anselmi’s trick [@Anselmi:1998ms] to reduce the calculation to two loops, led to the same result. The calculation in [@Sieg:2010tz] also confirms the result and furthermore fixes the planar three-loop spectrum of all composite single-trace operators of the flavour $SU(2)$ subsector from field theory by a direct Feynman diagram calculation of the dilatation operator. The previously mentioned twist-two operators of the $SL(2)$ sector are very important for tests of the ${\text{AdS}}/\text{CFT}$ correspondence and the underlying integrability. These tests are reviewed in chapter [@chapTwist]. In particular, the results in the strict $S\to\infty$ limit are not modified by wrapping interactions. At finite $S$ such modifications occur. The simplest example is $S=2$, i.e.the operator which appears in the Konishi multiplet. Its anomalous dimension is affected by wrapping interactions at four loops and beyond. Wrapping interactions {#sec:wrapint} ===================== In the following we briefly summarize the calculations of the previously mentioned wrapping interactions. A more detailed review is given by [@Fiamberti:2010fw]. The Bethe ansätze or the dilatation operator yield reliable results for the anomalous dimensions in the asymptotic limit only. The origin and precise form of this restriction can be understood by recalling the construction from Feynman diagrams. In Section \[sec:Dop\] it was argued that at a given loop order $K$ the dilatation operator is determined from Feynman diagrams with range $R\le K+1$, which lead to flavour permutations with range $\kappa\le R$. For the construction of the diagrams, it is thereby implicitly assumed that the length $L$ of the involved composite operators is at least as big as the maximal interaction range $K+1$. Therefore, an application of the dilatation operator to composite operators of length $L$ can in general only yield the correct anomalous dimensions in the asymptotic limit, i.e. to a loop order $K\le L-1$. At $K\ge L$ loops, the assumption of a sufficient length of the involved composite operators becomes invalid, and therefore contributions from diagrams with interaction range $R>L$ should be removed from the dilatation operator. Instead, there are contributions from new diagrams that are built with the operators of the respective lower length $L$. The new diagrams are called wrapping diagrams since, due to the insufficient length of the composite operators, the interactions wrap around them. Two examples of such diagrams are depicted in Figure \[fig:hightranswrap\]. Beyond the asymptotic limit, the dilatation operator explicitly depends on the length $L$ of the composite operators it is applied to. More precisely, the coefficients of the chiral functions in the expression of the dilatation operator become functions of $L$ at loop orders $K\ge L$, while in the asymptotic limit they are constants, and the dilatation operator depends on the length only via the permutation structures . The appearance of wrapping interactions is closely connected to the truncation of the genus $h$ expansion of the dilatation operator beyond the planar $h=0$ contribution [@Sieg:2005kd]. If in a planar wrapping diagram the composite operator is replaced by a longer operator, the additional fields lines cannot leave the diagram without crossing any other lines, i.e.it becomes a diagram of genus $h=1$. The appearing wrapping diagrams hence come from certain genus $h=1$ contributions to the dilatation operator, which become planar when it is applied to a sufficiently short composite operator. Wrapping diagrams appear at all orders in the genus expansion of the dilatation operator. They are of genus $h+1$ in the asymptotic regime and encode the finite size effects at genus $h$. The planar wrapping diagrams are special since they can be projected out of all genus one contributions by introducing spectator fields [@Sieg:2005kd]. While in general for higher genus diagrams the notion of the range of the interaction is not meaningful, it is still well defined for the subset of genus one diagrams when they become the planar wrapping diagrams. Integrability seems to persist, even if in general at higher genus its breakdown is expected [@Beisert:2003tq].[^10] In order to obtain the anomalous dimensions beyond the asymptotic regime, one should not abandon the dilatation operator as obtained from the underlying integrability at loop orders $K\ge 4$ and compute all Feynman diagrams. Instead, the considerations at the beginning of this section imply that the dilatation operator is still useful, since it can be corrected for an application to composite operators of shorter length $L$. First, at each loop order $K$ all contributions from Feynman graphs of longer range $K+1\ge R> L$ have to be removed. Then, contributions from the wrapping interactions have to be added. This procedure is particularly powerful at the critical order $K=L$ where wrapping arises for the first time, since only relatively few Feynman diagrams of restricted topology have to be computed explicitly. Most diagrams are captured automatically by those terms in the dilatation operator that are not removed in the modification process. Also, the only contributions that one has to remove from the dilatation operator are the ones that come from Feynman diagrams with maximum range $R=K+1$. It is convenient to divide these diagrams according to their range of interaction in flavour space $\kappa$ into two classes. The first class contains diagrams with $\kappa=R=K+1$, i.e.according to the definition of $\kappa$ in their range $R$ is encoded within the list of arguments of their chiral functions. The second class collects all the remaining diagrams with $\kappa<R=K+1$. Such Feynman diagrams contain a chiral structure with interaction range $\kappa$, and the remaining $R-\kappa$ neighbouring field lines are connected with it and with each other only by vector fields. Since the latter are flavour neutral, the range $R$ of these diagrams is not captured by the chiral functions. It was shown in [@Fiamberti:2008sh] in the $\mathcal{N}=1$ superfield formalism that the diagrams of the second class do not contribute to the dilatation operator: either they are finite or their overall UV divergences cancel against each other. This is also an implication of the generalized finiteness conditions derived in [@Sieg:2010tz]. In Section \[subsec:twoloops\] we have already used the results when we disregarded the two-loop diagrams with $R=3$ but $\kappa<3$ in the first two rows of the last column of Table \[tab:2loopdiagrams\]. The diagrams of the first class that have $\kappa=R=K+1$ are the only maximum range diagrams that contribute with their overall UV divergences. These contributions can be easily identified and removed from the expression of the dilatation operator, since their chiral functions are of maximum range. The subtraction procedure becomes almost trivial: one just has to remove all contributions with chiral functions that have $1$ and $K$ within their list of arguments. This does not require the calculation of any Feynman diagrams. For example, in the four-loop expression one removes the last contribution in the first line and the ones in the fifth, sixth and the last four lines. The eigenvalues of the subtracted dilatation operator are no longer independent of the scheme coefficients $\epsilon_i$, which have to be fixed by calculating at least some of the diagrams with range $R=K+1$. If one could compute the wrapping interactions that have to be added to the subtracted dilatation operator also as functions of $\epsilon_i$, the eigenvalues of the resulting operator should not depend on the $\epsilon_i$. However, the calculation of the wrapping interactions takes place in a scheme fixed by the use of $\mathcal{N}=1$ supergraphs, and therefore the $\epsilon_i$ in the subtracted dilatation operator have to assume the respective values. Finally, it is important to remark that the simplicity of the subtraction procedure is only guaranteed if chiral functions are used as basis elements. If, instead, the basis of permutation structures is used, the subtraction of the contribution from a Feynman diagram with $R=K+1$ affects the coefficients of several permutation structures also with different flavour interaction ranges $\kappa\le R$ in the dilatation operator.[^11] The aforementioned method was first introduced and used in [@Fiamberti:2007rj], with the details given in [@Fiamberti:2008sh], in the case $K=L=4$, i.e. for the four-loop anomalous dimension of the Konishi operator. In $\mathcal{N}=4$ SYM theory it is the simplest case where wrapping arises. The calculation starts from the four-loop asymptotic dilatation operator and modifies it for an application to the length four Konishi descendant of the flavour $SU(2)$ subsector in order to determine the correct eigenvalue [@Fiamberti:2007rj; @Fiamberti:2008sh]. Including also the lower orders, the anomalous dimension of the Konishi operator to four-loops was then determined as $$\label{gammaK} \gamma=12g^2-48g^4+336g^6+(-2496+576\zeta(3)-1440\zeta(5))g^8 {~,}$$ where the full conformal dimension is obtained as $\Delta=\Delta_0+\gamma$ with the bare scaling dimension $\Delta_0$ as described in Section \[subsubsec:checks\]. The four-loop contribution has also been obtained from a generalized Lüscher formula [@Bajnok:2008bm]. This approach is reviewed in chapter [@chapLuescher]. Furthermore, it was later also found in a computer-based calculation in component formalism [@Velizhanin:2008jd]. The matching of the Feynman diagram and Lüscher based calculations provides the first test of ${\text{AdS}}/{\text{CFT}}$ and the underlying integrability beyond the asymptotic limit. It is also reproduced by the recently proposed $Y$-system [@Gromov:2009bc; @Gromov:2009tv], which is derived from the thermodynamic Bethe ansatz (TBA) [@Bombardelli:2009ns; @Arutyunov:2007tc; @Arutyunov:2009zu] and is a candidate to capture the full planar spectrum of $\mathcal{N}=4$ SYM theory. The TBA and Y-system are reviewed, respectively, in chapters [@chapTBA] and [@chapTrans]. Earlier attempts to describe the wrapping effects in terms of integrable systems are included in chapter [@chapLR]. In [@Bajnok:2008qj] the result which also holds for the earlier mentioned twist-two operator with conformal spin $S=2$ has been generalized to arbitrary $S$. When analytically continued to $S=-1$, it yields the correct pole structure as predicted from the BFKL equation. A result for the five-loop anomalous dimension of the Konishi operator has been obtained in impressive calculations on the basis of the generalized Lüscher formula [@Bajnok:2009vm] and the TBA [@Arutyunov:2010gb; @Balog:2010xa]. Also this result has been generalized to arbitrary spin $S$, and it is in accord with the pole structure from the BFKL equation [@Lukowski:2009ce]. To obtain the five-loop result for the Konishi multiplet from a Feynman diagram calculation is very difficult, even with the universal cancellation mechanisms discovered in [@Sieg:2010tz]. Instead, a five-loop result for the $L=5$ operator $\operatorname{tr}\big({\left[\phi\smash[b]{\mathbin{,}}Z\right]}{\left[\phi\smash[b]{\mathbin{,}}Z\right]}Z\big)$ which is in the same supermultiplet as certain twist-three operators has been computed [@Fiamberti:2009jw], and it agrees with the result from the generalized Lüscher formula [@Beccaria:2009eq]. The six-loop results for the twist-three operators with generic conformal spin $S$ has recently become available [@Velizhanin:2010cm]. Beyond the asymptotic limit, the contributions of highest transcendentality, i.e. which contain the $\zeta$-function with biggest argument, are generated entirely by the wrapping interactions. In the four-loop result in this is the term with $\zeta(5)$. Its generalization to twist-two operators with generic conformal spin $S$ has been obtained from a Feynman diagram calculation in component formalism in [@Velizhanin:2008pc]. At generic loop and critical wrapping order $K=L$ the highest transcendentality degree of the wrapping diagrams is $2K-3$ compared to $2K-5$ of the dressing phase in the asymptotic Bethe ansatz. A clean setup that allows one to study the transcendentality structure without admixtures from the dressing phase is provided by single-impurity operators in the $\beta$-deformed $\mathcal{N}=4$ SYM theory.[^12] The leading wrapping corrections have been calculated up to $11$ loops in [@Fiamberti:2008sm; @Fiamberti:2008sn] and were confirmed in [@Beccaria:2009hg; @Gromov:2010dy; @Arutyunov:2010gu]. A clear pattern emerges also for the terms of lower transcendentality. The diagrams in Figure \[fig:hightranswrap\] are responsible for the highest transcendentality contribution involving $\zeta(2K-3)$. The respective term can be traced back to a component $\frac{1}{2}P_K$ in the decomposition of the integrals, where $P_K$ is the $K$-loop cake integral given in . Conclusions {#sec:concl} =========== We have reviewed the explicit Feynman diagram calculations which at small ’t Hooft coupling determine the planar spectrum of composite operators in the flavour $SU(2)$ subsector of $\mathcal{N}=4$ SYM theory and test the underlying integrability. We have presented the calculations up to two loops in detail and summarized the calculations and partial checks at higher loops. The use of $\mathcal{N}=1$ superspace techniques and of chiral functions as operators in flavour space allowed us to directly interpret the Feynman diagrams in terms of the dispersion relation and the scattering matrix that appear in the integrability-based Bethe ansatz. Then, we reviewed how anomalous dimensions beyond the asymptotic limit can be obtained by computing the leading wrapping corrections and which properties and interpretation these interactions have. The existing tests in these setups have been summarized. Acknowledgements {#acknowledgements .unnumbered} ================ I am very grateful to Francesco Fiamberti and Alberto Santambrogio for reading parts of the manuscript. I also want to thank Francesco Fiamberti, Matias Leoni, Andrea Mauri, Joseph Minahan, Alberto Santambrogio, Olof Ohlsson Sax, Gabriele Tartaglino-Mazzucchelli, Alessandro Torielli and Daniela Zanon for very pleasant collaborations on some of the papers reviewed here and in other chapters of this review. D-algebra {#app:Dalg} ========= The propagators and vertices of superfields depend not only on the bosonic, but also on the fermionic coordinates $\theta^\alpha$, $\bar\theta^{\dot\alpha}$, of superspace and carry covariant spinor derivatives $\operatorname{D}_\alpha$, $\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}_{\dot\alpha}$. By the $\operatorname{D}$-algebra manipulation which consists of transfers, partial integrations and the use of (anti)-commutation relations for products of these spinor derivatives, the underlying expression is transformed into the final result that is localized at a single point in the coordinates $\theta^\alpha$, $\bar\theta^{\dot\alpha}$. We refer the reader to [@Gates:1983nr] for an introduction to the $\mathcal{N}=1$ superfield formalism in the adopted conventions and to [@Sieg:2010tz] for an explicit presentation of the relevant Feynman rules. Here, we only recall that two $\operatorname{D}_\alpha$ and two $\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}_{\dot\alpha}$ have to remain in each loop in order to obtain a non-vanishing result. The loop is then localized in the fermionic coordinates. We indicate this by filling it grey. Also, we recall two simple relations, $\operatorname{D}^2\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2\operatorname{D}^2=\Box\operatorname{D}^2$ and $\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2\operatorname{D}^2\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2=\Box\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$, which transform spinor derivatives into spacetime derivatives $\Box=\partial^\mu\partial_\mu$. The one-loop diagram requires no $\operatorname{D}$-algebra manipulations, and one directly obtains $$\begin{aligned} \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{p4} \end{fmfchar*}}}= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfcmd{fill p4--reverse(p5)--vloc(__v3)--cycle withcolor {0.75 white};} \end{fmfchar*}}}\to -I_1 {~,}\end{aligned}$$ where the loop integral $I_1$, given in for $K=1$, is the one extracted from the grey-scaled region. Its UV pole is listed in . There appears an additional factor $-1$ in front of $I_1$: we have to transform the full fermionic measure in the algebraic expression of the diagram into the chiral measure of the term that adds the chiral composite operator with a chiral source to the action. This means, we replace ${\operatorname{d}\!}^4\theta\to{\operatorname{d}\!}^2\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$ and combine the extra derivatives $\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$ with the remaining $\operatorname{D}^2$ in the above diagram to $\Box$, such that the propagator that connects the chiral and anti-chiral cubic vertex is cancelled, thereby yielding the factor $-1$. In the result we have not considered any other non-trivial prefactors of the propagators and vertices. They are contained within the color- and flavour factors (chiral functions) of the complete result given in . The one-loop correction to the chiral vertex that enters is easily evaluated $$\begin{aligned}\label{ccconeloop} {\settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-13\unitlength}\raisebox{\eqoff}{\fmfframe(2,1)(2,1){\begin{fmfchar*}(21,24) \fmftop{v3} \fmfbottom{v2} \fmfforce{(w,h)}{v3} \fmfforce{(w,0)}{v2} \fmfpoly{phantom}{v1,v2,v3} \fmf{phantom,tension=2}{v1,vg1} \fmf{phantom}{v2,vg1} \fmf{phantom}{v3,vg1} \fmffreeze \fmf{plain,tension=1}{vg2,v2} \fmf{plain,tension=1}{vg3,v3} \fmf{phantom}{vg1,v1} \fmf{phantom,tension=0.5}{vg1,vg2} \fmf{phantom,tension=0.5}{vg1,vg3} \fmffreeze \fmfposition \fmf{phantom}{vg2,vg3} \fmfipath{p[],pca} \fmfipair{vm[],vo[],vi[]} \fmfiset{p1}{vpath(__vg1,__v1)} \fmfiset{p2}{vpath(__vg2,__v2)} \fmfiset{p3}{vpath(__vg3,__v3)} \fmfiset{p4}{vpath(__vg2,__vg3)} \fmfiset{p5}{vpath(__vg3,__vg1)} \fmfiset{p6}{vpath(__vg1,__vg2)} {\fmfcmd{fill fullcircle scaled 10 shifted vloc(__vg1) withcolor black ;} \fmfiv{plain,label=\small$\textcolor{white}{1}$,l.dist=0}{vloc(__vg1)}} \fmfis{plain,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{reverse(p1)} \fmfis{plain,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p6} \fmfis{plain,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p5} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p2} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p3} \end{fmfchar*}}}} &= {\settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-13\unitlength}\raisebox{\eqoff}{\fmfframe(2,1)(2,1){\begin{fmfchar*}(21,24) \fmftop{v3} \fmfbottom{v2} \fmfforce{(w,h)}{v3} \fmfforce{(w,0)}{v2} \fmfpoly{phantom}{v1,v2,v3} \fmf{phantom,tension=2}{v1,vg1} \fmf{phantom}{v2,vg1} \fmf{phantom}{v3,vg1} \fmffreeze \fmf{plain,tension=1}{vg2,v2} \fmf{plain,tension=1}{vg3,v3} \fmf{phantom}{vg1,v1} \fmf{phantom,tension=0.5}{vg1,vg2} \fmf{phantom,tension=0.5}{vg1,vg3} \fmffreeze \fmfposition \fmf{photon}{vg2,vg3} \fmfipath{p[],pca} \fmfipair{vm[],vo[],vi[]} \fmfiset{p1}{vpath(__vg1,__v1)} \fmfiset{p2}{vpath(__vg2,__v2)} \fmfiset{p3}{vpath(__vg3,__v3)} \fmfiset{p4}{vpath(__vg2,__vg3)} \fmfiset{p5}{vpath(__vg3,__vg1)} \fmfiset{p6}{vpath(__vg1,__vg2)} {} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{reverse(p1)} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p6} \fmfis{plain,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p5} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p2} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p3} \end{fmfchar*}}}} +\dots = \left({\settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-13\unitlength}\raisebox{\eqoff}{\fmfframe(2,1)(2,1){\begin{fmfchar*}(21,24) \fmftop{v3} \fmfbottom{v2} \fmfforce{(w,h)}{v3} \fmfforce{(w,0)}{v2} \fmfpoly{phantom}{v1,v2,v3} \fmf{phantom,tension=2}{v1,vg1} \fmf{phantom}{v2,vg1} \fmf{phantom}{v3,vg1} \fmffreeze \fmf{plain,tension=1}{vg2,v2} \fmf{plain,tension=1}{vg3,v3} \fmf{phantom}{vg1,v1} \fmf{phantom,tension=0.5}{vg1,vg2} \fmf{phantom,tension=0.5}{vg1,vg3} \fmffreeze \fmfposition \fmf{plain}{vg2,vg3} \fmfipath{p[],pca} \fmfipair{vm[],vo[],vi[]} \fmfiset{p1}{vpath(__vg1,__v1)} \fmfiset{p2}{vpath(__vg2,__v2)} \fmfiset{p3}{vpath(__vg3,__v3)} \fmfiset{p4}{vpath(__vg2,__vg3)} \fmfiset{p5}{vpath(__vg3,__vg1)} \fmfiset{p6}{vpath(__vg1,__vg2)} {\fmfiset{p10}{p4--p5--p6--cycle}\fmfcmd{fill p10 withcolor {0.75 white};}} \fmfis{plain,l.side=left,l.dist=2,label=$\scriptstyle\Box$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{reverse(p1)} \fmfis{plain,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p6} \fmfis{plain,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p5} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=3,ptext.hout=3,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p2} \fmfis{plain,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=6,ptext.oout=6,ptext.sep=;}{p3} \end{fmfchar*}}}} +\dots\right)i\lambda g_{\text{YM}}\epsilon_{ijk}\operatorname{tr}\big(T^a{\left[T^b\smash[b]{\mathbin{,}}T^c\right]}\big) {~,}\end{aligned}$$ where the ellipsis denote the remaining two diagrams obtained by cyclic permutations of the external legs, and we have included the color and flavour factors. Also in this case, the $\Box$ is produced after reducing the full fermionic measure to the chiral measure as mentioned above. When $\Box$ cancels the propagator a factor $-1$ is produced. The $\operatorname{D}$-algebra manipulations for the diagrams contributing to the two-loop dilatation operator are $$\begin{aligned} & \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=12,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{p4} \fmfcmd{fill fullcircle scaled 10 shifted vloc(__vc2) withcolor black ;} \fmfiv{plain,label=\small$\textcolor{white}{1}$,l.dist=0}{vloc(__vc2)} \end{fmfchar*}}}= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfiset{p6}{subpath (0,length(p3)/2) of p3} \fmfiset{p7}{subpath (length(p3)/2,length(p3)) of p3} \fmfiset{p8}{subpath (0,length(p4)/2) of p4} \fmfiset{p9}{subpath (length(p4)/2,length(p4)) of p4} \fmfiset{p10}{subpath (0,length(p5)/2) of p5} \fmfiset{p11}{subpath (length(p5)/2,length(p5)) of p5} \fmfcmd{fill reverse(p11)--vm5{dir 45}..{dir 180}vm3--p7--cycle withcolor {0.75 white};} \fmfi{plain}{vm5{dir 45}..{dir 180}vm3} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p6)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=5,ptext.sep=;}{p4} \fmfis{phantom,l.dist=2,label=$\scriptstyle\Box$}{p4} \end{fmfchar*}}}+ \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfiset{p6}{subpath (0,length(p3)/2) of p3} \fmfiset{p7}{subpath (length(p3)/2,length(p3)) of p3} \fmfiset{p8}{subpath (0,length(p4)/2) of p4} \fmfiset{p9}{subpath (length(p4)/2,length(p4)) of p4} \fmfiset{p10}{subpath (0,length(p5)/2) of p5} \fmfiset{p11}{subpath (length(p5)/2,length(p5)) of p5} \fmfcmd{fill reverse(p9)--vm4{dir 135}..{dir 0}vm3--p7--cycle withcolor {0.75 white};} \fmfi{plain}{vm4{dir 135}..{dir 0}vm3} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p6)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=5,ptext.sep=;}{p8} \fmfis{phantom,l.dist=2,label=$\scriptstyle\Box$}{p5} \end{fmfchar*}}}+ \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfiset{p6}{subpath (0,length(p3)/2) of p3} \fmfiset{p7}{subpath (length(p3)/2,length(p3)) of p3} \fmfiset{p8}{subpath (0,length(p4)/2) of p4} \fmfiset{p9}{subpath (length(p4)/2,length(p4)) of p4} \fmfiset{p10}{subpath (0,length(p5)/2) of p5} \fmfiset{p11}{subpath (length(p5)/2,length(p5)) of p5} \fmfcmd{fill reverse(p9)--vm4--vm5--p11--cycle withcolor {0.75 white};} \fmfi{plain}{vm4--vm5} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,l.side=left,l.dist=2,label=$\scriptstyle\Box$,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=5,ptext.sep=;}{p8} \end{fmfchar*}}}= 2 \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-27,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v3} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v3} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v3,v4} \fmf{plain,tension=0.5,right=0.25,fore=black}{v1,vc1} \fmf{plain,tension=0.5,left=0.25,fore=black}{v2,vc1} \fmf{plain,tension=1.25,fore=black}{vc1,vc2} \fmf{plain,tension=0.5,left=0.25,fore=black}{v3,vc2} \fmf{plain,tension=0.5,right=0.25,fore=black}{v4,vc2} \fmf{plain,tension=0.5,right=0,width=1mm,fore=black}{v3,v4} \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p3}{vpath(__vc1,__vc2)} \fmfiset{p4}{vpath(__v3,__vc2)} \fmfiset{p5}{vpath(__v4,__vc2)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vu1}{point length(p1)/3 of p1} \fmfiequ{vd1}{point 2length(p1)/3 of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vu2}{point length(p2)/3 of p2} \fmfiequ{vd2}{point 2length(p2)/3 of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vu3}{point length(p3)/3 of p3} \fmfiequ{vd3}{point 2length(p3)/3 of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vd4}{point length(p4)/3 of p4} \fmfiequ{vu4}{point 2length(p4)/3 of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vd5}{point length(p5)/3 of p5} \fmfiequ{vu5}{point 2length(p5)/3 of p5} } }\fmfiset{p6}{subpath (0,length(p3)/2) of p3} \fmfiset{p7}{subpath (length(p3)/2,length(p3)) of p3} \fmfiset{p8}{subpath (0,length(p4)/2) of p4} \fmfiset{p9}{subpath (length(p4)/2,length(p4)) of p4} \fmfiset{p10}{subpath (0,length(p5)/2) of p5} \fmfiset{p11}{subpath (length(p5)/2,length(p5)) of p5} \fmfcmd{fill reverse(p11)--vm5{dir 45}..{dir 180}vm3--p7--cycle withcolor {0.75 white};} \fmfcmd{fill p4--reverse(p5)--vloc(__v3)--cycle withcolor {0.75 white};} \fmfi{plain}{vm5{dir 45}..{dir 180}vm3} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p6)} \fmfis{phantom,l.dist=2,label=$\scriptstyle\Box$}{p4} \end{fmfchar*}}}\to 2I_2 {~,}\\ & \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-15,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v4} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v4} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v2,v3} \fmffixed{(0.25w,0)}{v4,v5} \fmffixed{(0.25w,0)}{v5,v6} \fmffixed{(0,whatever)}{vc1,vc3} \fmffixed{(0,whatever)}{vc2,vc4} \fmf{plain,tension=0.5,right=0.25}{v1,vc1} \fmf{plain,tension=0.5,left=0.25}{v2,vc1} \fmf{phantom,tension=0.5,right=0.25}{v2,vc2} \fmf{plain,tension=0.5,left=0.25}{v3,vc2} \fmf{plain,tension=0.5,left=0.25}{v4,vc3} \fmf{phantom,tension=0.5,right=0.25}{v5,vc3} \fmf{plain,tension=0.5,left=0.25}{v5,vc4} \fmf{plain,tension=0.5,right=0.25}{v6,vc4} \fmf{plain,tension=1.25,left=0}{vc1,vc3} \fmf{plain,tension=1.25,left=0}{vc2,vc4} \fmffreeze \fmf{plain,tension=1,left=0}{vc2,vc3} \fmf{plain,tension=0.5,right=0,width=1mm}{v4,v6} \fmffreeze \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p6}{vpath(__v3,__vc2)} \fmfiset{p4}{vpath(__v4,__vc3)} \fmfiset{p8}{vpath(__v5,__vc4)} \fmfiset{p9}{vpath(__v6,__vc4)} \fmfiset{p3}{vpath(__vc1,__vc3)} \fmfiset{p7}{vpath(__vc2,__vc4)} \fmfiset{p5}{vpath(__vc2,__vc3)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vm6}{point 0.5*length(p6) of p6} } {\fmfiequ{vm7}{point 0.5*length(p7) of p7} } {\fmfiequ{vm8}{point 0.5*length(p8) of p8} } {\fmfiequ{vm9}{point 0.5*length(p9) of p9} } }\fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{p4} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p5)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p6)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p7)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.out=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{p8} \end{fmfchar*}}}= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-27\unitlength}\raisebox{\eqoff}{\fmfframe(-2,2)(-15,2){\begin{fmfchar*}(50,50) {\fmftop{v1} \fmfbottom{v4} \fmfforce{(0.125w,h)}{v1} \fmfforce{(0.125w,0)}{v4} \fmffixed{(0.25w,0)}{v1,v2} \fmffixed{(0.25w,0)}{v2,v3} \fmffixed{(0.25w,0)}{v4,v5} \fmffixed{(0.25w,0)}{v5,v6} \fmffixed{(0,whatever)}{vc1,vc3} \fmffixed{(0,whatever)}{vc2,vc4} \fmf{plain,tension=0.5,right=0.25}{v1,vc1} \fmf{plain,tension=0.5,left=0.25}{v2,vc1} \fmf{phantom,tension=0.5,right=0.25}{v2,vc2} \fmf{plain,tension=0.5,left=0.25}{v3,vc2} \fmf{plain,tension=0.5,left=0.25}{v4,vc3} \fmf{phantom,tension=0.5,right=0.25}{v5,vc3} \fmf{plain,tension=0.5,left=0.25}{v5,vc4} \fmf{plain,tension=0.5,right=0.25}{v6,vc4} \fmf{plain,tension=1.25,left=0}{vc1,vc3} \fmf{plain,tension=1.25,left=0}{vc2,vc4} \fmffreeze \fmf{plain,tension=1,left=0}{vc2,vc3} \fmf{plain,tension=0.5,right=0,width=1mm}{v4,v6} \fmffreeze \fmfposition \fmfipath{p[]} \fmfipair{vd[],vm[],vu[]} \fmfiset{p1}{vpath(__v1,__vc1)} \fmfiset{p2}{vpath(__v2,__vc1)} \fmfiset{p6}{vpath(__v3,__vc2)} \fmfiset{p4}{vpath(__v4,__vc3)} \fmfiset{p8}{vpath(__v5,__vc4)} \fmfiset{p9}{vpath(__v6,__vc4)} \fmfiset{p3}{vpath(__vc1,__vc3)} \fmfiset{p7}{vpath(__vc2,__vc4)} \fmfiset{p5}{vpath(__vc2,__vc3)} {\fmfiequ{vm1}{point 0.5*length(p1) of p1} } {\fmfiequ{vm2}{point 0.5*length(p2) of p2} } {\fmfiequ{vm3}{point 0.5*length(p3) of p3} } {\fmfiequ{vm4}{point 0.5*length(p4) of p4} } {\fmfiequ{vm5}{point 0.5*length(p5) of p5} } {\fmfiequ{vm6}{point 0.5*length(p6) of p6} } {\fmfiequ{vm7}{point 0.5*length(p7) of p7} } {\fmfiequ{vm8}{point 0.5*length(p8) of p8} } {\fmfiequ{vm9}{point 0.5*length(p9) of p9} } }\fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p1)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p2)} \fmfis{phantom,ptext.in=$\scriptstyle\operatorname{D}^2$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p3)} \fmfis{phantom,l.side=left,l.dist=2,label=$\scriptstyle\Box$,ptext.clen=7,ptext.hin=2,ptext.hout=2,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p5)} \fmfis{phantom,ptext.out=$\scriptstyle\operatorname{\vphantom{\operatorname{D}}\smash[t]{\bar{\mathrm{D}}}}^2$,ptext.clen=7,ptext.hin=-10,ptext.hout=-10,ptext.oin=8,ptext.oout=8,ptext.sep=;}{reverse(p6)} \fmfcmd{fill reverse(p4)--reverse(p5)--p7--p8--vloc(__v5)--vloc(__v4)--cycle withcolor {0.75 white};} \fmfcmd{fill reverse(p8)--vloc(__v5)--vloc(__v6)--reverse(p9)--cycle withcolor {0.75 white};} \end{fmfchar*}}}\to I_2 {~,}\end{aligned}$$ where equalities hold up to disregarded finite contributions, and the final expressions in terms of the integral $I_2$ consider the aforementioned factor $-1$. =0.875mm Integrals {#app:integrals} ========= Using the scalar $G$-function defined as $$G(\alpha,\beta)= \frac{\Gamma(\tfrac{D}{2}-\alpha)\Gamma(\tfrac{D}{2}-\beta)\Gamma(\alpha+\beta-\tfrac{D}{2})}{(4\pi)^{\frac{D}{2}}\Gamma(\alpha)\Gamma(\beta)\Gamma(D-\alpha-\beta)}{~,}$$ in $D$-dimensional Euclidean space, the following integrals can be found exactly to all loop orders $$\label{IK} \begin{aligned} I_K= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-13\unitlength}\raisebox{\eqoff}{\fmfframe(-6,-2)(-1,-2){\begin{fmfchar*}(30,30) \fmfleft{in} \fmfright{out1} \fmf{phantom}{in,v1} \fmf{phantom}{out,v2} \fmfforce{(0,0.5h)}{in} \fmfforce{(w,0.5h)}{out} \fmfpoly{phantom}{v1,va4,va3,v2,va2,va1} \fmffixed{(0.75w,0)}{v1,v2} \fmf{phantom}{vc,v1} \fmf{plain}{vc,v2} \fmffreeze \fmf{phantom}{v1,va1} \fmf{plain}{va1,va2} \fmf{plain}{va2,v2} \fmf{dashes}{v2,va3} \fmf{plain}{va3,va4} \fmf{plain}{vc,va1} \fmf{plain}{vc,va2} \fmf{dashes}{vc,va3} \fmf{plain}{vc,va4} \fmffreeze \fmf{plain,left=0.5}{vc,va1} \fmfv{l=$\scriptscriptstyle 1$,l.dist=2}{va1} \fmfv{l=$\scriptscriptstyle 2$,l.dist=2}{va2} \fmfv{l=$\scriptscriptstyle 3$,l.dist=2}{v2} \fmfv{l=$\scriptscriptstyle K-1$,l.dist=2}{va3} \fmfv{l=$\scriptscriptstyle K$,l.dist=2}{va4} \end{fmfchar*}}}=\prod_{k=0}^{K-1}G(1-(\tfrac{D}{2}-2)k,1) {~.}\end{aligned}$$ They are logarithmically divergent in $D=4-2\varepsilon$ dimensions, and their overall UV divergence is obtained with the operations $\operatorname{K}$ to extract the pole part and $\operatorname{R}$ to subtract subdivergences as $$\operatorname{K}\operatorname{R}(I_K)=\operatorname{K}\Big(I_K-\sum_{k=1}^{K-1}\operatorname{K}\operatorname{R}(I_k)I_{K-k}\Big) {~.}$$ To the first few loop orders, one finds $$\label{IKpoles} \begin{aligned} \operatorname{K}\operatorname{R}(I_1)&=\frac{1}{(4\pi)^2}\frac{1}{\varepsilon} {~,}\\ \operatorname{K}\operatorname{R}(I_2)&=\frac{1}{(4\pi)^4}\Big(-\frac{1}{2\varepsilon^2}+\frac{1}{2\varepsilon}\Big) {~,}\\ \operatorname{K}\operatorname{R}(I_3)&=\frac{1}{(4\pi)^6}\Big(\frac{1}{6\varepsilon^3}-\frac{1}{2\varepsilon^2}+\frac{2}{3\varepsilon}\Big) {~,}\\ \operatorname{K}\operatorname{R}(I_4)&=\frac{1}{(4\pi)^8}\Big(-\frac{1}{24\varepsilon^4}+\frac{1}{4\varepsilon^3}-\frac{19}{24\varepsilon^2}+\frac{5}{4\varepsilon}\Big) {~,}\\ \operatorname{K}\operatorname{R}(I_5)&=\frac{1}{(4\pi)^{10}}\Big(\frac{1}{120\varepsilon^5}-\frac{1}{12\varepsilon^4}+\frac{11}{24\varepsilon^3}-\frac{19}{12\varepsilon^2}+\frac{14}{5\varepsilon}\Big) {~,}\\ \operatorname{K}\operatorname{R}(I_6)&=\frac{1}{(4\pi)^{12}}\Big(-\frac{1}{720\varepsilon^6}+\frac{1}{48\varepsilon^5}-\frac{25}{144\varepsilon^4}+\frac{47}{48\varepsilon^3}-\frac{1313}{360\varepsilon^2}+\frac{7}{\varepsilon}\Big) {~.}\\ \end{aligned}$$ The pole parts of the integrals that appear in the calculations of the four-loop dressing phase or of the wrapping interactions at critical wrapping order can very efficiently be computed by using a modified and extended version of the Gegenbauer polynomial $x$-space technique [@Chetyrkin:1980pr; @Kotikov:1995cw; @Fiamberti:2008sh]. The integral of the simplest contribution that allows us to determine the leading four-loop coefficient of the dressing phase reads $$\begin{aligned}\label{Ipoles} I_\beta&= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-7.5\unitlength}\raisebox{\eqoff}{\begin{fmfchar*}(20,15) \fmfleft{in} \fmfright{out} \fmf{plain}{in,v1} \fmf{plain,left=0.25}{v1,v2} \fmf{plain,left=0.25}{v2,v3} \fmf{plain,left=0.25}{v3,v4} \fmf{plain,left=0.25}{v4,v1} \fmf{plain,tension=0.5,right=0.25}{v1,v0,v1} \fmf{phantom}{v0,v3} \fmf{plain}{v2,v0} \fmf{plain}{v0,v4} \fmf{plain}{v3,out} \fmffixed{(0.9w,0)}{v1,v3} \fmffixed{(0,0.45w)}{v4,v2} \fmffreeze \end{fmfchar*}} {~,}\quad \operatorname{K}\operatorname{R}(I_\beta) =\frac{1}{(4\pi)^8}\Big( -\frac{1}{12\varepsilon^4}+\frac{1}{3\varepsilon^3} -\frac{5}{12\varepsilon^2} -\frac{1}{\varepsilon}\Big(\frac{1}{2}-\zeta(3)\Big)\Big) {~.}\end{aligned}$$ The terms of highest transcendentality from wrapping corrections at critical order are determined by the cake integral. This integral is logarithmically divergent for $K\ge3$ loops and reads $$\label{PK} P_K= \settoheight{\eqoff}{$\times$}\setlength{\eqoff}{0.5\eqoff}\addtolength{\eqoff}{-13\unitlength}\raisebox{\eqoff}{\fmfframe(-1,-2)(-1,-2){\begin{fmfchar*}(30,30) \fmfleft{in} \fmfright{out1} \fmf{phantom}{in,v1} \fmf{phantom}{out,v2} \fmfforce{(0,0.5h)}{in} \fmfforce{(w,0.5h)}{out} \fmfpoly{phantom}{v1,va4,va3,v2,va2,va1} \fmffixed{(0.75w,0)}{v1,v2} \fmf{phantom}{vc,v1} \fmf{plain}{vc,v2} \fmffreeze \fmf{plain}{v1,va1} \fmf{plain}{va1,va2} \fmf{plain}{va2,v2} \fmf{plain}{v2,va3} \fmf{dashes}{va3,va4} \fmf{plain}{va4,v1} \fmf{plain}{vc,va1} \fmf{plain}{vc,va2} \fmf{plain}{vc,va3} \fmf{dashes}{vc,va4} \fmf{plain}{vc,v1} \fmffreeze \fmfv{l=$\scriptscriptstyle 1$,l.dist=2}{va1} \fmfv{l=$\scriptscriptstyle 2$,l.dist=2}{va2} \fmfv{l=$\scriptscriptstyle 3$,l.dist=2}{v2} \fmfv{l=$\scriptscriptstyle 4$,l.dist=2}{va3} \fmfv{l=$\scriptscriptstyle K-1$,l.dist=2}{va4} \fmfv{l=$\scriptscriptstyle K$,l.dist=2}{v1} \end{fmfchar*}}}{~,}\qquad \operatorname{K}(P_K) =\frac{1}{(4\pi)^{2K}}\frac{1}{\varepsilon} \frac{2}{K}\binom{2K-3}{K-1}\zeta(2K-3) {~,}$$ where the pole part has been obtained in [@Broadhurst:1985vq] at generic loop order. [IV.3]{} J. Minahan, *“Review of AdS/CFT Integrability, Chapter I.1: Spin Chains in $\mathcal{N}$ = 4 SYM”*, `arxiv:1012.3983`. A. Rej, *“Review of AdS/CFT Integrability, Chapter I.3: Long-range spin chains”*, `arxiv:1012.3985`. A. Tseytlin, *“Review of AdS/CFT Integrability, Chapter II.1: Classical $AdS_5\times S^5$ string solutions”*, `arxiv:1012.3986`. T. McLoughlin, *“Review of AdS/CFT Integrability, Chapter II.2: Quantum Strings in $AdS_5\times S^5$”*, `arxiv:1012.3987`. M. Staudacher, *“Review of AdS/CFT Integrability, Chapter III.1: Bethe Ansätze and the R-Matrix Formalism”*, `arxiv:1012.3990`. P. Vieira and D. Volin, *“Review of AdS/CFT Integrability, Chapter III.3: The dressing factor”*, `arxiv:1012.3992`. L. Freyhult, *“Review of AdS/CFT Integrability, Chapter III.4: Twist states and the cusp anomalous dimension”*, `arxiv:1012.3993`. R. Janik, *“Review of AdS/CFT Integrability, Chapter III.5: Lüscher corrections”*, `arxiv:1012.3994`. Z. Bajnok, *“Review of AdS/CFT Integrability, Chapter III.6: Thermodynamic Bethe Ansatz”*, `arxiv:1012.3995`. V. Kazakov and N. Gromov, *“Review of AdS/CFT Integrability, Chapter III.7: Hirota Dynamics for Quantum Integrability”*, `arxiv:1012.3996`. C. Kristjansen, *“Review of AdS/CFT Integrability, Chapter IV.1: Aspects of Non-planarity”*, `arxiv:1012.3997`. K. Zoubos, *“Review of AdS/CFT Integrability, Chapter IV.2: Deformations, Orbifolds and Open Boundaries”*, `arxiv:1012.3998`. T. Klose, *“Review of AdS/CFT Integrability, Chapter IV.3: $\mathcal{N}$ = 6 Chern-Simons and Strings on $AdS_4 \times CP^3$”*, `arxiv:1012.3999`. J. M. Maldacena, *“[The large $N$ limit of superconformal field theories and supergravity]{}”*, , `hep-th/9711200`. S. S. Gubser, I. R. Klebanov and A. M. Polyakov, *“[Gauge theory correlators from non-critical string theory]{}”*, , `hep-th/9802109`. E. Witten, *“[Anti-de Sitter space and holography]{}”*, , `hep-th/9802150`. D. Serban and M. Staudacher, *“[Planar $\mathcal{N}=4$ gauge theory and the Inozemtsev long range spin chain]{}”*, , `hep-th/0401057`. N. Beisert, V. Dippel and M. Staudacher, *“[A novel long range spin chain and planar $\mathcal{N}=4$ super Yang-Mills]{}”*, , `hep-th/0405001`. C. Sieg and A. Torrielli, *“[Wrapping interactions and the genus expansion of the $2$- point function of composite operators]{}”*, , `hep-th/0505071`. J. Ambjorn, R. A. Janik and C. Kristjansen, *“[Wrapping interactions and a new source of corrections to the spin-chain / string duality]{}”*, , `hep-th/0510171`. W. Siegel, *“[Supersymmetric Dimensional Regularization via Dimensional Reduction]{}”*, . M. T. Grisaru, M. Rocek and W. Siegel, *“[Superloops 3, beta 0: A Calculation in $\mathcal{N}=4$ Yang-Mills Theory]{}”*, . M. T. Grisaru, M. Rocek and W. Siegel, *“[Zero Three Loop beta Function in $\mathcal{N}=4$ Superyang-Mills Theory]{}”*, . W. E. Caswell and D. Zanon, *“[Vanishing Three Loop beta Function in $\mathcal{N}=4$ Supersymmetric Yang-Mills Theory]{}”*, . W. E. Caswell and D. Zanon, *“[Zero Three Loop beta Function in the $\mathcal{N}=4$ Supersymmetric Yang-Mills Theory]{}”*, . S. Mandelstam, *“[Light Cone Superspace and the Ultraviolet Finiteness of the $\mathcal{N}=4$ Model]{}”*, . L. Brink, O. Lindgren and B. E. W. Nilsson, *“[The Ultraviolet Finiteness of the $\mathcal{N}=4$ Yang-Mills Theory]{}”*, . P. S. Howe, K. S. Stelle and P. K. Townsend, *“[The Relaxed Hypermultiplet: An Unconstrained $\mathcal{N}=2$ Superfield Theory]{}”*, . P. S. Howe, K. S. Stelle and P. K. Townsend, *“[Miraculous Ultraviolet Cancellations in Supersymmetry Made Manifest]{}”*, . S. Ferrara and B. Zumino, *“[Supergauge Invariant Yang-Mills Theories]{}”*, . C. Sieg, *“[Superspace calculation of the three-loop dilatation operator of $\mathcal{N}=4$ SYM theory]{}”*, `arxiv:1008.3351`. S. J. Gates, M. T. Grisaru, M. Rocek and W. Siegel, *“[Superspace, or one thousand and one lessons in supersymmetry]{}”*, , `hep-th/0108200`. J. A. Minahan, *“[The $SU(2)$ sector in AdS/CFT]{}”*, , `hep-th/0503143`. N. Beisert, C. Kristjansen and M. Staudacher, *“[The dilatation operator of $\mathcal{N}=4$ super Yang-Mills theory]{}”*, , `hep-th/0303060`. N. Beisert and T. Klose, *“[Long-range $gl(n)$ integrable spin chains and plane-wave matrix theory]{}”*, , `hep-th/0510124`. F. Fiamberti, A. Santambrogio, C. Sieg and D. Zanon, *“[Wrapping at four loops in $\mathcal{N}=4$ SYM]{}”*, , `arxiv:0712.3522`. F. Fiamberti, A. Santambrogio, C. Sieg and D. Zanon, *“[Anomalous dimension with wrapping at four loops in $\mathcal{N}=4$ SYM]{}”*, , `arxiv:0806.2095`. D. E. Berenstein, J. M. Maldacena and H. S. Nastase, *“[Strings in flat space and pp waves from $\mathcal{N}=4$ super Yang Mills]{}”*, , `hep-th/0202021`. J. A. Minahan and K. Zarembo, *“[The Bethe-ansatz for $\mathcal{N}=4$ super Yang-Mills]{}”*, , `hep-th/0212208`. D. J. Gross, A. Mikhailov and R. Roiban, *“[A calculation of the plane wave string Hamiltonian from $\mathcal{N}=4$ super-Yang-Mills theory]{}”*, , `hep-th/0208231`. B. Eden, C. Jarczak and E. Sokatchev, *“[A three-loop test of the dilatation operator in $\mathcal{N}=4$ SYM]{}”*, , `hep-th/0409009`. D. Anselmi, *“[The $\mathcal{N}=4$ quantum conformal algebra]{}”*, , `hep-th/9809192`. N. Beisert, *“[The $su(2|2)$ dynamic S-matrix]{}”*, , `hep-th/0511082`. D. J. Gross, A. Mikhailov and R. Roiban, *“[Operators with large R charge in $\mathcal{N}=4$ Yang-Mills theory]{}”*, , `hep-th/0205066`. N. Beisert, B. Eden and M. Staudacher, *“[Transcendentality and crossing]{}”*, , `hep-th/0610251`. G. Arutyunov, S. Frolov and M. Staudacher, *“[Bethe ansatz for quantum strings]{}”*, , `hep-th/0406256`. R. Hernandez and E. Lopez, *“[Quantum corrections to the string Bethe ansatz]{}”*, , `hep-th/0603204`. Z. Bern, M. Czakon, L. J. Dixon, D. A. Kosower and V. A. Smirnov, *“[The Four-Loop Planar Amplitude and Cusp Anomalous Dimension in Maximally Supersymmetric Yang-Mills Theory]{}”*, , `hep-th/0610248`. A. Santambrogio and D. Zanon, *“[Exact anomalous dimensions of $\mathcal{N}=4$ Yang-Mills operators with large R charge]{}”*, , `hep-th/0206079`. B. Eden, *“[A two-loop test for the factorised S-matrix of planar $\mathcal{N}=4$]{}”*, , `hep-th/0501234`. N. Beisert, T. McLoughlin and R. Roiban, *“[The Four-Loop Dressing Phase of $\mathcal{N}=4$ SYM]{}”*, , `arxiv:0705.0321`. F. Fiamberti, A. Santambrogio and C. Sieg, *“[Five-loop anomalous dimension at critical wrapping order in $\mathcal{N}=4$ SYM]{}”*, `arxiv:0908.0234`. K. Konishi, *“[Anomalous Supersymmetry Transformation of Some Composite Operators in SQCD]{}”*, . D. Anselmi, M. T. Grisaru and A. Johansen, *“[A Critical Behaviour of Anomalous Currents, Electric- Magnetic Universality and CFT$_4$]{}”*, , `hep-th/9601023`. D. Anselmi, D. Z. Freedman, M. T. Grisaru and A. A. Johansen, *“[Universality of the operator product expansions of SCFT(4)]{}”*, , `hep-th/9608125`. M. Bianchi, S. Kovacs, G. Rossi and Y. S. Stanev, *“[On the logarithmic behavior in $\mathcal{N}=4$ SYM theory]{}”*, , `hep-th/9906188`. M. Bianchi, S. Kovacs, G. Rossi and Y. S. Stanev, *“[Anomalous dimensions in $\mathcal{N}=4$ SYM theory at order g\*\*4]{}”*, , `hep-th/0003203`. G. Arutyunov, B. Eden, A. C. Petkou and E. Sokatchev, *“[Exceptional non-renormalization properties and OPE analysis of chiral four-point functions in $\mathcal{N}=4$ SYM(4)]{}”*, , `hep-th/0103230`. M. Bianchi, S. Kovacs, G. Rossi and Y. S. Stanev, *“[Properties of the Konishi multiplet in $\mathcal{N}=4$ SYM theory]{}”*, , `hep-th/0104016`. A. V. Kotikov, L. N. Lipatov and V. N. Velizhanin, *“[Anomalous dimensions of Wilson operators in $\mathcal{N}=4$ SYM theory]{}”*, , `hep-ph/0301021`. A. V. Kotikov, L. N. Lipatov, A. I. Onishchenko and V. N. Velizhanin, *“[Three-loop universal anomalous dimension of the Wilson operators in $\mathcal{N}=4$ SUSY Yang-Mills model]{}”*, , `hep-th/0404092`. S. Moch, J. A. M. Vermaseren and A. Vogt, *“[The three-loop splitting functions in QCD: The non-singlet case]{}”*, , `hep-ph/0403192`. A. V. Kotikov and L. N. Lipatov, *“[DGLAP and BFKL evolution equations in the $\mathcal{N}=4$ supersymmetric gauge theory]{}”*, , `hep-ph/0208220`. F. Fiamberti, A. Santambrogio and C. Sieg, *“[Superspace methods for the computation of wrapping effects in the standard and beta-deformed $\mathcal{N}=4$ SYM]{}”*, `arxiv:1006.3475`. T. Fischbacher, T. Klose and J. Plefka, *“[Planar plane-wave matrix theory at the four loop order: Integrability without BMN scaling]{}”*, , `hep-th/0412331`. Z. Bajnok and R. A. Janik, *“[Four-loop perturbative Konishi from strings and finite size effects for multiparticle states]{}”*, , `arxiv:0807.0399`. V. N. Velizhanin, *“[The Four-Loop Konishi in $\mathcal{N}=4$ SYM]{}”*, `arxiv:0808.3832`. N. Gromov, V. Kazakov, A. Kozak and P. Vieira, *“[Integrability for the Full Spectrum of Planar AdS/CFT II]{}”*, `arxiv:0902.4458`. N. Gromov, V. Kazakov and P. Vieira, *“[Exact Spectrum of Anomalous Dimensions of Planar $\mathcal{N}=4$ Supersymmetric Yang-Mills Theory]{}”*, , `arxiv:0901.3753`. D. Bombardelli, D. Fioravanti and R. Tateo, *“[Thermodynamic Bethe Ansatz for planar AdS/CFT: a proposal]{}”*, , `arxiv:0902.3930`. G. Arutyunov and S. Frolov, *“[On String S-matrix, Bound States and TBA]{}”*, , `arxiv:0710.1568`. G. Arutyunov and S. Frolov, *“[String hypothesis for the $AdS_5 \times S^5$ mirror]{}”*, , `arxiv:0901.1417`. Z. Bajnok, R. A. Janik and T. Lukowski, *“[Four loop twist two, BFKL, wrapping and strings]{}”*, , `arxiv:0811.4448`. Z. Bajnok, A. Hegedus, R. A. Janik and T. Lukowski, *“[Five loop Konishi from AdS/CFT]{}”*, , `arxiv:0906.4062`. G. Arutyunov, S. Frolov and R. Suzuki, *“[Five-loop Konishi from the Mirror TBA]{}”*, , `arxiv:1002.1711`. J. Balog and A. Hegedus, *“[5-loop Konishi from linearized TBA and the XXX magnet]{}”*, , `arxiv:1002.4142`. T. Lukowski, A. Rej and V. N. Velizhanin, *“[Five-Loop Anomalous Dimension of Twist-Two Operators]{}”*, , `arxiv:0912.1624`. M. Beccaria, V. Forini, T. Lukowski and S. Zieme, *“[Twist-three at five loops, Bethe Ansatz and wrapping]{}”*, , `arxiv:0901.4864`. V. N. Velizhanin, *“[Six-Loop Anomalous Dimension of Twist-Three Operators in $\mathcal{N}=4$ SYM]{}”*, `arxiv:1003.4717`. V. N. Velizhanin, *“[Leading transcedentality contributions to the four-loop universal anomalous dimension in $\mathcal{N}=4$ SYM]{}”*, `arxiv:0811.0607`. F. Fiamberti, A. Santambrogio, C. Sieg and D. Zanon, *“[Finite-size effects in the superconformal beta-deformed $\mathcal{N}=4$ SYM]{}”*, , `arxiv:0806.2103`. F. Fiamberti, A. Santambrogio, C. Sieg and D. Zanon, *“[Single impurity operators at critical wrapping order in the beta-deformed $\mathcal{N}=4$ SYM]{}”*, , `arxiv:0811.4594`. M. Beccaria and G. F. De Angelis, *“[On the wrapping correction to single magnon energy in twisted N=4 SYM]{}”*, `arxiv:0903.0778`. N. Gromov and F. Levkovich-Maslyuk, *“[Y-system and beta-deformed $\mathcal{N}=4$ Super-Yang-Mills]{}”*, `arxiv:1006.5438`. G. Arutyunov, M. de Leeuw and S. J. van Tongeren, *“[Twisting the Mirror TBA]{}”*, , `arxiv:1009.4118`. K. G. Chetyrkin, A. L. Kataev and F. V. Tkachov, *“[New Approach to Evaluation of Multiloop Feynman Integrals: The Gegenbauer Polynomial $x$ Space Technique]{}”*, . A. V. Kotikov, *“[The Gegenbauer Polynomial Technique: the evaluation of a class of Feynman diagrams]{}”*, , `hep-ph/9512270`. D. J. Broadhurst, *“[Evaluation of a class of Feynman diagrams for all numbers of loops and dimensions]{}”*, . [^1]: The one-loop results are reviewed in chapter [@chapChain]. [^2]: This holds apart from gauge artefacts that are not relevant here. [^3]: See chapter [@chapObserv] for a review concerning effects of non-planarity. [^4]: The flavor $SO(6)$ subsector is only closed to one loop. [^5]: The explicit Feynman diagram calculation in [@Sieg:2010tz] confirms that this is correct to three loops. It is non-trivial in the ${\text{AdS}}_4/{\text{CFT}}_3$ correspondence that is reviewed in chapter [@chapN6]. [^6]: The dressing phase is reviewed in chapter [@chapSProp]. [^7]: This breakdown is independent of the general restriction of the Bethe ansatz to the asymptotic regime that requires a termination of the expansion at a loop order $K\le L-1$ to avoid the wrapping corrections. [^8]: A two-loop test of the S-matrix of the $SL(2)$ subsector can be found in [@Eden:2005bt]. [^9]: At four and higher loops this is no longer the case since the wrapping interactions have to be considered. This will be discussed in Section \[sec:wrapint\]. [^10]: In chapter [@chapObserv] the analyses of higher genus contributions are reviewed. [^11]: In the context of the BMN matrix model a subtraction attempt was made in [@Fischbacher:2004iu]. It does not lead to the correct result, since the necessary modifications of the contributions with permutation structures of lower range and the addition of the wrapping diagrams was not performed. [^12]: Among other deformations the $\beta$-deformation is reviewed in chapter [@chapDeform].

\#1[\#1]{} \#1 =21.5cm =15.8cm plus 1pt makefntext\#1[to 1em[$^{\@thefnmark}$]{}\#1]{} [**Constraints on left–right symmetric models\ from the process ${b \rightarrow s \gamma}$**]{}\ [K.S. Babu]{}\ [*Bartol Research Institute, University of Delaware\ Newark, DE 19716, U.S.A.*]{} [Kazuo Fujikawa]{} and [Atsushi Yamada]{}\ [**ABSTRACT**]{} In left-right symmetric models, large contributions to the decay amplitude ${b \rightarrow s \gamma}$ can arise from the mixing of the $W_L$ and $W_R$ gauge bosons as well as from the charged Higgs boson. These amplitudes are enhanced by the factor ${m_{t}}/{m_{b}}$ compared to the contributions in the standard model. We use the recent CLEO results on the radiative $B$ decay to place constraints on the $W_L-W_R$ mixing angle $\zeta$ and the mass of the charged Higgs boson $m_{H^\pm}$. Significant departures from the standard model predictions occur when $|\zeta| \stackrel{_>}{_\sim} 0.003$ and/or when $m_{H^\pm} \stackrel{_<}{_\sim}$ a few TeV. = 0.7cm Introduction: {#introduction .unnumbered} ============= Left–right symmetric theories of the weak interactions based on the gauge group $SU(2)_L \times SU(2)_R \times U(1)_Y$ are attractive extensions of the standard model possessing manifest parity invariance [@pati]. These theories also have greater quark–lepton symmetry than the standard model since they require the existence of the right–handed partner of the neutrino $\nu_R$, leading naturally to non–zero neutrino masses. The observed (V-A) nature of the weak interactions is explained by the spontaneous breaking of parity along with the breaking of $SU(2)_R$ gauge symmetry at a scale $v_R \gg m_W$. If the scale $v_R$ of $SU(2)_R$ breaking is not much above the weak scale, observable deviations from the predictions of the standard model are possible. Flavor changing neutral current processes have proven in the past to be powerful probes of physics beyond the standard model. For example, in the context of the left–right symmetric models, the mass of the charged $W_R$ gauge boson should exceed about $1.6$ TeV, or else it would contribute to the $K^0-\overline{K}^0$ mass difference at an unacceptable level [@beall]. In this paper we study the constraints on the parameters of the left–right symmetric model arising from the process $b \rightarrow s\gamma$. Recently the CLEO collaboration has reported the first observation of the exclusive decay $B \rightarrow K^*\gamma$ with a branching ratio [@cleo] $$Br(B \rightarrow K^* \gamma) = (4.5 \pm 1.5 \pm 0.9) \times 10^{-5}~. \label{eqn:bound}$$ Eq. (\[eqn:bound\]) implies both lower and an upper limits on the inclusive decay $B\rightarrow X_s\gamma$ \[e.g. $Br (B \rightarrow X_s\gamma) < 5.4 \times 10^{-5}$ at 95% C.L.\]. These numbers are in good agreement with the standard model predictions and as such, are sensitive to new physics. In renormalizable gauge theories, the radiative decay ${b \rightarrow s \gamma}$ proceeds through the magnetic moment operators $\overline{b}_R\sigma_{\mu \nu} s_L F^{\mu \nu}$ and $\overline{b}_L\sigma_{\mu \nu} s_R F^{\mu \nu}$, where $F^{\mu \nu}$ is the electromagnetic field strength tensor. In the standard model, the ${b \rightarrow s \gamma}$ amplitude is proportional to $m_b$ or $m_s$, the mass of the bottom quark or the strange quark, because the pure (V-A) structure of the charged currents requires the chirality-flip to proceed only through the mass of the initial or the final state quark. In contrast, in left-right symmetric models, the mixing of the $W_L$ and $W_R$ gauge bosons leads also to (V+A) interactions between the $W_1$ boson and the quarks, where $W_1$ is the lighter mass eigenstate formed by $W_L$ and $W_R$. In this case, the ${b \rightarrow s \gamma}$ amplitude can be proportional to the top quark mass ${m_{t}}$ rather than ${m_{b}}$ or ${m_{s}}$ since chirality flip can now occur with the top quark mass in the intermediate state. This enhancement of the amplitude gives rise to significant departure of the decay rate $Br(b \rightarrow s\gamma)$ from the prediction in the standard model, if the $W_L-W_R$ mixing angle $\zeta$ exceeds about $10^{-3}$. Left–right symmetric models also predict the existence of a charged Higgs boson that couples to the quarks. Its contributions to the ${b \rightarrow s \gamma}$ amplitude are also proportional to the top quark mass, and the experimental result (\[eqn:bound\]) already probes the charged Higgs mass of a few TeV. This feature should be compared to the charged Higgs contributions to the ${b \rightarrow s \gamma}$ amplitude in the minimal supersymmetric standard model (MSSM) which are proportional to $m_b$ or $m_s$. The experimental result (\[eqn:bound\]) excludes a charged Higgs boson lighter than a few hundred GeV in this case [@berger]. The enhancement of the charged Higgs boson contributions in left-right symmetric models stems from the absence of natural flavor conservation in the Higgs sector of the model. In spite of the absence of flavor conservation in left–right symmetric models, the interactions of the charged Higgs boson to the quarks are determined in terms of the quark masses, Cabibbo-Kobayashi-Maskawa (CKM) mixing angles and the ratio of vacuum expectation values, just as in the MSSM. Radiative $b$–decays have been studied in the context of left–right symmetric models in the past in Refs. [@coco; @asa]. The effects of the $W_L-W_R$ mixing on the ${b \rightarrow s \gamma}$ amplitude were studied in Ref. [@coco], but the contributions from the charged Higgs boson were not examined there. Moreover, our result on the contributions of the $W_L-W_R$ mixing disagrees with that in Ref. [@coco]. The charged Higgs contributions were analyzed in Ref. [@asa], but the effects of the $W_L-W_R$ mixing and the leading QCD corrections were not included. In realistic left-right models, large contributions to ${b \rightarrow s \gamma}$ amplitude arising from the $W_L-W_R$ mixing and those from the charged Higgs boson are closely related to each other and they can be simultaneously sizable. Here we present a comprehensive analysis of both of these contributions taking into account their correlations and clarify the implications of the recent ${b \rightarrow s \gamma}$ experiment on the parameters in the left–right symmetric models. Left-right symmetric models: {#left-right-symmetric-models .unnumbered} ============================ Left-right symmetric models of weak interactions are based on the gauge group $SU(2)_L \times SU(2)_R \times U(1)_{B-L}$. The quarks ($q$) and leptons ($l$) transform under the gauge group as $$\begin{aligned} q_L(2,1,{1 \over 3}) &=& \left(\matrix{u \cr d}\right)_L;~~ q_R(1,2,{1 \over 3}) = \left(\matrix{u \cr d}\right)_R \nonumber \\ l_L(2,1,-1) &=& \left(\matrix{\nu \cr e}\right)_L;~~ l_R(1,2,-1) = \left(\matrix{\nu \cr e}\right)_R\end{aligned}$$ where generation indices have been suppressed. The minimal Higgs sector compatible with the see–saw mechanism for small neutrino masses consists of the multiplets [@rnm] $\Delta_L(3,1,2)$, $\Delta_R(1,3,2)$ and $\Phi(2,2,0)$ which in component form read as $$\begin{aligned} \Delta_{L,R} = \left( \begin{array}{cc} \delta^+/\sqrt{2} & \delta^{++} \\ \delta^0 & -\delta^+/\sqrt{2} \end{array} \right)_{L,R}, \Phi= \left( \begin{array}{cc} \phi^0_1 & \phi^+_2 \\ \phi^-_1 & \phi^0_2 \end{array} \right)~. \label{eqn:higgs}\end{aligned}$$ The field $\Delta_R$ is needed for breaking the gauge symmetry $SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ down to the gauge symmetry in the standard model and to give Majorana masses to the right-handed neutrinos. The field $\Phi$ is required for generating the quark and lepton masses. The field $\Delta_L$ is present in the theory to maintain the discrete parity invariance. Under parity transformation, $q_L \rightarrow q_R, l_L \rightarrow l_R, \Delta_L \rightarrow \Delta_R, \Phi \rightarrow \Phi^{\dagger}$ and $W_L \rightarrow W_R$. Spontaneous breaking of the gauge symmetry $SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ down to $U(1)_{EM}$ is achieved by the vacuum expectation values (VEV) of the neutral Higgs fields denoted by $$\begin{aligned} \VEV{\Delta_{L,R}} = \left( \begin{array}{cc} 0 & 0 \\ v_{L,R} & 0 \end{array} \right),~~ \VEV{\Phi}= \left( \begin{array}{cc} k & 0 \\ 0 & k' \end{array} \right) \label{eqn:vev}~~.\end{aligned}$$ Among the vacuum expectation values $k$, $k'$ and $v_{L,R}$, the hierarchy $k,k^\prime \ll v_R$ is needed to preserve the success of the standard (V-A) theory. In this case, another hierarchy $v_L \ll k,k^\prime$ follows from a detailed analysis of the Higgs potential [@rnm] which yields the relation $v_L \sim \gamma k^2/v_R$, where $\gamma$ is some combination of the Higgs quartic coupling constants. This is a welcome result since the analysis of the electroweak $\rho$–parameter leads to the constraint $v_L {\hbox{ \raise3pt\hbox to 0pt{$<$}\raise-3pt\hbox{$\sim$} }}10$ GeV [@rev] and a natural realization of the see–saw mechanism for small neutrino masses requires $v_L {\hbox{ \raise3pt\hbox to 0pt{$<$}\raise-3pt\hbox{$\sim$} }}$ a few MeV. In what follows, we shall work in the limit $v_L \rightarrow 0$, which is justified for the above reasons. The VEVs $k$ or $k^\prime$ can in general have a phase, but we shall assume this phase to be small. This is also justified from the detailed analysis of the Higgs potential [@wolf]. Small non–zero values of $v_L$ or the relative phase will not alter our conclusions. The Yukawa Lagrangian involving the quark fields is given by $${\cal L}_Y = \overline{q}_L h \Phi q_R + \overline{q}_L \tilde{h} \Tilde{\Phi}q_R + h.c., \label{eqn:Yuk}$$ where $\Tilde{\Phi} \equiv \tau_2 \Phi^* \tau_2$, $h$ and $\tilde{h}$ are $3 \times 3$ hermitian matrices in generation space. Eq. (5) leads to the following mass matrices for the up–type and down–type quarks: $$M_u = h k + \Tilde{h} k^\prime~,~~M_d = h k^\prime+\Tilde{h} k~.$$ In the charged gauge boson sector, the $W_L^\pm$ and $W_R^\pm$ mix with their mass–squared matrix given by $$\begin{aligned} {\cal M}^2 = {{g^2}\over 2} \left(\matrix{k^2+k^{\prime^2} & -2 k k^\prime \cr -2kk^\prime & 2v_R^2+k^2+k^{\prime^2}}\right)~~.\end{aligned}$$ The two mass eigenstates are $$\begin{aligned} W_1^\pm &=& c_\zeta W_L^\pm + s_\zeta W_R^\pm ~,\nonumber \\ W_2^\pm &=& -s_\zeta W_L^\pm + c_\zeta W_R^\pm~,\end{aligned}$$ where $s_\zeta = {\rm sin}\zeta$, $c_\zeta={\rm cos}\zeta$, respectively, and $${\rm tan}2\zeta = {{2kk^\prime}\over {v_R^2}}~. \label{eqn:tanzeta}$$ We have defined $m_{W_1} \le m_{W_2}$ with $m_{W_1} \simeq 80$ GeV. The mass eigenvalues of $W_1^\pm$ and $W_2^\pm$ are given by $$\begin{aligned} {m_{W_1}}^2 = \frac{g^2}{2}(k^2+k'^2 - 4 kk' {s_\zeta}{c_\zeta}+ 2 v^2_R {s_\zeta}^2), \,\,\,\,\,\,\, {m_{W_2}}^2 = \frac{g^2}{2}(k^2+k'^2 + 4 kk' {s_\zeta}{c_\zeta}+ 2 v^2_R {c_\zeta}^2). \label{eqn:mwlh}\end{aligned}$$ The coupling of the lighter charged $W_1$–boson to the quarks is given by $$\begin{aligned} {\cal L}_{W_1}=\frac{g}{2\sqrt{2}} \left( \begin{array}{ccc} \bar{u}, & \bar{c}, & \bar{t} \end{array} \right) \{ {c_\zeta}\gamma^\mu (1-\gamma_5)+ {s_\zeta}\gamma^\mu (1+\gamma_5) \} {W_1}^+_\mu V \left( \begin{array}{c} d \\ s \\ b \end{array} \right) + h.c., \label{eqn:lw1}\end{aligned}$$ where $V$ is the Cabibbo-Kobayashi-Maskawa (CKM) matrix. Note that since the up and down mass matrices are hermitian (as the VEVs $k,k^\prime$ are taken to be real), the right–handed CKM matrix and the left–handed CKM matrix are equal ($V_L = V_R = V$), which is reflected in Eq. (11). The interaction (5) leads to the following coupling of the corresponding (unphysical) Nambu-Goldstone boson $G_1$ to the quarks, $$\begin{aligned} {\cal L}_{G_1}= \frac{g}{2 \sqrt{2} {m_{W_1}}} \left( \begin{array}{ccc} \bar{u}, & \bar{c}, & \bar{t} \end{array} \right) {c_\zeta}\{ (1-\gamma_5) D_u V - (1+\gamma_5) V D_d \} {G^+_1} \left( \begin{array}{c} d \\ s \\ b \end{array} \right) \nonumber \\ + \frac{g}{2 \sqrt{2} {m_{W_1}}} \left( \begin{array}{ccc} \bar{u}, & \bar{c}, & \bar{t} \end{array} \right) {s_\zeta}\{ (1+\gamma_5) D_u V - (1-\gamma_5) V D_d \} {G^+_1} \left( \begin{array}{c} d \\ s \\ b \end{array} \right) + h.c., \label{eqn:lg1}\end{aligned}$$ where we define the diagonal mass matrices $D_u=diag(m_u, m_c, {m_{t}})$ and $D_d=diag(m_d, m_s, {m_{b}})$. The couplings of the heavier $W_2^\pm$ gauge boson and the analogous Nambu–Goldstone boson $G_2$ can be obtained from Eqs. (11)-(12) by the replacement $s_\zeta \rightarrow c_\zeta, c_\zeta \rightarrow -s_\zeta$ and $\gamma_5 \rightarrow -\gamma_5$, but are not necessary for the present analysis. The model has two physical charged Higgs bosons. In the limit of $v_L \rightarrow 0$, one of them, $\delta_L^+$, becomes mass eigenstate by itself. The Higgs boson $\delta_L^+$ has couplings only to the leptons, and does not enter into the discussion of the ${b \rightarrow s \gamma}$ amplitude. The second physical Higgs boson, $H^\pm$, which is the linear combination orthogonal to $G_{1,2}^\pm$, $$\begin{aligned} H^\pm = N_{H^+}\left[k^\prime \phi_1^\pm + k \phi_2^\pm + {{(k^2-k^{\prime^2})}\over {\sqrt{2}v_R}} \delta_R^\pm\right],\end{aligned}$$ has the following Yukawa coupling to the quarks: $$\begin{aligned} & & {\cal L}_{{H^{+}}}= -\frac{\sin(2\beta)N_{{H^{+}}}}{2\cos(2\beta)} \left( \begin{array}{ccc} \bar{u}, & \bar{c}, & \bar{t} \end{array} \right) \{ (1-\gamma_5) D_u V - (1+\gamma_5) V D_d \} {H^{+}}\left( \begin{array}{c} d \\ s \\ b \end{array} \right) \nonumber \\ & & -\frac{N_{{H^{+}}}}{2\cos(2\beta)} \left( \begin{array}{ccc} \bar{u}, & \bar{c}, & \bar{t} \end{array} \right) \{ (1+\gamma_5) D_u V - (1-\gamma_5) V D_d \} {H^{+}}\left( \begin{array}{c} d \\ s \\ b \end{array} \right) + h.c., \label{eqn:lch}\end{aligned}$$ where $$\begin{aligned} N_{{H^{+}}}=1/\sqrt{k^2+k'^2 +\frac{(k^2-k'^2)^2}{2v^2_R}},\,\,\,\,\, \tan\beta=k/k'.\end{aligned}$$ The mass of this charged Higgs boson is dependent on the detailed structure of the Higgs potential, and we leave it as a free parameter in our analysis. The process [${b \rightarrow s \gamma}$]{}: {#the-process-b-rightarrow-s-gamma .unnumbered} =========================================== We now investigate the effective Hamiltonian describing ${b \rightarrow s \gamma}$ in the left-right symmetric model. Using the Lagrangians (\[eqn:lw1\]), (\[eqn:lg1\]) and (\[eqn:lch\]), the effective Hamiltonian for ${b \rightarrow s \gamma}$ decay can be written as $$H_{eff} = \frac{e}{16\pi^2} \frac{2 G_F}{\sqrt{2}} V_{tb} V^*_{ts} m_b (A_L \bar{s}_L \sigma^{\mu\nu} b_R + A_R\bar{s}_R \sigma^{\mu\nu} b_L) F_{\mu\nu}, \label{eqn:heff}$$ with $$\begin{aligned} A_L&=&A_{SM}(x) + \zeta \frac{{m_{t}}}{{m_{b}}} A_{RH}(x) + \frac{{m_{t}}{s_{2\beta}}}{{m_{b}}{c_{2\beta}}^2} A^1_{{H^{+}}}(y) + \tan^2(2\beta) A^2_{{H^{+}}}(y), \label{eqn:al} \\ A_R&=&\zeta \frac{{m_{t}}}{{m_{b}}} A_{RH}(x) + \frac{{m_{t}}{s_{2\beta}}}{{m_{b}}{c_{2\beta}}^2} A^1_{{H^{+}}}(y) + \frac{1}{{c_{2\beta}}^2} A^2_{{H^{+}}}(y) {}~. \label{eqn:ar}\end{aligned}$$ Here the masses of the light quarks $u$, $d$, $s$ and $c$ have been neglected and the approximations ${c_\zeta}\simeq 1$, ${s_\zeta}\simeq \zeta$ and $N_{{H^{+}}} \simeq g/(\sqrt{2}{m_{W}}) $ have been used. In eq. (\[eqn:heff\]), $F_{\mu\nu}$ is the electromagnetic field strength tensor, $x={m_{t}}^2/{m_{W}}^2$, $y={m_{t}}^2/{m_{{H^{+}}}}^2$, $s_{2\beta} = {\rm sin}2\beta$ and $c_{2\beta}={\rm cos}2\beta$. The functions $A_{SM}$, $A_{RH}$, $A^1_{{H^{+}}}(y)$ and $A^2_{{H^{+}}}(y)$ are found to be $$\begin{aligned} A_{SM}(x) &=& {\frac{1}{(1-x)^4}}{Q_t}\{ \frac{x^4}{4} - \frac{3}{2}x^3 + \frac{3}{4}x^2 + \frac{x}{2} + \frac{3}{2}x^2 \log(x) \} \label{eqn:asm} \nonumber\\ &+& {\frac{1}{(1-x)^4}}\{ \frac{x^4}{2} + \frac{3}{4}x^3 - \frac{3}{2}x^2 + \frac{x}{4} - \frac{3}{2}x^3 \log(x) \}, \nonumber\\ A_{RH}(x) &=& {\frac{1}{(1-x)^3}}{Q_t}\{ - \frac{x^3}{2} - \frac{3}{2}x + 2 + 3x\log(x) \} \nonumber\\ &+& {\frac{1}{(1-x)^3}}\{ - \frac{x^3}{2} + 6x^2 - \frac{15}{2}x + 2 - 3x^2 \log(x) \}. \label{eqn:arh}\nonumber\\ A^1_{{H^{+}}}(y) &=& {\frac{1}{(1-y)^3}}{Q_t}\{ - \frac{y^3}{2} +2y^2 - \frac{3}{2}y - y\log(y) \} + {\frac{1}{(1-y)^3}}\{ - \frac{y^3}{2} + \frac{y}{2} + y^2 \log(y) \}, \label{eqn:ach}\nonumber \\ A^2_{{H^{+}}}(y) &=& {1 \over 3}A_{SM}(y)-A^1_{{H^{+}}}(y),\end{aligned}$$ where $Q_t=2/3$ is the electric charge of the top–quark. $A_{SM}$ in eq. (\[eqn:al\]) is the contribution given by the standard model [@inamilim]. The right-handed coupling in eq. (\[eqn:lw1\]) leads to the contributions $({m_{t}}/{m_{b}})\zeta A_{RH}$ [@fy] in eqs. (\[eqn:al\]) and (\[eqn:ar\]). These contributions arise from the chirality flip induced by the top quark mass in the intermediate state, and they are enhanced by the factor ${m_{t}}/{m_{b}}$ compared to the standard model. Chirality-flip inside the loop is forbidden in the standard model because of the purely left-handed nature of the W-boson coupling to the quarks. The last two terms in eqs. (\[eqn:al\]) and (\[eqn:ar\]) are the contributions from the charged Higgs boson ${H^{+}}$. Note that these are also enhanced by the factor ${m_{t}}/{m_{b}}$ compared to the contributions in the standard model. This is in contrast with the charged Higgs contributions in the minimal supersymmetric standard model (MSSM), which are proportional to ${m_{b}}$ or ${m_{s}}$, without any large enhancement factor [@berger]. It is worthwhile to emphasize the correlation between the $W_L-W_R$ mixing contributions and the charged Higgs contributions in eqs. (\[eqn:al\]) and (\[eqn:ar\]). From the expressions (\[eqn:tanzeta\]) and (\[eqn:mwlh\]), the mixing angle $\zeta$ can be written as $\zeta \simeq \sin(2\beta)\cdot(m_{W_1}^2/m_{W_2}^2)$. Therefore the large contributions coming from the $W_L-W_R$ mixing and the charged Higgs boson are proportional to each other and they can be sizable if $k$ and $k'$ are of the same order. The effective Hamiltonian $H_{eff}$ given in eq. (\[eqn:heff\]) has been evaluated at the electroweak scale ($\mu \sim m_W$). To make contact with the ${b \rightarrow s \gamma}$ decay, $H_{eff}$ should be evolved down to lower momentum scale ($\mu \sim m_b$) by the renormalization group analysis. The leading QCD corrections to the Hamiltonian (\[eqn:heff\]) during its evolution turn out to be significant [@fran; @qcd]. These QCD corrections have been computed in the standard model in Refs. [@gsw; @ita] by analyzing the operator mixing between the magnetic moment operator $\bar{s}_L \sigma^{\mu\nu} b_R F_{\mu\nu}$ in eq. (\[eqn:heff\]) and the four Fermi operators involving the quarks lighter than the $W$ bosons. In left-right models, because of the $W_L-W_R$ mixing, there exists some new four-Fermi operators which mix with the magnetic moment operator $\bar{s}_L \sigma^{\mu\nu} b_R F_{\mu\nu}$. However, the effects of these new operators are simply order $\zeta$ [*without the enhancement factor of*]{} ${m_{t}}/{m_{b}}$, and are negligible in our analysis. The running of the strong coupling constants and the effects of the operator mixing are also negligible in the momentum region above the $W_1$ boson mass, because of the asymptotic freedom of the strong interactions, and consequently the effects of new scales characterized by ${m_{t}}$ and ${m_{{H^{+}}}}$ are ignored. The QCD corrections to the other magnetic moment operator $\bar{s}_R \sigma^{\mu\nu} b_L F_{\mu\nu}$ can be computed in analogy to the case of the operator $\bar{s}_L \sigma^{\mu\nu} b_R F_{\mu\nu}$ because the strong interactions respect parity. Therefore, we compute the QCD corrections to the Hamiltonian (\[eqn:heff\]) following the procedure of Ref. [@gsw; @ita] established in the standard model. Here we use the simplified analytical results which are exact to within a few percent [@gsw]. Including the QCD corrections, $A_L$ and $A_R$ in the effective Hamiltonian (\[eqn:heff\]) are renormalized as $$\begin{aligned} A^{eff}_L &=& \eta^{-32/23} \{ A_L + \frac{3}{10}X(\eta^{10/23}-1) + \frac{3}{28}X(\eta^{28/23}-1)\}, \nonumber \\ \,\,\,A^{eff}_R &=& \eta^{-32/23}A_R. \label{eqn:alreff}\end{aligned}$$ with $X=208/81$ [@ita] and $\eta=\alpha_s({m_{b}}^2)/\alpha_s({m_{W}}^2) \simeq 1.8$. The last two terms in $A^{eff}_L$ come from the operator mixing in the standard model. Analogous contributions to $A^{eff}_R$ in the standard model are proportional to the mass of the strange quarks and are neglected in our analysis. The branching fraction ${Br(b \rightarrow s \gamma)}$ is computed following the procedure of Ref. [@fran] by normalizing the decay width ${\Gamma(b \rightarrow s \gamma)}$ to the semileptonic decay width ${\Gamma(b \rightarrow ce\bar{\nu})}$, $$\begin{aligned} {Br(b \rightarrow s \gamma)}= \frac{{\Gamma(b \rightarrow s \gamma)}}{{\Gamma(b \rightarrow ce\bar{\nu})}} {Br(b \rightarrow ce\bar{\nu})}, \label{eqn:bbsg}\end{aligned}$$ and using ${Br(b \rightarrow ce\bar{\nu})}\simeq 11 \%$ [@pd]. Using the Hamiltonian (\[eqn:heff\]) with renormalized quantities $A^{eff}_{L,R}$, the rate for ${\Gamma(b \rightarrow s \gamma)}$ normalized to the semileptonic rate is given by, $$\begin{aligned} & &\frac{{\Gamma(b \rightarrow s \gamma)}}{{\Gamma(b \rightarrow ce\bar{\nu})}} = \frac{3\alpha}{2\pi\rho({m_{c}}/{m_{b}})(1-\delta_{QCD})} ( |A^{eff}_L|^2 + |A^{eff}_R|^2 ) \label{eqn:rate}\end{aligned}$$ In eq. (\[eqn:rate\]), $\rho({m_{c}}/{m_{b}})$ and $\delta_{QCD}$ are the phase space suppression factor and the QCD corrections to the semileptonic decay, respectively. These factors are evaluated as $\rho({m_{c}}/{m_{b}})=0.447$ and $\delta_{QCD} = (2 \alpha_s({m_{b}}^2)/3\pi)f({m_{c}}/{m_{b}})$, where $f({m_{c}}/{m_{b}}) = 2.41$ [@lpqcd]. (We take $\alpha_s({m_{b}}^2) =0.23$.) We investigate implications of eqs. (\[eqn:bbsg\]) and (\[eqn:rate\]) numerically. In Fig. 1, we plot the branching ratio as a function of the left–right mixing angle $\zeta$ for three different values of the top–quark mass, $m_t$=110 GeV, 140 GeV and 170 GeV. The value of $m_b$ appearing in the enhancement factor $m_t/m_b$ in eqs. (\[eqn:al\]) and (\[eqn:ar\]) is the one evaluated at $\mu \sim m_t$. We choose it to be 3 GeV corresponding to a pole mass of $4.8$ GeV. Here we have kept the charged Higgs boson mass $m_{H^\pm}$ rather high ($m_{H^\pm}$ = 20 TeV). In this case the contributions of the charged Higgs boson are negligible. The CLEO limit on the branching ratio, $Br({b \rightarrow s \gamma}) \le 5.4 \times 10^{-5}$ implies a limit $-0.015 \le \zeta \le 0.003$, with the region around $\zeta = -0.005$ disfavored. This limit should be compared to the existing bounds on $\zeta$ in left–right models. A limit $|\zeta| \le 0.035$ has been inferred from the measurement of the $\xi$–parameter in polarized $\mu$ decay [@pd], but it assumes the neutrino to be a Dirac particle, which is not the case in the popular see–saw mechanism. A bound $|\zeta| \le 0.004$ has been derived from non–leptonic $K$ decays using the MIT bag model and assuming current algebra and PCAC [@bag], but this bound is clouded by traditional strong interaction uncertainties. The bound derived here from the ${b \rightarrow s \gamma}$ process holds regardless of the nature of the neutrino and has considerably less strong interaction uncertainty. In Fig. 2, we have plotted the branching ratio as a function of $\sin(2\beta)$ for various Higgs masses ($m_{H^\pm} = 0.5, 1,3,5$ and $10$ TeV) with $m_t = 140$ GeV kept fixed. Here we have set $\zeta=0$, so that the entire non–standard contribution arises from the charged Higgs sector. When $\sin(2\beta)$ approaches $\pm 1$ (keeping the CKM matrix elements fixed), the coupling of the charged Higgs boson to the quarks diverges. By requiring that these couplings should be perturbatively controllable, i.e., less than $4\pi$, we find that $|s_{2\beta}| {\hbox{ \raise3pt\hbox to 0pt{$<$}\raise-3pt\hbox{$\sim$} }}0.98$. Charged Higgs boson with the mass below a few TeV will contradict the CLEO results on ${b \rightarrow s \gamma}$ if $k$ and $k^\prime$ are of the same order. In Fig. 3, we plot contours of branching ratios for various Higgs boson masses ($m_{H^\pm} = 0.5,1,3,5,10$ TeV) for non–zero values of $\zeta$ and $m_t=140$ GeV. We have fixed $m_{W_2} = 1.6$ TeV for this graph (which satisfies the indirect bound from $K^0-\overline{K}^0$ mass difference) and used the relation $\zeta \simeq \sin(2\beta)\cdot m_{W_1}^2/m_{W_2}^2$ to determine $\zeta$ for a given $\sin(2\beta)$. The contributions from the mixing angle $\zeta$ and the charged Higgs boson [*act additively*]{} to the effective Hamiltonian (\[eqn:heff\]). Consequently, the charged Higgs boson lighter than a few TeV are excluded for a wide range of parameter space even after the inclusion of the effects of $\zeta$. In particular, the charged Higgs boson with a mass ${m_{{H^{+}}}}{\hbox{ \raise3pt\hbox to 0pt{$<$}\raise-3pt\hbox{$\sim$} }}1$ TeV is allowed only if $k$ and $k'$ differ by an order of magnitude or so. In Fig. 4 we plot analogous contours, but corresponding to $m_{W_2}=800$ GeV. A wider range of parameters are excluded by ${b \rightarrow s \gamma}$ in this case. Conclusions: {#conclusions .unnumbered} ============ We have examined the decay ${b \rightarrow s \gamma}$ in left-right symmetric models. Large contribution to ${b \rightarrow s \gamma}$ amplitude arises from $W_L-W_R$ mixing. The physical charged Higgs boson present in the minimal Higgs sector of the model also yields significant contribution to the decay amplitude. Both these contributions are enhanced by the factor ${m_{t}}/{m_{b}}$ compared to the standard model or the minimal supersymmetric standard model. This enhancement stems from the chirality-flip induced by the right-handed coupling of the $W_1$-boson to the quarks in the case of the the $W_L-W_R$ mixing. In the case of the charged Higgs contribution, the enhancement is closely related to the absence of flavor conservation in the Higgs sector. Because of these enhanced contributions, the decay ${b \rightarrow s \gamma}$ can serve as a sensitive probe to possible signals of left-right symmetric models. The recent CLEO results on the radiative $B$ decay lead to the most stringent and essentially model–independent bound on the $W_L-W_R$ mixing angle $\zeta$ in a general class of left-right models: $-0.015 \le \zeta \le 0.003$. The mass of the charged Higgs boson, which has not been probed so far by other experiments, is also stringently constrained by the ${b \rightarrow s \gamma}$ experiments. In particular, the charged Higgs boson mass lighter than about a few TeV is excluded for a wide range of parameter space. Acknowledgments: {#acknowledgments .unnumbered} ================ One of us (K.S.B) thanks the Theory Group at the University of Tokyo for the warm hospitality extended to him during a visit when this work was initiated. K.S.B is supported by a grant from the U.S. Department of Energy. [**Note added**]{}\ While preparing this manuscript, we received a preprint by P. Cho and M. Misiak (CALT-68-1893, hep-ph 9310332). They examine the effect of the $W_L-W_R$ mixing on the ${b \rightarrow s \gamma}$ amplitude, however, the charged Higgs contributions are not considered there. Our result on the contributions of the $W_L-W_R$ mixing are in agreement with theirs. We also found that in the calculation of the QCD corrections to the effective Hamiltonian (\[eqn:heff\]), the effects of the new operators ($O_{9,10}$ in their paper) on the analysis of the operator mixing are small due to the reasons we explained in the text. We would like to thank M. Misiak for a clarifying discussion on the QCD corrections. [99]{} J.C. Pati and A. Salam, [Phys. Rev. [**D10**]{}, (1974) 275]{}; R.N. Mohapatra and J.C. Pati, [Phys. Rev. [**D11**]{}, (1975) 566, 2558]{}; G. Senjanovic and R.N. Mohapatra, [Phys. Rev. [**D12**]{}, (1975) 1502]{}. G. Beall, M. Bender and A. Soni, [Phys. Rev. Lett. [**48**]{}, (1982) 848]{}. CLEO Collaboration, R. Ammer [*et. al*]{}, [Phys. Rev. Lett. [**71**]{}, (1993) 674]{}. V. Barger, J. Hewett and R. Phillips have discussed the charged Higgs contributions to ${b \rightarrow s \gamma}$ amplitude in four classes of two–Higgs–doublet models, including the Higgs sector of the MSSM, imposing natural flavor conservation at the tree–level. None of the models considered there yields the enhancement factor ${m_{t}}/{m_{b}}$. See V. Barger, J. Hewett and R. Phillips, [Phys. Rev. [**D41**]{}, (1990) 3421]{}. D. Cocolicchio [*et. al*]{}, [Phys. Rev. [**D40**]{}, (1989) 1477]{}. G. M. Asatrayn and A. N. Ioannisyan, Sov. J. Nucl. Phys. [**51**]{}, 858 (1990). R.N. Mohapatra and G. Senjanovic, [Phys. Rev. Lett. [**44**]{}, (1980) 912]{}. U. Amaldi [*et. al*]{}, [Phys. Rev. [**D36**]{}, (1987) 1385]{}. J. Basecq [*et. al*]{}, Nucl. Phys. [**B272**]{}, 145 (1986). T. Inami and C. S. Lim, [Prog. Theor. Phys. [**65**]{}, (1981) 297]{}. K. Fujikawa and A. Yamada, Tokyo preprint UT-656, September 1993. S. Bertolini, F. Borzumati and A. Masiero, [Phys. Rev. Lett. [**59**]{}, (1987) 180]{}. N. G. Deshpande [*et al.*]{}, [Phys. Rev. Lett. [**59**]{}, (1987) 183]{}. B. Grinstein, R. Springer and M. Wise, [Phys. Lett. [**B202**]{}, (1988) 138]{}. M. Ciuchini [*et. al*]{}, Rome preprint 93/958; M. Misiak, Nucl. Phys. [**B393**]{}, 23 (1993) and preprint TUM-T-31-46/93; K. Adel and Y.P. Yao, UM-TH-92/32 (1992). Particle Data Group, K Hikasa [*et al.*]{}, Phys. Rev. D[**45**]{}, 1 (1992). N. Cabibbo and L. Maiani, [Phys. Lett. [**B79**]{}, (1978) 109]{}. J. F. Donoghue and B. R. Holstein, [Phys. Lett. [**B113**]{}, (1982) 382]{}. Figure Captions {#figure-captions .unnumbered} =============== 1. The branching fraction $Br(b \rightarrow s\gamma)$ in the left–right symmetric model as a function of the $W_L^+-W_R^+$ mixing angle $\zeta$. The three curves correspond to $m_t$=110 GeV, 140 GeV, and 170 GeV. The charged Higgs boson mass is set equal to 20 TeV so that its effects are negligible for these curves. 2. The branching ratio $Br(b \rightarrow s\gamma)$ as a function of the mixing angle $\sin(2\beta)$ for various values of the charged Higgs boson mass. $m_{H^+}=$ 0.5 TeV (inner solid), 1 TeV (inner dotdash), 3 TeV (dash), 5 TeV (outer dotdash) and 10 TeV (outer solid). The top mass is fixed as $m_t$=140 GeV. This graph corresponds to $\zeta=0$. 3. $Br(b \rightarrow s\gamma)$ versus $\sin(2\beta)$ for $m_t$=140 GeV and $m_{W_2}$=1.6 TeV for the same set of $m_{H^+}$ values as in Fig. 2. 4. Same as in Fig. 3, but for $m_{W_2}=800$ GeV.

--- abstract: 'Recently, we proposed an state model (compartment model) to describe the progression of a chronic disease with an pre-clinical (“undiagnosed”) state before clinical diagnosis. It is an open question, if a sequence of cross-sectional studies with mortality follow-up is sufficient to estimate the true incidence rate of the disease, i.e. the incidence of the undiagnosed and diagnosed disease. In this note, we construct a counterexample and show that this cannot be achieved in general.' author: - Ralph Brinks bibliography: - 'references.bib' title: An identifiability problem in a state model for partly undetected chronic diseases --- Introduction ============ Compartment model ----------------- Recently, we introduced a compartment model with a pre-clinical stage preceding the clinical stage [@Bri15b]. The model involves calendar time $t$, and the different ages $a$ of the subjects in the population. The transition rates between the states are denoted as in Figure \[fig:CompModel\]. ![Chronic disease model with four states and the corresponding transition rates. People in the state *Normal* are healthy with respect to the disease under consideration. After onset of the disease, they change to state *Undiagnosed* and maybe later to the state *Diagnosed*. The absorbing state *Dead* can be reached from all other states. The numbers of persons in the states and the transition rates depend on calendar time $t$ and age $a.$[]{data-label="fig:CompModel"}](StateModel.eps){width="90.00000%"} Using the definition $N(t, a) = S(t, a) + U(t, a) + C(t, a)$ and setting $$\begin{aligned} p_0(t, a) &= \frac{S(t, a)}{N(t, a)} \\ p_1(t, a) &= \frac{U(t, a)}{N(t, a)} \\ p_2(t, a) &= \frac{C(t, a)}{N(t, a)},\end{aligned}$$ the compartment model in Figure \[fig:CompModel\] is governed by a system of partial differential equations (PDEs): $$\begin{aligned} (\partial_t + \partial_a) p_1 &= - \bigl ( \lambda_0 + \lambda_1 + \mu_1 - \mu \bigr ) \, p_1 - \lambda_0 \, p_2 + \lambda_0 \label{e:pde1}\\ (\partial_t + \partial_a) p_2 &= \lambda_1 \, p_1 - \bigl ( \mu_2 - \mu \bigr ) \, p_2 \label{e:pde2}.\end{aligned}$$ For brevity we have written $\partial_x = \frac{\partial}{\partial x}, ~x\in \{t, a\}.$ In Eq. – the general mortality $\mu$ is given by $$\label{e:genMort} \mu = p_0 \mu_0 + p_1 \mu_1 + p_2 \mu_2.$$ Together with the initial conditions $p_1(t , 0) = p_2(t , 0) = 0$ for all $t,$ the system – completely describes the temporal dynamics of the disease in the considered population. The quantity $p_0$ can be obtained by $$\label{e:p0} p_0 = 1 - p_1 - p_2.$$ Direct and inverse problem -------------------------- Assumed that the functions $\lambda_0, \lambda_1, \mu_1, \mu_2, \mu$ on the right-hand sides of system – are suffiently smooth, then the associated initial value problem $p_1(t , 0) = p_2(t , 0) = 0$ for all $t$ has a unique solution. This means that together with the initial condition, there is a function $$\label{e:opera} \Phi: \Theta = \bigl ( \lambda_0, \lambda_1, \mu_1, \mu_2, \mu \bigr ) \mapsto P = (p_1, p_2).$$ Given the initial conditions, the operator $\Phi$ maps the transition rates $\Theta$ onto the uniquely associated prevalence functions $\Phi(\Theta) = P = (p_1, p_2).$ This problem is called the *direct problem* or *forward problem* [@Ast11]. Similar to the simpler compartment model in [@Bri15], the question arises if and under which circumstances the opposite way is possible. Does a series prevalence studies $P$ allow to estimate the transition rates $\Theta$? Mathematically, this problem is expressed as inversion of the function $\Phi$. Given $P,$ the question is if there is a unique $\Theta$ such that $\Phi(\Theta) = P?$ The problem of estimating the rates from prevalence data, is called an *inverse problem* [@Ast11]. It is not guaranteed that the inverse problem has a solution. Examination of conditions such that the inverse problem has a solution is called the analysis of *identifiability* [@Eis86]. Under certain circumstances, the operator $\Phi$ is indeed invertible. Assumed that the mortality rates $\mu_1, \mu_2,$ and $\mu$ are known, then for given $P = (p_1, p_2)$ the system – can be solved for $\lambda_0$ and $\lambda_1.$ Thus, in these cases $\Phi$ is invertible. In the next section, we will show that is not always the case. Identifiability problem ======================= We consider two prevalence studies at calendar times $t_1 < t_2$ with *mortality follow-up.* This means, on the one hand we have estimates for the age courses of the prevalences $p_1$ and $p_2$ at $t_1$ and $t_2.$ On the other hand, we have additional information if and when any participant at $t_1$ has died before $t_2.$ Let us assume that for any participant who deceased between $t_1$ and $t_2$, we do not have information about what state the person was in at the time of death. For example, a person who was in the *Normal* state at $t_1$ and died before $t_2$ could have deceased when he was still in the *Normal* state, in the *Undiagnosed* state or in the *Diagnosed* state. An exception is someone dying between $t_1$ and $t_2$, who was in the *Diagnosed* state. As the *Diagnosed* state can only be left via the transition to *Dead* state, the information from the mortality follow-up helps to estimate $\mu_2.$ Thus, the mortality follow-up contributes to estimate the general mortality $\mu$ or occasionally the mortality $\mu_2$, but not to estimate $\mu_0$ or $\mu_1.$ The question arises: Given $p_k(t_j, \cdot), ~j, k=1, 2, ~\mu(t^\star, \cdot)$ and $\mu_2(t^\star, \cdot)$ for some $t^\star$ with $t_1 < t^\star < t_2,$ are we able to estimate the rates $\lambda_0, \lambda_1, \mu_0,$ and $\mu_1$ at $t^\star?$ In the following we will show that this is not the case. This is done by constructing a counterexample with given $p_1, p_2, \mu, \mu_2$ but different $\lambda_0, \lambda_1, \mu_0,$ and $\mu_1.$ Consider the system – being in equilibrium such that $\partial_t \, p_k(t^\star, a) = \partial_a \, p_k(t^\star, a) = 0, ~k=1,2,$ for all $a.$ Furthermore, let $p_0 = 0.5, p_1 = 0.3$ and $p_2 = 0.2, \mu = 0.6, \mu_2 = 0.8.$ Obviously, it holds $p_0 + p_1 + p_2 = 1.$ From $\partial_x p_2 = 0, ~x\in \{t, a\}$ it follows that $\lambda_1 = (\mu_2 - \mu) \tfrac{p_2}{p_1} = \tfrac{4}{30}.$ If we choose $\mu_1^{(1)} = 0.5$ and $\mu_1^{(2)} = 0.6,$ then from $\mu = p_0 \mu_0 + p_1 \mu_1 + p_2 \mu_2$ it follows that $\mu_0^{(1)} = 0.58$ and $\mu_0^{(2)} = 0.52.$ In addition, $\partial_x p_1 = 0, ~x\in \{t, a\}$ implies $\lambda_0 = (\lambda_1 + \mu_1 - \mu) \tfrac{p_1}{p_0}.$ Thus, it holds $\lambda_0^{(1)} = 0.02$ and $\lambda_0^{(2)} = 0.08.$ The results are summarized in Table \[tab:res\]. Variable Value 1 Value 2 ------------- --------- --------- $p_0$ $p_1$ $p_2$ $\mu$ $\mu_2$ $\lambda_1$ $\mu_1$ 0.5 0.6 $\mu_0$ 0.58 0.52 $\lambda_0$ 0.02 0.08 : Example for non-identifiability of the system – . In an equilibrium state ($\partial_x p_k = 0, ~k=1,2, ~x\in \{t, a\}$), measured values in the upper half of the table are consistent with the values in the lower half.[]{data-label="tab:res"} Hence, from given $p_1, p_2, \mu, \mu_2,$ in equilibrium, we were able to construct *different* $\lambda_0, \lambda_1, \mu_0,$ and $\mu_1,$ which are consistent with the system – . This implies that two cross-sections at $t_1$ and $t_2$ with mortality follow-up are not sufficient to make the system identifiable. Conclusion ========== In this technical note it was shown by a counterexample that two cross-sectional studies with mortality follow-up are not sufficient to make the system – identifiable. This means, from two cross-sectional studies and measured $p_k, ~k=0,1,2,$ and known $\mu, \mu_2$ it is not possible to estimate the incidence rates $\lambda_0$ and $\lambda_1.$ The counterexample was constructed by the system – being in equilibrium. This is not a loss of generalizability. It is sufficient to find one example of non-identifiability to prove non-existence of a solution of the inverse problem. Note that from measured $p_k, ~k=0,1,2,$ and known $\mu, \mu_2,$ the rate $\lambda_1$ is estimable. This can be seen by solving Eq. for $\lambda_1.$ *Contact:*\ Ralph Brinks\ German Diabetes Center\ Auf’m Hennekamp 65\ D- 40225 Duesseldorf\ `ralph.brinks@ddz.uni-duesseldorf.de`

--- abstract: 'Very recently the most general ensemble of qubits are identified using the notion of linearity; any of these qubits gets accepted by a Hadamard gate to generate the equal superposition of the qubit and its orthogonal. Towards more generalization, we investigate the possibility and impossibility results related to Discrete Fourier Transform (DFT) type of operations for a more general set up of qutrits.' author: - | Arpita Maitra\ Residential Campus\ Indian Statistical Institute\ 205 B T Road, Kolkata 700 108, India\ arpita76b@rediffmail.com title: Some Possibility and Impossibility Results related to Discrete Fourier type transforms in Quantum Information --- [**Keywords:**]{} Unitary operations, Fourier Transform, Qubits, Qutrits, Qudits, Hadamard Transformation. Introduction ============ The two very important basic building blocks in the area of quantum computation and quantum information are qubits (in general qudits for higher dimensions) and the quantum gates. These basic building blocks of quantum computers are believed to be computationally stronger than their classical counterparts. One of the most important quantum gates in this paradigm is the Hadamard gate (that introduces equal superposition of the input state and its orthogonal) which has found wide applications in computer and communication sciences [@qNC02]. One may refer to a number of seminal papers in quantum computation and information theory where Hadamard transform has been used [@qDJ92; @BV93; @qGR96; @DD89]. Shor’s polynomial time algorithm for factoring and discrete logarithm [@Sh94] is based on Fourier transform which is a generalization of the Hadamard transform in higher dimensions. One should also note that the Toffoli and Hadamard gates comprise the simplest quantum universal set of gates [@Shi02; @DA03]. Thus, the role played by the Hadamard gate (and more generally Fourier transform) in quantum information theory is indeed significant. Pati [@PT02] has observed that one can not design a universal Hadamard gate for an arbitrary unknown qubit as linear superposition of an arbitrary unknown state $|\psi\rangle$ with its orthogonal complement $|\psi_{\perp}\rangle$ is not achievable. However, it was noted that if one considers qubit states from the polar or equatorial great circles on a Bloch sphere, then it is possible to design Hadamard type of gates. By a Hadamard ‘type’ gate we mean a unitary matrix that is not exactly a Hadamard matrix, but it still creates an equal superposition (up to a sign or a phase) of a qubit and its complement to produce two orthogonal states. Later Song et. al. [@Son04] have tried to implement the Hadamard gate in a probabilistic manner for any unknown state chosen from a set of linearly independent states. Further to Pati’s work, very recently Maitra and Parashar [@AR05] identified the most general class of qubit states, for which the Hadamard gate works in a deterministic fashion. Further, it is also shown in [@AR05] that certain Hadamard ‘type’ transformations are indeed possible for arbitrary states when partial information is available. Based on long standing interest in possibility and impossibility results in the field of quantum information, it is motivating to study these results in higher dimension. To be more precise one may study how the results of [@PT02; @AR05] can be generalized for qudits when we consider the Fourier transformation. It has been clearly commented in [@PT02] that one may try to extend the limitations and possible operations for higher dimensional quantum systems. We study these issues for qutrits in Section \[qt\] and towards the end present a generalized result for qudits too. Qutrit results {#qt} ============== To study the possibility and impossibility results in higher dimensions, we mainly concentrate on qutrits in this section. The most general form of a qutrit is [@PT03] $$|\psi \rangle = \cos \gamma_1 |0\rangle + \sin \gamma_1 \cos \gamma_2 \, e^{i\delta} |1\rangle + \sin \gamma_1 \sin \gamma_2 \, e^{i\phi} |2\rangle,$$ where $0 \leq \gamma_1, \gamma_2 \leq \frac{\pi}{2}$ and $0 \leq \delta, \phi \leq 2\pi$. The Discrete Fourier Transform (DFT) can be seen as the Hadamard transform over higher dimensional space. It is defined as $$|j\rangle = \frac{1}{\sqrt{n}} \sum_{k=0}^{n-1} e^{\frac{2\pi i j k}{n}} |k\rangle = \frac{1}{\sqrt{n}} \sum_{k=0}^{n-1} \Gamma^{j k} |k\rangle,$$ where $\Gamma = e^{\frac{2\pi i}{n}}$, $n$ being the dimension of the Hilbert space. For a qutrit this can be seen as follows when $\Gamma = e^{\frac{2\pi i}{3}}$. $$\begin{aligned} U(|0\rangle) & = & \frac{1}{\sqrt{3}} (|0\rangle + |1\rangle + |2\rangle), \nonumber \\ U(|1\rangle) & = & \frac{1}{\sqrt{3}} (|0\rangle + \Gamma |1\rangle + \Gamma^2 |2\rangle), \nonumber \\ U(|2\rangle) & = & \frac{1}{\sqrt{3}} (|0\rangle + \Gamma^2 |1\rangle + \Gamma |2\rangle).\end{aligned}$$ That is $U = \frac{1}{\sqrt{3}} \left[\begin{array}{crr} 1 & 1 & 1\\ 1 & \Gamma & \Gamma^2 \\ 1 & \Gamma^2 & \Gamma \\ \end{array}\right]$. The question is “is it possible to get a unitary transformation for a set of three arbitrary qutrits which are orthogonal", i.e., is it possible to get a unitary transformation $U$ such that $$\begin{aligned} \label{eqt1} U(|\psi_0\rangle) & = & \frac{1}{\sqrt{3}} (|\psi_0\rangle + |\psi_1\rangle + |\psi_2\rangle), \nonumber \\ U(|\psi_1\rangle) & = & \frac{1}{\sqrt{3}} (|\psi_0\rangle + \Gamma |\psi_1\rangle + \Gamma^2 |\psi_2\rangle), \nonumber \\ U(|\psi_2\rangle) & = & \frac{1}{\sqrt{3}} (|\psi_0\rangle + \Gamma^2 |\psi_1\rangle + \Gamma |\psi_2\rangle).\end{aligned}$$ We will show that it is not true in general. Let us consider two sets of mutually orthogonal qutrits $\{v_k, v_k^{\prime}, v_k^{\prime\prime}\}$ and $\{w_j, w_j^{\prime}, w_j^{\prime\prime}\}$. If DFT is possible in general, then $U(v_k) = \frac{1}{\sqrt{3}} (v_k + v_k^{\prime} + v_k^{\prime\prime})$, $U(w_j) = \frac{1}{\sqrt{3}} (w_j + w_j^{\prime} + w_j^{\prime\prime})$, $U(v_k^{\prime}) = \frac{1}{\sqrt{3}} (v_k + \Gamma v_k^{\prime} + \Gamma^2 v_k^{\prime\prime})$, $U(w_j^{\prime}) = \frac{1}{\sqrt{3}} (w_j + \Gamma w_j^{\prime} + \Gamma^2 w_j^{\prime\prime})$, $U(v_k^{\prime\prime}) = \frac{1}{\sqrt{3}} (v_k + \Gamma^2 v_k^{\prime} + \Gamma v_k^{\prime\prime})$, $U(w_j^{\prime\prime}) = \frac{1}{\sqrt{3}} (w_j + \Gamma^2 w_j^{\prime} + \Gamma w_j^{\prime\prime})$, and hence $\langle v_k | w_j \rangle = \langle v_k | U^\dagger U | w_j \rangle = \frac{1}{\sqrt{3}}( \langle v_k | w_j \rangle + \langle v_k | w_j^{\prime} \rangle + \langle v_k | w_j^{\prime\prime} \rangle + \langle v_k^{\prime} | w_j \rangle + \langle v_k^{\prime} | w_j^{\prime} \rangle + \langle v_k^{\prime} | w_j^{\prime\prime} \rangle + \langle v_k^{\prime\prime} | w_j \rangle + \langle v_k^{\prime\prime} | w_j^{\prime} \rangle + \langle v_k^{\prime\prime} | w_j^{\prime\prime} \rangle)$, which is not true in general. As example, one can take $v_k = \frac{1}{\sqrt{6}} (|0\rangle + 2e^{i\delta_k} |1\rangle + e^{i\phi_k} |2\rangle)$, $w_j = \frac{1}{\sqrt{3}} (|0\rangle + e^{i\delta_j} |1\rangle + e^{i\phi_j} |2\rangle)$, $v_k^{\prime} = \frac{1}{\sqrt{11}} (|0\rangle + e^{i\delta_k} |1\rangle - 3 e^{i\phi_k} |2\rangle)$, $w_j^{\prime} = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma e^{i\delta_j} |1\rangle + \Gamma^2 e^{i\phi_j} |2\rangle)$, $v_k^{\prime\prime} = \frac{1}{\sqrt{66}} (7|0\rangle -4 e^{i\delta_k} |1\rangle + e^{i\phi_k} |2\rangle)$, $w_j^{\prime\prime} = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma^2 e^{i\delta_j} |1\rangle + \Gamma e^{i\phi_j} |2\rangle)$, and then check that $\langle v_k | w_j \rangle \neq \frac{1}{\sqrt{3}}( \langle v_k | w_j \rangle + \langle v_k | w_j^{\prime} \rangle + \langle v_k | w_j^{\prime\prime} \rangle + \langle v_k^{\prime} | w_j \rangle + \langle v_k^{\prime} | w_j^{\prime} \rangle + \langle v_k^{\prime} | w_j^{\prime\prime} \rangle + \langle v_k^{\prime\prime} | w_j \rangle + \langle v_k^{\prime\prime} | w_j^{\prime} \rangle + \langle v_k^{\prime\prime} | w_j^{\prime\prime} \rangle)$ in this case. Thus we have the following result. \[tt1\] It is not possible to get a generalized Discrete Fourier Transform for qutrits as given by $U$ in Equation \[eqt1\]. After getting the impossibility result in general, we now consider the well known equatorial ensembles as a restricted case and try to obtain a “DFT like transformation" for this particular ensemble. For the equatorial qutrits the inner product laws are: $\langle v_k | w_j \rangle = \langle v_k^{\prime} | w_j^{\prime} \rangle = \langle v_k^{\prime\prime} | w_j^{\prime\prime} \rangle$, $\langle v_k | w_j^{\prime} \rangle = \langle v_k^{\prime\prime} | w_j \rangle = \langle v_k^{\prime} | w_j^{\prime\prime} \rangle$, and $\langle v_k | w_j^{\prime\prime} \rangle = \langle v_k^{\prime} | w_j \rangle = \langle v_k^{\prime\prime} | w_j^{\prime} \rangle$. Let $U(|\psi_0\rangle) = \frac{1}{\sqrt{3}} (|\psi_0\rangle + \alpha |\psi_1\rangle + \beta |\psi_2\rangle)$. Now for two sets of mutually orthogonal equatorial qutrits $\{v_k, v_k^{\prime}, v_k^{\prime\prime}\}$ and $\{w_j, w_j^{\prime}, w_j^{\prime\prime}\}$ we get the following. Let $U(v_k) = \frac{1}{\sqrt{3}} (v_k + \alpha v_k^{\prime} + \beta v_k^{\prime\prime})$ and $U(w_j) = \frac{1}{\sqrt{3}} (w_j + \alpha w_j^{\prime} + \beta w_j^{\prime\prime})$. Thus $\langle v_k | w_j \rangle = \frac{1}{3}( \langle v_k | w_j \rangle + \alpha \langle v_k | w_j^{\prime} \rangle + \beta \langle v_k | w_j^{\prime\prime} \rangle + \alpha^* \langle v_k^{\prime} | w_j \rangle + \alpha \alpha^* \langle v_k^{\prime} | w_j^{\prime} \rangle + \beta \alpha^* \langle v_k^{\prime} | w_j^{\prime\prime} \rangle + \beta^* \langle v_k^{\prime\prime} | w_j \rangle + \alpha \beta^* \langle v_k^{\prime\prime} | w_j^{\prime} \rangle + \beta \beta^* \langle v_k^{\prime\prime} | w_j^{\prime\prime} \rangle)$. Further $\langle v_k | w_j \rangle = \frac{1}{3}((1 + \alpha \alpha^* + \beta \beta^*) \langle v_k | w_j \rangle + (\alpha + \beta^* + \beta \alpha^*) \langle v_k | w_j^{\prime} \rangle + (\beta + \alpha^* + \alpha \beta^*) \langle v_k | w_j^{\prime\prime} \rangle)$. For the left hand and right hand side to be equal, $1 + \alpha \alpha^* + \beta \beta^* = 3$, i.e., $\alpha \alpha^* + \beta \beta^* = 2$, $\alpha + \beta^* + \beta \alpha^* = 0$ and $(\beta + \alpha^* + \alpha \beta^*) = 0$. Note that $\alpha = \beta = \Gamma = e^{\frac{2\pi i}{3}}$ satisfy the above three equations. Thus $U(|\psi_0\rangle) = \frac{1}{\sqrt{3}} (|\psi_0\rangle + \Gamma |\psi_1\rangle + \Gamma |\psi_2\rangle)$. Consider the equatorial qutrits $|\psi_0\rangle = \frac{1}{\sqrt{3}} (|0\rangle + e^{i\delta}|1\rangle + e^{i\phi}|2\rangle)$, $|\psi_1\rangle = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma e^{i\delta} |1\rangle + \Gamma^2 e^{i\phi} |2\rangle)$, and $|\psi_2\rangle = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma^2 e^{i\delta} |1\rangle + \Gamma e^{i\phi} |2\rangle)$. Now consider the unitary transformation $U = \frac{1}{\sqrt{3}} \left[\begin{array}{crr} 2\Gamma+1 & 0 & 0\\ 0 & 2+\Gamma^2 & 0 \\ 0 & 0 & 2+\Gamma^2 \\ \end{array}\right]$. One can check that, $U(|\psi_0\rangle) = \frac{1}{\sqrt{3}} (|\psi_0\rangle + \Gamma |\psi_1\rangle + \Gamma |\psi_2\rangle)$, $U(|\psi_1\rangle) = \frac{1}{\sqrt{3}} (|\psi_1\rangle + \Gamma |\psi_0\rangle + \Gamma |\psi_2\rangle)$ and $U(|\psi_2\rangle) = \frac{1}{\sqrt{3}} (|\psi_2\rangle + \Gamma |\psi_0\rangle + \Gamma |\psi_1\rangle)$. Renaming, $U(|\psi_0\rangle)$, $U(|\psi_1\rangle)$ and $U(|\psi_2\rangle)$ as $|\psi_A\rangle$, $|\psi_B\rangle$ and $|\psi_C\rangle$, one can check that $\langle \Psi_A | \Psi_B \rangle = \langle \Psi_A | \Psi_C \rangle = \langle \Psi_B | \Psi_C \rangle = 0$, i.e., they are orthogonal to each other. If we consider the other solution $\alpha = \beta = \Gamma^2 = e^{\frac{4\pi i}{3}}$, then the form of the matrix is $U = \frac{1}{\sqrt{3}} \left[\begin{array}{crr} 1+2\Gamma^2 & 0 & 0\\ 0 & 2+\Gamma & 0 \\ 0 & 0 & 2+\Gamma \\ \end{array}\right]$. Thus we get a possibility result for equatorial qutrits which provides a DFT kind of transformation. In case of equatorial qubits, the transformation is $U|\psi\rangle \rightarrow |\psi\rangle + \Gamma^{\frac{1}{2}} |{\overline \psi}\rangle$, where $\Gamma = e^{\frac{2\pi i}{2}} = e^{\pi i}$, where $U = \frac{1}{\sqrt{2}} \left[\begin{array}{cr} 1+\Gamma^{\frac{1}{2}} & 0\\ 0 & 1-\Gamma^{\frac{1}{2}}\\ \end{array}\right]$. This has been referred in [@PT02]. Here $|\psi \rangle = \frac{1}{\sqrt{2}} (|0\rangle + e^{i\phi} |1\rangle)$ and $|{\overline \psi} \rangle = \frac{1}{\sqrt{2}} (|0\rangle + \Gamma e^{i\phi} |1\rangle) = \frac{1}{\sqrt{2}} (|0\rangle - e^{i\phi} |1\rangle)$. Thus our result clearly extends the work of [@PT02] for higher dimension (qutrits). Pairwise Hadamard Type Operation -------------------------------- We like to extend the idea for qutrits where an equatorial qutrit will be the input and the output will be the superposition of the qutrit and one of its orthogonals in rotational manner. Let us consider the equatorial qutrits $|\psi_0\rangle = \frac{1}{\sqrt{3}} (|0\rangle + e^{i\delta}|1\rangle + e^{i\phi}|2\rangle)$, $|\psi_1\rangle = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma e^{i\delta} |1\rangle + \Gamma^2 e^{i\phi} |2\rangle)$, and $|\psi_2\rangle = \frac{1}{\sqrt{3}} (|0\rangle + \Gamma^2 e^{i\delta} |1\rangle + \Gamma e^{i\phi} |2\rangle)$. Now consider the unitary transformation $U = \frac{1}{\sqrt{2}} \left[\begin{array}{crr} 1+i & 0 & 0\\ 0 & 1+i\Gamma & 0 \\ 0 & 0 & 1+i\Gamma^2 \\ \end{array}\right]$. One can check that $U(|\psi_0\rangle) = \frac{1}{\sqrt{2}}(|\psi_0\rangle + i |\psi_1\rangle)$, $U(|\psi_1\rangle) = \frac{1}{\sqrt{2}}(|\psi_1\rangle + i |\psi_2\rangle)$ and $U(|\psi_2\rangle) = \frac{1}{\sqrt{2}}(|\psi_2\rangle + i |\psi_0\rangle)$. Note that the states $U(|\psi_0\rangle), U(|\psi_1\rangle), U(|\psi_2\rangle)$ are not orthogonal to each other. The similar technique can be extended for any dimensions, i.e., for the qudits too. Note that if $\Gamma = e^{\frac{2\pi i}{2}} = e^{\pi i} = -1$, i.e., in the qubit case, $U$ reduces to $U = \frac{1}{\sqrt{2}} \left[\begin{array}{cr} 1+i & 0\\ 0 & 1-i\\ \end{array}\right]$. This means that the qubit matrix for equatorial great circle is embedded in the qutrit one. This implies that remote state preparation is also valid for equatorial qutrits but not for polar qutrits, as there is no such similarity in the polar case. Towards the end we like to present a result for the most general set up of qudits. For qudits, the cyclic transformations are $U(|\psi_0\rangle) = \frac{1}{\sqrt{2}}(|\psi_0\rangle + i |\psi_1\rangle)$, $\ldots$, $U(|\psi_r\rangle) = \frac{1}{\sqrt{2}}(|\psi_r\rangle + i |\psi_{r+1}\rangle)$, $\ldots$, $U(|\psi_{n-1}\rangle) = \frac{1}{\sqrt{2}}(|\psi_{n-1}\rangle + i |\psi_0\rangle)$, where $|\psi_r \rangle = \frac{1}{\sqrt{n}} \sum_{k = 0}^{n-1} \Gamma^{rk} |k\rangle$. So the form of $U$ in this case will be $U = \frac{1}{\sqrt{2}} \left[\begin{array}{crrrr} 1+i & 0 & 0 & \ldots & 0\\ 0 & 1+i\Gamma & 0 & \ldots & 0\\ 0 & 0 & 1+i\Gamma^2 & \ldots & 0\\ \ldots & \ldots & \ldots & \ldots & \ldots\\ 0 & 0 & 0 & \ldots & 1 + i\Gamma^{n-1}\\ \end{array}\right]$. However, one should note that the states $U(|\psi_0\rangle), \ldots, U(|\psi_r\rangle), \ldots, U(|\psi_{n-1}\rangle)$ are non orthogonal. Conclusion ========== In this paper we have studied the possibility and impossibility results related to Discrete Fourier Transform type operations. We study the possibility and impossibility results for DFT type of operations in the case of qutrits. Towards the end we present a result on qudits too. [10]{} P. Agrawal, P. Parashar and A. K. Pati, Exact remote state preparation for multiparties using dark states. International Journal of Quantum Information, Pages 301–319, Volume 1, No. 3, 2003. arXiv:quant-ph/0304006 v1, 1 April, 2003. D. Aharonov, A Simple Proof that Toffoli and Hadamard are Quantum Universal. quant-ph/0301040. E. Bernstein and U. Vazirani, Quantum complexity theory, In Proceedings of 25th Annual ACM Symposium on Theory of Computing, pp 11 (1993). D. Deutsch, Quantum computational networks, , A [**425**]{}, 73 (1989). D. Deutsch and R. Jozsa, Rapid solution of problems by quantum compuation, , A [**439**]{}, 553 (1992). L. Grover, A fast quantum mechanical algorithm for database search, In [*Proceedings of 28th Annual Symposium on the Theory of Computing (STOC)*]{}, pp 212 (1996). Available at xxx.lanl.gov/quant-ph/9605043. A. Maitra and P. Parashar, Hadamard type operations for qubits. International Journal of Quantum Information, Vol 4, No. 4, (2006), 653-664. Also available at arXiv:quant-ph/0505068 v1, 10 May, 2005. M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, Cambridge University Press, 2002. A. K. Pati, General impossible operations in quantum information. Phys. Rev. A 66, 062319, 2002. arXiv:quant-ph/0111153 v3, 17 September, 2002. Y. Shi, Both Toffoli and controlled-NOT need little help to do universal quantum computation. Quantum Information and Computation, 3(1):84-92, 2002. Also available at arXiv:quant-ph/0205115. P. W. Shor, Algorithms for quantum computation: discrete logarithms and factoring, , IEEE Press, Los Alamitos, CA (1994). W. Song, M. Yang and Z-L Cao, Probabilistic implementation of universal Hadamard and Unitary gates. Phys. Lett. A [**330**]{} (2004) 155.

--- abstract: 'We first discuss the mass range of type IIP SN progenitors and how the upper and lower limits impose interesting constraints on stellar evolution. Then we discuss the possible implications of two SNe, 2002ap and 2006jc, for Wolf-Rayet star mass-loss rates and long Gamma-ray bursts.' author: - John Eldridge bibliography: - 'eldridge2.bib' title: Mass loss and supernova progenitors --- [ address=[Astrophysics Research Centre, Physics Building, Queen’s University, Belfast, BT7 1NN, UK.]{} ]{} Introduction ============ Core-collapse supernovae (SNe) are the violent deaths of stars more massive than $\approx~7.5 {\rm M}_{\odot}$. They occur when nuclear burning reactions or electron degeneracy-pressure can no longer support the core against gravitational collapse. Either a neutron star or black hole is formed from the collapsing iron core and the outer layers of the star explode in a violent display. The nature of this display depends strongly on the final state of the progenitor star and the circumstellar medium; because there are many paths of stellar evolution the are many types of SNe. SNe are classified according to their observed spectra and lightcurves. The first differentiation is made by the absence or presence of hydrogen in a spectrum. If hydrogen emission lines exist then a SN is type II and type I otherwise. For type II SNe there are four subgroups: IIP when there is a plateau to the lightcurve lasting a few months. These are the most common type of SN. Type IIL have a linear decay to their lightcurve, IIn have narrow hydrogen emission lines in their spectrum and IIb initially appear to be type II until after a short time the hydrogen lines disappear and the SN becomes type Ib. The hydrogen free type I SNe have three separate subtypes. Type Ia are thought to be thermonuclear explosions of Chandrasekhar mass white dwarfs and are not considered further here, type Ib have helium lines in emission while type Ic have no helium lines. Single star models predicts that the type II SNe will be the result of stars between about $7$ to $27M_{\odot}$ [@H03; @ETsne; @arend] as these retain their hydrogen envelopes. Stars more massive than this lose all their hydrogen via stellar winds and therefore give rise to type Ib/c SNe. The only way to confirm this mapping is to observe the progenitors of SN. This is achieved by searching telescope archives to discover pre-explosion images. While the progenitors of SNe 1987A and 1993J where detected both were in nearby galaxies and both rare and unusual SNe. It was not until 1999 when the HST archive covered enough galaxies at sufficient depth and resolution that it was possible to start looking for progenitors of SN in a large number of galaxies at distances up to 20 megaparsecs. The first detection for the most common, type IIP, SN was for 2003gd [@2004Sci...303..499S]. This confirmed that their progenitors were red supergiants. With eight years of observations there are currently 32 SNe with useful pre-explosion images, 18 of these are type IIP (6 detections) and the remainder being types IIb and Ib/c. There are no detections of type Ib/c progenitors and with the growing number of the non-detections there is growing evidence that our standard view of mass loss during the late stages of evolution may be incorrect. In this proceedings we briefly highlight some of the main conclusions the mass range inferred for type IIP progenitors. We then discuss two interesting cases of two type Ib/c SNe, 2002ap and 2006jc that provide very stringent limits on the evolution of the most massive stars. The population of type IIPs. ============================ With the sample of 18 type IIP detections and non-detections it is possible to begin to characterise the population of the progenitors @compilation. The main important result for their study is that by using a maximum likelihood analysis it is possible to infer the minimum and maximum initial masses for type IIP SN progenitors. The initial mass range is between $7.5$ to $16.5M_{\odot}$, however the error bars are large and the range could be as large as $6$ to $18{\rm M}_{\odot}$. The initial mass depends strongly on the mixing and mass-loss scheme of the stellar models use to estimate an initial mass from a final luminosity. To remove this systematic it is better to work out the range of final helium core masses which is approximately $1.9$ to $6M_{\odot}$ in the STARS stellar models [@ETsne]. The minimum mass can be used to constrain models of convection in stellar models. Most models with helium cores at the lower end of this range experience second dredge-up and become AGB stars. In fact the stellar models used in this work do experience second dredge-up and we use the pre-dredge up models. Therefore something is required to prevent these stars from becoming AGB stars. It is possible for the most massive AGB stars to undergo SN, however their observational signature in pre-explosion images would be quite different to the red supergiants observed to date as they are cooler and therefore more luminous at infra-red wavelengths [@EMS07]. The maximum mass is due to one of two factors, because stars above this limit have lost most (or all) of their hydrogen and produce another type of SN or it is because the cores are massive enough to form a black hole and this also leads to another type of SN. In reality it is probably a combination of these factors. However the black hole explanation could be favoured as the inferred helium core mass is similar to that which is expected to produce a black hole rather than a neutron star at core-collapse [@H03; @ETsne]. But what about other types? =========================== With only one detection for the other SN types there is little that can be said. If the upper limits that have been obtained on the progenitors’ luminosity are compared to the luminosity of observed Wolf-Rayet stars, the suspected progenitors, it is apparent that the pre-explosion in all but one cases are not deep enough to have revealed the progenitors. For a detection it is a waiting game to determine for type type Ib/c SN to occur nearby and have deep pre-explosion images. The culprits for type Ib/c progenitors are Wolf-Rayet (WR) stars. These are evolved massive stars that have completed core hydrogen burning and lost (or in the process of losing) their hydrogen envelopes and are naked helium stars. These stars are also the preferred progenitors of long Gamma-ray bursts [@grbreview]. Despite this there are two interesting pre-explosion images for type Ib/c SNe. SN 2002ap, because the limit is so low that we were only just unable to detect the progenitor and are able to rule out single stars for the progenitor. While before SN 2006jc a luminous outburst transient was detected two years prior to the SN and this may have interesting consequences for our understanding of stellar-wind mass-loss. 2002ap ------ @crockett used previously unused deep pre-explosion images to reexamine the progenitor of SN 2002ap. The limit is the most stringent to date on any type Ib/c progenitor. They were able to rule out all standard single-star models are suggest that the most likely progenitor was a binary. Figure \[02ap1\] shows the luminosity limit derived from the pre-explosion images on a theoretical Hertzsprung-Russell diagram. The grey stellar track is of a $17{\rm M}_{\odot}$ star that has its hydrogen envelope stripped by interaction via Roche-Lobe Overflow in a binary. The final stellar model has a final mass or $5{\rm M}_{\odot}$ that agrees with the mass inferred from modelling the SN lightcurve by @Mazzali. Also shown on the figure are the tracks of possible companion stars the mass of any companion star can also be limited by the pre- and post- explosion images. The mass of the companion star must have been less than $\approx 14M_{\odot}$. Any mass transfer must have been quite inefficient otherwise the secondary would have accreted a large amount of mass and be visible in the pre-explosion image. However there is a problem with this model, there is a large amount of helium in the progenitor model. While the amount of helium required for a type Ib SN is uncertain, over $0.5M_{\odot}$ should produce the signature of helium lines in the SN spectrum. The binary model has around $1M_{\odot}$ of helium in the envelope. Therefore there is some uncertainty in whether this is a reasonable progenitor model. There are solutions, one is that the star was more massive and underwent more severe stripping during the binary interaction. Another is that we underestimate the strength of stellar winds of Wolf-Rayet stars. Increase the mass-loss rates of a star by a factor of 2 - 3 would reduce the final mass of helium and therefore produce the required type Ic SN. A source of increased mass-loss rates could be rapid rotation \[REF\]. An alternative is that the mass-loss rates of WR stars are underestimated. This would be possible if the mass-loss rates are enhanced for only a short time of evolution. A possible mechanism has been suggested by @petrovic who discuss how helium star envelopes can become inflated. Inflated helium envelopes are caused by a bump in the opacity which causes the envelopes to become extended and the density profile of the envelope to invert so density increases with radius from the core. @petrovic suggest that this inflation may not be physical and note that by increasing the mass-loss rates it is possible to remove the inflated envelope and density inversion. Figure \[02ap2\] shows a standard stellar model with the mass-loss rates of @NL00 and a model which replaces these with the mass loss rates of @petrovic when the helium envelope becomes inverted. The figure shows that during the nitrogen rich WN evolution much more mass is lost than in the standard case. Further more most of the lost material is helium. The final model retains only $0.25M_{\odot}$ and the mass again agrees with the analysis of @Mazzali. With one non-detection it is not possible to decide if all WR mass-loss rates need to be increased but if the number of non-detections increase then the situation may begin to become more serious and the mass-loss rates will have to be closely examined. The only other scenario is that WR stars produce large amounts of dust a few years before core-collapse. This would reduce their apparent luminosity in pre-explosion images. ![Theoretical Hertzsprung-Russell diagram. The dashed line is the luminosity limit from the non-detection of any object in the pre-explosion image, the grey box represents the uncertainty in the limit for WR stars [@crockett]. The solid grey line is the evolution for a $17M_{\odot}$ star which has its hydrogen envelope removed in a binary interaction. The solid black line is the evolution of a $11.9M_{\odot}$ binary companion and the dotted line is the evolution of a $15.3M_{\odot}$ binary companion. At the time of the primary SN the latter companion would have been observed while the former would have remained undetected.[]{data-label="02ap1"}](fig02ap1.ps){height=".3\textheight"} ![Theoretical Hertzsprung-Russell diagram of the evolution of initially $50M_{\odot}$ stars. The dashed line is the luminosity limit from the non-detection of a WR star the pre-explosion image, the grey box represents the uncertainty. The solid grey line is for a standard WR model [@EV06] and the solid black line a model using the mass-loss rates of [@petrovic]. []{data-label="02ap2"}](fig02ap2.ps){height=".3\textheight"} 2006jc ------ @pastorello and @folye describe a most unusual SN. SN 2006jc was discovered on 9th October 2006 in the galaxy, UGC4904. @pastorello found that it was spatially coincident with a bright optical transient that occurred in 2004. The SN itself was classified as a type Ib-n due to narrow helium lines in the spectra. The current interpretation is that the narrow helium lines are due to helium-rich material ejected by the star in a dramatic mass-loss episode that was observed as the optical transient. Then two years after this episode the star exploded as a type Ic SN and therefore the progenitor star had be stripped of helium, otherwise broad helium emission lines would have been observed in combination with the narrow lines. The problem with this interpretation is that if the progenitor was a single star then our understanding of Wolf-Rayet stars and their winds must be revised. The only single stars known to produce similar bright optical transients are LBV stars [@lbv1; @lbv2]. However these stars tend to retain some hydrogen which would have been observed in the SN spectrum. It is conceivable that there are transition objects between LBVs and WR stars that would lead to the observed evolution for the progenitor of SN2006jc. These objects may be rare to get the right amount of mass-loss to remove all helium just before core-collapse. Although such objects would be more common at lower metallicities. An alternative to the LBV-like WR star is that the transient and mass-loss was due to a pair-production instability causing a dramatic mass-loss episode a few years before a normal core-collapse SN [@langerinprep]. The models of @heger02 show that massive stars with little mass loss will experience such outbursts prior to core-collapse. The upper metallicity limit is uncertain. It is possible to estimate from stellar models to be below SMC metallicity. There is final a possibility with an observed analogue, the may have been progenitor a binary with a WR star and an evolved LBV star. The outburst was produced by the LBV star while the WR star exploded to produce the type Ic SN. There is one similar system in the SMC that has been observe to undergo outburst and at some point in the future may lead to a similar SN [@smclbvwrbin]. These systems are not as rare as might be first thought. If we assume LBV evolution happens after core-hydrogen burning is complete then any star which has a secondary companion that has completed core-hydrogen burning could be a possible LBV-WR system. Figure \[06jc\] shows how the hydrogen burning lifetime compares to the total lifetime for stars with different initial masses. It is possible to see that the more massive stars can have a wider range of secondary masses for this kind of evolution. If we assume the mass ratio of binaries ($q=M_{\rm secondary}/M_{\rm primary}$) has a flat distribution and is independent of primary mass then 60% of binaries with a $200M_{\odot}$ primary might be LBV-WR systems, while this reduces to only 20% for a primary initially $50M_{\odot}$. The only way to distinguish between these three plausible models will be the rate of these events and the metallicities of the host galaxies. The binary scenario will be only weakly metallicity dependent, the pair-production outburst will be concentrated to low metallicities while the single star LBV/WR has an unknown metallicity dependence but if it is related to inflation of the WR star then is will be concentrated at higher metallicities. The total rate of type Ib-n SN is $< 4$% of all type Ib/c SN. The rate of GRBs as a fraction of all Ib/c SNe is between 0.1 to 1% \[REF\]. It is possible therefore that some GRBs may occur with 2006jc-like SNe. If this is the case then the circumburst medium inferred from the optical afterglow would be a constant density medium due to the changes in mass-loss rate and wind speed. There are a number of GRBs where this has been observed and is another possible solution when the CBM is a constant density rather than that expected of a free-wind density profile [@vanmarle; @eldridge]. ![The lifetimes for massive single stars. The solid line is the total lifetime versus initial mass. The dashed line is the time of the end of core hydrogen burning and the dotted-dashed line is the time of the end of core helium burning. The three shaded regions show the range of secondary masses which have completed core hydrogen burning and are LBV candidates by the time the primary (the high mass edge of the region) experiences a SN.[]{data-label="06jc"}](fig06jc1.ps){height=".3\textheight"} Conclusions =========== While the progenitors of type IIP SNe are becoming well understood there is still great uncertainty over the progenitors of other SN types. It is becoming apparent that our understanding of WR mass-loss maybe incorrect and one possible reason that we do not observe more WR stars is they lose more mass than currently thought and that some of this mass-loss may be in luminous outburst such as the one that proceeded SN 2006jc. Or they produce copious amount of dust in the last few years before core-collapse. JJE would like to thank Stephen Smartt, Andrea Pastorello, Seppo Matilla, Mark Crockett and Dave Young for many discussions and making his time at Queen’s University Belfast so enjoyable.

--- abstract: | We prove that a nonzero idempotent is zero-diagonal if and only if it is not a Hilbert–Schmidt perturbation of a projection, along with other useful equivalences. Zero-diagonal operators are those whose diagonal entries are identically zero in some basis. We also prove that any bounded sequence appears as the diagonal of some idempotent operator, thereby providing a characterization of inner products of dual frame pairs in infinite dimensions. Furthermore, we show that any absolutely summable sequence whose sum is a positive integer appears as the diagonal of a finite rank idempotent. address: - 'Jireh Loreaux, Department of Mathematics, McMicken College of Arts and Sciences, University of Cincinnati, Cincinnati, Ohio, 45212, USA' - 'Gary Weiss, Department of Mathematics, McMicken College of Arts and Sciences, University of Cincinnati, Cincinnati, Ohio, 45212, USA' author: - Jireh Loreaux Gary Weiss title: | Diagonality and Idempotents\ with applications to problems in\ operator theory and frame theory --- INTRODUCTION ============ Throughout this paper the underlying space is either a finite dimensional or separable infinite dimensional complex Hilbert space. We first establish some terminology. A “basis” herein is an *orthonormal* basis of the underlying Hilbert space. When a basis $\mathfrak{e} = \{e_j\}_{j=1}^N$ (possibly $N=\infty$) is specified, “diagonal” or “diagonal sequence” of an operator $T$ is the sequence ${\ensuremath{\left\langle (Te_j,e_j) \right\rangle}}$, that is, the diagonal sequence of the matrix representation for $T$ with respect to the basis $\mathfrak{e}$. Sometimes we will say that a sequence is “a diagonal” of $T$, by which we mean that there exists some basis with respect to which $T$ has this sequence as its diagonal. Diagonality is a term coined by the authors for the study of: 1. \[item:4\] properties that the diagonal sequences can possess for a fixed operator and in all bases, and characterizations of those sequences; 2. \[item:5\] properties that the diagonal sequences can possess for a class of operators and in all bases, and characterizations of those sequences. Such information is used ubiquitously throughout operator theory. With this term we here attempt to bring these phenomena under a unifying umbrella in the hope this will stimulate bridges of insight connecting them. This paper focuses mainly on \[item:5\], but some results also have the flavor of \[item:4\]. Starting with the most basic and then on to current active areas of research, we give some examples that pervade our work. 1. Which numbers can appear on the diagonal of an operator? Clearly these numbers constitute precisely its numerical range. And which operators have only positive diagonal entries? Clearly these are the positive operators. 2. Well-known highly useful diagonality example: every trace-class operator in every basis has an absolutely summable diagonal sequence and those sums are invariant; likewise every compact operator has diagonal sequences tending to zero in every basis. In contrast, finite rank operators fail to always have finite rank diagonals, witness any nonzero rank-one projection $\xi\otimes\xi$, $\xi\in\ell^2$ of infinite support. This phenomenon for the trace-class ideal and the ideal of compact operators is subsumed under the more general notion of diagonal invariance. Given a basis $\mathfrak{e}$, we let $E_{\mathfrak{e}}(T)$ be the conditional expectation of $T$ with respect to $\mathfrak{e}$ which replaces the off-diagonal entries with zeros. An ideal $\mathcal{I}$ is said to be diagonally invariant if for every $\mathfrak{e}$ and every $T\in\mathcal{I}$, $E_{\mathfrak{e}}(T)\in\mathcal{I}$. Diagonal invariance is equivalent to the ideal being arithmetic mean-closed ($_a(\mathcal{I}_a) = \mathcal{I}$, am-closed for short). For details see [@KW-2011-IUMJ Theorem 4.5], but for now, $\mathcal{I}_a$ and $_a\mathcal{I}$ are the arithmetic mean and pre-arithmetic mean ideals generated respectively by: operators with $s$-numbers the arithmetic means of the $s$-numbers of operators from ideal $\mathcal{I}$, and operators whose arithmetic means of their $s$-numbers are $s$-numbers of operators in $\mathcal{I}$. The converses seem to us to be less well-known: if in every basis an operator’s diagonal sequence is absolutely summable, then the operator is trace-class; and likewise if in every basis the operator’s diagonal sequence tends to zero, then it is a compact operator. This phenomenon is totally general. That is, a sufficient test for membership in an arbitrary ideal $\mathcal{I}$ is: $$\label{eq:2} E_{\mathfrak{e}}(T)\in\mathcal{I}, \forall\mathfrak{e} \implies T\in\mathcal{I}.$$ Although not immediate, this follows easily from the contrapositive by considering the real and imaginary parts of $T$, and by considering separately the compact and non-compact cases. 3. \[item:6\] What diagonal sequences can arise for a specific operator? The study of \[item:4\]. We think of this subject as Schur–Horn theory, although traditionally Schur–Horn theory refers to the study of the diagonals of selfadjoint operators, a study almost a century old that continues today and is beginning to extend into operator algebras. Much of this work focuses on diagonals of positive compact operators. A fundamental tool used is majorization theory, including new types of majorization such as $\infty$- and approximate $\infty$-majorization defined using $p$- and approximate $p$-majorization. Convexity also plays a central role. Some 1923–1964 contributors are Schur [@Sch-1923-SBMG], Horn [@Hor-1954-AJoM], Markus [@Mar-1964-UMN], Gohberg–Markus [@GM-1964-MSN], and in the last 10 years — Arveson–Kadison [@AK-2006-OTOAaA], Antezana–Massey–Ruiz–Stojanoff [@AMRS-2007-IJoM], Kaftal–Weiss [@KW-2010-JoFA] and Loreaux–Weiss [@LW-2015-JFA]. Others for operator algebra Schur–Horn theory include Argerami and Massey [@AM-2007-IUMJ], [@AM-2008-JMAA], [@AM-2013-PJM] and most recently Ravichandran [@Rav-2012] and Kennedy–Skoufranis [@KS-2014]. Schur–Horn theory for finite spectrum selfadjoint operators was studied extensively by Kadison [@Kad-2002-PotNAoS], [@Kad-2002-PotNAoSa] (the carpenter problem for projections, or equivalently 2-point spectrum normal operators), Arveson [@Arv-2007-PotNAoSotUSoA] (a necessary condition on diagonals of certain finite spectrum normal operators), Jasper [@Jas-2013-JoFA] (3-point spectrum selfadjoint operators), and Bownik–Jasper [@BJ-2012], [@BJ-2013-TAMS] (finite spectrum selfadjoint operators), and along with [@Neu-1999-JoFA] are the only non-compact operator results known to the authors. In [@Neu-1999-JoFA], A. Neumann obtained a Schur–Horn type theorem for general selfadjoint operators. However, it should be noted that his results are *approximate* in the sense that he identified the $\ell^\infty$-closure of the diagonal sequences of a selfadjoint operator with a certain convex set. In contrast, the aforementioned results of Kaftal–Weiss, Loreaux–Weiss, Kadison, Jasper and Bownik–Jasper are all *exact* in the sense that they describe precisely the diagonals of certain classes of selfadjoint operators. 4. \[item:7\] What diagonal sequences can arise for a class of operators? The study of \[item:5\]. There is a variety of material on this subject. We reference only that which we know, but there are almost certainly results we have inadvertently overlooked. In the same paper [@Hor-1954-AJoM] in which he characterizes the diagonal sequences of a fixed selfadjoint matrix in $M_n(\mathbb{C})$, Horn identifies the diagonals of the class of rotation matrices. He then uses this to identify the diagonals of the classes of orthogonal matrices and of unitary matrices. See [@Hor-1954-AJoM Theorems 8-11]. Fong shows in [@Fon-1986-PotEMSSI] that any bounded sequence of complex numbers appears as the diagonal of a nilpotent operator in $B(H)$ of order four ($N^4=0$), thus seamlessly characterizing diagonals of the broader classes of nilpotent and also quasinilpotent operators. In this paper Fong remarks that a finite complex-valued sequence appears as the diagonal of a nilpotent matrix in $M_n(\mathbb{C})$ if and only if its sum is zero. More recently, Giol, Kovalev, Larson, Nguyen and Tener [@GKL+-2011-OaM] classified the diagonals of idempotent matrices in $M_n(\mathbb{C})$ as those whose sum is a positive integer less than $n$, along with the constant sequences $\langle 0,\ldots,0 \rangle$ and $\langle 1,\ldots,1 \rangle$ (see Theorem \[thm:idempotent-matrix-diagonals\] below). 5. In this context, J. Jasper posed to us a frame theory question which for us evolved into questions below on diagonal sequences of idempotents (operators for which $D^2=D$) and gave rise to this paper: Questions \[que:ell1-diagonal-finite-rank\]–\[que:zero-diagonal-finite-rank\] below and the immediately preceding comment on the frame theory connection. As mentioned above, a good deal of work concerning diagonal sequences of operators deals with the selfadjoint case. Here we study diagonal sequences of idempotents, and so diagonals of projections (selfadjoint idempotents) are of particular relevance to us. These were characterized by Kadison in [@Kad-2002-PotNAoS], [@Kad-2002-PotNAoSa] in the following theorem. We find this theorem especially interesting because it straddles the fence between \[item:6\] and \[item:7\]. Indeed, although it is stated as a characterization of the diagonals of the class of projections, it can easily be adapted to identify the diagonals of any fixed projection. This is because two projections $P,P'\in B(H)$ are unitarily equivalent if and only if $\operatorname{Tr}P = \operatorname{Tr}P'$ and $\operatorname{Tr}(1-P) = \operatorname{Tr}(1-P')$. And so for ${\ensuremath{\left\langle d_k \right\rangle}}$ an admissible diagonal sequence for $P$, these trace quantities are precisely the sum of the diagonal entries $d_k$ and the sum of $1-d_k$, respectively. Then one can apply the four finite/infinite cases in the next theorem. \[thm:projection-diagonals\] Given an infinite sequence $\langle d_k \rangle\in [0,1]^{\mathbb{N}}$ with $$a = \sum_{d_k<\nicefrac{1}{2}} d_k \quad \text{and} \quad b = \sum_{d_k\ge \nicefrac{1}{2}} (1-d_k),$$ then there is a projection $P\in B(H)$ (i.e., $P^2=P=P^*$) with diagonal $\langle d_k \rangle$ if and only if one of the following mutually exclusive conditions holds: 1. either $a$ or $b$ is infinite; 2. $a,b<\infty$ and $a-b\in\mathbb{Z}$. The requirement that $0\le d_k\le 1$ for all $k\in\mathbb{N}$ is clearly necessary since $P\ge 0$, ${\ensuremath{\lVert P \rVert }}=1$ and the diagonal entries of $P$ are elements of its numerical range. The second condition, that $a-b\in \mathbb{Z}$, is less obvious but can viewed as a kind of index obstruction to an arbitrary sequence in $[0,1]^{\mathbb{N}}$ appearing as the diagonal of a projection. Indeed, in [@Arv-2007-PotNAoSotUSoA], Arveson provided details on this index obstruction and showed that it applies more generally to any normal operator with finite spectrum that consists of the vertices of a convex polygon. Since we study diagonals of idempotents in $B(H)$, which when not projections are non-selfadjoint, we are interested in diagonals of non-selfadjoint operators. One particularly relevant result in this direction is the aforementioned characterization of diagonals of idempotent matrices in $M_n(\mathbb{C})$ by Giol, Kovalev, Larson, Nguyen and Tener [@GKL+-2011-OaM]. \[thm:idempotent-matrix-diagonals\] A finite sequence $\langle d_k \rangle\in\mathbb{C}^n$ appears as the diagonal of an idempotent $D\in M_n(\mathbb{C})$ if and only if one of the following three mutually exclusive conditions holds. 1. $d_k=0$ for all $k$ (in which case $D=0$); 2. $d_k=1$ for all $k$ (in which case $D=I$); 3. $\sum d_k \in \{1,\ldots,n-1\}$. Since $\operatorname{Tr}D\in\{1,\ldots,n-1\}$ for any nonzero, non-identity idempotent matrix (as is well-known, see for instance Lemma \[lem:idempotent-decomp\]), this theorem says that this is the only requirement for a sequence to appear as the diagonal of some idempotent. Giol, Kovalev, Larson, Nguyen and Tener were interested in this result because of its relevance to frame theory. Because of a similar frame-theoretic question (characterizing inner products of dual frame pairs) Jasper asked for a characterization of diagonals of idempotents in $B(H)$. Such a result would simultaneously be an extension of the previous two theorems. For a key test case, Jasper posed to us the following two operator-theoretic questions (private communication, May 2013 [@Jas-2013]): \[que:ell1-diagonal-finite-rank\] If an idempotent has a basis in which its diagonal is absolutely summable, is it finite rank? \[que:zero-diagonal-finite-rank\] If an idempotent has a basis in which its diagonal consists solely of zeros (i.e., is a zero-diagonal operator in the terminology of Fan [@Fan-1984-TotAMS]), is it finite rank? If we restrict the idempotents to be selfadjoint (i.e., projections), then they are positive operators and the answer to each question is certainly affirmative since the trace is preserved under conjugation by a unitary operator (i.e., a change of basis). In fact, for projections, having an absolutely summable (or even summable) diagonal is a characterization of those projections with finite rank since $\operatorname{rank}P = \operatorname{Tr}P$. Moreover, the only projection with a zero diagonal is the zero operator for this same reason. Hence, a negative answer to either of these questions for the entire class of idempotents would be a notable departure from the case of projections, and would therefore suggest that the classification of their diagonals is potentially harder than one might naïvely expect. As it turns out, Larson constructed a nonzero (and even necessarily infinite rank) idempotent that lies in a continuous nest algebra which has zero diagonal with respect to this nest [@Lar-1985-AoMSS Proof of Theorem 3.7]. An operator $T$ has zero diagonal *with respect to the nest* if $P_{\lambda}TP_{\lambda}=0$ for some linearly ordered set of projections $\{P_{\lambda}\}_{\lambda\in\Lambda}$ inside the nest such that with respect to the decomposition of the identity $I=\sum_{\Lambda}P_{\lambda}$ every element of the nest is block upper-triangular. However, the existence of an idempotent with zero diagonal with respect to a nest algebra certainly depends on the order type of the nest to some extent. For example, the nest algebra consisting of the upper triangular matrices with respect to some basis $\{e_n\}_{n\in\mathbb{N}}$ for $H$ has order type $\omega$ (the first infinite ordinal), and simple computations show that the only idempotent with zero diagonal inside this nest algebra is the zero operator. Once we leave the realm of nest algebras, we can ask two questions: - Which idempotents are zero-diagonal? - Which idempotents have an absolutely summable diagonal? As it turns out, both of these questions have the same answer, which we provide in Theorem \[thm:idempotent-zero-diagonal\]. Before we state this theorem, we expound slightly on the methods involved. The techniques for analyzing diagonals of non-selfadjoint operators seem to differ greatly from those used for selfadjoint operators. For example, the techniques used in determining diagonals of selfadjoint operators often rely heavily on majorization and keeping track of the explicit changes of the basis (or equivalently, the unitary operators) involved in the construction. In contrast, the Toeplitz–Hausdorff Theorem, that the numerical range $W(T)$ of a bounded operator $T$ is convex, is one of the central tools in the work of Fan, Fong and Herrero [@Fan-1984-TotAMS], [@FF-1994-PotAMS], [@FFH-1987-PotAMS] to determine diagonals of non-selfadjoint operators. Indeed, they frequently use the nonconstructive version of the Toeplitz–Hausdorff Theorem despite the existence of constructive versions in which a formula is specified for the vector yielding the prescribed value of the quadratic form. The Fan, Fong and Herrero results relevant to us here are restated below. The first is an infinite dimensional analogue of the finite dimensional result that an $n\times n$ matrix has trace zero if and only if it is zero-diagonal. \[thm:subsequence-zero-diagonalizability\] If $T\in B(H)$ and there exists some basis $\{e_j\}_{j=1}^\infty$ for $H$ for which the partial sums $$s_n \coloneqq{} \sum_{j=1}^n (Te_j,e_j)$$ have a subsequence converging to zero, then $T$ is zero-diagonal. \[def:trace-set\] Let $T\in B(H)$ and let $\mathfrak{e}=\{e_j\}_{j=1}^\infty$ be a basis for $H$. Suppose the partial sums $s_n = \sum_{j=1}^n (Te_j,e_j)$ converge to some value $s\in\mathbb{C}$. Then we say that $\operatorname{Tr}_{\mathfrak{e}} T:=s$ is the *trace of $T$ with respect to the basis $\mathfrak{e}$*. The *set of traces of $T$*, denoted $R\{\operatorname{Tr}T\}$, is then the set of all such traces $\operatorname{Tr}_{\mathfrak{e}} T$ as $\mathfrak{e}$ ranges over all orthonormal bases for which $\operatorname{Tr}_{\mathfrak{e}} T$ is defined. Observe that in order to make sense of this definition it is essential both that these trace values are finite and that we must order $\mathfrak{e}$ by $\mathbb{N}$. A curious fact about the set $R\{\operatorname{Tr}T\}$ from Definition \[def:trace-set\] is that it may take on only four different shapes: the plane, a line, a point or the empty set. It is no coincidence that these shapes coincide with those obtainable as the limits of convergent rearrangements of a series of complex numbers (i.e., the Lévy–Steinitz Theorem extending the Riemann Rearrangement Theorem to complex numbers). \[thm:trace-range-shape\] Suppose $T\in B(H)$. Then there are four possible shapes that $R\{\operatorname{Tr}T\}$ can acquire. More specifically, $R\{\operatorname{Tr}T\}$ is: 1. the plane $\mathbb{C}$ if and only if for all $\theta\in\mathbb{R}$, $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_+\notin\mathcal{C}_1$ (the trace-class); 2. a line if and only if for some $\theta\in\mathbb{R}$,\ $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_\pm\notin\mathcal{C}_1$ but $({\ensuremath{\mathfrak{Im}}}\, \mathrm{e}^{\mathrm{i}\theta}T)\in\mathcal{C}_1$; 3. a point if and only if $T\in\mathcal{C}_1$; 4. the empty set $\emptyset$ if and only if for some $\theta\in\mathbb{R}$,\ $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_+\notin\mathcal{C}_1$ but $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_-\in\mathcal{C}_1$. In fact, their proof of Theorem \[thm:trace-range-shape\] shows that given $T\in B(H)$ there exists a basis $\mathfrak{e}$ for $H$, $\mathfrak{e}$ ordered by $\mathbb{N}$, for which every element of $R\{\operatorname{Tr}T\}$ can be obtained from a basis which is a permutation of $\mathfrak{e}$. For the next theorem Fan–Fong utilize the previous two theorems to provide intrinsic (i.e., basis independent) criteria for when a bounded operator is zero-diagonal. \[thm:zero-diagonal-trace-theta\] An operator $T$ is zero-diagonal if and only if for all $\theta\in\mathbb{R}$, $$\operatorname{Tr}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_+=\operatorname{Tr}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}T)_-.$$ We neither use nor cite this theorem elsewhere in the paper. However, it seems interesting to include it because it shares its intrinsic nature with our Theorem \[thm:idempotent-zero-diagonal\](i). Later we will use Theorem \[thm:trace-range-shape\] to prove our first main theorem: [theorem]{}[firstmaintheorem]{} \[thm:idempotent-zero-diagonal\] For $D\in B(H)$ an infinite rank idempotent the following are equivalent: 1. \[item:mthm1\] $D$ is not a Hilbert–Schmidt perturbation of a projection; 2. \[item:mthm2\] the nilpotent part of $D$ is not Hilbert–Schmidt; 3. \[item:mthm3\] $R\{\operatorname{Tr}D\}=\mathbb{C}$; 4. \[item:mthm4\] $D$ is zero-diagonal; 5. \[item:mthm5\] $D$ has an absolutely summable diagonal; 6. \[item:mthm6\] $D$ has a summable diagonal (i.e., $R\{\operatorname{Tr}D\}\not=\emptyset$). We have not yet defined the *nilpotent part* of an idempotent $D$, but it is a natural object defined in Lemma \[lem:idempotent-decomp\] that gives a canonical decomposition for idempotents. It turns out that \[item:mthm5\] and \[item:mthm6\] of Theorem \[thm:idempotent-zero-diagonal\] are actually equivalent for *any* bounded operator, not merely idempotents (see Proposition \[prop:sum-iff-abs-sum-diag\]). Our next main theorem answers Jasper’s frame theory problem which, as equivalently stated earlier, characterizes diagonals of the class of idempotents. The equivalence of these two problems was originally described to us by Jasper, but a fairly concise explanation can be found on the MathOverflow post: <http://mathoverflow.net/q/132592>. [theorem]{}[secondmaintheorem]{} \[thm:idempotent-diagonals-ell-infty\] Every $\langle d_n \rangle\in\ell^\infty$ admits an idempotent $D\in B(H)$ whose diagonal is $\langle d_n \rangle$ with respect to a basis $\mathfrak{b}$. Herein we use $\mathfrak{b}$ to denote a target basis, whereas we use $\mathfrak{e}$ to denote an arbitrary basis. While Theorem \[thm:idempotent-diagonals-ell-infty\] can be viewed as an extension of Theorem \[thm:idempotent-matrix-diagonals\], so also can our last main theorem. [theorem]{}[thirdmaintheorem]{} \[thm:finite-rank-idempotent-diagonals\] The diagonals of the class of nonzero finite rank idempotents consist precisely of those absolutely summable sequences whose sum is a positive integer. ZERO-DIAGONAL IDEMPOTENTS ========================= We begin with a canonical decomposition of idempotents into $2\times 2$ operator matrices. \[lem:idempotent-decomp\] Let $D^2=D\in B(H)$ be an idempotent. Then with respect to the decomposition $H=\ker^\perp D\oplus \ker D$, $D$ has the following block matrix form: $$D = \begin{pmatrix} I & 0 \\ T & 0 \\ \end{pmatrix},$$ where $I \in B(\ker^\perp D)$ is the identity operator and $T\in B(\ker^\perp D,\ker D)$ is a bounded operator which we call the **nilpotent part** of the idempotent $D$, short for the corner of the nilpotent operator $ \begin{psmallmatrix} 0 & 0 \\ T & 0 \\ \end{psmallmatrix} $. Note that the term ‘nilpotent part’ is a natural slight abuse of language in that $T$ itself is not nilpotent; $T^2$ is not even defined. The only non-obvious fact we must prove is that the upper left-hand corner of $D$ is the identity on the compression to $\ker^\perp D$. To verify this let $x \oplus 0 \in \ker^\perp D$ be arbitrary and let $D(x \oplus 0) = y \oplus z$. Then because $D$ is idempotent one has $$D(x \oplus 0) = D^2(x \oplus 0) = D(y \oplus z) = D(y \oplus 0).$$ Since $x\oplus 0, y\oplus 0\in \ker^\perp D$ on which $D$ acts one-to-one, $x=y$. An important stepping stone to our first main theorem is the following proposition in which the idempotent acts on $H\oplus H$ and its nilpotent part is normal. \[prop:stepping-stone\] Suppose $H$ is separable infinite dimensional and the idempotent $D\in B(H \oplus H)$ has the respective block matrix form $$D = \begin{pmatrix} I & 0 \\ T & 0 \\ \end{pmatrix}$$ where $T\in B(H)$ is normal. Then 1. \[item:1\] both $({\ensuremath{\mathfrak{Im}}}\, D)_\pm = ({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\nicefrac{\mathrm{i}\pi}{2}}D)_\pm \in \mathcal{C}_1$ if and only if $T \in \mathcal{C}_1$; 2. \[item:2\] $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)_+ \notin \mathcal{C}_1$ for $\nicefrac{-\pi}{2}<\theta<\nicefrac{\pi}{2}$; 3. \[item:3\] $({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)_- \in \mathcal{C}_1$ for $\nicefrac{-\pi}{2}<\theta<\nicefrac{\pi}{2}$ if and only if $T \in \mathcal{C}_2$. The core of the proof is an analysis of the $2\times 2$ case followed by a straightforward application of the Borel functional calculus to the operator case. For $z\in \mathbb{C}$, let $A_z\in M_2(\mathbb{C})$ be given by $$A_z \coloneqq{} \begin{pmatrix} 1 & 0 \\ z & 0 \\ \end{pmatrix}.$$ Then fixing $-\nicefrac{\pi}{2}<\theta\le\nicefrac{\pi}{2}$, $$2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}A_z) = \mathrm{e}^{\mathrm{i}\theta}A_z + \mathrm{e}^{-\mathrm{i}\theta}A_z^* = \begin{pmatrix} 2\cos\theta & \mathrm{e}^{-\mathrm{i}\theta}\bar z \\ \mathrm{e}^{\mathrm{i}\theta}z & 0 \\ \end{pmatrix},$$ which has characteristic polynomial $\det(\lambda-2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}A_z))=\lambda^2-2\cos\theta\lambda -{\ensuremath{\lvert z \rvert}}^2$. Hence the selfadjoint matrix $2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}A_z)$ has eigenvalues which depend on $z$ by $$\label{eq:eigenvalues} \lambda_\pm(z) = \cos\theta\pm\sqrt{\cos^2\theta+{\ensuremath{\lvert z \rvert}}^2}.$$ When $z\not=0$, normalized eigenvectors corresponding to these eigenvalues are $$\label{eq:eigenvectors} x_+(z) = \begin{pmatrix} \frac{\lambda_+(z)}{\sqrt{\lambda_+^2(z)+{\ensuremath{\lvert z \rvert}}^2}} \\ \frac{\mathrm{e}^{\mathrm{i}\theta}z}{\sqrt{\lambda_+^2(z)+{\ensuremath{\lvert z \rvert}}^2}} \\ \end{pmatrix} \quad \text{and} \quad x_-(z) = \begin{pmatrix} \frac{\lambda_-(z)}{\sqrt{\lambda_-^2(z)+{\ensuremath{\lvert z \rvert}}^2}} \\ \frac{\mathrm{e}^{\mathrm{i}\theta}z}{\sqrt{\lambda_-^2(z)+{\ensuremath{\lvert z \rvert}}^2}} \\ \end{pmatrix} .$$ On the other hand, when $z=0$, the normalized eigenvectors are just the standard basis $x_+(0) = \begin{psmallmatrix} 1 \\ 0 \\ \end{psmallmatrix} $ and $x_-(0) = \begin{psmallmatrix} 0 \\ 1 \\ \end{psmallmatrix} $. We now return to the operator case. Since $T\in B(H)$ is normal, the Borel functional calculus provides a $*$-homomorphism $\Phi:\mathcal{B}(\operatorname{sp}(T))\to W^*(T)$ from the bounded Borel functions on the spectrum of $T$ to the abelian von Neumann algebra generated by $T$ for which $\cramped{\operatorname{Id}_{\operatorname{sp}T} \xmapsto{\Phi} T}$, where the identity function on $\operatorname{sp}T$ is $\cramped{\operatorname{Id}_{\operatorname{sp}T}}(z) = z$ [@Kadison.Ringrose-1997a Theorem 5.2.9]. Moreover, since $\Phi$ is a $*$-homomorphism, it preserves the partial order on selfadjoint elements. Let $\mathbf{1}\in\mathcal{B}(\operatorname{sp}(T))$ denote the identity element (the map $z\mapsto 1$) of the algebra $\mathcal{B}(\operatorname{sp}(T))$, and $x_\pm^i$ ($i=1,2$) the coordinate functions of the eigenvectors obtained in (\[eq:eigenvectors\]), which are bounded Borel functions on $\mathbb{C}$. Define $$U \coloneqq{} \begin{pmatrix} \Phi(x_+^1) & \Phi(x_-^1) \\ \Phi(x_+^2) & \Phi(x_-^2) \\ \end{pmatrix},$$ which is unitary on $H\oplus H$ because $\Phi$ is a $*$-homomorphism and $\{x_\pm(z)\}$ form a basis for $\mathbb{C}^2$ for every $z\in\mathbb{C}$. That is, because the $z$-functions $x_+^1x_-^1+x_+^2x_-^2 \equiv 0$ and ${\ensuremath{\left\lvert x_{\pm}^1 \right\rvert}}^{2}+{\ensuremath{\left\lvert x_{\pm}^2 \right\rvert}}^{2} \equiv 1$ and $\Phi(\mathbf{1})=I$ is the identity on $H$. And for what follows, recall $\Phi(\operatorname{Id}_{\operatorname{sp}T}) = T$. Furthermore, because $$D = \begin{pmatrix} \Phi(\mathbf{1}) & \Phi(0\cdot\mathbf{1}) \\ \Phi(\operatorname{Id}_{\operatorname{sp}T}) & \Phi(0\cdot\mathbf{1}) \\ \end{pmatrix},$$ where here $\cdot$ denotes multiplication by scalars in the algebra $\mathcal{B}(\operatorname{sp}(T))$ and hence $0\cdot\mathbf{1}$ is simply the zero function, and so also $$2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D) = \begin{pmatrix} \Phi(2\cos\theta\cdot\mathbf{1}) & \Phi(\mathrm{e}^{-\mathrm{i}\theta}\cdot\overline{\operatorname{Id}_{\operatorname{sp}T}}) \\ \Phi(\mathrm{e}^{\mathrm{i}\theta}\cdot\operatorname{Id}_{\operatorname{sp}T}) & \Phi(0\cdot\mathbf{1}) \\ \end{pmatrix},$$ one obtains $$U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U = \begin{pmatrix} \Phi(\lambda_+) & \Phi(0\cdot\mathbf{1}) \\ \Phi(0\cdot\mathbf{1}) & \Phi(\lambda_-) \\ \end{pmatrix}.$$ When $-\nicefrac{\pi}{2}<\theta<\nicefrac{\pi}{2}$ one has $\cos\theta>0$, and therefore $\lambda_+\ge 0$ and $\lambda_-\le 0$. Hence $$\label{eq:1} (U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U)_+ = \Phi(\lambda_+)\oplus 0 \quad\text{and}\quad (U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U)_- = 0\oplus\Phi(-\lambda_-).$$ Moreover, for all $z\in\mathbb{C}$, $$\lambda_+(z) = \cos\theta + \sqrt{\cos^2\theta+{\ensuremath{\lvert z \rvert}}^2} \ge 2\cos\theta.$$ Furthermore, for the same range of $\theta$, and for all $z$ lying inside the closed ball $\bar{B}(0;{\ensuremath{\lVert T \rVert }})\supseteq \operatorname{sp}T$, $$\begin{aligned} -\lambda_-(z) = \sqrt{\cos^2\theta+{\ensuremath{\lvert z \rvert}}^2} - \cos\theta & = \frac{{\ensuremath{\left\lvert z \right\rvert}}^2}{\cos\theta + \sqrt{\cos^2\theta+{\ensuremath{\lvert z \rvert}}^2}} \\ & \ge \frac{{\ensuremath{\left\lvert z \right\rvert}}^2}{\cos\theta + \sqrt{\cos^2\theta+{\ensuremath{\lVert T \rVert }}^2}}, \end{aligned}$$ and $$-\lambda_-(z) = \frac{{\ensuremath{\left\lvert z \right\rvert}}^2}{\cos\theta + \sqrt{\cos^2\theta+{\ensuremath{\lvert z \rvert}}^2}} \le \frac{{\ensuremath{\lvert z \rvert}}^2}{2\cos\theta}.$$ From these inequalities, as Borel functions on the spectrum of $T$, we have the following $z$-function inequalities for $-\nicefrac{\pi}{2}<\theta<\nicefrac{\pi}{2}$: $$\label{eq:10} \lambda_+ \ge 2\cos\theta\cdot\mathbf{1} \quad\text{and}\quad C_1\cdot{\ensuremath{\left\lvert \operatorname{Id}_{\operatorname{sp}T} \right\rvert}}^2 \le -\lambda_- \le C_2\cdot{\ensuremath{\left\lvert \operatorname{Id}_{\operatorname{sp}T} \right\rvert}}^2,$$ where $C_1,C_2$ are the positive constants given by $C_1 \coloneqq{} \frac{1}{\cos\theta + \sqrt{\cos^2\theta+{\ensuremath{\lVert T \rVert }}^2}}$ and $C_2 \coloneqq{} \frac{1}{2\cos\theta}$. After applying $\Phi$ to these inequalities, one has $\Phi(\lambda_+) \ge (2\cos\theta) I$ and $C_1{\ensuremath{\left\lvert T \right\rvert}}^2 \le \Phi(-\lambda_-) \le C_2{\ensuremath{\left\lvert T \right\rvert}}^2$. Applying the trace yields $$\label{eq:11} \begin{aligned} \operatorname{Tr}_{H\oplus H}(U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U)_+ &= \operatorname{Tr}_{H\oplus H}(\Phi(\lambda_+)\oplus 0) \\ &= \operatorname{Tr}_H \Phi(\lambda_+) \ge (2\cos\theta) \operatorname{Tr}_H I = \infty, \end{aligned}$$ and $$\label{eq:13} C_1 \operatorname{Tr}_H{\ensuremath{\left\lvert T \right\rvert}}^2 \le \operatorname{Tr}_H \Phi(-\lambda_-) \le C_2 \operatorname{Tr}_H{\ensuremath{\left\lvert T \right\rvert}}^2.$$ Because $$\label{eq:12} \operatorname{Tr}_{H\oplus H}(U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U)_- = \operatorname{Tr}_{H\oplus H}(0\oplus\Phi(-\lambda_-)) = \operatorname{Tr}_H \Phi(-\lambda_-)$$ and $$\label{eq:14} \operatorname{Tr}_{H\oplus H}(U^*2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)U)_\pm = \operatorname{Tr}_{H\oplus H}(2({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D))_\pm,$$ inequalities (\[eq:11\])–(\[eq:14\]) prove \[item:2\] and \[item:3\]. To prove \[item:1\], simply notice that when $\theta=\nicefrac{\pi}{2}$, we have $\lambda_+ = -\lambda_- = {\ensuremath{\lvert \operatorname{Id}_{\operatorname{sp}T} \rvert}}$ and apply the same arguments as above in and along with the fact that $\Phi({\ensuremath{\lvert \operatorname{Id}_{\operatorname{sp}T} \rvert}}) = {\ensuremath{\left\lvert T \right\rvert}}$ The following remark shows that idempotents can be decomposed even further than the $2\times 2$ matrix of Lemma \[lem:idempotent-decomp\]. \[rem:finer-idempotent-decomp\] With the same notation as Lemma \[lem:idempotent-decomp\], we may further decompose the underlying space as $\ker^{\perp}D=\ker T\oplus \ker^{\perp}T$ and $\ker D = \operatorname{ran}^{\perp}T\oplus \overline{\operatorname{ran}T}$, where $\ker^{\perp}T\coloneqq{} \ker^{\perp}D\ominus \ker T$ and $\operatorname{ran}^{\perp}T\coloneqq{}\ker D\ominus \overline{\operatorname{ran}T}$. With respect to the ordering of subspaces $H = \ker T \oplus \operatorname{ran}^\perp T \oplus \ker^\perp T \oplus \overline{\operatorname{ran}T}$ one can write $$D = \begin{pmatrix} I & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & I & 0 \\ 0 & 0 & \tilde T & 0 \\ \end{pmatrix},$$ where $\tilde T \in B(\ker^\perp T,\overline{\operatorname{ran}T})$, and the identity operators act on the appropriate spaces. In the decomposition above we have used the ordering of subspaces $\ker T \oplus \operatorname{ran}^\perp T \oplus \ker^\perp T \oplus \overline{\operatorname{ran}T}$, which makes it clear that $D$ can be written as the direct sum of a projection and another idempotent. It is possible for this decomposition to degenerate into simpler ones if, say, $\ker T = \{0\}$, in which the first row and column would disappear. Other rows and columns would disappear if their corresponding subspaces were zero, but none of this is problematic. If $Q_3:\ker^{\perp} T\to H$ denotes the (linear) inclusion operator and $Q_4: H\to\overline{\operatorname{ran}T}$ the projection operator, then $\tilde T = Q_4TQ_3$. From this it is clear that $\tilde T$ is injective and has dense range. Furthermore, if $\tilde T = U{\ensuremath{\lvert \tilde T \rvert}}$ is the polar decomposition for $\tilde T$, then $U:\ker^\perp T \to \overline{\operatorname{ran}T}$ is unitary (i.e., a surjective isometry, see [@Halmos1982 Problem 134 and corollaries]). Conjugating $D$ by the unitary $V \coloneqq{} I \oplus I \oplus I \oplus U^* \in B(H,H')$, where $H' = \ker T \oplus \operatorname{ran}^\perp T \oplus \ker^\perp T \oplus \ker^\perp T$, one obtains $$D' \coloneqq{} VDV^* = \begin{pmatrix} I & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & I & 0 \\ 0 & 0 & {\ensuremath{\lvert \tilde T \rvert}} & 0 \\ \end{pmatrix}.$$ We need one more lemma before we can prove our main theorem for this section. \[lem:positive-ideal-perturb\] Let $\mathcal{I}$ be a two-sided ideal of $B(H)$ and let $B=B^*\in\mathcal{I}$ and $A=A^*\in B(H)$. Then $A_+\in\mathcal{I}$ if and only if $(A+B)_+\in\mathcal{I}$. Similarly, $A_-\in\mathcal{I}$ if and only if $(A+B)_-\in\mathcal{I}$ Let $R_{A_+}$ be the range projection of the positive part $A_+$ of $A$. Then since $A+B \le (A+B)_+$, one has $A \le (A+B)_+ - B$. Therefore $$A_+ = R_{A_+}AR_{A_+} \le R_{A_+}((A+B)_+ - B)R_{A_+},$$ and hence $A_+\in\mathcal{I}$ whenever $(A+B)_+\in\mathcal{I}$. Here we are using the fact that two-sided ideals of $B(H)$ are hereditary, which is a well-known consequence of Calkin’s characterization of ideals of $B(H)$ in terms of their $s$-numbers in [@Cal-1941-AoM2]. For the other implication, make the substitutions $A \mapsto A+B$, $B \mapsto - B$ and apply the result just proved. More precisely, one obtains $$(A+B)_+ \le P((A+B)-B)_+ + B)P = P(A_+ + B)P,$$ where $P \coloneqq{} R_{(A+B)_+}$. Hence $(A+B)_+\in\mathcal{I}$ if $A_+\in\mathcal{I}$. To see that $A_-\in\mathcal{I}$ if and only if $(A+B)_-\in\mathcal{I}$, note that $A_- = (-A)_+$ and apply the result just proved. We are now in a position to prove our first main theorem. The implication \[item:mthm1\]$\implies$\[item:mthm2\] is clear, as are the implications \[item:mthm4\]$\implies$\[item:mthm5\]$\implies$\[item:mthm6\]. The implication \[item:mthm3\]$\implies$\[item:mthm4\] is a direct consequence of Theorem \[thm:subsequence-zero-diagonalizability\], for if $R\{\operatorname{Tr}T\}=\mathbb{C}$, then there exists a basis $\mathfrak{e}$ with respect to which $\operatorname{Tr}_\mathfrak{e} T = 0$, and thus by Theorem \[thm:subsequence-zero-diagonalizability\], $T$ is zero-diagonal. Hence the main thrust of this theorem is proving the implications \[item:mthm6\]$\implies$\[item:mthm2\]$\implies$\[item:mthm3\] and \[item:mthm2\]$\implies$\[item:mthm1\]. For the remainder of the proof we use Lemma \[lem:positive-ideal-perturb\], Remark \[rem:finer-idempotent-decomp\], Proposition \[prop:stepping-stone\], and Theorem \[thm:trace-range-shape\] to prove the implications \[item:mthm6\]$\implies$\[item:mthm2\]$\implies$\[item:mthm3\] which, with the above paragraph, establishes the equivalences \[item:mthm2\]–\[item:mthm6\]. Having demonstrated these equivalences, we prove \[item:mthm4\]$\implies$\[item:mthm1\] in lieu of \[item:mthm2\]$\implies$\[item:mthm1\]. We prove the contrapositive, that the nilpotent part of $D$ is Hilbert–Schmidt implies $R\{\operatorname{Tr}D\}=\emptyset$. Suppose the nilpotent part of $D$ is Hilbert–Schmidt. *Case 1: The nilpotent part of $D$ has finite rank.* By Lemma \[lem:idempotent-decomp\], $D$ has the form $$D = \begin{pmatrix} I & 0 \\ T & 0 \\ \end{pmatrix}.$$ Set $$A = \begin{pmatrix} I & 0 \\ 0 & 0 \\ \end{pmatrix} \quad\text{and}\quad B = \frac{1}{2} \begin{pmatrix} 0 & T^* \\ T & 0 \\ \end{pmatrix},$$ and so ${\ensuremath{\mathfrak{Re}}}D=A+B$. By hypothesis, $T$ has finite rank hence $B$ has finite rank. Since $A=A_+\notin\mathcal{C}_1$ and $B\in\mathcal{C}_1$ because $B$ has finite rank, $({\ensuremath{\mathfrak{Re}}}\, D)_+=(A+B)_+\notin\mathcal{C}_1$ by Lemma \[lem:positive-ideal-perturb\]. However, $A_-=0\in\mathcal{C}_1$ and so again Lemma \[lem:positive-ideal-perturb\] ensures $({\ensuremath{\mathfrak{Re}}}\, D)_-=(A+B)_-\in\mathcal{C}_1$. Therefore, $$\label{eq:theta-zero-1} ({\ensuremath{\mathfrak{Re}}}\, D)_+\notin\mathcal{C}_1 \quad\text{and}\quad ({\ensuremath{\mathfrak{Re}}}\, D)_-\in\mathcal{C}_1.$$ Then Theorem \[thm:trace-range-shape\](iv) with $\theta=0$ ensures $R\{\operatorname{Tr}D\}=\emptyset$. *Case 2: The nilpotent part of $D$ has infinite rank.* By Remark \[rem:finer-idempotent-decomp\], write $$D' = \begin{pmatrix} I & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & I & 0 \\ 0 & 0 & {\ensuremath{\lvert \tilde T \rvert}} & 0 \\ \end{pmatrix},$$ and from $\tilde T = Q_4TQ_3$ we know that ${\ensuremath{\lvert \tilde T \rvert}}$ is Hilbert–Schmidt, and since $\tilde T$ has dense range in $\overline{\operatorname{ran}T}$ which is infinite dimensional $\tilde T$, and hence also ${\ensuremath{\lvert \tilde T \rvert}}$, have infinite rank. Define $J:=\ker T\oplus \operatorname{ran}^\perp T$ and $K:=\ker^\perp T$, then set $P\in B(J)$ and $\tilde D\in B(K\oplus K)$ to $$P \coloneqq{} \begin{pmatrix} I & 0 \\ 0 & 0 \\ \end{pmatrix} \quad\text{and}\quad \tilde D \coloneqq{} \begin{pmatrix} I & 0 \\ {\ensuremath{\lvert \tilde T \rvert}} & 0 \\ \end{pmatrix}$$ Then $\tilde D$ satisfies the conditions of Proposition \[prop:stepping-stone\] and so $({\ensuremath{\mathfrak{Re}}}\, D')_+ = P \oplus ({\ensuremath{\mathfrak{Re}}}\, \tilde D)_+ \notin\mathcal{C}_1$ because $$\operatorname{Tr}_H(P \oplus ({\ensuremath{\mathfrak{Re}}}\, \tilde D)_+) = \operatorname{Tr}_J P + \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Re}}}\, \tilde D)_+ \ge \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Re}}}\, \tilde D)_+ \underset{\ref{prop:stepping-stone}\ref{item:mthm2}}{=} \infty.$$ Furthermore, $({\ensuremath{\mathfrak{Re}}}\, D')_- \in \mathcal{C}_1$ because $$({\ensuremath{\mathfrak{Re}}}\, D')_- = 0 \oplus ({\ensuremath{\mathfrak{Re}}}\, \tilde D)_-$$ and $({\ensuremath{\mathfrak{Re}}}\, \tilde D)_-\in\mathcal{C}_1$ by Proposition \[prop:stepping-stone\](iii) since the nilpotent part ${\ensuremath{\lvert \tilde T \rvert}}$ of $\tilde D$ is Hilbert–Schmidt. Therefore $({\ensuremath{\mathfrak{Re}}}\, D')_+\notin \mathcal{C}_1$ and $({\ensuremath{\mathfrak{Re}}}\, D')_-\in \mathcal{C}_1$, and also via unitary equivalence $$\label{eq:theta-zero-2} ({\ensuremath{\mathfrak{Re}}}\, D)_+\notin \mathcal{C}_1 \quad\text{and}\quad ({\ensuremath{\mathfrak{Re}}}\, D)_-\in \mathcal{C}_1.$$ Thus by Theorem \[thm:trace-range-shape\](iv), one has that $R\{\operatorname{Tr}D\} = \emptyset$. Suppose the nilpotent part of $D$ is not Hilbert–Schmidt. Then just like in Case 2 above use Remark \[rem:finer-idempotent-decomp\] to decompose $D' = P \oplus \tilde D$, with $\tilde D$ satisfying the conditions of Proposition \[prop:stepping-stone\](ii). Then for $-\nicefrac{\pi}{2}<\theta<\nicefrac{\pi}{2}$ $$\operatorname{Tr}_H({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D')_+ = \operatorname{Tr}_J(\cos\theta P) + \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}\tilde D)_+ \ge \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}\tilde D)_+ = \infty.$$ Furthermore, since the nilpotent part $T$ of $D$ is not Hilbert–Schmidt, and hence ${\ensuremath{\lvert \tilde T \rvert}}$ is not Hilbert–Schmidt, one has $$\operatorname{Tr}_{H}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D')_- = 0 + \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}\tilde D)_- = \infty$$ by Proposition \[prop:stepping-stone\](iii). Finally, since ${\ensuremath{\lvert \tilde T \rvert}}$ is not Hilbert–Schmidt, neither is it trace-class. Therefore, by Proposition \[prop:stepping-stone\](i) $$\operatorname{Tr}_H({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\nicefrac{\mathrm{i}\pi}{2}}D')_{\pm} = \operatorname{Tr}_H({\ensuremath{\mathfrak{Im}}}\, D')_\pm = 0 + \operatorname{Tr}_{K\oplus K}({\ensuremath{\mathfrak{Im}}}\, \tilde D)_\pm = \infty.$$ Thus we have proven that $\operatorname{Tr}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D)_\pm=\operatorname{Tr}({\ensuremath{\mathfrak{Re}}}\, \mathrm{e}^{\mathrm{i}\theta}D')_\pm=\infty$ for all $-\nicefrac{\pi}{2}<\theta\le\nicefrac{\pi}{2}$ and hence also for all $\theta\in\mathbb{R}$, and so by Theorem \[thm:trace-range-shape\](iv) one has $R\{\operatorname{Tr}D\}=\mathbb{C}$. Having established the equivalence of \[item:mthm2\]–\[item:mthm6\] and the implication \[item:mthm1\]$\implies$\[item:mthm2\], it suffices to prove \[item:mthm4\]$\implies$\[item:mthm1\]. We will in fact prove the contrapositive. To this end, suppose $D$ is a Hilbert–Schmidt perturbation of a projection. That is, $D=P+K$ where $P$ is a projection and $K\in\mathcal{C}_2$. Because $D$ is idempotent one has $$P+K = D = D^2 = P^2 + PK + KP + K^2 = P + PK + KP + K^2,$$ and so $$\label{eq:K=PK+KP+K^2} K = PK + KP + K^2 \qquad\text{and}\qquad PKP = 2PKP + PK^2P,$$ so $PKP=-PK^2P\in\mathcal{C}_1$. Similarly for $P^{\perp}$ one has $P^\perp KP^\perp = P^\perp K^2P^\perp\in\mathcal{C}_1$. Therefore, with respect to the decomposition $H = PH \oplus P^\perp H$, one has $$K = \begin{pmatrix} K_1 & K_2 \\ K_3 & K_4 \\ \end{pmatrix},$$ where $K_1,K_4\in\mathcal{C}_1$ and $K_2,K_3\in\mathcal{C}_2$. A technical note is that $P$ must have infinite rank. Otherwise, if $P$ were finite rank, then so also $K_2,K_3$ would be finite rank. Hence $K$ would be trace-class, and so also would $D=P+K$, which contradicts the fact that $D$ is an infinite rank idempotent because of Lemma \[lem:idempotent-decomp\]. Thus relative to $H = PH \oplus P^\perp H$ we may write $$D = \overbrace{ \begin{pmatrix} I & K_2 \\ K_3 & 0 \\ \end{pmatrix} }^{D_1} + \overbrace{ \begin{pmatrix} K_1 & 0 \\ 0 & K_4 \\ \end{pmatrix} }^{D_2}.$$ Moreover, because $${\ensuremath{\mathfrak{Re}}}\, D_1 = {\ensuremath{\mathfrak{Re}}}\overbrace{ \begin{pmatrix} I & 0 \\ K_2^*+K_3 & 0 \\ \end{pmatrix} }^{\tilde D_1}$$ and $K_2^*+K_3\in\mathcal{C}_2$, by the proof of \[item:mthm6\]$\implies$\[item:mthm2\] (see and for Cases 1 and 2), $({\ensuremath{\mathfrak{Re}}}\, D_1)_+=({\ensuremath{\mathfrak{Re}}}\, \tilde D_1)_+\notin\mathcal{C}_1$ but $({\ensuremath{\mathfrak{Re}}}\, D_1)_-=({\ensuremath{\mathfrak{Re}}}\, \tilde D_1)_-\in\mathcal{C}_1$. So by Theorem \[thm:trace-range-shape\](iv), $R\{\operatorname{Tr}D_1\}=\emptyset$ and hence $D_1$ does not have an absolutely summable diagonal in *any* basis. Because $D_2\in\mathcal{C}_1$, its diagonal in *any* basis is absolutely summable. Therefore, there is no basis in which $D=D_1+D_2$ has a zero diagonal, which completes the proof. The following corollary answers Question \[que:zero-diagonal-finite-rank\] due to Jasper. \[cor:answer-q2\] A nonzero idempotent $D$ is zero-diagonal if and only if it is not a Hilbert–Schmidt perturbation of a projection. If $D$ has infinite rank, this is handled by Theorem \[thm:idempotent-zero-diagonal\]. If $D$ has finite rank, then so does the nilpotent part of $D$. Thus $D$ is a finite rank (and hence Hilbert–Schmidt) perturbation of the zero projection. Furthermore, by Lemma \[lem:idempotent-decomp\] $\operatorname{Tr}D = \operatorname{rank}D >0$ for finite rank idempotents, and so $D$ is not zero-diagonal. In the case of infinite rank projections with infinite dimensional kernel, the next corollary is a strengthening of the result due to Fan [@Fan-1984-TotAMS Theorem 3] that an operator $T$ is a norm limit of zero-diagonal operators if and only if $0\in W_e(T)$, the essential numerical range. For $P$ a projection, $0\in W_e(P)$ if and only if $\operatorname{Tr}P^\perp = \infty$, and thus Fan’s result guarantees such projections are a norm limit of zero-diagonal operators. However, we take this a step further by proving these zero-diagonal operators may be taken to be idempotent so long as $\operatorname{Tr}P = \infty$ as well. Every projection $P$ with $\operatorname{Tr}P =\operatorname{Tr}P^\perp =\infty$ is a norm limit of zero-diagonal idempotents. For $P=I\oplus 0$ consider idempotents $ \begin{psmallmatrix} I & 0 \\ T & 0 \\ \end{psmallmatrix} $ whose nilpotent part has arbitrarily small norm but is not Hilbert–Schmidt and apply Theorem 2.5 (ii)$\iff$(iv). Constructing bases to achieve zero-diagonality {#constructing-bases-to-achieve-zero-diagonality .unnumbered} ---------------------------------------------- The proof of Theorem \[thm:idempotent-zero-diagonal\] was existential in the sense that it did not explicitly construct a basis in which a given idempotent has zero diagonal. The remainder of this section is devoted to providing an algorithm for constructing such a basis when it exists (i.e., when the idempotent is not a Hilbert–Schmidt perturbation of a projection, which is included in the case when $\dim\ker D = \infty = \dim\ker^{\perp} D$). As with the proof of Proposition \[prop:stepping-stone\], a careful consideration first of the $2\times 2$ case is in order. \[rem:2x2-theta-minimum\] Consider a $2\times 2$ idempotent matrix, $D$, and the counterclockwise rotation matrix through an angle $\theta$, $R_\theta$, given by the formulas $$D = \begin{pmatrix} 1 & 0 \\ d & 0 \\ \end{pmatrix} \quad \text{and}\quad R_\theta = \begin{pmatrix} \cos\theta & -\sin\theta \\ \sin\theta & \cos\theta \\ \end{pmatrix},$$ where $d\ge 0$. Conjugating $D$ by $R_\theta$ is equivalent to changing the basis for $\mathbb{C}^2$: $$\begin{aligned} R_{-\theta}DR_\theta &= \begin{pmatrix} \cos\theta & \sin\theta \\ -\sin\theta & \cos\theta \\ \end{pmatrix} \begin{pmatrix} 1 & 0 \\ d & 0 \\ \end{pmatrix} \begin{pmatrix} \cos\theta & -\sin\theta \\ \sin\theta & \cos\theta \\ \end{pmatrix} \\ &= \begin{pmatrix} \cos^2\theta+d\sin\theta\cos\theta & -\sin\theta\cos\theta-d\sin^2\theta \\ -\sin\theta\cos\theta+d\cos^2\theta & \sin^2\theta-d\sin\theta\cos\theta \\ \end{pmatrix} \\ &= \begin{pmatrix} \frac{1+\cos 2\theta+d\sin 2\theta}{2} & -\sin\theta\cos\theta-d\sin^2\theta \\ -\sin\theta\cos\theta+d\cos^2\theta & \frac{1-\cos 2\theta-d\sin 2\theta}{2} \\ \end{pmatrix}. \end{aligned}$$ Elementary calculus shows that the minimum diagonal entry occurs when $\theta=\frac{\arctan d}{2}$ and corresponds to a negative value of $$-d^- \coloneqq{} \frac{1}{2}\left(1-\sqrt{1+d^2}\right) = \frac{-d^2}{2\left(1+\sqrt{1+d^2}\right)}.$$ Since the trace is basis independent, the other diagonal entry is necessarily $1+d^-$. Furthermore, by continuity of the diagonal entries as a function of $\theta$, for any value $x$ with $-d^-\le x\le 0$, there is some $\theta$ for which one of the diagonal entries is $x$. We require the following elementary result in linear algebra [@HJ-1990-MA Page 77, Problem 3]. It’s proof by induction is straightforward and we include it here for completeness. \[lem:zero-diagonal-trace-zero\] Let $X\in M_n(\mathbb{C})$. Then $\operatorname{Tr}X = 0$ if and only if there is a basis in which $X$ has zero diagonal. One direction is clear, so suppose $\operatorname{Tr}X = 0$. We proceed by induction on the size $n$ of the $n \times n$ matrix $X$. The case $n=1$ is clear. Given any basis $\{e_j\}_{j=1}^n$, one has $$0 = \operatorname{Tr}X = \sum_{j=1}^n (Xe_j,e_j),$$ and therefore also $0 = \sum_{j=1}^n \frac{(Xe_j,e_j)}{n}$. Thus zero is in the convex hull of $\{(Xe_j,e_j)\} \subseteq W(X)$. But the Toeplitz–Hausdorff Theorem, that the numerical range $W(X)$ is convex, ensures $0 \in W(X)$. So there is some unit vector $f_1$ for which $(Xf_1,f_1) = 0$. Let $P$ be the projection onto the orthogonal complement of $f_1$. Then we find $$0 = \operatorname{Tr}X = (Xf_1,f_1) + \operatorname{Tr}(PXP) = \operatorname{Tr}(PXP).$$ The matrix $PXP$ can be viewed as being of size $(n-1) \times (n-1)$ by expressing it in a basis which contains $f_1$ and deleting the row and column corresponding to $f_1$ (which consist solely of zeros). By applying the inductive hypothesis to $PXP$ we obtain orthonormal vectors $f_2,\ldots,f_n$ which are orthogonal to $f_1$ and satisfy $(Xf_j,f_j) = (XPf_j,Pf_j) = (PXPf_j,f_j) = 0$ for $2 \le j \le n$. Therefore $\{f_j\}_{j=1}^n$ is a basis with respect to which $X$ has zero diagonal. We will use the following obvious corollary of Lemma \[lem:zero-diagonal-trace-zero\] extensively in the next section. \[cor:trace-nlambda-diagonal-lambda\] Let $X\in M_n(\mathbb{C})$. Then $\operatorname{Tr}X = n\lambda$ if and only if there is a basis in which $X$ has constant diagonal sequence $\lambda$. More generally, if $X\in B(H)$ with basis $\{e_n\}_{n\in\mathbb{N}}$ and $\langle n_k \rangle_{k=1}^m$ a finite subsequence of $\mathbb{N}$ with restricted trace $\sum_{k=1}^m (Xe_{n_k},e_{n_k}) = m\lambda$, then there is an orthonormal set $\{f_{n_k}\}_{k=1}^m$ for which $(Xf_{n_k},f_{n_k})=\lambda$ for $k=1,\ldots,m$ and $\operatorname{span}\{f_{n_k}\}_{k=1}^m = \operatorname{span}\{e_{n_k}\}_{k=1}^m$. For $X\in M_n(\mathbb{C})$ apply Lemma \[lem:zero-diagonal-trace-zero\] to $X-\lambda I$ and note that $\lambda I$ has constant diagonal sequence $\lambda$ with respect to *any* basis. For the general case $X\in B(H)$, let $P$ be the projection on $\operatorname{span}\{e_{n_k}\}_{k=1}^m$ and apply the matrix result to $PXP$. Then simply notice that $(PXPf_{n_k},f_{n_k}) = (XPf_{n_k},Pf_{n_k}) = (Xf_{n_k},f_{n_k})$. We are now ready to provide our algorithm. It requires an elementary theoretical first step with all succeeding steps algorithmic. Suppose that $D\in B(H)$ is not a Hilbert–Schmidt perturbation of a projection. Then the following explicitly constructs (i.e., gives an algorithm for producing) a basis in which $D$ is zero-diagonal. By Theorem \[thm:idempotent-zero-diagonal\], the nilpotent part of $D$ is not Hilbert–Schmidt. Then by introduction ((ii)-, contrapositive), there exists a basis in which the diagonal of the nilpotent part is not square-summable. That is, there exists a basis for $H$ for which $$\label{eq:nilpotent-part-diag-not-ell-2} D = \left( \begin{array}{ccc|ccc} 1 & \cdots & 0 & 0 & \cdots & 0 \\ \vdots & 1 & \vdots & \vdots & \ddots & \vdots \\ 0 & \cdots & \ddots & 0 & \cdots & \ddots \\ \hline d_1 & \cdots & {\Huge *} & 0 & \cdots & 0 \\ \vdots & d_2 & \vdots & \vdots & \ddots & \vdots \\ {\Huge *} & \cdots & \ddots & 0 & \cdots & \ddots \\ \end{array} \right)$$ with $\langle d_n \rangle\in\ell^\infty\setminus\ell^2$. Furthermore, by conjugating by a unitary $U$ of the form $U = I \oplus \operatorname{diag}\langle u_n \rangle$, we may even assume without loss of generality that $d_n\ge 0$. Let the basis which gives the form equation be $\mathfrak{e}\coloneqq{}\{e_n,e'_n\}_{n\in\mathbb{N}}$. We will transform these into a new basis $\mathfrak{f}\coloneqq{}\{f_n,f'_n\}_{n\in\mathbb{N}}$ for which $\operatorname{span}\{e_n,e'_n\} = \operatorname{span}\{f_n,f'_n\}$ for each $n\in\mathbb{N}$. Specifically, $f_n = \cos\theta_ne_n+\sin\theta_ne'_n$ and $f'_n = -\sin\theta_ne_n+\cos\theta_ne'_n$ form a rotation of the pair $e_n,e'_n$ through an angle $\theta_n$ which we will choose momentarily. Recall and notice that $$\sum_{n=1}^\infty d^-_n = \sum_{n=1}^\infty \frac{d_n^2}{2\big(1+\sqrt{1+d_n^2}\big)} \ge \frac{1}{2\big(1+\sqrt{1+{\ensuremath{\lVert \langle d_n \rangle \rVert }}_\infty^2}\big)} \sum_{n=1}^\infty d_n^2 = \infty.$$ Let $m_1$ be the smallest integer for which $\sum_{n=1}^{m_1} d^-_n \ge 1+d^-_1$. Necessarily $m_1 \ge 2$. Now define $\theta_n = \frac{\arctan d_n}{2}$ for $1\le n< m_1$, hence by , $-d_n^-=(Df'_n,f'_n)$ and $1+d_n^-=(Df_n,f_n)$. Our choice of $m_1$ guarantees $$\sum_{n=1}^{m_1-1} d^-_n < 1+d^-_1 \le \sum_{n=1}^{m_1} d^-_n, \quad\text{and thus}\quad -d^-_{m_1} \le -1-d^-_1 + \sum_{n=1}^{m_1-1} d^-_n < 0.$$ For the latter, using the continuity described in (last sentence) choose $\theta_{m_1}$ so that $$(Df'_{m_1},f'_{m_1}) = -1-d^-_1 + \sum_{n=1}^{m_1-1} d^-_n,$$ and therefore $$\sum_{n=1}^{m_1} -(Df'_n,f'_n) = \sum_{n=1}^{m_1-1} d^-_n - (Df'_{m_1},f'_{m_1}) = 1+d^-_1 = (Df_1,f_1).$$ We will now inductively define the sequences $\langle m_k \rangle$ and $\langle \theta_n \rangle$ in the following interwoven fashion. Suppose that these sequences are already defined up to $m_{k-1}$ and $\theta_{m_{k-1}}$. Let $m_k$ be the smallest positive integer for which $$\sum_{n=m_{k-1}+1}^{m_k} d^-_n \ge 1 - (Df'_k,f'_k) = (Df_k,f_k).$$ Then for $m_{k-1} < n < m_k$, let $\theta_n = \frac{\arctan d_n}{2}$, and as above let $\theta_{m_k}$ be chosen so as to satisfy $$\sum_{n=m_{k-1}+1}^{m_k} -(Df'_n,f'_n) = \sum_{n=m_{k-1}+1}^{m_k-1} d^-_n - (Df'_{m_k},f'_{m_k}) = 1 - (Df'_k,f'_k) = (Df_k,f_k).$$ Finally observe from this that with respect to the basis $\{f_n,f'_n\}_{n\in\mathbb{N}}$ the diagonal sequence of $D$ can be partitioned into finite subsets $\{A_k\}_{k\in\mathbb{N}}$ for which the sum over each subset is zero. Indeed, let $A_k$ consist of the diagonal entries corresponding to the basis elements $\mathfrak{f}_k \coloneqq{}\{f_k,f'_{m_{k-1}+1},\ldots,f'_{m_k}\}$. So for each $k\in\mathbb{N}$ we may apply Lemma \[cor:trace-nlambda-diagonal-lambda\] to the collection $\mathfrak{f}_k$ to obtain a new collection of orthonormal vectors $\mathfrak{g}_k$ with $\operatorname{span}\mathfrak{f}_k = \operatorname{span}\mathfrak{g}_k$ and the diagonal of $D$ with respect to $\mathfrak{g}_k$ is constantly zero. Thus $D$ has a zero diagonal with respect to the basis $\mathfrak{g} \coloneqq{} \bigcup_k \mathfrak{g}_k$. We stated in the introduction that Theorem \[thm:idempotent-zero-diagonal\]\[item:mthm5\] and \[item:mthm6\] are equivalent for *any* bounded operator, not merely idempotents, which we now prove. \[prop:sum-iff-abs-sum-diag\] An operator $T\in B(H)$ has an absolutely summable diagonal in some basis if and only if it has a summable diagonal in some basis. One direction is trivial. For the other direction, suppose that $T\in B(H)$ and $\mathfrak{e} \coloneqq{} \{e_n\}_{n\in\mathbb{N}}$ is a basis with respect to which the corresponding diagonal $\langle d_n \rangle$ is summable with sum $s$. Then there exists a strictly increasing sequence of positive integers $\langle n_k \rangle$ with the property that ${\ensuremath{\left\lvert s_{n_k}-s \right\rvert}} \le 2^{-k}$, where $s_m$ denotes the partial sum $\sum_{n=1}^m d_j$. Since $\sum_{j=1}^{n_1} d_j = s_{n_1}$, by Corollary \[cor:trace-nlambda-diagonal-lambda\] there is an orthonormal set $\{b_j\}_{j=1}^{n_1}$ for which $\operatorname{span}\{b_j\}_{j=1}^{n_1} = \operatorname{span}\{e_j\}_{j=1}^{n_1}$ and $(Tb_j,b_j) = \nicefrac{s_{n_1}}{n_1}$. Similarly for each $k\in\mathbb{N}$, because $\sum_{j=n_k+1}^{n_{k+1}} d_j = s_{n_{k+1}}-s_{n_k}$ there is an orthonormal set $\{b_j\}_{j=n_k+1}^{n_{k+1}}$ for which $\operatorname{span}\{b_j\}_{j=n_k+1}^{n_{k+1}}=\operatorname{span}\{e_j\}_{j=n_k+1}^{n_{k+1}}$ and $(Tb_j,b_j) = \nicefrac{(s_{n_{k+1}}-s_{n_k})}{n_{k+1}-n_k}$. Thus $\mathfrak{b} \coloneqq{} \{b_j\}_{j=1}^\infty$ is a basis since $\operatorname{span}\mathfrak{b} = \operatorname{span}\mathfrak{e}$. For convenience of notation, set $n_0=0=s_{n_0}$. Then with respect to the basis $\mathfrak{b}$, the diagonal sequence is absolutely summable since $$\begin{aligned} \sum_{j=1}^\infty {\ensuremath{\left\lvert (Tb_j,b_j) \right\rvert}} &= \sum_{k=0}^\infty \sum_{j=n_k+1}^{n_{k+1}} {\ensuremath{\left\lvert (Tb_j,b_j) \right\rvert}} \\ &= \sum_{k=0}^\infty \sum_{j=n_k+1}^{n_{k+1}} \frac{{\ensuremath{\lvert s_{n_{k+1}}-s_{n_k} \rvert}}}{n_{k+1}-n_k} \\ &= \sum_{k=0}^\infty {\ensuremath{\left\lvert s_{n_{k+1}}-s_{n_k} \right\rvert}} \\ &= {\ensuremath{\left\lvert s_{n_1} \right\rvert}} + \sum_{k=1}^\infty \left({\ensuremath{\left\lvert s_{n_{k+1}}-s \right\rvert}} + {\ensuremath{\left\lvert s-s_{n_k} \right\rvert}}\right) \\ &\le {\ensuremath{\left\lvert s_{n_1} \right\rvert}} + \sum_{k=1}^\infty \left(2^{-(k+1)} + 2^{-k}\right) \\ &= {\ensuremath{\left\lvert s_{n_1} \right\rvert}} + \frac{3}{2}. \qedhere\end{aligned}$$ DIAGONALS OF THE CLASS OF IDEMPOTENTS AND APPLICATIONS ====================================================== In this section we investigate Jasper’s initial frame theory problem concerning dual frame pairs via its equivalent operator-theoretic formulation: Characterize the diagonals of the class of idempotent operators. In particular, we prove that every bounded sequence appears as the diagonal of some idempotent (Theorem \[thm:idempotent-diagonals-ell-infty\]). We prove this result in stages. First we consider diagonals of idempotents in $M_2(\mathbb{C})$ (Lemma \[lem:2x2-lemma\]). Then we give a direct sum construction of an idempotent with constant diagonal (Proposition \[prop:idempotent-constant-diagonal\]). From this we show that any bounded sequence with at least one value repeated infinitely many times appears as the diagonal of some idempotent (Proposition \[prop:infinite-multiplicity-diagonal\]). And we conclude by showing that we may obtain any bounded sequence as the diagonal of an idempotent. The following technical lemma is a trivial corollary of Theorem \[thm:idempotent-matrix-diagonals\] except for its norm bound which we require for the forthcoming results. \[lem:2x2-lemma\] If $d\in\mathbb{C}$, then there is a $2\times 2$ idempotent $D\in M_2(\mathbb{C})$ with norm ${\ensuremath{\lVert D \rVert }} \le 6{\ensuremath{\left\lvert d \right\rvert}}+4$ which takes the values $3d-1,-3d+2$ on its diagonal. Start with the idempotent $$D_z = \begin{pmatrix} 1 & 0 \\ z & 0 \\ \end{pmatrix},$$ with $z\in\mathbb{C}$ to be chosen later. Conjugating by the (unitary) rotation matrix $R_{\nicefrac{\pi}{4}}$, one obtains $$\begin{aligned} R_{\nicefrac{\pi}{4}}D_zR_{-\nicefrac{\pi}{4}} &= \begin{pmatrix} \frac{1}{\sqrt{2}} & -\frac{1}{\sqrt{2}} \\ \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \end{pmatrix} \begin{pmatrix} 1 & 0 \\ z & 0 \\ \end{pmatrix} \begin{pmatrix} \frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ -\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{2}} \\ \end{pmatrix} = \begin{pmatrix} \frac{1-z}{2} & \frac{1-z}{2} \\ \frac{1+z}{2} & \frac{1+z}{2} \\ \end{pmatrix} \end{aligned}$$ Choosing $z = 6d-3$ gives the correct diagonal values. Furthermore, ${\ensuremath{\lVert D_z \rVert }} \le 1 + {\ensuremath{\left\lvert z \right\rvert}} \le 6d + 4$. Then $D = R_{\nicefrac{\pi}{4}}D_zR_{-\nicefrac{\pi}{4}}$ gives our required idempotent. In the next proposition we exhibit an idempotent with constant diagonal $d$. The idea is to take an infinite direct sum of the $2\times 2$ matrix $D$ from Lemma \[lem:2x2-lemma\] (whose diagonal entries $d_1,d_2$ satisfy $2d_1 + d_2 = 3d$), regroup the diagonal entries and apply Corollary \[cor:trace-nlambda-diagonal-lambda\] repeatedly. \[prop:idempotent-constant-diagonal\] Given $d\in\mathbb{C}$, there is an idempotent $D_d\in B(H)$ with norm ${\ensuremath{\left\lVert D_d \right\rVert }} \le 6{\ensuremath{\left\lvert d \right\rvert}}+4$ with constant diagonal $d$ in some basis. Let $D'$ be the $2\times 2$ idempotent matrix obtained from Lemma \[lem:2x2-lemma\] and set $D = \bigoplus_{i=1}^\infty D'$. Then the diagonal of $D$ consists of the values $d_1 = 3d-1$ and $d_2 = -3d+2$, each repeated infinitely many times. With respect to the basis $\mathfrak{e} \coloneqq{} \{e_j\}_{j\in\mathbb{N}}$, the diagonal entries are $$(De_j,e_j) = \begin{cases} d_1 & \text{if $j$ is odd,} \\ d_2 & \text{if $j$ is even,} \\ \end{cases}$$ and these diagonal entries satisfy $2d_1 + d_2 = 3d$. Let $\pi$ be any permutation of $\mathbb{N}$ which sends $2\mathbb{N}$ onto $3\mathbb{N}$ (i.e., maps the even positive integers to positive multiples of three). Create a new basis $\mathfrak{f}:=\{f_j\}_{j\in\mathbb{N}}$ by $f_j := e_{\pi^{-1}(j)}$. Then we have $$(Df_j,f_j) = (De_{\pi^{-1}(j)},e_{\pi^{-1}(j)}) = \begin{cases} d_1 & \text{if $j\in\mathbb{N}\setminus 3\mathbb{N}$,} \\ d_2 & \text{if $j\in 3\mathbb{N}$.} \\ \end{cases}$$ For each $j\in 3\mathbb{N}$, the sum of the diagonal entries corresponding to $f_{j-2},f_{j-1},f_j$ is $2d_1+d_2 = 3d$. Thus for each $j\in 3\mathbb{N}$ we may apply Corollary \[cor:trace-nlambda-diagonal-lambda\] to obtain new orthonormal vectors $g_{j-2},g_{j-1},g_j$ with $\operatorname{span}\{f_{j-2},f_{j-1},f_j\} = \operatorname{span}\{g_{j-2},g_{j-1},g_j\}$ (hence $\mathfrak{g}:=\{g_k\}_{k\in\mathbb{N}}$ is a basis) and $(Dg_k,g_k) = d$ for any $k\in\mathbb{N}$. Taking $D_d:= D$ with respect to the basis $\mathfrak{g}$ is the required idempotent. Using Proposition \[prop:idempotent-constant-diagonal\] we will now prove that any bounded sequence with at least one value repeated infinitely many times appears as the diagonal of some idempotent. \[prop:infinite-multiplicity-diagonal\] Suppose $d := \langle d_j \rangle \in \ell^\infty$ and for some $m$ one has $d_m=d_k$ for infinitely many $k\in\mathbb{N}$. Then there exists an idempotent $D\in B(H)$ with diagonal $d$ for which ${\ensuremath{\left\lVert D \right\rVert }} \le 18{\ensuremath{\left\lVert d \right\rVert }}_\infty+4$. Observe that the direct sum of idempotents from Proposition \[prop:idempotent-constant-diagonal\]: $$D := \bigoplus_{j=1}^\infty (D_{d_j} \oplus D_{-d_j+2d_m}),$$ is a bounded operator whose norm satisfies $$\begin{aligned} {\ensuremath{\left\lVert D \right\rVert }} &= \sup_{j} \{{\ensuremath{\lVert D_{d_j} \rVert }},{\ensuremath{\lVert D_{-d_j+2d_m} \rVert }} \} \\ &\le \sup_{j} \{6{\ensuremath{\left\lvert d_j \right\rvert}}+4,6{\ensuremath{\left\lvert -d_j+2d_m \right\rvert}}+4 \} \\ &\le 18{\ensuremath{\lVert d \rVert }}_\infty+4. \end{aligned}$$ The idempotent $D$ comes with an associated basis $\mathfrak{e}:=\{e_{i,j,k} \mid i=1,2;\ j,k\in\mathbb{N}\}$ with respect to which the diagonal is $$(De_{i,j,k},e_{i,j,k}) = \begin{cases} d_j & \text{if $i=1$,} \\ -d_j+2d_m & \text{if $i=2$.} \\ \end{cases}$$ Create a new basis by the following procedure. Set $f_j := e_{1,j,1}$, so that $(Df_j,f_j) = d_j$. Then for each $j,k\in\mathbb{N}$, apply Corollary \[cor:trace-nlambda-diagonal-lambda\] to the pair $e_{1,j,k+1},e_{2,j,k}$ to obtain orthonormal vectors $g_{1,j,k},g_{2,j,k}$ with the same span and corresponding diagonal entries $d_m = \frac{1}{2}(d_j+(-d_j+2d_m))$. Then $\mathfrak{g} := \{f_j\}_{j\in\mathbb{N}} \cup \{g_{i,j,k} \mid i=1,2;\ j,k\in\mathbb{N}\}$ is a basis with diagonal entries $d=\langle d_j \rangle$ (from the $f_j$) along with $d_m$ with infinite multiplicity (from the $g_{i,j,k}$). Since the value $d_m$ is repeatedly infinitely many times in the sequence $d$, after a suitable relabeling (permutation of the basis), the diagonal is precisely the sequence $d$. Before we prove our main result for this section we need Fan’s quantitative version of the Toeplitz–Hausdorff Theorem on the convexity of the numerical range. As a matter of notation, throughout the remainder of this paper we will use $[a,b]$ to denote the complex line segment joining $a,b \in \mathbb{C}$. Then each $d \in [a,b]$ has a *convexity coefficient* $\lambda$ defined by $d = \lambda a + (1-\lambda)b$ for $0 \le \lambda \le 1$, with the convention that $\lambda = 0$ when $a=b$. Equivalently, $\lambda = \frac{b-d}{b-a}$ if $b \not= a$ and $\lambda = 0$ if $b = a$. \[lem:quantitative-toeplitz-hausdorff\] Let $$A = \begin{pmatrix} d_1 & {\Huge *} \\ {\Huge *} & d_2 \\ \end{pmatrix} \in M_2(\mathbb{C})$$ be a matrix with respect to the basis $\{e_1,e_2\}$ and let $d \in [d_1,d_2]$ with convexity coefficient $\lambda$. Then there exists a basis $\{b,f\}$ for which $(Af,f)=d$, $(Ab,b)=d_1+d_2-d$ and ${\ensuremath{\left\lvert (e_1,f) \right\rvert}}^2 \le \lambda$. We bootstrap this lemma to modify diagonals in an interesting useful way in . A main tool is to use the following lemma in the case all convexity coefficients $\lambda_n \equiv \nicefrac{1}{2}$ to prove both and . \[lem:bootstrap-fan\] Suppose that $T$ is an operator and $\mathfrak{e} = \{e_n\}_{n=0}^{\infty}$ an orthonormal set. Let $r_n \coloneqq{} (Te_n,e_n)$ and suppose ${\ensuremath{\left\langle d_n \right\rangle}}_{n=1}^{\infty}$ is a sequence such that for $n \ge 1$, $d_n \in [d_{n-1},r_n]$ with convexity coefficient $\lambda_n$ and where $d_0 \coloneqq{} r_0$. Then there is an orthonormal set $\mathfrak{b} = \{b_n\}_{n=1}^{\infty}$ for which $(Tb_n,b_n) = r_n + d_{n-1} - d_n$. Moreover, if $\prod_{i=n}^{\infty} \lambda_i = 0$ for all $n \in \mathbb{N}$, then $\operatorname{span}\mathfrak{b} = \operatorname{span}\mathfrak{e}$. Set $f_0 \coloneqq{} e_0$. Since $d_1 \in [d_0,r_1] = [r_0,r_1]$, by with diagonal entries $r_0,r_1$, there exist orthonormal $f_1,b_1$ for which $\operatorname{span}\{f_1,b_1\} = \operatorname{span}\{f_0,e_1\}$ and $(Tf_1,f_1) = d_1$ and $(Tb_1,b_1) = r_1+r_0-d_1 = r_1+d_0-d_1$ and ${\ensuremath{\left\lvert (f_1,f_0) \right\rvert}}^2 \le \lambda_1$. Iterating this procedure produces an orthonormal set $\mathfrak{b} = \{b_n\}_{n=1}^{\infty}$ and a sequence of unit vectors $\{f_n\}_{n=0}^{\infty}$ satisfying, for each $n\in\mathbb{N}$, 1. $\operatorname{span}\{f_n,b_n\} = \operatorname{span}\{f_{n-1},e_{n}\}$; \[item:2x2-span-3\] 2. $(Tf_n,f_n) = d_n$ and $(Tb_n,b_n) = r_n+d_{n-1}-d_n$; \[item:desired-diagonal-3\] 3. ${\ensuremath{\left\lvert (f_n,f_{n-1}) \right\rvert}}^2 \le \lambda_n$; \[item:small-inner-product-3\] 4. $\{b_1,\ldots,b_n,f_n\}$ is an orthonormal set; \[item:orthonormality-3\] 5. $\operatorname{span}\{b_1,\ldots,b_n,f_n\} = \operatorname{span}\{e_0,\ldots,e_n\}$. \[item:same-span-3\] We prove this via induction. The case $n=1$ is handled in the first paragraph. Suppose that \[item:2x2-span-3\]$_n$–\[item:same-span-3\]$_n$ hold for some fixed $n \in \mathbb{N}$. Then by hypothesis and \[item:desired-diagonal-3\]$_n$ one has $d_{n+1} \in [d_n,r_{n+1}] = [(Tf_n,f_n),(Te_{n+1},e_{n+1})]$, so we may apply to obtain orthonormal $f_{n+1},b_{n+1}$ for which \[item:2x2-span-3\]$_{n+1}$–\[item:small-inner-product-3\]$_{n+1}$ hold. By \[item:orthonormality-3\]$_n$ we know that $f_n$ is orthogonal to $\operatorname{span}\{b_1,\ldots,b_n\}$, and by \[item:same-span-3\]$_n$ we know $e_{n+1}$ is orthogonal to $\operatorname{span}\{b_1,\ldots,b_n\}$. Thus we obtain $\operatorname{span}\{b_1,\ldots,b_n\}$ is orthogonal to $\operatorname{span}\{f_n,e_{n+1}\} = \operatorname{span}\{b_{n+1},f_{n+1}\}$ by \[item:2x2-span-3\]$_{n+1}$, thereby establishing \[item:orthonormality-3\]$_{n+1}$. Finally, by \[item:2x2-span-3\]$_{n+1}$ and \[item:same-span-3\]$_n$ we find $$\operatorname{span}\{b_1,\ldots,b_{n_+1},f_{n+1}\} = \operatorname{span}\{b_1,\ldots,b_n,f_n,e_{n+1}\} = \operatorname{span}\{e_0,\ldots,e_{n+1}\},$$ proving \[item:same-span-3\]$_{n+1}$. Hence by induction we have shown \[item:2x2-span-3\]–\[item:same-span-3\] for all $n \in \mathbb{N}$. Suppose now that $\prod_{i=n}^{\infty} \lambda_i = 0$ for each $n\in\mathbb{N}$. Let $P_n$ be the projection on $\{b_1,\ldots,b_n\}$ and let $P$ be the projection onto $\operatorname{span}\mathfrak{e}$. Observe $\operatorname{span}\mathfrak{b} \subseteq \operatorname{span}\mathfrak{e}$ by , and so to prove $\operatorname{span}\mathfrak{b} = \operatorname{span}\mathfrak{e}$ it suffices to show that $(P-P_{n+k})e_n\to 0$ in norm as $k\to\infty$ for each $n\in\mathbb{Z}_{\ge 0}$. Since $f_j \in \{f_{j-1},e_j\}$ for all $j \in \mathbb{N}$ by \[item:2x2-span-3\], one has $$\label{eq:inner-product-split} \begin{aligned} (e_n,f_{n+k}) &= \Big(e_n,(f_{n+k},f_{n+k-1})f_{n+k-1}+(f_{n+k},e_{n+k})e_{n+k}\Big) \\ &= (e_n,f_{n+k-1})\cdot (f_{n+k},f_{n+k-1}), \end{aligned}$$ and from \[item:orthonormality-3\]–\[item:same-span-3\], $P-P_{n+k}$ is the projection onto $\operatorname{span}\{f_{n+k},e_{n+k+1},e_{n+k+2},\ldots\}$. This, along with \[item:small-inner-product-3\] and repeated use of proves $$\begin{aligned} {\ensuremath{\lVert (P-P_{n+k})e_n \rVert }}^2 & = {\ensuremath{\left\lvert (e_n,f_{n+k-1}) \right\rvert}}^2 \cdot {\ensuremath{\left\lvert (f_{n+k},f_{n+k-1}) \right\rvert}}^2 \\ & = {\ensuremath{\left\lvert (e_n,f_n) \right\rvert}}^2 \cdot \prod_{i=1}^k {\ensuremath{\left\lvert (f_{n+i},f_{n+i-1}) \right\rvert}}^2 \le {\ensuremath{\left\lvert (e_n,f_n) \right\rvert}}^2 \cdot \prod_{i=n+1}^{n+k} \lambda_i. \end{aligned}$$ As $k\to\infty$ the latter product converges to zero by hypothesis. Our main result for this section characterizes the diagonals of the class of idempotents to be $\ell^{\infty}$. This, according to Jasper, also characterizes all inner products of dual frame pairs. Let $\mathbb{N} = \bigsqcup_{j\in\mathbb{N}} \mathbb{N}_j$ be any partition of $\mathbb{N}$ such that each $\mathbb{N}_j$ is infinite. Let $\phi_j:\mathbb{N}\to\mathbb{N}_j$ be any bijection. Then for each $j$ define $d_{j,n} \coloneqq{} d_{\phi_j(n)}$; in this way we partition the desired sequence into infinitely many infinite sequences. By Proposition \[prop:infinite-multiplicity-diagonal\] there is an idempotent $D\in B(H)$ and a basis $\mathfrak{e} = \bigsqcup_j \mathfrak{e}_j$ where $\mathfrak{e}_j \coloneqq{} \{e_{j,n}\}_{n\in\mathbb{Z}_{\ge 0}}$ for which $$d_{j,0} \coloneqq{} 0 = (De_{j,0},e_{j,0}) \quad\text{and}\quad 2d_{j,n}-d_{j,n-1} = (De_{j,n},e_{j,n})\ \text{for $n\in\mathbb{N}$}.$$ In the above we have assigned $d_{j,0}=0$, and since there are infinitely many zeros, we can apply . Note however that $d_{j,0}$ bears no relation to the sequence $\langle d_n \rangle$, unlike $d_{j,n}$ when $n>0$. The remainder of the argument is independent of $j$. For each $j$ we will employ a judicious use of . Our initial orthonormal set will be $\mathfrak{e}_j$ with diagonal entries $r_{j,n} = (De_{j,n},e_{j,n}) = 2d_{j,n}-d_{j,n-1}$. We then note that $d_{j,n} \in [d_{j,n-1},2d_{j,n}-d_{j,n-1}]$ with convexity coefficient $\lambda_{j,n} = \nicefrac{1}{2}$ since $d_{j,n} = \frac{1}{2}(d_{j,n-1} + (2d_{j,n}-d_{j,n-1}))$. Thus for any $n \in \mathbb{N}$. $$\prod_{i=n+1}^{\infty} \lambda_{j,i} = \prod_{i = n+1}^{\infty} \frac{1}{2} = 0.$$ By there exists an orthonormal set $\mathfrak{b}_j = \{b_{j,n}\}_{n=1}^{\infty}$ for which $$(Db_{j,n},b_{j,n}) = r_{j,n} + d_{j,n-1} - d_{j,n} = (2d_{j,n} - d_{j,n-1}) + d_{j,n-1} - d_{j,n} = d_{j,n},$$ and $\operatorname{span}\mathfrak{b}_j = \operatorname{span}\mathfrak{e}_j$. Thus $\mathfrak{b} \coloneqq{} \bigcup_j \mathfrak{b}_j$ is a basis because $\mathfrak{e} = \bigcup_j \mathfrak{e}_j$ is a basis. With respect to the basis $\mathfrak{b}$ the idempotent $D$ has diagonal ${\ensuremath{\left\langle d_{j,n} \right\rangle}}$ which is precisely ${\ensuremath{\left\langle d_n \right\rangle}}$ after a suitable relabeling. DIAGONALS OF THE CLASS OF FINITE RANK IDEMPOTENTS ================================================= Recall that Lemma \[lem:idempotent-decomp\] is valid for both finite and infinite dimensional $H$. As a result, for $D\in M_n(\mathbb{C})$ with $0\not=D\not=I$, $\operatorname{Tr}D = \operatorname{rank}D \in \{1,\ldots,n-1\}$. Theorem \[thm:idempotent-matrix-diagonals\] shows that this trace condition is the only restriction for a given sequence to be the diagonal of a nonzero non-identity idempotent matrix. Because not all idempotent operators $D\in B(H)$ ($H$ infinite dimensional) are trace-class, it is unnatural to expect there to be any sort of trace restriction on the diagonals of idempotent operators in $B(H)$. In this light, Theorem \[thm:idempotent-diagonals-ell-infty\] is naturally expected: if the only restriction in the $n\times n$ matrix case was the trace, there should be no restrictions in $B(H)$. However, there is another perfectly reasonable class to consider: the trace-class idempotents. Again, Lemma \[lem:idempotent-decomp\] ensures that trace-class idempotents are actually finite rank idempotents. The restriction that $\operatorname{Tr}D = \operatorname{rank}D \in \mathbb{N}$ is still applicable for finite rank idempotents $D\in B(H)$. In this section we prove that, as for $M_n(\mathbb{C})$, this trace condition is the only restriction for an $\ell^1$ (absolutely summable) sequence to be the diagonal of a finite rank idempotent, which is Theorem \[thm:finite-rank-idempotent-diagonals\] below. A corollary of the next lemma verifies Theorem \[thm:finite-rank-idempotent-diagonals\] when restricted to rank-one idempotents. That is, the diagonals of the class of rank-one idempotents are precisely those absolutely summable sequences which sum to one. \[lem:T\^2=Tr(T)T\] If $T\in B(H)$ is a rank-one operator then $T^2 = \operatorname{Tr}(T)T$, hence $T$ is idempotent if and only if $\operatorname{Tr}T = 1$. We may write any rank-one operator as an infinite matrix with entries $a_ib_j$ where $\langle a_i \rangle,\langle b_j \rangle\in\ell^2$. Since the trace is independent of the choice of basis, $\operatorname{Tr}T = \sum_{k=1}^\infty a_kb_k$. Finally, $$T^2 = \left( \sum_{k=1}^\infty (a_ib_k)(a_kb_j) \right) = \left( a_i\left(\sum_{k=1}^\infty a_kb_k\right) b_j\right) = \operatorname{Tr}(T) (a_ib_j) = \operatorname{Tr}(T)T.$$ Another proof which is less, but not entirely, coordinate free: since $T$ is rank-one, there are $x,y \in H$ for which $Tz = (z,x)y$. By expanding $T$ in a basis for $H$ which contains $\nicefrac{y}{{\ensuremath{\lVert y \rVert }}}$, it is clear that $\operatorname{Tr}T = (y,x)$. Thus $$T^2z = T(z,x)y = (z,x)(y,x)y = (y,x)Tz = \operatorname{Tr}(T)Tz. \qedhere$$ \[cor:rank-1-idempotent-diags\] An absolutely summable sequence $\langle d_j \rangle\in\ell^1$ is the diagonal of some rank-one idempotent $D$ if and only if $\sum_j d_j = 1$. One direction is trivial since $\sum_j d_j = \operatorname{Tr}D = \operatorname{rank}D = 1$ by Lemma \[lem:T\^2=Tr(T)T\]. For the other direction, let $\langle d_j \rangle \in\ell^1$ be any absolutely summable sequence which sums to one. Write $d_j = r_j\mathrm{e}^{\mathrm{i}\theta_j}$ with $r_j\ge 0$ and $j\in\mathbb{R}$. Then define $\sqrt{d}\in\ell^2$ as $(\sqrt{d})_j \coloneqq{} \sqrt{r_j}\mathrm{e}^{\nicefrac{\mathrm{i}\theta_j}{2}}$. Then define $D = \left((\sqrt{d})_i(\sqrt{d})_j\right) = \sqrt{d}\otimes\sqrt{d}$. By Lemma \[lem:T\^2=Tr(T)T\], $D$ is idempotent since its diagonal is $\langle d_n \rangle$ which sums to one. We now prove Theorem \[thm:finite-rank-idempotent-diagonals\] by two distinct methods. The first uses Theorem \[thm:idempotent-matrix-diagonals\], Corollary \[cor:rank-1-idempotent-diags\], and . The second proof is an inductive argument analogous to the proof of Theorem \[thm:idempotent-matrix-diagonals\] by Giol, Kovalev, Larson, Nguyen and Tener in [@GKL+-2011-OaM]. It uses Corollary \[cor:rank-1-idempotent-diags\] as the base case and exploits the fact that the class of finite rank idempotents is similarity invariant. Lemma \[lem:idempotent-decomp\] makes this sum condition obviously necessary, so sufficiency is all that is needed. Let $d \coloneqq{} \langle d_n \rangle\in\ell^1$ be an absolutely summable sequence whose sum $\sum_n d_n = m$ is a positive integer. If $m=1$, then $\langle d_n \rangle$ is the diagonal of a rank-one idempotent by Corollary \[cor:rank-1-idempotent-diags\]. So suppose $m>1$, in which case $m-1\in\mathbb{N}$. Set $d'_m \coloneqq{} (m-1) - \sum_{n=1}^{m-1} d_n$. By Theorem \[thm:idempotent-matrix-diagonals\], there is an idempotent matrix $D_1 \in M_m(\mathbb{C})$ with diagonal $d^{(1)} \coloneqq{} \langle d_1,\ldots,d_{m-1},d'_m \rangle$. Now consider the sequence $d^{(2)} \coloneqq{} \langle 2d_m-d'_m,2d_{m+1}-d_m,2d_{m+2}-d_{m+1},\ldots \rangle$. It is clear that $d^{(2)}\in\ell^1$ since $d\in\ell^1$. Furthermore, $$\sum_{n=1}^\infty d^{(2)}_n = 2d_m-d'_m + \sum_{n=m}^\infty (2d_{n+1}-d_n) = -d'_m + \sum_{n=m}^\infty d_n = \sum_{n=1}^\infty d_n - (m-1) = 1.$$ Therefore, by Corollary \[cor:rank-1-idempotent-diags\], there is a rank-one idempotent $D_2$ with diagonal sequence $d^{(2)}$. Defining $D = D_1 \oplus D_2$, we find that $D$ has a basis $\mathfrak{e} \coloneqq\{e_n\}_{n\in\mathbb{N}}$ in which its diagonal is $$\langle d_1, \ldots, d_{m-1}, d'_m, 2d_m-d'_m, 2d_{m+1}-d_m, 2d_{m+2}-d_{m+1}, \ldots \rangle.$$ That is, $(De_n,e_n) = d_n$ for $1\le n<m$; $(De_m,e_m) = d'_m$; $(De_{m+1},e_{m+1}) = 2d_m-d'_m$; and $(De_n,e_n) = 2d_{n-1}-d_{n-2}$ for $n>m+1$. We will now apply to the orthonormal set $\{e_m,e_{m+1},\ldots\}$. So $$r_n \coloneqq{} (De_n,e_n) = \begin{cases} d'_m & \text{if $n=m$} \\ 2d_m - d'_m & \text{if $n=m+1$} \\ 2d_{n-1} - d_{n-2} & \text{if $n>m+1$} \\ \end{cases}$$ Since $d_m \in [r_m,r_{m+1}]$ and $d_n \in [d_{n-1},r_{n+1}]$ for $n>m$ (with convexity coefficients all $\lambda_n \equiv \nicefrac{1}{2}$), after a suitable relabeling of the sequences involved ($r_n \mapsto r_{n-m}; d_n \mapsto d_{n-m+1}$) we may apply to obtain an orthonormal set $\{b_m,b_{m+1},\ldots\}$ satisfying $$(Db_n,b_n) = \begin{cases} r_{m+1}+r_m-d_m & \text{if $n=m$} \\ r_{n+1}+d_{n-1}-d_n & \text{if $n>m$} \\ \end{cases} \Bigg\} = d_n.$$ Moreover, by , since the convexity coefficients are $\lambda_n = \nicefrac{1}{2}$, we have $\operatorname{span}\{b_n\}_{n=m}^{\infty} = \operatorname{span}\{e_n\}_{n=m}^{\infty}$. Setting $b_n \coloneqq{} e_n$ for $n<m$, we find that $\mathfrak{b} = \{b_n\}_{n=1}^{\infty}$ is a basis with respect to which $D$ has diagonal ${\ensuremath{\left\langle d_n \right\rangle}}$. We proceed by induction on the sum $m\coloneqq{} \sum_{n=1}^\infty d_n$ where $\langle d_n \rangle\in\ell^1$ is an absolutely summable sequence whose sum is a positive integer. The base case $m=1$ is handled by Corollary \[cor:rank-1-idempotent-diags\]. Now suppose $m>1$ and for any absolutely summable sequence whose sum is $m-1$, there is a finite rank idempotent with that sequence on its diagonal. By possibly permuting the sequence $d_n$, we may assume without loss of generality that $d_1+d_2\not=2$. Since $\sum_{n=1}^\infty d_n = m$, then $(d_1+d_2-1)+\sum_{n=3}^\infty d_n = m-1$. So by the induction hypothesis there exists a finite rank (in fact, rank-$(m-1)$) idempotent $\tilde{D}$ with diagonal sequence $\langle d_1+d_2-1,d_3,d_4,\ldots \rangle$. Then consider the rank-$m$ operator $$D' = \begin{pmatrix} 1 & 0_{1\times\infty} \\ 0_{\infty\times 1} & \tilde{D} \\ \end{pmatrix},$$ which is obviously idempotent. With respect to the basis $\mathfrak{e} = \{e_j\}_{j=1}^\infty$, $D'$ has diagonal $\langle 1,d_1+d_2-1,d_3,d_4,\ldots \rangle$. Then consider the invertible $S$ which is the identity on $\operatorname{span}\{e_j\}_{j=3}^\infty$ and whose compression to $\operatorname{span}\{e_1,e_2\}$ has the matrix representation $$\begin{pmatrix} \lambda & \lambda-1 \\ 1 & 1 \\ \end{pmatrix},$$ where $\lambda \coloneqq{} \nicefrac{(d_2-1)}{(d_1+d_2-2)}$. Conjugating $D'$ by $S$ produces an idempotent $D \coloneqq{} SD'S^{-1}$ whose diagonal with respect to $\mathfrak{e}$ is precisely the sequence $d$. The reader should note that although conjugating by a similarity can be viewed as changing the *linear* basis (as opposed to conjugating by a unitary which changes the *orthonormal* basis) we are not using the similarity in this context. Instead, we only use the similarity to produce a new idempotent $D$ (which still has finite rank) and has the desired diagonal with respect to the *orthonormal* basis $\mathfrak{e}$. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== The authors are indebted to John Jasper who at GPOTS 2013 alerted us to his frame theory and equivalent operator theory problems and their history. And special thanks to Daniel Beltiţă for the reference to the work of Fan, Fong and Herrero [@Fan-1984-TotAMS], [@FF-1994-PotAMS], [@FFH-1987-PotAMS]. Jean-Christophe Bourin deserves special mention for communicating to us that our is a corollary of his pinching theorem [@Bou-2003-JOT Theorem 2.1]. In particular, to apply his theorem one only needs to produce idempotents whose essential spectra contain disks centered at the origin. This is especially interesting because his methods of proof are completely different from ours. The first author was supported by the Charles Phelps Taft Dissertation Fellowship. The second author was partially supported by the Simons Foundation Collaboration Grant for Mathematicians \#245014 and the Charles Phelps Taft Research Center. [10]{} <span style="font-variant:small-caps;">J. [Antezana]{}, P. [Massey]{}, M. [Ruiz]{}, D. [Stojanoff]{},</span> [The Schur–Horn theorem for operators and frames with prescribed norms and frame operator,]{} *[Ill. J. Math.]{}*, **51.2**(2007), 537–560. <span style="font-variant:small-caps;">M. Argerami, P. Massey,</span> A [S]{}chur-[H]{}orn theorem in [${\rm II}_1$]{} factors, *Indiana Univ. Math. J.*, **56.5**(2007), 2051–2059. <span style="font-variant:small-caps;">M. Argerami, P. Massey,</span> A contractive version of a [S]{}chur-[H]{}orn theorem in [$\rm II_1$]{} factors, *J. Math. Anal. Appl.*, **337.1**(2008), 231–238. <span style="font-variant:small-caps;">M. Argerami, P. Massey,</span> Schur-[H]{}orn theorems in [${\rm II}_\infty$]{}-factors, *Pacific J. Math.*, **261.2**(2013), 283–310. <span style="font-variant:small-caps;">W. [Arveson]{},</span> [Diagonals of normal operators with finite spectrum]{}, *[Proc. Natl. Acad. Sci. USA]{}*, **104.4**(2007), 1152–1158. <span style="font-variant:small-caps;">W. Arveson, R. V. Kadison,</span> Diagonals of self-adjoint operators, *Operator Theory, Operator Algebras, and Applications*, **414**(2006), 247–263. <span style="font-variant:small-caps;">J.-C. Bourin,</span> Compressions and Pinchings, *J. Oper. Theory*, **50.2**(2003), 211–220. <span style="font-variant:small-caps;">M. Bownik, J. Jasper,</span> Diagonals of self-adjoint operators with finite spectrum, preprint, arXiv:1302.5106 \[math.FA\], June 2012. <span style="font-variant:small-caps;">M. Bownik, J. Jasper,</span> [The Schur–Horn Theorem for Operators with Finite Spectrum]{}, *Trans. Amer. Math. Soc.* (to appear). <span style="font-variant:small-caps;">J. W. Calkin,</span> [Two-sided ideals and congruences in the ring of bounded operators in Hilbert space]{}, *Ann. of Math. (2)*, **42**(1941), 839–873. <span style="font-variant:small-caps;">P. [Fan]{},</span> [On the diagonal of an operator]{}, *[Trans. Amer. Math. Soc.]{}*, **283**(1984), 239–251. <span style="font-variant:small-caps;">P. [Fan]{}, C.-K. [Fong]{},</span> [An intrinsic characterization for zero-diagonal operators]{}, *Proc. Amer. Math. Soc.*, **121.3**(1994), 803–805. <span style="font-variant:small-caps;">P. Fan, C.-K. Fong, D. A. Herrero,</span> On zero-diagonal operators and traces, *Proc. Amer. Math. Soc.*, **99.3**(1987), 445–451. <span style="font-variant:small-caps;">C.-K. [Fong]{},</span> [Diagonals of nilpotent operators]{}, *[Proc. Edinb. Math. Soc. (2)]{}*, **29**(1986), 221–224. <span style="font-variant:small-caps;">J. [Giol]{}, L. V. [Kovalev]{}, D. [Larson]{}, N. [Nguyen]{}, J. E. [Tener]{},</span> [Projections and idempotents with fixed diagonal and the homotopy problem for unit tight frames]{}, *[Oper. Matrices]{}*, **5.1**(2011), 139–155. <span style="font-variant:small-caps;">I. C. Gohberg, A. Markus,</span> Some relations between eigenvalues and matrix elements of linear operators, *Mat. Sb. (N.S.)*, **64**(1964), 481–496. <span style="font-variant:small-caps;">P. R. [Halmos]{},</span> *A Hilbert space problem book. 2nd ed., rev. and enl.*, Graduate Texts in Mathematics, vol. 19, Springer–Verlag, New York – Heidelberg – Berlin 1982. <span style="font-variant:small-caps;">A. Horn,</span> Doubly stochastic matrices and the diagonal of a rotation matrix, *Amer. J. Math.*, **76**(1954), 620–630. <span style="font-variant:small-caps;">R. A. Horn, C. R. Johnson,</span> *Matrix analysis*, Cambridge University Press, Cambridge, 1990. Corrected reprint of the 1985 original. <span style="font-variant:small-caps;">J. Jasper,</span> private communication, GPOTS – UC Berkeley, May 2013. <span style="font-variant:small-caps;">J. Jasper,</span> The [S]{}chur-[H]{}orn theorem for operators with three point spectrum, *J. Funct. Anal.*, **265.8**(2013), 1494–1521. <span style="font-variant:small-caps;">R. V. Kadison,</span> The [P]{}ythagorean [T]{}heorem [I]{}: the finite case, *Proc. Natl. Acad. Sci. USA*, **99.7**(2002), 4178–4184. <span style="font-variant:small-caps;">R. V. Kadison,</span> The [P]{}ythagorean [T]{}heorem [II]{}: the infinite discrete case, *Proc. Natl. Acad. Sci. USA*, **99.8**(2002), 5217–5222. <span style="font-variant:small-caps;">R. V. [Kadison]{}, J. R. [Ringrose]{},</span> *[Fundamentals of the theory of operator algebras. Vol. I: Elementary theory]{}*, American Mathematical Society, Providence 1997. <span style="font-variant:small-caps;">V. Kaftal, G. Weiss,</span> An infinite dimensional [S]{}chur-[H]{}orn [T]{}heorem and majorization theory, *J. Funct. Anal.*, **259.12**(2010), 3115–3162. <span style="font-variant:small-caps;">V. Kaftal, G. Weiss,</span> [Majorization and arithmetic mean ideals]{}, *[Indiana Univ. Math. J.]{}*, **60.5**(2011), 1393–1424. <span style="font-variant:small-caps;">M. Kennedy, P. Skoufranis,</span> [Thompson’s Theorem for II$_1$ factors]{}, preprint, arXiv:1407.1564, July 2014. <span style="font-variant:small-caps;">D. R. [Larson]{},</span> [Nest algebras and similarity transformations]{}, *Ann. of Math. (2)*, **121**(1985), 409–427. <span style="font-variant:small-caps;">J. Loreaux, G. Weiss,</span> [Majorization and a Schur–Horn Theorem for positive compact operators, the nonzero kernel case]{}, *J. Funct. Anal.*, **268.3**(2015), 703–731. <span style="font-variant:small-caps;">A. Markus,</span> Eigenvalues and singular values of the sum and product of linear operators, *Uspekhi Mat. Nauk*, **19.4**(1964), 93–123. <span style="font-variant:small-caps;">A. Neumann,</span> An infinite dimensional version of the [S]{}chur-[H]{}orn convexity theorem, *J. Funct. Anal.*, **161.2**(1999), 418–451. <span style="font-variant:small-caps;">M. Ravichandran,</span> The [S]{}chur-[H]{}orn [T]{}heorem in von [N]{}eumann algebras, preprint, arXiv:1209.0909 \[math.OA\], November 2014. <span style="font-variant:small-caps;">I. Schur,</span> Über eine klasse von mittelbildungen mit anwendungen auf der determinantentheorie, *[S]{}itzungsber. [B]{}erliner [M]{}at. [G]{}es.*, **22**(1923), 9–29. <span style="font-variant:small-caps;">User: 1015</span>, *Unicity of a vector space frame’s dual frame*, MathOverflow, URL: `http://mathoverflow.net/q/132592` (version: 2013-06-02).

--- abstract: 'A complex Hadamard matrix is a square matrix $W$ with complex entries of absolute value $1$ satisfying $WW^\ast= nI$, where $\ast$ stands for the Hermitian transpose and $I$ is the identity matrix of order $n$. In this paper, we give constructions of complex Hadamard matrices in the Bose–Mesner algebra of a certain 4-class symmetric association scheme. Moreover, we determine the Nomura algebras to show that the resulting matrices are not decomposable into nontrivial generalized tensor products.' author: - Takuya Ikuta and Akihiro Munemasa title: Complex Hadamard matrices attached to even orthogonal schemes of class $4$ --- Introduction {#sec:Intro} ============ A complex Hadamard matrix is a square matrix $W$ with complex entries of absolute value $1$ satisfying $WW^\ast= nI$, where $\ast$ stands for the Hermitian transpose and $I$ is the identity matrix of order $n$. They are the natural generalization of real Hadamard matrices. Complex Hadamard matrices appear frequently in various branches of mathematics and quantum physics. A type-II matrix, or an inverse orthogonal matrix, is a square matrix $W$ with nonzero complex entries satisfying $W{W^{(-)}}^\top=nI$, where $W^{(-)}$ denotes the entrywise inverse of $W$. Obviously, a complex Hadamard matrix is a type-II matrix. In [@MI], we gave a method to find a complex Hadamard matrix in the Bose–Mesner algebra of a symmetric association scheme. Applying this result, we classified complex Hadamard matrices in the Bose–Mesner algebra of a certain 3-class association scheme. In this paper, we construct certain complex Hadamard matrices in the Bose–Mesner algebra of a 4-class association scheme $(X,\{R_i\}_{i=0}^4)$ with the first eigenmatrix: $$\label{P} P=\begin{bmatrix} 1&\frac12(q-2)q^{2m-1}&\frac12q^{2m}&q(q^{2m-2}-1)&q-2 \\ 1&\frac12(q-2)q^{m-1}&\frac12q^{m}&-(q-1)(q^{m-1}+1)&q-2 \\ 1&-\frac12(q-2)q^{m-1}&-\frac12q^{m}&-(q-1)(q^{m-1}-1)&q-2 \\ 1&\frac12q^m&-\frac12q^m&0&-1 \\ 1&-\frac12q^m&\frac12q^m&0&-1 \end{bmatrix},$$ where $q$ and $m$ are positive integers with $q\geq 4$ and $m\geq2$. Then $|X|=q^{2m}-1$, $R_4$ is a disconnected relation, and $R_2$ defines a strongly regular graph. If $q$ is a power of $2$, an even orthogonal scheme is an example of an association scheme with the first eigenmatrix (\[P\]) (see [@BCN Chapter 12.1]). If $m=1$, then $R_3=\emptyset$, and this scheme reduces to an even orthogonal scheme of class $3$ which we considered in [@MI]. For a type-II matrix $W\in M_X({\mathbb{C}})$ and $a,b\in X$, we define column vectors $Y_{ab}$ by setting $$(Y_{ab})_x= \frac{W_{xa}}{W_{xb}}\quad(x\in X).$$ The [*Nomura algebra*]{} $N(W)$ of $W$ is the algebra of matrices in $M_n({\mathbb{C}})$ such that $Y_{ab}$ is an eigenvector for all $a,b\in X$. It is shown in [@JMN Theorem 1] that the Nomura algebra is a Bose–Mesner algebra. Throughout this paper, we denote by $\frak{X}=(X,\{R_i\}_{i=0}^4)$ a symmetric association scheme with the first eigenmatrix (\[P\]). Let $A_0,A_1,A_2,A_3,A_4$ be the adjacency matrices of $\frak{X}$. Let $w_0=1,w_1,w_2,w_3,w_4$ be nonzero complex numbers, and set $$\label{W} W=\sum_{j=0}^4w_j A_j,$$ $$\label{aij0802} a_{i,j}=\frac{w_i}{w_j}+\frac{w_j}{w_i} \quad (0\leq i<j\leq 4).$$ The main purpose of this paper is to prove the following: \[thm:main\] Assume that $$\label{assump} w_4=1.$$ 1. Assume $w_1=1$. Then, the matrix $W$ in [(\[W\])]{} is a complex Hadamard matrix if and only if $$w_2^2+\frac{2(q^{2m}-2)}{q^{2m}}w_2+1=0 \quad \text{and} \quad w_3=1.$$ 2. Assume $$\label{eq:1024} a_{0,1}=\frac{2(q^{4m-2}-(q+2)q^{2m-1}+2)}{(q^{2m-1}+q-2)q^{2m-1}}.$$ Then, the matrix $W$ in [(\[W\])]{} is a complex Hadamard matrix if and only if $$\begin{aligned} w_2=&-\frac{(q-1)q^{2m-1}w_1+q^{2m-1}+q-2}{(q^{2m-1}-1)q}, \label{eq:3-2} \\ w_3=&1. \nonumber \end{aligned}$$ \[thm:main2\] Let $W$ be a complex Hadamard matrix given in [(i)]{} and [(ii)]{} of Theorem [\[thm:main\]]{}. The algebra $N(W)$ coincides with the linear span of $I$ and $J$. In particular, $W$ is not equivalent to a nontrivial generalized tensor product. The reason for the assumption (\[assump\]) is as follows: Calculating the conditions under which the matrix (\[W\]) becomes a complex Hadamard matrix experimentally for small $q$ and $m$, we find that (\[assump\]) is fulfilled, or 3. $a_{0,4}=2(q^{2m}-6)/(q^{2m}-4)$, or 4. $a_{0,4}$ is a zero of a polynomial of degree $9$. For the case (iii) with $m=2,3$, we have $w_1=w_3=w_4$. Therefore, this case reduces to the case in which the matrix $W$ given in (\[W\]) belongs to the Bose–Mesner algebra of the strongly regular graph defined by $R_2$. However, it seems to be difficult to prove $w_1=w_3=w_4$ for arbitrary $m\geq 4$. For the case (iv), we verified that the polynomial in $a_{0,4}$ of degree $9$ is an irreducible polynomial for $m=2,\ldots,9$ and $q=2^s$ with $2\leq s\leq 10000$. However, it seems difficult to determine the polynomial of degree $9$ satisfied by $a_{0,4}$ in general. For example, for $(q,m)=(4,2)$, if the matrix (\[W\]) is a complex Hadamard matrix, then $a_{0,4}$ is a zero of the polynomial $$\begin{aligned} p(x)=&x^9-\frac{235721}{1785}x^8-\frac{17957726593}{62475}x^7+\frac{33219815829811}{937125}x^6 \\ &-\frac{12554318926285933}{4685625}x^5+\frac{29740292638491103}{312375}x^4 \\ &-\frac{696525696876795217}{187425}x^3+\frac{851886544261448041}{37485}x^2 \\ &-\frac{124583919439776136}{2499}x+\frac{30888835313436500}{833}.\end{aligned}$$ It can be shown that $p(x)$ has only one real root in $(-2,2)$ by using Sturm’s theorem. Then, by using Lemmas \[thm:1\] and \[thm:2\] below, there exist $w_1,w_2,w_3,w_4$ such that (\[W\]) is a complex Hadamard matrix. Under the hypothesis of (\[assump\]), we find that $a_{0,1}=2$ or $a_{0,1}$ is given by (\[eq:1024\]), or 5. $a_{0,1}$ is a zero of a polynomial of degree $4$. It seems to be difficult to determine $w_1,w_2,w_3$ for the case (vi) for arbitrary $q$ and $m$. For example, for $(q,m)=(4,2)$, if the matrix (\[W\]) is a complex Hadamard matrix, then $w_1,w_2,w_3$ are given by the following: $$\begin{aligned} & w_1^2+\frac{21s-7140\pm 85t}{176}w_1+1=0, \\ & w_2=-\frac{64(w_1^2-1)}{127w_1+64a_{0,2}}, \\ & w_3=\frac{90(w_1^2-1)}{90a_{1,3}w_1-4s+1117}, \\ & a_{0,2}=\frac{43s-14620\pm 85t}{352}, \\ & a_{1,3}=\frac{21s-1848\mp(4s+1253)t}{2640}, \\ & s=\sqrt{104899}, \\ & t^2=\frac{8s-2591}{3}.\end{aligned}$$ Preliminaries {#sec:2} ============= We define a polynomial in three indeterminates $X,Y,Z$ as follows: $$g(X,Y,Z)=X^2+Y^2+Z^2-XYZ-4.$$ We define a polynomial in six indeterminates $X_{0,1},X_{0,2},X_{0,3},X_{1,2},X_{1,3},X_{2,3}$ as follows: $$h(X_{0,1},X_{0,2},X_{0,3},X_{1,2},X_{1,3},X_{2,3}) =\det\begin{bmatrix} 2&X_{0,1}&X_{0,2}\\ X_{0,1}&2&X_{1,2}\\ X_{0,3}&X_{1,3}&X_{2,3} \end{bmatrix}.$$ For a finite set $N$ and a positive integer $k$, we denote by $\binom{N}{k}$ the collection of all $k$-element subsets of $N$. \[thm:1\] Let $N=\{0,1,\dots,d\}$, $N_3=\binom{N}{3}$ and $N_4=\binom{N}{4}$. Let $a_{i,j}$ $(0\leq i,j\leq d,\;i\neq j)$ be complex numbers satisfying $$\begin{aligned} a_{i,j}&=a_{j,i}\quad(0\leq i<j\leq d),\label{aij1}\\ g(a_{i,j},a_{j,k},a_{i,k})&=0\quad(\{i,j,k\}\in N_3),\label{N1}\\ h(a_{i,j},a_{i,k},a_{i,\ell},a_{j,k},a_{j,\ell},a_{k,\ell})&=0 \quad(\{i,j,k,\ell\}\in N_4).\label{N4}\end{aligned}$$ Assume $$\label{a1} a_{i_0,i_1}\neq\pm2\quad\text{for some $i_0,i_1$ with $0\leq i_0<i_1\leq d$.}$$ Let $w_{i_0}$, $w_{i_1}$ be nonzero complex numbers satisfying $$\label{az1d} \frac{w_{i_0}}{w_{i_1}}+\frac{w_{i_1}}{w_{i_0}}=a_{{i_0,i_1}}.$$ Then for complex numbers $w_i$ $(0\leq i\leq d,\;i\neq i_0,i_1)$, the following are equivalent: 1. for all $i,j$ with $0\leq i,j\leq d$ and $i\neq j$, $$\label{az2} \frac{w_j}{w_i}+\frac{w_i}{w_j}=a_{i,j}$$ 2. for all $i,j$ with $0\leq i\leq d$, $i\neq i_0,i_1$, $$\label{azid} w_i= \frac{w_{i_1}^2-w_{i_0}^2}{a_{{i_1},i}w_{i_1}-a_{{i_0},i}w_{i_0}}.$$ Moreover, if one of the two equivalent conditions [(i), (ii)]{} is satisfied, $a_{i,j}$ $(0\leq i<j\leq d)$ are all real and $$\label{-22} -2<a_{i_0,i_1}<2,$$ then $|w_i|=|w_j|$ for $0\leq i<j\leq d$. We let $\mathcal{A}$ denote a symmetric Bose–Mesner algebra with adjacency matrices $A_0,A_1,\dots,A_d$. Let $n$ be the size of the matrices $A_i$, and we denote by $$P=(P_{i,j})_{\substack{0\leq i\leq d\\ 0\leq j\leq d}}$$ the first eigenmatrix of $\mathcal{A}$. Then the adjacency matrices are expressed as $$A_j=\sum_{i=0}^d P_{i,j}E_i\quad(j=0,1,\dots,d),$$ where $E_0=\frac{1}{n}J,E_1,\dots,E_d$ are the primitive idempotents of $\mathcal{A}$. Let $w_0,w_1,\dots,w_d$ be nonzero complex numbers, and set $$\label{eq:W} W=\sum_{j=0}^dw_j A_j\in\mathcal{A}.$$ \[thm:2\] Let $X_{i,j}$ $(0\leq i<j\leq d)$ be indeterminates and let $e_k$ be the polynomial defined by $$\label{eq:04} e_k= \sum_{0\leq i<j\leq d} P_{k,i}P_{k,j}X_{i,j} +\sum_{i=0}^d P_{k,i}^2-n \quad(k=1,\dots,d).$$ Let $a_{i,j}$ $(0\leq i,j\leq d,\;i\neq j)$ and $w_i$ $(0\leq i\leq d)$ be complex numbers. Assume that $w_i\neq0$ for all $i$ with $0\leq i\leq d$ and that [(\[az2\])]{} holds. Then the following statements are equivalent: 1. the matrix $W$ given by [(\[eq:W\])]{} is a type-II matrix, 2. $(a_{i,j})_{0\leq i<j\leq d}$ is a common zero of $e_k$ $(k=1,\dots,d)$. Moreover, if one of the two equivalent conditions [(i), (ii)]{} is satisfied, $a_{i,j}\in{\mathbb{R}}$ for all $i,j$ with $0\leq i<j\leq d$, and [(\[-22\])]{} holds for some $i_0, i_1$ with $0\leq i_0<i_1\leq d$, then $W$ is a scalar multiple of a complex Hadamard matrix. We now describe the proof of Theorem \[thm:main\] briefly. Let $A_0,A_1,A_2,A_3,A_4$ be the adjacency matrices of an association scheme $\frak{X}$ with the first eigenmatrix (\[P\]). Let $w_0=1,w_1,w_2,w_3,w_4$ be nonzero complex numbers, and $W$ be the matrix defined by (\[W\]). For $i,j\in\{0,1,2,3,4\}$, define $a_{i,j}$ by (\[aij0802\]). We write $$\label{aa} \boldsymbol{a}=(a_{0,1},a_{0,2},a_{0,3},a_{0,4},a_{1,2},a_{1,3},a_{1,4},a_{2,3},a_{2,4},a_{3,4})$$ for brevity. Consider the polynomial ring $$\label{RC} R=\mathbb{C}[X_{0,1},X_{0,2},X_{0,3},X_{0,4},X_{1,2},X_{1,3},X_{1,4},X_{2,3},X_{2,4},X_{3,4}].$$ In Section 3, we first assume that $W$ is a complex Hadamard matrix. Then by Lemmas \[thm:1\] and \[thm:2\], $\boldsymbol{a}$ is a common zero of the polynomials $$\begin{aligned} & g(X_{i,j},X_{i,k},X_{j,k})\quad(\{i,j,k\}\in\binom{\{0,1,2,3,4\}}{3}),\label{gg} \\ & h(X_{i,j},X_{i,k},X_{i,l},X_{j,k},X_{j,l},X_{k,l}) \quad(\{i,j,k,l\}\in\binom{\{0,1,2,3,4\}}{4}), \label{hh} \\ & e_k\quad (k\in\{1,2,3,4\}). \label{ee}\end{aligned}$$ Let $\mathcal{I}$ be the ideal of $R$ generated by these polynomials. Calculating the ideal generated by $\mathcal{I}$ and $X_{0,4}-2$, we find (i) and (ii) of Theorem \[thm:main\]. Conversely, we assume that $w_4=1$ and $w_1,w_2,w_3$ are given in Theorem \[thm:main\]. Then, to show that the matrix $W$ given in (\[W\]) is a complex Hadamard matrix, we check that $\boldsymbol{a}$ defined by (\[aij0802\]), (\[aa\]) is a zero of the polynomials (\[gg\]), (\[hh\]), (\[ee\]). Moreover, we check that $-2<a_{i_0,i_1}<2$ holds for some $i_0,i_1$ with $0\leq i_0<i_1\leq 4$. All the computer calculations in this paper were performed with the help of Magma [@magma]. In order to facilitate computations covering all possible values of the integer $q$, we perform the computations in the polynomial ring with $12$ variables $q,r=q^m$ and $X_{i,j}$ over the field of rational numbers, rather than the ring . The results valid for this generic setting are also valid for arbitrary integers $q$ and $m$. Proof of Theorem \[thm:main\] {#sec:3} ============================= Recall $q\geq 4$ and $m\geq 2$, and $\mathcal{I}$ is the ideal of the polynomial ring $R$ generated by the polynomials (\[gg\]), (\[hh\]), and (\[ee\]). For the remainder of this section, we assume that $a_{0,4}=2$, that is, $w_4=1$. Let $\mathcal{I}_1$ denote the ideal generated by $\mathcal{I}$ and $X_{0,4}-2$. For Lemmas \[lem:a01\]–\[lem:a02d\] we assume that $\boldsymbol{a}$ defined in (\[aa\]) is a common zero of the polynomials in $\mathcal{I}_1$. \[lem:a01\] We have $$\label{a12} a_{1,2}=-\frac{2(q^{2m}-2)}{q^{2m}}.$$ We can verify that $\mathcal{I}_1$ contains $X_{1,2}+2(q^{2m}-2)/q^{2m}$. Hence we have (\[a12\]). \[lem:a02eq2\] Assume $a_{0,1}=2$. Then, $(w_1,w_2,w_3)$ is given in [(i)]{} of Theorem [\[thm:main\]]{}. Let $\mathcal{I}_2$ denote the ideal generated by $\mathcal{I}_1$ and $X_{0,1}-2$. Then we can verify that $\mathcal{I}_2$ contains $(X_{0,3}-2)^2$, that is, $a_{0,3}=2$. Hence $w_1=w_3=1$. Since $a_{1,2}$ is given in (\[a12\]), the matrix $W$ given in (\[W\]) belongs to the Bose–Mesner algebra of the strongly regular graph defined by $R_2$. From [@ChanGodsil] we have the condition of $w$ given in [(i)]{} of Theorem [\[thm:main\]]{}. \[lem:a02d\] Assume that $a_{0,1}$ is given in [(\[eq:1024\])]{}. Then, $(w_1,w_2,w_3)$ is given in [(ii)]{} of Theorem [\[thm:main\]]{}. Let $\mathcal{I}_3$ denote the ideal generated by $\mathcal{I}_1$ and $X_{0,1}-a_{0,1}$. Then we can verify that $\mathcal{I}_3$ contains $q(q^{2m-1}+q-2)X_{0,2}+2(q^{2m}-q^2+2q-2)$, that is, $$\label{a02inLem7} a_{0,2}=-\frac{2(q^{2m}-q^2+2q-2)}{q(q^{2m-1}+q-2)}$$ Let $\mathcal{I}_4$ denote the ideal generated by $\mathcal{I}_3$ and $p_1(X_{0,2})$. Then we can verify that $\mathcal{I}_4$ contains $X_{0,3}-2$, that is, $w_3=1$. From (\[azid\]) we obtain $$w_2=\frac{w_1^2-1}{a_{1,2}w_1-a_{0,2}}.$$ Since $w_1^2-a_{0,1}w_1+1=0$, we have (\[eq:3-2\]) from (\[a12\]), (\[a02inLem7\]). Suppose that the matrix (\[W\]) is a complex Hadamard matrix. For $i,j\in\{0,1,2,3,4\}$, define $a_{i,j}$ by (\[aij0802\]). Let $\boldsymbol{a}$ be given in (\[aa\]). Then $\boldsymbol{a}$ is a common zero of the polynomials in $\mathcal{I}_1$ by Lemma \[thm:2\]. From Lemmas \[lem:a02eq2\], \[lem:a02d\] we have (i) and (ii) of Theorem \[thm:main\]. Conversely, assume that $w_1,w_2,w_3$, and $w_4$ are given in Theorem \[thm:main\]. Then, we show that the matrix given in (\[W\]) is a complex Hadamard matrix. To do this, we check that $\boldsymbol{a}$ defined by (\[aij0802\]) is a zero of the polynomials (\[gg\]), (\[hh\]), and (\[ee\]), and $(w_1,w_2,w_3)$ are complex numbers of absolute value $1$. The latter condition is satisfied if $-2<a_{i_0,i_1}<2$ holds for some $i_0,i_1$ with $0\leq i_0<i_1\leq 4$. Case (i) is done by [@ChanGodsil]. Next consider Case (ii). From (\[aij0802\]), (\[eq:1024\]), and (\[eq:3-2\]) we have (\[a12\]) and (\[a02inLem7\]). Then we have $$\boldsymbol{a}=\left( a_{0,1},a_{0,2},2,2,a_{1,2},a_{0,1},a_{0,1},a_{0,2},a_{0,2},2\right).$$ This is a zero of the polynomials (\[gg\]), (\[hh\]), and (\[ee\]). It is easy to check that $0<a_{0,1}<2$. Proof of Theorem \[thm:main2\] {#sec:4} ============================== Since $q^{2m}-1$ is a composite, there are uncountably many inequivalent complex Hadamard matrices of order $q^{2m}-1$ by [@Craigen]. Indeed, such matrices can be constructed using generalized tensor products [@HS]. We show that none of our complex Hadamard matrices is equivalent to a nontrivial generalized tensor product. This is done by showing that the Nomura algebra of our complex Hadamard matrices has dimension $2$. According to [@HS], the Nomura algebra of a nontrivial generalized tensor product of type-II matrices is imprimitive, and this is never the case when it has dimension $2$. Recall $q\geq 4$ and $m\geq 2$. The intersection matrices $B_i=(p_{ij}^k)$ ($i=0,\ldots,4$) of $\frak{X}$ are given by the following: $$\begin{aligned} B_0&=\begin{bmatrix} 1&0&0&0&0 \\ 0&1&0&0&0 \\ 0&0&1&0&0 \\ 0&0&0&1&0 \\ 0&0&0&0&1 \end{bmatrix}, \\ B_1&=\begin{bmatrix} 0&1&0&0&0 \\ \frac{(q-2)q^{2m-1}}{2}&\frac{(q-2)^2q^{2m-2}}{4}&\frac{(q-2)^2q^{2m-2}}{4}&\frac{(q-2)^2q^{2m-2}}{4}&\frac{(q-4)q^{2m-1}}{4} \\ 0&\frac{(q-2)q^{2m-1}}{4}&\frac{(q-2)q^{2m-1}}{4}&\frac{(q-2)q^{2m-1}}{4}&\frac{q^{2m}}{4} \\ 0&\frac{(q-2)(q^{2m-2}-1)}{2}&\frac{(q-2)(q^{2m-2}-1)}{2}&\frac{(q-2)q^{2m-2}}{2}&0 \\ 0&\frac{q-4}{2}&\frac{q-4}{2}&0&0 \end{bmatrix}, \\ B_2&=\begin{bmatrix} 0&0&1&0&0 \\ 0&\frac{(q-2)q^{2m-1}}{4}&\frac{(q-2)q^{2m-1}}{4}&\frac{(q-2)q^{2m-1}}{4}&\frac{q^{2m}}{4} \\ \frac{q^{2m}}{2}&\frac{q^{2m}}{4}&\frac{q^{2m}}{4}&\frac{q^{2m}}{4}&\frac{q^{2m}}{4} \\ 0&\frac{q(q^{2m-2}-1)}{2}&\frac{q(q^{2m-2}-1)}{2}&\frac{q^{2m-1}}{2}&0 \\ 0&\frac{q}{2}&\frac{q-2}{2}&0&0 \end{bmatrix}, \\ B_3&=\begin{bmatrix} 0&0&0&1&0 \\ 0&\frac{(q-2)(q^{2m-2}-1)}{2}&\frac{(q-2)(q^{2m-2}-1)}{2}&\frac{(q-2)q^{2m-2}}{2}&0\\ 0&\frac{q(q^{2m-2}-1)}{2}&\frac{q(q^{2m-2}-1)}{2}&\frac{q^{2m-1}}{2}&0 \\ q(q^{2m-2}-1)&q^{2m-2}-1&q^{2m-2}-1&q^{2m-2}-2q+1&q(q^{2m-2}-1) \\ 0&0&0&q-2&0 \end{bmatrix}, \\ B_4&=\begin{bmatrix} 0&0&0&0&1 \\ 0&\frac{q-4}{2}&\frac{q-2}{2}&0&0 \\ 0&\frac{q}{2}&\frac{q-2}{2}&0&0 \\ 0&0&0&q-2&0 \\ q-2&0&0&0&q-3 \end{bmatrix}.\end{aligned}$$ \[lem:NS\] The algebra $N(W)$ is symmetric. Suppose that $N(W)$ is not symmetric. Then by [@JMN Proposition 6(i)], there exists $(b,c)\in X^2$ with $b\neq c$ such that $$\sum_{x\in X}\frac{W_{x,b}^2}{W_{x,c}^2}=0.$$ This is equivalent to $$\sum_{j,k} p_{jk}^i \frac{w_j^2}{w_k^2}=0$$ for some $i\in\{1,2,3,4\}$. Using the notation (\[aij0802\]), we have $$\begin{aligned} \sum_{j,k}p_{jk}^i \frac{w_j^2}{w_k^2} &= \sum_{j<k}p_{jk}^i \left(\frac{w_j^2}{w_k^2} +\frac{w_k^2}{w_j^2}\right) +\sum_{j=0}^4 p_{jj}^i {\notag\displaybreak[0]\\ &=}\sum_{j<k}p_{jk}^i \left(\left(\frac{w_j}{w_k} +\frac{w_k}{w_j}\right)^2-2\right) +\sum_{j=0}^4 p_{jj}^i {\notag\displaybreak[0]\\ &=}\sum_{j<k}p_{jk}^i (a_{j,k}^2-2) +\sum_{j=0}^4 p_{jj}^i. \label{eq:NS}\end{aligned}$$ It can be verified by computer that (\[eq:NS\]) is nonzero for each of the cases (i)–(ii) in Theorem \[thm:main\]. Since $N(W)$ is symmetric, the adjacency matrices of $N(W)$ are the $(0,1)$-matrices representing the connected components of the Jones graph defined as follows (see [@JMN Sect. 3.3]). The [*Jones graph*]{} of a type-II matrix $W\in M_X({\mathbb{C}})$ is the graph with vertex set $X^2$ such that two distinct vertices $(a,b)$ and $(c,d)$ are adjacent whenever $\langle Y_{ab}, Y_{cd}\rangle\neq0$, where $\langle \;,\;\rangle$ denotes the ordinary (not Hermitian) scalar product. We claim that $(x,y)$ and $(x,z)$ belong to the same connected component in the Jones graph whenever $(x,y),(y,z),(z,x)\in R_4$. Indeed, if $(x,y)$ and $(x,z)$ belong to different connected components, then $(y,x)$ and $(z,x)$ belong to different connected components by Lemma \[lem:NS\]. In particular, $$\langle Y_{xy},Y_{xz}\rangle=\langle Y_{yx},Y_{zx}\rangle=0.$$ Let $$c_{i,j,k}=|\{u\in X\mid (x,u)\in R_i,\;(y,u)\in R_j,\;(z,u)\in R_k\}|.$$ Since $p_{1,3}^4=p_{2,3}^4=0$, we have $$\label{cijk} c_{1,j,k}+c_{2,j,k}= c_{j,1,k}+c_{j,2,k}= c_{j,k,1}+c_{j,k,2}=p_{j,k}^4$$ for $j,k\in\{1,2\}$. Then we have $$\label{eq:1026} \sum_{i,j,k=0}^4 c_{i,j,k}\frac{w_i^2}{w_jw_k}= \sum_{i,j,k=0}^4 c_{i,j,k}\frac{w_jw_k}{w_i^2}=0.$$ Since the rank of the coefficient matrix in (\[cijk\]) is $7$, we have one degree of freedom in (\[cijk\]). Combining (\[cijk\]) and (\[eq:1026\]), it can be verified by computer that these conditions give rise to a polynomial equation in $q$ which has no solution in positive integers $q\geq 4$ for each of the cases (i) and (ii) in Theorem \[thm:main\]. Therefore, we have proved the claim. This, together with Lemma \[lem:NS\], implies that, for each equivalence class $C$ of the equivalence relation $R_0\cup R_4$, $(C\times C)\cap R_4$ belongs to the same connected component in the Jones graph. Let $C$ and $C'$ be two distinct equivalence classes of $R_0\cup R_4$. We claim that, for any $(x,z)\in C\times C'$, there exist $y\in C$ such that $\langle Y_{xy},Y_{xz}\rangle\neq0$, and there exist $y'\in C'$ such that $\langle Y_{y'z},Y_{xz}\rangle\neq0$. Suppose $(x,z)\in R_1$ and $\langle Y_{xy},Y_{xz}\rangle=0$ for all $y\in R_4(x)$. Then $$\begin{aligned} 0&=\sum_{y\in R_4(x)}\langle Y_{xy},Y_{xz}\rangle {\displaybreak[0]\\ &=}\sum_{y\in R_4(x)}\sum_{u\in X} (Y_{xy})_u(Y_{xz})_u {\displaybreak[0]\\ &=}\sum_{y\in R_4(x)}\sum_{u\in X} \frac{W_{xu}^2}{W_{yu}W_{zu}} {\displaybreak[0]\\ &=}\sum_{y\in R_4(x)}\sum_{i,j=0}^4 \sum_{u\in R_i(x)\cap R_j(z)} \frac{W_{xu}^2}{W_{yu}W_{zu}} {\displaybreak[0]\\ &=}\sum_{i,j=0}^4 \sum_{u\in R_i(x)\cap R_j(z)} \sum_{k=0}^4 \sum_{y\in R_4(x)\cap R_k(u)} \frac{w_i^2}{w_kw_j} {\displaybreak[0]\\ &=}\sum_{i,j=0}^4 \sum_{u\in R_i(x)\cap R_j(z)} \sum_{k=0}^4 p_{4k}^i \frac{w_i^2}{w_kw_j} {\displaybreak[0]\\ &=}\sum_{i,j,k=0}^4 p_{ij}^1 p_{4k}^i \frac{w_i^2}{w_kw_j}.\end{aligned}$$ It can be verified by computer that this leads to a polynomial equation in $q$ which has no solution in positive integers $q\geq4$. Set $\ell\in\{2,3\}$. Similarly, suppose $(x,z)\in R_{\ell}$ and $\langle Y_{xy},Y_{xz}\rangle=0$ for all $y\in C$. Then $$\sum_{i,j,k=0}^4 p_{ij}^{\ell} p_{4k}^i \frac{w_i^2}{w_kw_j}=0,$$ and again this leads to a contradiction. Thus, there exists $y\in C$ such that $\langle Y_{xy},Y_{xz}\rangle\neq0$. Switching the role of $x$ and $z$, we see that there exists $y'\in C'$ such that $\langle Y_{y'z},Y_{xz}\rangle\neq0$. Therefore, we have proved the claim. Since $C$ and $C'$ are arbitrary, the claim shows that, in the Jones graph, $R_4$ is contained in a single connected component, and that every element $(x,z)\in R_1\cup R_2\cup R_3$ is adjacent to an element of $R_4$. Thus, $R_1\cup R_2\cup R_3\cup R_4$ is a connected component of the Jones graph. Therefore, $\dim N(W)=2$. [9]{} E. Bannai and T. Ito, [*Algebraic Combinatorics I: Association Schemes*]{}, Benjamin/Cummings, Menlo Park, 1984. W. Bosma, J. Cannon, and C. Playoust, *The Magma algebra system. I. The user language*, J. Symbolic Comput., 24 (1997), 235–265. A. E. Brouwer, A. M. Cohen, and A. Neumaier, [*Distance-Regular Graphs*]{}, Springer-Verlag, Berlin, Heidelberg, 1989. A. Chan and C. Godsil, *Type-II matrices and combinatorial structures*, Combinatorica, **30** (2010), 1–24. A. Chan and R. Hosoya, *Type-II matrices attached to conference graphs*, J. Algebraic Combin. **20** (2004), 341–351. R. Craigen, *Equivalence classes of inverse orthogonal and unit Hadamard matrices*, Bull. Austral. Math. Soc. **44** (1991), no. 1, 109–115. T. Ikuta and A. Munemasa, *Complex Hadamard matrices contained in a Bose–Mesner Algebra*, Spec. Matrices, **3** (2015), 91–110. R. Hosoya and H. Suzuki, *Type II matrices and their Bose–Mesner algebras*, J. Algebraic Combin. **17** (2003), 19–37. F. Jaeger, M. Matsumoto, and K. Nomura, *Bose–Mesner algebras related to type II matrices and spin models*, J. Algebraic Combin. **8** (1998), 39–72.

--- abstract: 'We continue the study of the lower central series of a free associative algebra, initiated by B. Feigin and B. Shoikhet [@FS]. We generalize via Schur functors the constructions of the lower central series to any symmetric tensor category; specifically we compute the modified first quotient $\bar{B}_1,$ and second and third quotients $B_2,$ and $B_3$ of the series for a free algebra $T(V)$ in any symmetric tensor category, generalizing the main results of [@FS] and [@AJ]. In the case $A_{m|n}:=T(\CC^{m|n})$, we use these results to compute the explicit Hilbert series. Finally, we prove a result relating the lower central series to the corresponding filtration by two-sided associative ideals, confirming a conjecture from [@EKM], and another one from [@AJ], as corollaries.' author: - Asilata Bapat - David Jordan title: Lower central series of free algebras in symmetric tensor categories --- Introduction ============ The *lower central series* of an associative algebra $A$ is the descending filtration by Lie ideals $A=L_1(A)\supset L_2(A)\supset\cdots$ defined inductively by $L_{k+1} := [A,L_k]$ for $k\geq 2$. The corresponding associated graded Lie algebra is denoted by $B(A) = \bigoplus_{i=0}^\infty B_k$, where $B_k:= L_k/L_{k+1}$. We let $M_k:=AL_k$ denote the two-sided associative ideal generated by $L_k$, and let $N_k:=M_k/M_{k+1}$. We let $Z$ denote the image of $M_3$ in $B_1$, and $\bar{B}_1:=B_1/Z$. The notions of associative algebras and Lie algebras make sense in any symmetric tensor category $\cC$, as does the lower central series filtration. These constructions are given in Section 3; in particular, we consider the lower central series of the tensor algebra $T(V)$ of any object $V\in\cC$. In this generality, the lower central series quotients $B_k$ are functorial in $V$, and may be expressed in the basis of Schur functors $\mathbb{S}_\lambda$, as is explained in Section 4. It follows that there exists a universal decomposition of the functor $V\mapsto B_k(T(V))$ into a sum of Schur functors $\mathbb{S}_\lambda$, which holds in any symmetric tensor category. Moreover, to compute this decomposition it suffices to consider only the cases $\cC=\Vect, V=\CC^n$, as $n\to\infty$. As an application, we have: [theorem]{}[Bthm]{} \[B1B2B3\]For any symmetric tensor category $\cC$, and any $V\in\cC$, the maps $\phi$ from Theorem \[thm:fsisomorphism\], and $f_3$ from Theorem \[thm:AJisomorphism\], induce natural isomorphisms: $$\begin{aligned} \bar{B}_1(T(V)) &\cong \Omega^{ev}(V)/\Omega^{ev}_{ex}(V),\\ B_2(T(V)) &\cong \Omega^{ev}_{ex}(V),\\ B_3(T(V)) &\cong S(V)\ot (\oplus_{k=0}^{\infty} \mathbb{S}_{(2,1^{2k+1})}(V)).\end{aligned}$$ In Section 5, we consider the free algebra $A=A_{m|n}=T(\CC^{m|n})$ on even generators $x_1,\ldots x_m$ and odd generators $y_1,\ldots y_n$; we study the lower central series with respect to the usual super commutator $[a,b]:=ab - (-1)^{|a||b|}ba$. The algebra $A$ and thus each of the $B_k$ are graded by the degree in the generators. Let us write $h_{B_k}(\mathbf{u},\mathbf{v})$ for the multi-graded Hilbert series of the $B_k$, where $\mathbf{u}=(u_1,\ldots,u_m), \mathbf{v}=(v_1,\ldots,v_n)$ record the degree in the even and odd variables, respectively. Our first main result is a computation of the Hilbert series for $\bar{B}_1$, $B_2$ and $B_3$, based on the above decomposition: [theorem]{}[BthmHS]{} \[B1B2B3HS\] We have the following Hilbert series: $$\begin{aligned} h_{\bar{B}_1}(\mathbf{u},\mathbf{v})&= \frac{1}{4}(\mathbf{P}_{m|n} + \sum_{i=1}^m\frac{u_i}{2(1-u_i)} + \sum_{j=1}^n\frac{v_j}{2(1+v_j)} + 3), \\ h_{B_2}(\mathbf{u},\mathbf{v})&= \frac{1}{4}(\mathbf{P}_{m|n} - \sum_{i=1}^m\frac{u_i}{2(1-u_i)} - \sum_{j=1}^n\frac{v_j}{2(1+v_j)} - 1),\\ h_{B_3}(\mathbf{u},\mathbf{v})&=\frac12((\sum_i u_i + \sum_j v_j)(\mathbf{P}_{m|n} +1) - (\mathbf{P}_{m|n} -1)),\end{aligned}$$ where $\mathbf{P}_{m|n}:=\prod_{i=1}^m\frac{(1+u_i)}{(1-u_i)}\cdot\prod_{j=1}^n\frac{(1+v_j)}{(1-v_j)}$, Our second main result, presented in Section 6, relates the Lie ideals $L_k$ for any algebra $A$ to the two-sided associative ideals $M_k:=A L_k$. More precisely, we prove the following theorem (for any algebra in any symmetric tensor category $\cC$): [theorem]{}[mainident]{}\[mainident\] $[M_j,L_k]\subset L_{k+j}$, whenever $j$ is odd. In the case $A=A_n$, this may be seen as a strengthening of a special case of “Key-Lemma" of [@FS], which asserts that $[M_3,L_1]\subset L_3$. While the proof of this theorem is largely by direct computation, there are many implications. Firstly, we prove the “if" direction of [@EKM], Conjecture 3.6 as a corollary: [corollary]{}[mjmkcor]{} $M_jM_k\subset M_{j+k-1}$, whenever $j$ or $k$ is odd. As another application of Theorem \[mainident\], we have: [theorem]{}[Bmcor]{}\[Bmcor\] $B_k= [A_{\leq 2},B_{k-1}],$ for all $k$. A central observation of [@FS] was that the lower central series quotients $B_k, k\geq 2$ each carry an action of the Lie algebra $W_n$ of vector fields on $\CC^n$ and that they are iterated extensions of so-called “tensor field modules” $\flambda$ associated to a Young diagram $\lambda$. Theorem \[Bmcor\] was conjectured in [@AJ], Remark 1.4, where it was explained (see also the proof of Lemma 5.4, [*loc. cit.*]{}) how to derive the following: [corollary]{}[strongerlambdabound]{} \[thm:strongerlambdabound\] Let $k\geq 3$. For $\flambda$ in the Jordan-Hölder series of $B_k(A_n)$, we have: $$|\lambda|\leq 2k-3+2\left\lfloor \frac{n-2}{2}\right\rfloor.$$ Let $\overline{\lambda}$ denote the Young diagram obtained by deleting the first column of $\lambda$. [corollary]{}[lambdabound]{} \[cor:lambdabound\] Let $k\geq 3$. For $\flambda$ in the Jordan-Hölder series of $B_k(A_n)$, we have: $$|\overline{\lambda}|\leq 2k-5.$$ Let $\mathcal{G}_\lambda$ denote the direct Schur functor, $$\mathcal{G}_\lambda(V) := S(V)\ot (\oplus_{k=0}^\infty \mathbb{S}_{(\lambda,1^{2k})}(V)).$$ For example, $A/M_3\cong \mathcal{G}_{(0)}$, and $B_3\cong \mathcal{G}_{(2,1)}$, by Theorem \[B1B2B3\]. As another consequence of Theorem \[Bmcor\], we have: [proposition]{}[glambdabound]{}\[glambdabound\] Let $\cC$ be a symmetric tensor category, let $V\in\cC$, and let $A=T(V)$. There exists a finite collection $\Lambda_k$ of Young diagrams $\lambda$, and a surjection $\oplus_{\lambda\in\Lambda_k}\mathcal{G}_{\lambda} \to B_k$. For a graded vector space $M$, let $M[d]$ denote the $d$th graded component. We have: [corollary]{}[weakbound]{}\[weakbound\]Let $k\geq 3$, and let $A:=A_{m|n}$. There exists $C_k(m,n)$ such that $\dim B_k(A_{m|n})[d]\leq C_k(m,n)d^{m+n-1}$. In Section 8, we state some open questions and conjectures, and suggest possible approaches to their solution. Acknowledgments {#acknowledgments .unnumbered} --------------- We are deeply grateful to P. Etingof for his guidance and advice. We thank Darij Grinberg for pointing out an error in the formulation of Conjecture \[glambdaconj\], in a previous version of the paper. The work of the first author was supported by MIT’s SPUR and UROP programs for undergraduate research. The work of the second author was supported by NSF grant DMS-0504847. Preliminaries ============= In this section, we recall definitions for the lower central series of an associative algebra $A$, its associated graded Lie algebra $B(A)$, the Lie algebras $W_n$, and the tensor field modules $\flambda$. Let $W_n:=Der(\C[x_1,\ldots,x_n])$ denote the Lie algebra of polynomial vector fields on $\CC^n$. Then $W_n\cong\bigoplus_i\C[x_1,\ldots,x_n]\partial_i$, with bracket $$[p\partial_i,q\partial_j] = p\frac{\partial q}{\partial x_i}\partial_j - q\frac{\partial p}{\partial x_i}\partial_j.$$ Let $W_n^0$ denote the Lie subalgebra of vector fields vanishing at the origin, and let $W_n^{00}\subset W_n^0$ denote its Lie ideal of vector fields vanishing at the origin to at least second order. Then $W_n^0/W_n^{00}$ is isomorphic to $\gl_n$, via the map $x_i\partial_j\mapsto E_{ij}$. Let $\lambda = (\lambda_1\geq\cdots\geq\lambda_n)$ be a Young diagram, and let $V_\lambda$ be the corresponding irreducible $\gl_n$-module. Then $V_\lambda$ is also a representation of $W_n^0$ via the projection $W_n^0\to W_n^0/W_n^{00}\cong gl_n$. Let $\widetilde{\ff}_\lambda= \Hom^{fin.}_{\uu(W_n^0)}(\uu(W_n),V_\lambda)$, denote the finite part of the coinduced module $\Hom_{\uu(W_n^0)}(\uu(W_n),V_\lambda)$, spanned by homogeneous vectors. As a vector space, $\widetilde{\ff}_\lambda$ is isomorphic to $\C[x_1,\ldots,x_n]\otimes V_\lambda$. \[flambdairrep\] If $\lambda_1\geq 2$ or if $\lambda = (1^n)$, then $\widetilde{\ff}_\lambda$ is an irreducible $W_n$-module. Otherwise, $\lambda = (1^k,0^{n-k})$, and $\widetilde{\ff}_\lambda$ is $\Omega^k(\C^n)$, the space of polynomial differential $k$-forms on $\C^n$. In this case $\widetilde{\ff}_\lambda$ contains a unique irreducible submodule consisting of the closed $k$-forms. Let $\flambda$ denote the unique irreducible submodule of $\widetilde{\ff}_\lambda$, which is equal to $\widetilde{\ff}_\lambda$ if $\lambda_1\geq 2$ or $\lambda = (1^n)$. Any $W_n$-module on which the operators $x_i\partial_i$, for $i = 1,\ldots , n$, act semisimply with nonnegative integer eigenvalues and finite dimensional common eigenspaces has a Jordan-Hölder series whose composition factors are $\flambda$, each occurring with finite multiplicity. \[thm:rudakovwn\] In [@FS], it was realized that for $A=A_n$, each $B_k$, $k\geq 2$ carries a $W_n$-action, arising from an identification of $A/M_3$ with the algebra $\Omega^{ev}$ of even degree differential forms on $\CC^n$. Let $\Omega^{ev}_{ex}$ denote the subspace of even degree exact forms. Let us denote by “$\wedge$” the wedge product of differential forms. We let “$\ast$” denote the *Fedosov product*, $$a\ast b := a\wedge b + (-1)^{\degree a}da\wedge db,$$ which defines an associative multiplication on $\Omega,$ preserving $\Omega^{ev}$. The space $\Omega$ (resp. $\Omega^{ev}$) equipped with the Fedosov product is denoted by $\Omega_\ast$ (resp. $\Omega^{ev}_\ast$). We have: The map $\phi: x_i\mapsto x_i$ extends to an isomorphism of algebras $\Omega^{ev}_\ast\cong A/M_3(A)$, restricts to an isomorphism $\Omega^{ev}_{ex}\cong B_2(A)$, and descends to an isomorphism of $\Omega^{ev}/\Omega^{ev}_{ex}\cong \bonebar(A)$.\[thm:fsisomorphism\] The action of $W_n$ on $\bonebar\cong\Omega^{ev}/\Omega^{ev}_{ex}$ by Lie derivatives extends uniquely to an action of $W_n$ on $B_k$, for $k\geq 2$. This action of $W_n$ clearly satisfies the conditions of Theorem \[thm:rudakovwn\]. It follows that $\bar{B}_1$ and each $B_k, k\geq 2$ have a finite Jordan-Hölder series with respect to this action, such that each composition factor is isomorphic to $\flambda$ for some Young diagram $\lambda$. By [@DE], Corollary 3, the Jordan-Hölder series is of finite length. In [@AJ], similar techniques were used to give a description of $B_3(A_n)$. We have: \[thm:AJisomorphism\] The map $f_3:\bar{B}_1\otimes B_2\to B_3, a\ot b\mapsto [a,b]$ restricts to a surjection $(\Omega^0/\CC)\otimes B_2\to B_3$, and induces an isomorphism $$B_3\cong \oplus_{k=0}^{\infty} \mathcal{F}_{(2,1^{2k+1})}.$$ Lower series filtrations in symmetric tensor categories ======================================================= Let $(\cC,\sigma)$ be a symmetric tensor category. For the sake of clarity, we suppress explicit mention of associators in all formulas and commutative diagrams in this section. A (unital) associative algebra $(A,m,\eta)$ in $\cC$ is an object $A\in\cC$, together with morphisms $\eta:\mathbf{1}\to A$ and $m:A\ot A\to A$ such that $m\circ(\eta\ot\id)=m\circ(\id\ot\eta)=\id$ and $m\circ(m\ot\id)=m\circ(\id\ot m)$. Algebra morphisms are defined in the usual way, yielding the category $\cAlg$. We will say that $A\in\cAlg$ is commutative if $m\circ(\id_{A\ot A}-\sigma_{A,A})=0$. A Lie algebra $(\A,\mu)$ in $\cC$ is an object $\A$ and a morphism $\mu:\A\ot\A\to\A$ satisfying: 1. Skew symmetry: $\mu\circ(\id_{\A\ot\A} + \sigma_{\A,\A})=0$. 2. Jacobi identity: $\mu\circ(\id_{\A}\ot\mu)\circ(\id + (123) + (132))=0$. Lie algebra morphisms are defined in the usual way, yielding the category $\cC$-Lie. As with the example $\cC=\Vect$, we have the functor, $$\operatorname{Lie}:\cAlg\to\cC\text{-Lie}$$ $$(A,m,\eta)\mapsto (A,m\circ(\id - \sigma_{A,A})).$$ The *commutator morphisms* $l_k:\A^{\ot k}\to \A$ are defined inductively as follows: $l_1:=\id_{\A}, l_2:=\mu, l_m:=l_{k-1}\circ (\id^{k-2}\ot l_2)$. The *lower central series filtration* is the collection of $L_k(\A):=\operatorname{Im}(l_k)\subset \A$. For each $k\in \NN$, we have injections $i_k: L_{k+1}\to L_k$ induced by the natural injections $i_1:L_2\to \A$. We let $B_k:=\operatorname{coker} i_k$ denote the associated graded components. The following is a straightforward generalization of the corresponding statement for $\cC=\Vect$, which is proved by repeated use of Jacobi identity: Let $f:\A^{\ot k}\to \A$ be any morphism obtained by iterating $\mu$. Then $\operatorname{Im}(f)\subset L_k$. We will consider the following examples of algebras and their associated Lie algebras. Let $V\in\cC$. 1. The tensor algebra $T(V):=\oplus_i V^{\ot i}$ with the usual product. 2. The symmetric algebra $S(V):=T(V)/\langle\operatorname{Im}(\sigma_{V,V}-\id_{V\ot V})\rangle$. 3. The exterior algebra $\Lambda(V):=T(V)/\langle\operatorname{Im}(\sigma_{V,V} + \id_{V\ot V})\rangle$. 4. The algebra of differential forms $\Omega(V):=S(V)\ot\Lambda(V)$, and its commutative subalgebra of even forms (relative to grading on $\Lambda(V)$). Recall that we have canonical isomorphisms $s:V\to S^1(V)$ and $\lambda:V\to \Lambda^1(V)$. The algebra $\Omega(V)$ has a unique derivation $d:\Omega^*(V)\to\Omega^{* + 1}(V)$, such that $d|_{S^1(V)}=\lambda\circ s^{-1}$, and $d|_{\Lambda^1(V)}=0$. As with ordinary differential forms, $d$ defines a differential: $d^2=0$. We let $\Omega(V)_{ex}=\operatorname{Im}(d)$ denote the exact forms, and $\Omega(V)_{cl}=\ker(d)$ the closed forms. By the Poincaré lemma, we have an isomorphism $\Omega(V)_{cl}\cong \mathbf{1} \oplus \Omega(V)_{ex}$. From $\text{Vect}$ to any symmetric tensor category =================================================== All definitions in this section are over a field $k$ of characteristic zero. The abelian category of abstract Schur functors is: $$\mathbf{Sch}:=\oplus_{j\geq 0} \textrm{Rep}(S_j).$$ $\mathbf{Sch}$ is a tensor category, where for $V\in \textrm{Rep}(S_j)$ and $W\in\textrm{Rep}(S_k)$, we set: $$V\ot W:=\text{Ind}_{S_j\times S_k}^{S_{j+k}} V\bt W,$$ The isomorphisms $S_j\times S_k\cong S_k\times S_j$ induced by $(1,\ldots, j, j+1,\ldots j+k)\mapsto (j+1,\ldots j+k,1,\ldots j)$ make $\mathbf{Sch}$ a symmetric tensor category. As an abelian category, $\mathbf{Sch}$ is filtered by its full subcategories, $$\mathbf{Sch}_N:=\oplus_{j\leq N}\textrm{Rep}(S_j),$$ and is semi-simple, with simple objects $(k,V)$, where $k\in \ZZ_{\geq 0}$ and $V\in\textrm{Irrep}(S_k)$. Let $\cC$ be a symmetric tensor category. We have a bi-functor $F:\mathbf{Sch}\times \cC\to\cC$, sending $(W\in \textrm{Rep}(S_k),V\in\cC)$ to $F_{V}(W):=(W\ot V^{\ot k})^{S_k}\in\cC$. For a Young diagram $\lambda$ of size $k$, we let $\mathbb{S}_\lambda:=V\mapsto F_{V}(W_\lambda)$, called the irreducible Schur functor in $\cC$ of type $\lambda$. $F_V:\mathbf{Sch}\to\cC$ is a symmetric tensor functor. For $\cC=\textrm{GL}(n)$, we have functors $F_n:=F_{\CC^n}:\mathbf{Sch}\to \textrm{Rep}(GL_n)$. Then Schur-Weyl duality implies the following: \[fullyfaithful\] The family of functors $F_n$ is asymptotically fully faithful. More precisely, $F_n|_{\mathbf{Sch}_N}$ is fully faithful for $n\geq N$. We will apply this claim in the form of the following: \[meta\] 1. Suppose we have an identity of Schur functors (up to an isomorphism) that holds in $\textrm{Rep}(GL_n)$, for all $n$. Then it holds for abstract Schur functors, and hence in any symmetric tensor category. 2. Any morphism of abstract Schur functors which yields an isomorphism (epi, mono, or zero morphism, resp.) in $\textrm{Rep}(GL_n)$ for all $n$ is itself an isomorphism (epi, mono, or zero morphism, resp.), and hence yields an isomorphism (epi, mono, or zero morphism, resp.) in any symmetric tensor category. A simple, yet important, observation is that the members $L_k:=L_k(T(V))$ of the lower central series, and the quotients $B_k$, as well as all the algebras and Lie algebras constructed in the previous sections, are defined purely in terms of the symmetric group action on various tensor products of an object $V$. Thus each of these constructions is in fact functorial in $V$, and moreover can be expressed in the basis of Schur functors $\mathbb{S}_\lambda$. An immediate consequence is that if we can provide a consistent decomposition of $L_i(T(\CC^n))$ into a sum of $\mathbb{S}_\lambda(\CC^n)$, for all $n$, then that formula holds in any symmetric tensor category. As an application, we have: Both sides of each asserted isomorphism are expressed in terms of the Schur functors $\mathbb{S}_\lambda(V)$. It is easily checked that the morphisms from [@FS] and [@AJ] are natural in $V$, and defined in any symmetric tensor category; thus they define a natural transformation of the corresponding functors. Moreover it is shown that they are natural isomorphisms in the case $\cC=\textrm{Rep}(GL_n)$, for all $n$, and therefore, by Proposition \[meta\], in any tensor category. Super vector spaces and $A_{m|n}$ ================================= Now we apply the definitions and propositions of the previous section to the setting of super vector spaces. Let $\cC:=\text{SuperVect}$, and let $\CC^{m|n}\in\cC$ denote the super vector space on even basis $\{x_1,\ldots,x_m\}$ and odd basis $\{y_1,\ldots,y_n\}$. Let $A:=A_{m|n}$ denote the tensor algebra $T(\CC^{m|n})$, and consider the lower central series $L_k(A)$, and the associated graded quotients $B_k(A)$. We compute $h_{B_2}$ first. By Theorem \[B1B2B3\], we have that $B_2$ is the subspace $\Omega^{ev}_{ex}$ of exact super differential forms of even degree: $$h_{B_2} = h_{\Omega^{ev}_{ex}} = \sum_{i=0}^\infty h_{\Omega^{2i}_{ex}}.$$ The map $d\colon\Omega(\C^{m|n})\to\Omega(\C^{m|n})$ gives rise to the following complex: $$\cdots\stackrel{d}{\longrightarrow}\Omega^{i-1}\stackrel{d}{\longrightarrow}\Omega^{i}\stackrel{d}{\longrightarrow}\Omega^{i+1}\stackrel{d}{\longrightarrow}\cdots .$$ This sequence is exact except for a copy of $\C$ at $i=0$. Therefore we obtain the recursion relation: $$\begin{aligned} \label{eq:1} h_{\Omega^{i}_{ex}} &= h_{\Omega^{i-1}} - h_{\Omega^{i-1}_{ex}}, \,\,(i\geq 2) \\ h_{\Omega^1_{ex}} &= h_{\Omega^{0}} - 1,\nonumber\end{aligned}$$ which has solution: $$\label{eq:2} h_{\Omega^{i}_{ex}} = \sum_{j=0}^{i-1}(-1)^{i-j+1}h_{\Omega^j} + (-1)^{i+1}.$$ By the tensor decomposition, $\Omega(\C^{m|n}) = S(\C^{m|n})\otimes\Lambda(\C^{m|n})$, we may write: $$h_\Omega(\mathbf{u},\mathbf{v},t) = \frac{\prod_{j=1}^n(1+v_j)}{\prod_{i=1}^m(1-u_i)}\cdot \frac{\prod_{i=1}^m(1+tu_i)}{\prod_{j=1}^n(1-tv_j)},$$ where the variable $t$ is a counter for the degree of the form. Thus $h(\Omega^i)$ is the coefficient of $t^i$ in $h(\Omega)$. From equations and , we have: $$\begin{aligned} h_{\Omega^{2k}_{ex}} &= \sum_{i=0}^{2k-1}(-1)^{i+1}\text{Coeff}_{t^i}(h_\Omega) +1,\\ & = \text{Res}_{t=0}\left(h_\Omega\cdot\frac{1}{t}\cdot\frac{t^{-2k}-1}{1+t^{-1}}\right) +1.\end{aligned}$$ As the term $(t^{-2k}-1)/(1+t^{-1})$ has a convergent power series for $|t|>1$, we can compute this residue by taking a contour integral around a circle $\gamma$ with centre at the origin and radius greater than $1$: $$h_{\Omega^{2k}_{ex}} =\frac{1}{2\pi i}\int_{\gamma}\left(h_\Omega\cdot\frac{-1}{1+t}\right) +\frac{1}{2\pi i}\int_{\gamma}\left(h_\Omega\cdot\frac{t^{-2k}}{1+t}\right) +1.$$ An elementary computation shows that $$\frac{1}{2\pi i}\int_{\gamma}\left(h_\Omega\cdot\frac{-1}{1+t}\right) = -1,$$ so we find $$h_{\Omega^{2k}_{ex}} = \frac{1}{2\pi i}\int_{\gamma}\left(h_\Omega\cdot\frac{t^{-2k}}{1+t}\right),$$ and thus: $$h_{\Omega^{ev}_{ex}} =\sum_{k\geq 1}h_{\Omega^{2k}_{ex}}= \frac{1}{2\pi i}\int_{\gamma}\left(h_\Omega\cdot\frac{1}{(1+t)(1-t^{-2})}\right) -1.$$ The integrand has a simple pole at $t=1$ and a double pole at $t=-1$. Calculating the integral by the method of residues yields us the formula asserted for $h_{B_2}$. We can compute the Hilbert series of $\Omega^{ev}$ directly: $$\begin{aligned} h_{\Omega^{ev}} &= \frac12(h_{\Omega}(\mathbf{u},\mathbf{v},1) + h_{\Omega}(\mathbf{u},\mathbf{v},-1))\\ &=\frac12\cdot\frac{\prod_{j=1}^n(1+v_j)}{\prod_{i=1}^m(1-u_i)}\cdot (\frac{\prod_{i=1}^m(1+u_i)}{\prod_{j=1}^n(1-v_j)} + \frac{\prod_{i=1}^m(1-u_i)}{\prod_{j=1}^n(1+v_j)})\\ &=\frac12\cdot( \mathbf{P}_{m|n} + 1).\end{aligned}$$ Thus, we compute: $$\begin{aligned} h_{\bar{B}_1}= h_{\Omega^{ev}} - h_{B_2} = \frac{1}{4}(\mathbf{P}_{m|n} + \sum_{i=1}^m\frac{u_i}{2(1-u_i)} + \sum_{j=1}^n\frac{v_j}{2(1+v_j)} + 3),\end{aligned}$$ as asserted. Finally, we can compute the Hilbert series of $B_3$ from its tensor decomposition in Theorem \[B1B2B3\]. We have: $$h_{\mathbb{S}_{(2,1^k)}(V)} = h_V\cdot h_{\mathbb{S}_{(1^{k+1})}(V)} - h_{\mathbb{S}_{(1^{k+2})}(V)},$$ from which we compute: $$\begin{aligned} h_{B_3}&= h_{S(V)}\left(h_V \sum_{k=1}^\infty h_{\mathbb{S}_{(1^{2k})}} - \sum_{k=1}^\infty h_{\mathbb{S}_{(1^{2k+1})}}\right)\\ &= h_V(h_{\Omega^{ev}}-1) - (h_{\Omega^{odd}}-h_V)\\ &= \frac12((\sum_i u_i + \sum_j v_j)(\mathbf{P}_{m|n} +1) - (\mathbf{P}_{m|n} -1)).\end{aligned}$$ Relations between $M_j$ and $L_k$ ================================= All statements in this section, as well as Lemma \[B4cor\] and Theorem \[Bmcor\] apply to any symmetric tensor category, while proofs are carried out only in the category of vector spaces. The proofs in general follow by application of Proposition \[meta\]. Notational conventions {#notational-conventions .unnumbered} ---------------------- We have to write many expressions involving iterated commutators: when it is clearer, we omit brackets and adopt the convention of right-iterated bracketing: $$[x_1,\ldots,x_n] := [x_1,[x_2,[\cdots,[x_{n-1},x_n]\cdots].$$ We let “$\star$” denote the symmetric product: $a\star b = \frac12(ab + ba)$. Let $Sym(X)$ denote the symmetric group on the set $X$, and let $G:=Sym(\{x,y,z\})\times Sym(\{u,v\})\subset Sym(\{x,y,z,u,v\})$. Let $\textrm{Alt}_G = \sum_{g\in G} \sgn(g) g\in \CC[G]$. The proof is a double induction, first on $k$, then on $j$. The base of induction is $(j,k)=(3,1)$, as the trivial case $(1,1)$ is not sufficient for the induction step. \[keylemma\]$[M_3,L_1] \subset L_4$. We show that $[x[y,z,u],v]\in L_4$; letting $x,y,z,u,v$ range over all monomials, we span $[M_3,L_1]$, proving the lemma. We may replace $[x[y,z,u],v]$ by $[x\star[y,z,u],v]$, as they are congruent modulo $L_5$. The following two identities are straightforward: $$[x\star[y,z,u],v]+[y\star[x,z,u],v] = [[x\star y,z,u],v] \in L_4$$ $$[x\star[y,z,u],v]+[v\star[y,z,u],x]=-[x\star v,y,z,u]\in L_4.$$ Together they imply that $[x\star[y,z,u],v]$ is alternating for $G$, modulo $L_4$. Finally, we have the following identity, which may be verified directly on coefficients of the ${5\choose 2}=10$ monomials where $x$ is left of $y$ is left of $v$ and $z$ is left of $u$: $$\textrm{Alt}_G [x\star[y,z,u],v] = \textrm{Alt}_G[u,x,y,z\star v].$$ We conclude that $[x[y,z,u],v]\in L_4$. $[M_3,L_k]\subset L_{k+3}$. Consider $[a[b,c,d],l]$, with $l \in L_k$. Write $l=[x,y]$, where $x\in L_1, y\in L_{k-1}$. We have $$[a[b,c,d],[x,y]] = [[a[b,c,d],x],y]+ [x,[a[b,c,d],y]]\subset L_{k+3},$$ by induction. To conclude the proof of the theorem, we consider a general element $[m,l]$ with $m\in M_j$, and $l=[l_1,\ldots,l_k]\in L_k$. Write $m= a[b,c,d]$, where $d\in L_{j-2}$. Then by Lemma \[keylemma\], we have that $[a[b,c,d],l] \subset L_{k+3}(a,b,c,d,l_1,\ldots l_k)$, meaning that we regard $d$ as a variable instead of an iterated bracket. Each summand in $L_{k+3}$ is an iterated bracket of linear and quadratic expressions in the variables $a,b,c,d,l_1,\ldots l_k$. By repeated application of the Jacobi identity, we may assume that all summands are of the form $[y_1,\ldots,y_{k+2},y_{k+3}d],$ for various permutations $\{y_1,\ldots, y_{k+3}\}=\{a,b,c,l_1,\ldots,l_k\}$. We now plug in $d=[d_1,\ldots, d_{j-2}]$. Then $[y_{k+2},y_{k+3}d]\subset L_{j-1}$ by induction, which concludes the proof of the theorem. As a corollary, we can give an affirmative answer to the “if" direction of Conjecture 3.6 from [@EKM]. We may assume $k$ is odd. Clearly it is enough to show that $L_jL_k\subset M_{j+k-1}$. Let $a \in A, x\in L_{j-1}, y\in L_k$. Then we have $[x,a]y=[x,ay]-a[x,y]$. The LHS is a completely general generator of $L_jL_k$. By Theorem \[mainident\], both the terms on the RHS are in $M_{j+k-1}$. Applications of Theorem \[mainident\] to the modules $B_k$ ========================================================== \[B4cor\] $B_4 = [A_{\leq 2},B_3].$ In [@AJ], it has been shown that $$B_4 = [A_{\leq 2},B_3]+ \sum_{x,y,z\in A_1} [x[y,z],B_3].$$ Thus our task is to show that each summand $[x[y,z],B_3]$ is actually contained in the span of iterated brackets with only quadratic or linear outermost expression. We consider a general element of this form, $[x[y,z],u,v,w]$, where we assume that $u$ and $v$ are either linear or quadratic ($w$ is arbitrary). Then we have: $$[x[y,z],u,v,w]=[u,x[y,z],v,w]+[[x[y,z],u],v,w].$$ The Leibniz rule implies that $[x[y,z],u]-[x,y][z,u]\in M_3$. Thus by Theorem \[mainident\], we have: $$\begin{aligned} &=[[x,y][z,u],[v,w]]=[[z,u],[x,y][v,w]]+[[x,y],[z,u][v,w]]\nonumber\mod L_5\\ &=[[z,u],[x,y[v,w]]] + [[x,y],[z,u[v,w]]] \mod L_5.\label{B4eqn}\end{aligned}$$ We now observe that each of the above expressions has only a single term of degree higher than two. By repeated use of Jacobi identity, we can put that term in the innermost slot. We have established the corollary. **Case 1: m is odd.** Let us consider the element $[x[y,z],w,v] \in B_k$, where $v=[v_1,\ldots,v_{k-2}]$ is some iterated bracket expression. By Lemma 5.2 of [@AJ], we have: $$[x[y,z], [w, v]] - [x, [w[y, z], v]] + [y, [w[x, z], v]] - [z ,[w[x, y], v]] \in L_4(x,y,z,w,v),$$ meaning that we regard $v$ as a variable, rather than as an iterated commutator. Each summand in $L_4$ is necessarily an iterated bracket of linear and quadratic expressions in the variables $x,y,z,w,v$. By repeated application of the Jacobi identity, we may assume that all summands are of the form $[a_1,\ldots,a_4v]$ for various permutations $\{a_1,\ldots,a_4\} = \{x,y,z,w\}$. We now plug in $v=[v_1,\ldots,v_{k-2}]$. Theorem \[mainident\] implies $[a_3,a_4v]\in L_{k-1}$, which concludes the proof. **Case 2: m is even.** We have already addressed the case $k=4$. Let us consider the element $[x[y,z],u,v,w]$ of $B_k$, where $w=[w_1,\ldots,w_{k-3}]$. We may repeat the arguments of Lemma \[B4cor\] up to equation without significant modification. By Theorem \[mainident\], both terms on the RHS are in fact zero in $B_k$, which concludes the proof. The proof of [@AJ], Lemma 5.4, applies here without significant modification. First, let $n=2$; the first corollary implies that if $\mathcal{F}_\lambda$ occurs in $B_k(A_2)$, then $|\lambda|\leq 2k-3$. Moreover, the modules $\mathcal{F}_{(r)}$ do not appear in $B_k(A_2)$, for $k\geq 2$, as $\CC[z_1]$ is a commutative algebra. Thus $\lambda$ must have at least two rows, so that $|\overline{\lambda}|\leq 2k-5$. We proceed by induction on $n$. Consider some $\flambda$ which occurs in $B_k(A_n)$. If $\lambda$ has less than $n$ rows, let $\mu = (\lambda_1,\ldots,\lambda_{n-1})$, then it follows that $\mathcal{F}_\mu$ occurs in $B_k(A_{n-1})$, so that $|\overline{\lambda}|=|\overline{\mu}|\leq 2k-5$, by the induction hypothesis. Thus we may assume $\lambda$ has exactly $n$ rows. Then the first corollary implies: $$|\overline{\lambda}| = |\lambda|-n \leq 2k-3 + 2\left\lfloor \frac{n-2}{2}\right\rfloor -n \leq 2k-5.$$ Let $\lambda$ be a Young diagram, and recall that $\mathcal{G}_\lambda(V)=\mathbb{S}_\lambda(V)\ot\Omega^{ev}(V)$. Using the techniques of Section 3 and 4, we have: For a graded vector space $M$, recall that $M[d]$ denotes the $d$th graded component. Then we have: By Lemma \[B4cor\] and Theorem \[Bmcor\], we have a surjections $A_{\leq 2}^{\ot k-2}\ot\Omega^{ev}\twoheadrightarrow B_k$. $A_{\leq 2}^{\ot k-2}$ is finite dimensional, while $\dim\Omega^{ev}[d]$ is a polynomial of degree $m+n-1$ in $d$. Open questions and conjectures ============================== While Corollary \[weakbound\] states that the $B_k(A_{m|n})$ exhibit polynomial growth, computer experiments suggest something stronger: For all algebras $A_{m|n}$, the Hilbert series of $B_k$, $k\geq 3$ are rational functions with denominator $\prod_i(1-u_i)\prod(1-v_j)$. In the completely even case, this is a theorem of [@DE], following from the description of $B_k$ as a finite iterated extension of the modules $\flambda$. In general, this conjecture would follow as a corollary from the following strengthening of Proposition \[glambdabound\]: \[glambdaconj\] Let $k\geq 3$. There exists a finite collection $\Lambda$ of Young diagrams $\lambda$, and an isomorphism of abstract Schur functors $V\mapsto B_k(T(V))\cong \oplus_{\lambda\in\Lambda}\mathcal{G}_\lambda$. One approach to this conjecture is to consider the family $A=A_n$ as $n\to\infty$, along the lines of Sections 3 and 4. While each $B_k(A_n)$ has a finite length Jordan-Hölder series consisting of modules $\flambda$ with $|\bar{\lambda}|\leq 2k-5$, there is currently no control over the first column of $\lambda$ as $n$ varies. The conjecture states that as $n$ grows, the new diagrams to appear are precisely those obtained from an earlier diagram by adding two boxes in the first column. Towards a possible explanation, we observe that not only the Lie algebra $W_n$, but a larger Lie algebra acts on each $B_k$: Let $A=A_n$. We have $\textrm{Der}(A/M_3)\cong \widetilde{W_n}:=\oplus_i \Omega_\ast^{ev} \ot\partial_i$, with the Lie bracket: $$[a\partial_i,b\partial_j]=(a\ast\partial_i(b))\partial_j-(b\ast\partial_j(a))\partial_i,$$ where $\partial_i(a)$ is the Lie derivative, and $\ast$ is the Fedosov product. $A/M_3$ is an $A$-bimodule via the projection $\pi:A\to A/M_3$. Clearly, $$\textrm{Der}_A(A,A/M_3)\cong A/M_3\ot V^*\cong \Omega_{\ast}^{ev}(V)\ot V^*,$$ by Theorem \[thm:fsisomorphism\]. Since any derivation of $A$ preserves the ideal $M_3$, it descends to a derivation of $\textrm{Der}(A/M_3)$. $\widetilde{W_n}$ has a nilpotent ideal consisting of the forms of nonzero degree; the quotient by this ideal may be identified with $W_n$. In general, for a symmetric tensor category $\cC$, and $V\in\cC$, one can define the analogous Lie algebras $\widetilde{W(V)}$ and $W(V) \in\cC$-Lie. In particular, for $A=A_{m|n}$, one has the Lie algebras $\widetilde{W_{m|n}}$ and $W_{m|n}$; however the kernel of the map $\widetilde{W_{m|n}}\to W_{m|n}$ can be rather large; for instance in the completely odd setting, $W_{0|n}$ is finite dimensional, while $\widetilde{W_{0|n}}$ is infinite dimensional. It follows as in [@FS] that the Lie algebra $\widetilde{W_n}$ acts on each $B_k$. The following conjecture can be checked for $\bar{B}_1, B_2,$ and $B_3$ from their explicit descriptions. Let us denote the $p$th multi-graded part of a multi-graded vector space $M$ by $M[p]$. Let $E:=\sum x_i\partial_i$ denote the Euler operator. \[neweuler\] The operator $D:=[x_{n+1},x_{n+2}] E\in \widetilde{W_{n+2}}$ is an injection $D:B_k[d_1,\ldots,d_n,0,0]\hookrightarrow B_k[d_1,\ldots,d_n,1,1].$ Let $M$ be a $W_n$-module. We call $v\in M$ *singular* if $\partial_i v=0$ for $i=1,\ldots n$, and we denote by $M^{sing}$ the vector subspace of singular vectors. $M^{sing}$ is a $\gl_n$-submodule of $M$. Let $\mathcal{O}$ denote the category of $W_n$-modules satisfying the conditions of Theorem \[thm:rudakovwn\], and let $\mathcal{O}^\circ$ denote the full subcategory of $\mathcal{O}$ consisting of modules whose Jordan-Hölder series do not contain $\lambda=(1^k)$ for any $k$. The following proposition is a consequence of Theorem \[flambdairrep\], together with the character formula for $\flambda$: \[singprop\] Every $M\in\mathcal{O}^\circ$ is generated by $M^{sing}$ as a $W_n^0$-module. The multiplicity of $\flambda$ in the Jordan-Hölder series of $M$ is equal to the multiplicity of $V_\lambda$ in $M^{sing}$. For $k\geq 3$, and for all $n$, . The operator $D$ preserves the vector subspace $B_k^{sing}$, and moreover maps highest weight vectors to highest weight vectors for the $\gl_n$ action. Let $m_{k,\lambda}$ denote the multiplicity of $V_\lambda$ in $B_k$ (and thus by Proposition \[singprop\], the multiplicity of $\flambda$ in $B_k$). Then \[neweuler\] would imply that $m_{k,\lambda} \leq m_{k,(\lambda,1,1)}$. On the other hand, Conjecture \[glambdabound\] implies that the $m_{k,\lambda}$ are bounded from above, uniformly in $n$ and $\lambda$, since each $\mathcal{G}_\mu$ contains any $\flambda$ (regarded here as a $\gl_n$-module) as a summand at most once. Thus Conjecture \[neweuler\] would imply that the sequence $m_{k,(\lambda,1^{2j})}$ is both non-decreasing, and bounded from above, hence eventually constant, implying Conjecture \[glambdaconj\]. [99]{} N. Arbesfeld, D.Jordan. New results on the lower central series of a free algebra. J. Algebra, to appear. Wieb Bosma, John Cannon, and Catherine Playoust. The Magma algebra system. I. The user language. J. Symbolic Comput., 24(3-4):235-265, 1997 G. Dobrovolska and P. Etingof. An upper bound for the lower central series quotients of a free associative algebra. International Mathematics Research Notices, Vol. 2008, rnn039. G. Dobrovolska, J. Kim, X. Ma. On the lower central series of an associative algebra. Journal of Algebra Volume 320, Issue 1, 1 July 2008, Pages 213-237. P. Etingof, J. Kim, X. Ma. On universal Lie nilpotent associative algebras. arXiv/0805.1909 B. Feigin, D. B. Fuks. Cohomologies of Lie Groups and Lie Algebras, in A.L. Onishchik, E.B. Vinberg (Eds.), *Lie Groups and Lie Algebras II*. Springer, New York 2000. B. Feigin, B. Shoikhet. On $[A,A]/[A,[A,A]]$ and on a $W_{n}$-action on the consecutive commutators of free associative algebras. Math. Res. Lett. 14 (2007), no. 5, 781-795. A. N. Rudakov. Irreducible representations of infinite-dimensional Lie algebras of Cartan type. Math. USSR Izv. Vol. 8, pgs. 836-866.

ł ß Ł \#1[\#1|]{} \#1[|\#1]{} \#1[\#1]{} \#1[\#1|]{} \#1[|\#1]{} \#1[\#1|]{} \#1[|\#1]{} \#1\#2[\#1|\#2]{} \#1\#2[\#1|\#2]{} Introduction ============ Diffusion-limited chemical reactions are known in lower dimensions to exhibit anomalous kinetics [@Kuzovkov; @OTB]. That is, the evolution of the density depends strongly on fluctuations, and cannot be derived from mean-field rate equations. In this paper we apply renormalization group (RG) techniques to the two-species reaction $A+B\goto\inert$, with the goal of determining systematically the effects of these fluctuations. The model for the $A+B\goto\inert$ reaction involves two types of particles, both undergoing diffusive random walks, and reacting upon contact to form an inert particle. In the density rate equation approach it is assumed that the $A$ and $B$ particles densities $a$ and $b$ are uniform, and that reactions occur at a rate proportional to the product $ab$, giving $${da\over dt}={db\over dt}=-\Gamma ab,$$ with rate constant $\Gamma$. In the case of equal initial densities $a(0)=b(0)=n_0$, the solution goes as $a,b\sim (\Gamma t)^{-1}$ asymptotically, with an amplitude which is independent of the initial density. It was first suggested by Ovchinnikov and Zeldovich [@OZ], and later demonstrated by Toussaint and Wilczek [@TW], that relaxing the assumption of uniformity yields a slower density decay. In particular, Toussaint and Wilczek made the observation that if the two species have the same diffusion constants $D_A=D_B=D$, then the density difference $a-b$ obeys the diffusion equation. As a result they found, by using central limit arguments to calculate the fluctuations in $a-b$ due to equal density, random initial conditions, the asymptotic density[^1] $$\label{tw} a,b\sim{\sqrt{n_0}\over\pi^{1/2}(8\pi)^{d/4}}\,(Dt)^{-d/4},$$ where $d$ is the dimension of space. Comparing with the result of the rate equation, we see that for $d<4$ the asymptotically dominant process is the diffusive decay of the fluctuations in the initial conditions. Using a particular version of the model, Bramson and Lebowitz confirmed rigorously the decay exponent of Toussaint and Wilczek, finding for $d<4$ $$a,b\sim C_d\sqrt{n_0}\>t^{-d/4},$$ where $C_d$ is some constant which depends on the dimension $d$ [@BLi; @BLii]. In their treatment they demonstrated that the two species are asymptotically segregated for $d<4$. This segregation was assumed in reference [@TW] in deriving Eq. (\[tw\]). Numerical simulations have confirmed the value $-d/4$ for the decay exponent in one [@TW; @KRi], two [@TW; @CDCii], and three dimensions [@Leyvraz]. For all of these simulations restrictions were placed on the occupation number per site, and usually only single occupancy allowed. In the one-dimensional simulation of Toussaint and Wilczek the initial density was varied, and reasonable agreement was found with their analytic result, Eq. (\[tw\]) [@TW]. However, in higher dimensions the $\sqrt{n_0}$ amplitude dependence, when tested, has not been observed [@CDCii; @Leyvraz]. In the former case the initial average occupation number per site was kept low, whereas for the higher dimensional simulations it was necessary to start with a nearly full lattice in order to reach the asymptotic regime. This suggests that Eq. (\[tw\]) might not be a universal result, but rather a limit for small initial density $n_0$. While $d=4$ appears to be the upper critical dimension for homogeneous initial conditions, this is not the case when the two species are initially segregated, where instead the upper critical dimension is found to be $d=2$ [@CD]. That is, as a result of the segregation, a localized region forms in which nearly all reactions occur. This reaction zone exhibits scaling behavior, and the characteristic exponents are independent of the dimension $d$ when $d>2$, but crossover to dimension-dependent values for $d<2$. Hence, one of our goals in applying RG techniques is to better understand the role of the dimensions $d=2$ and $d=4$. The problem can be mapped to a field theory by starting from a master equation description of the model [@PelitiRev]. From an analysis of the field theory we find that there is an upper critical dimension $d_c=2$, which is associated with the stochastic processes of reaction and diffusion. Hence, for $d>2$ one can replace the full field theory with an effective theory, which is valid for asymptotically late times, while for $d\le 2$ one must instead perform an explicit renormalization group calculation. The effective theory for $d>2$ is equivalent to the deterministic partial differential equations $$\label{pde} \dee_t a=D_A\nabla^2 a-\Gamma ab\qquad\dee_t b=D_B\nabla^2 b-\Gamma ab$$ with stochastic, non-negative [*effective*]{} initial conditions. In deriving the effective theory we find that the initial distribution is finitely renormalized due to the presence of relevant initial terms, the analog of surface terms for a $t=0$ boundary in a $d+1$ dimensional theory. The resulting distribution can be characterized by a parameter $\Delta$ which depends [*nonuniversally*]{} on the initial density. We demonstrate explicitly that from these equations follows generally the asymptotic segregation of the $A$, $B$ particles when $d<4$, and subsequently the universal decay exponent $-d/4$. However, the amplitude of the density decay depends on the initial conditions, and is therefore nonuniversal. It is important to note that if one uses instead central limit arguments to calculate the initial distribution which should be fed into (\[pde\]), then one is implicitly making the assumption that these equations hold for all times, rather than just asymptotically. Such an assumption will get the exponent correct, but we claim that it does not, in general, predict the correct amplitude because it neglects the dynamics at short times. For $2<d<4$ and $D_A=D_B$ we find $$\label{den} \br a,\br b\sim{\sqrt{\D}\over\pi^{1/2}(8\pi)^{d/4}}\,(Dt)^{-d/4},$$ where the angular brackets denote averages over both the processes of reaction and diffusion and the initial conditions. Here $\D$ is the coupling constant of the induced initial terms, and can be calculated as an expansion in the initial density, giving $$\label{Delta} \D=n_0-{(d+2)(d+4)\over 384(8\pi)^{d/2-1}\sin(\pi(d-2)/2)}\l_{\rm eff}^{d/2} n_0^{d/2}+\dots,$$ where $\l_{\rm eff}$ is a nonuniversal effective rate constant, defined in section \[efftheory\] and used in (\[eom\]). Hence, in the small $n_0$ limit the amplitude is universal, and we recover the result of Toussaint and Wilczek, Eq. (\[tw\]). The higher order terms in $n_0$ are nonuniversal, and offer a possible explanation for the deviation from $\sqrt{n_0}$ behavior found in the simulations [@CDCii; @Leyvraz]. Our results (\[den\]) and (\[Delta\]) appear to disagree with those of Bramson and Lebowitz [@BLii], for $d=4$ as well as in the case above. However, we stress that since the dependence of the amplitude on the initial density is nonuniversal there is no explicit contradiction. Our model is defined by a continuous time master equation in which the reaction occurs at a rate $\l$, and multiple occupancy per site of each particle type is allowed. Bramson and Lebowitz also study a continuous time model with multiple occupancy allowed, but with an instantaneous reaction [@BLi; @BLii]. In this case a lattice site can only contain one type of particle. We use a finite reaction rate since this is convenient for mapping to the field theory, and because it allows one to determine better the extent of universality. However, we cannot directly relate our results to those of Bramson and Lebowitz, since the field theory techniques we use are no longer valid in the limit $\l\to\infty$, to which their model corresponds. We note that if our results should be valid for large but finite $\lambda$, then $\lf$, given by (\[Delta\]), goes to a limiting value of the order $h^{d-2}$, where $h$ is the short distance cutoff. For $d\le 2$ the full field theory and the subsequent renormalization must be considered. We find that the field theory may be exactly renormalized, as was shown previously by Peliti for the one-species reactions $A+A\goto A$ and $A+A\goto\inert$ [@Peliti]. However, the $\eps$-expansion calculation of observables requires non-perturbative sums over all orders of the initial density $n_0$ and the parameter $\D$, and while these may be carried out straightforwardly in the one-species reaction [@Lee], we are unable to apply these methods to the present case beyond the leading order in $\eps=2-d$. Thus, at least in this approach, we are not yet able to establish even the power law $t^{-d/4}$ for the density to all orders in $\eps$, although we believe it to be true. We also consider the case of unequal diffusion constants, $D_A\neq D_B$, and show from the effective theory that in the small $n_0$ limit $$\br{a(\d)}\sim\sqrt{Q(d,\d)}\>\br{a(\d=0)}$$ where $\d=(D_A-D_B)/(D_A+D_B)$, and $$Q(d,\d)={4\Bigl[(1+\d)^{2-d/2}+(1-\d)^{2-d/2}-2\Bigr]\over\d^2(d-2)(d-4)}.$$ Therefore this falls into the same universality class, with regard to the decay exponent, as the symmetric case. From the effective field theory for $2<d<4$ it follows that the density difference $a-b$ is at late times a gaussian random field. This, combined with the asymptotic segregation $a+b\sim|a-b|$ allows one to calculate any correlation function. We calculate exactly the equal time two-point correlation functions $\br{a(r)a(0)}$ and $\br{a(r)b(0)}$. The final topic we discuss is that of reaction zones, which form whenever $A$ and $B$ particles are segregated. One example of a reaction zone is that which results from opposing currents of $A$ and $B$ particles. We apply RG methods to this steady-state case, and show that the densities and the rate of reaction have universal scaling forms. The upper critical dimension for this system is $d_c=2$. These results can be extended to apply to reaction zones in initially segregated systems [@CD; @GR; @BR], and also homogeneous systems for $d<4$ [@LC]. The Model and the Corresponding Field Theory {#model} ============================================ The model is defined by a continuous time master equation for the probability $P(\{m\},\{n\},t)$. The set $\{m\}$ denotes the occupation numbers of $A$ particles on each lattice site, $\{n\}$ the occupation numbers of $B$ particles, and $P$ is the probability of a given configuration occurring at time $t$. The master equation for $P$ reads $$\begin{aligned} \label{me} \pt&& P(\{m\},\{n\},t)=\nonumber\\ &&{\ds D_A\over \ds h^2}\sum_{i,j}\biggl\{(m_j+1) P(\dots,m_i-1,m_j+1\dots,t)-m_iP\biggr\}\nonumber\\ +&&{\ds D_B\over \ds h^2}\sum_{<ij>}\biggl\{(n_j+1)P(\dots,n_i-1,n_j+1 \dots,t)-n_iP\biggr\}\nonumber\\ +&&\l\sum_i\biggl\{(m_i+1)(n_i+1)P(m_i+1,n_i+1,t)-m_in_iP\biggr\},\nonumber\\\end{aligned}$$ where $D_A$, $D_B$ are the diffusion constants for $A$ and $B$ particles, $h$ is the size of the hypercubic lattice, and $\l$ is the microscopic reaction rate constant. In the first two curly bracket terms, which describe the diffusion of $A$ and $B$ particles, $i$ is summed over all sites, and $j$ is summed over nearest neighbors to $i$. The initial conditions for $P$ are given by a Poissonian distribution, with the average occupation number per lattice site equal to $\bar n_0$ for each species. That is, $$\label{ic} P(\{m\},\{n\},0)=e^{-2n_0}\prod_i{\bar n_0^{m_i+n_i}\over m_i!\>n_i!}.$$ Mapping to Field Theory ----------------------- The first step in mapping the master equation to a field theory is to recast it in a ‘second quantized’ form, following a procedure developed by Doi [@Doi]. Two sets of creation and annihilation operators—$\ah$, $\ad$ for $A$ particles and $\bh$, $\bd$ for $B$ particles—are introduced at each lattice site. These obey the usual commutation relations: $$[\ah_i,\ad_j]=[\bh_i,\bd_j]=\d_{ij},$$ with all other commutators zero. The vacuum ket is defined by $\ah_i\vac=0$ and $\bh_i\vac=0$ for all $i$. In terms of these operators the state of the system at time $t$ is defined to be $$\ket{\phi(t)}=\sum_{\{m\},\{n\}}P(\{m\},\{n\},t)\prod_i\bigl( \ad_i\bigr)^{m_i}\bigl(\bd_i\bigr)^{n_i}\vac.$$ The master equation can be rewritten in terms of this state as $$\label{sqme} -\pt\ket{\phi(t)}=\hat H\ket{\phi(t)},$$ with the operator $$\begin{aligned} \label{defH} \hat H=&&\sum_{<ij>}\biggl\{{D_A\over h^2}(\ad_j-\ad_i)(\ah_j-\ah_i)+ {D_B\over h^2}(\bd_j-\bd_i)(\bh_j-\bh_i)\biggr\}\nonumber\\ &&+\l\sum_i(\ad_i\bd_i-1)\ah_i\bh_i.\end{aligned}$$ The formal solution of Eq. (\[sqme\]) is $$\label{sqsol} \ket{\phi(t)}=e^{-\hat H t}\ket{\phi(0)}.$$ The density and other averages, which are defined in the original occupation number representation, can be calculated from $\ket{\phi(t)}$ by introducing the projection state $$\label{proj} \stnd=\bra{0}\prod_ie^{\ah_i+\bh_i},$$ in terms of which the average is $$\begin{aligned} \label{average} \dl A(t)\dr&\equiv&\sum_{\{m\},\{n\}}A\Bigl(\{m\},\{n\}\Bigr)\> P\Bigl(\{m\},\{n\},t\Bigr)\nonumber\\ &=&\stnd\>\hat A\>e^{-\hat Ht}\ket{\phi(0)}.\end{aligned}$$ The operator analog $\hat A$ can be derived for any $A(\{m\},\{n\})$ by Taylor expanding the latter with respect to $m_i,n_i$, and then substituting $m_i\goto\ad_i\ah_i$, $n_i\goto\bd_i\bh_i$. Note that $$\label{projev} \stnd\ad_i=\stnd\bd_i=\stnd,$$ for all $i$, implying that $\hat A$ can be expressed solely in terms of annihilation operators by first writing it in normal ordered form. The second quantized version of the model is mapped to a field theory by the use of the coherent state representation [@PelitiRev; @Schulman]. The time evolution operator in Eq.(\[sqsol\]) is rewritten via the Trotter formula $$\label{trotter} \exp(-\hat Ht)=\lim_{\D t\goto 0}(1-\hat H\D t)^{t/\D t}.$$ The right-hand side, before the $\D t\goto 0$ limit is taken, can be regarded as a factorization of the operator into time slices, and a complete set states inserted between each factor. The identity is given in the coherent state basis by $$\label{identity} {\bf 1}=\int{d^2 z\over\pi}\ket z\bra z$$ where $\ket z$ is the normalized eigenstate of the annihilation operator with complex eigenvalue $z$: $$\ket z=e^{z\ad-|z|^2/2}\vac.$$ Equation (\[identity\]) is generalized to a product over all lattice sites and particle species, and then the operator $(1-\hat H\D t)$ is evaluated between successive time slices, resulting as $\D t\goto 0$ in a path integral representation of (\[average\]). The corresponding action is $$\begin{aligned} \label{Sab} S=\int d^dx\biggl[&&\int_0^{t_f}dt\biggl\{a^*\Bigl(\dee_t- {D_A\over\bar D}\nabla^2\Bigr)a\nonumber\\ \mbox{}&&+b^*\Bigl(\dee_t-{D_B\over\bar D}\nabla^2 \Bigr)b-\lo(1-a^*b^*)ab\biggr\}\nonumber\\ &&-n_0a^*(0)-n_0b^*(0)-a(t_f)-b(t_f)\biggr].\end{aligned}$$ where the fields $a,a^*, b, b^*$ originate from the complex variables $z$ for each particle type. Time has been rescaled by the average diffusion constant , and the coupling constant is given by $\lo=\l h^d/\bar D$. The time derivatives above come from the overlap between time slices, and the other the curly brace terms result from the operator $\hat H$. The remaining terms are not integrated over time and represent the random initial state, with $n_0\equiv\bar n_0/h^d$, and the final projection state (\[proj\]). Averages, as defined in (\[average\]), are given in terms of this action by $$\label{FTavg} \dl A(t)\dr={\int{\cal D}(a,a^*,b,b^*)\>A[a(t),b(t)]\>e^{-S}\over \int{\cal D}(a,a^*,b,b^*)\>e^{-S}}.$$ where the script ${\cal D}$ denotes functional integration. The functional $A[a,b]$ is found by directly substituting the fields $a,b$ for the annihilation operators $\ah,\bh$ in $\hat A$. The time $t_f$ of the projection state is arbitrary as long $t_f>t$, where $t$ is the time argument of the observable. This follows directly from the condition $\stnd\hat H=0$ for probability conservation. The final terms can be eliminated by making the field shifts $a^*=1+\ba$ and $b^*=1+\bb$. Then the reaction terms are $$-\lo(1-a^*b^*)ab\quad\goto\quad\lo(\ba+\bb)ab+\lo\ba\bb ab.$$ Since the conserved mode $a-b$ plays an important role in the dynamics, it is useful to transform (\[Sab\]) to the fields $\phi,\bphi,\psi,\bpsi$ defined by $$\phi={a+b\over\sqrt 2}\qquad\bphi={\ba+\bb\over\sqrt 2} \qquad\psi={a-b\over\sqrt 2}\qquad\bpsi={\ba-\bb\over\sqrt 2}.$$ The $\sqrt 2$ factors are included so that the derivative terms in (\[Sab\]) maintain a coefficient of unity. The subsequent action is $$\begin{aligned} \label{S} S=\int &&d^dx\biggl[\int dt \biggl\{\bphi(\dee_t-\nabla^2)\phi+\bpsi(\dee_t-\nabla^2)\psi -\d\bpsi\nabla^2\phi\nonumber\\ -&&\d\bphi\nabla^2\psi-\l_1\bphi(\phi^2-\psi^2) -\l_2(\bphi^2-\bpsi^2)(\phi^2-\psi^2)\biggr\}\nonumber\\ &&-n_{\phi}\bphi(0)\biggr],\end{aligned}$$ where $\d=(D_A-D_B)/(D_A+D_B)$, the couplings are $\l_1=\lo/\sqrt 2$ and , and the initial density is $n_{\phi}=\sqrt 2n_0$. This action is the starting point for our analysis. Note that, since we are considering only equal initial densities, $\br a=\br b=\br\phi/\sqrt 2$. The mapping outlined above is a general technique, which may be applied to many different reactions, for example, the general two-species annihilation reaction $mA+nB\rightarrow\inert$. However, in the present work we restrict ourselves to the case $m=n=1$. Diagrams and Power Counting {#power} --------------------------- A perturbation expansion for a given observable can be developed from (\[FTavg\]) and (\[S\]), and expressed in the usual diagrammatic fashion. The propagators for $\phi,\psi$ are given by the first two terms in (\[S\]), and are the diffusion equation Green’s function: $G_{\phi\bphi}(k,t)=G_{\psi\bpsi}(k,t)=e^{-k^2t}$ when , and $G_{\phi\bphi}=G_{\psi\bpsi}=0$ for . These propagators are represented by solid and dashed lines respectively. The three- and four-point vertices, which correspond to the annihilation reaction, are shown in Fig. \[verts\]. When there are also two-point vertices which connect a $\phi$ propagator to a $\psi$ propagator, and vice versa. These vertices are wave number dependent, with magnitude $k^2$. .2in .15in In addition there is a source term $e^{\np\bphi(t=0)}$. By Taylor expanding the exponential an expansion in powers of $\np$ is generated, where the diagrams giving the $\np^i$ coefficient have a source of $i$ $\phi$ propagators at $t=0$. It is useful to introduce the classical density and the classical response function, which both involve sums over all powers of $\np$. These quantities are important because it is found that under renormalization $\np$ flows to a strong coupling limit, and these sums are still meaningful in this limit [@Lee]. The term ‘classical’ refers to absence of loops in the diagrams. The classical density is defined to be the sum of all tree diagrams which contribute to the average $\br\phi$, as shown in Fig. \[treesum\]. Note that these diagrams contain only $\phi$ propagators, because of the three-point vertices in (\[S\]). This sum obeys an integral equation which can be solved exactly, giving $$\br\phi_{\rm cl}={\np\over 1+\np\l_1 t}.$$ .2in .15in The classical response function is defined to be the $\phi$ propagator with all possible tree diagrams branching off to $t=0$, as shown in Fig. \[response\]. Again this can be solved exactly, giving $$\label{resp} \br{\phi({\bf k},t_2)\bphi(-{\bf k},t_1)}_{\rm cl}=e^{-k^2(t_2-t_1)} \left(1+\np\l_1t_1\over 1+\np\l_1t_2\right)^2.$$ .2in .15in In order to renormalize the field theory we must first determine the primitive divergences, for which we consider the following dimensional analysis. There is a rigid constraint that $[\bphi\phi]=[\bpsi\psi]=k^d$, where $k$ has the dimensions of wave number. If we take the dimensions of the conjugate fields to be $[\bphi]=[\bpsi]=k^0$, as was done for the one-species reaction [@Lee], then a general vertex $\bpsi^i\bphi^j\psi^k\phi^\ell$ is found to be relevant only for $k+\ell\le 2$ and $d\le 2$. Next, we observe that it is not possible to generate any vertices with $k=1$ or $\ell=1$ from the vertices in (\[S\]). Therefore all relevant vertices are exactly those present in (\[S\]), with an upper critical dimension $d_c=2$. We will discuss the renormalization of the theory in section \[renormab\], and for now focus on the case for $d>2$. To elucidate the crossover which occurs for $d>2$ it is useful to consider rescaling the fields by dimensionful parameters, which is consistent as long as the conjugate fields are rescaled accordingly. Under such a rescaling the couplings $\l_1$ and $\l_2$ behave differently, although originally they are both proportional to $\lo$. In particular, we can take $$\label{rsc} \phi\goto\phi/\l_1\qquad\bphi\goto\l_1\bphi\qquad \psi\goto\psi/\l_1\qquad\bpsi\goto\l_1\bpsi,$$ which has the result of setting the $\bphi(\phi^2-\psi^2)$ coupling to unity, while leaving $\l_2$ unchanged. The rescaling (\[rsc\]) also results in $\np\goto\l_1\np$. This is the proper quantity to study when addressing issues of relevance and irrelevance, which can be seen by studying the diagrams generated by the action (\[S\]): whenever an additional $t=0$ line is added with weight $\np$, there is an additional $\l_1$ required to connect it. In this system of units, then, one finds that $$[\l_2]=k^{2-d}\qquad[\l_1\np]=k^2\qquad[\d]=k^0.$$ Therefore there exists a critical dimension $d_c=2$, above which $\l_2$ flows to zero. Doing the complete power counting method with the rescaled fields yields the same result. The initial density is a strongly relevant parameter for all $d$. The diffusion constant difference $\d$ is always marginal whenever it is not zero. Before turning to the consequences of the power counting for the case $d>2$, we mention another approach to this problem, which is to integrate out the conjugate fields $\bpsi$ and $\bphi$ in (\[S\]). This leads to the equations of motion (for $\d=0$) $$\label{phin} \pt\phi=\nabla^2\phi-\l_1\phi^2+\l_1\psi^2+\eta_\phi$$ $$\label{psin} \pt\psi=\nabla^2\psi+\eta_\psi,$$ where $\eta_\phi,\eta_\psi$ are multiplicative, complex noise terms [@Howard]. It is important to note that the physical density is not the field $\phi$, but rather the average of $\phi$ over the noise terms. These equations, without the noise terms included, are often taken as the starting point for analysis, but this approach is not generally valid. As we will show in the next section, one can neglect the noise terms only for $d>2$ and for asymptotically large times. Equation (\[psin\]) can be simplified in any dimension, since it is a linear equation, by averaging over the noise. This is an average over the stochastic process of diffusion, and not over the initial conditions. Then the averaged field $\br\psi$ obeys the simple diffusion equation for any given initial configuration. Effective Field Theory for $d>2$ {#efftheory} -------------------------------- From the dimensional analysis and power counting above it follows that for $d>2$ the full theory given by (\[S\]) can be replaced by an effective theory in which $\l_2=0$ and $\l_1\goto\l_{\rm eff}(\l_1,\l_2,\L)$, where $\L$ is a wave number cutoff, of the order of the inverse lattice spacing. However, in constructing such an effective theory one has to consider all possible relevant terms, consistent with the symmetry of the theory, which might be generated under renormalization. In order to identify these terms we note that this problem is analogous to that of a semi-infinite system in equilibrium statistical mechanics in $d+1$ dimensions, the analog of the boundary being the hyperplane $t=0$. While one finds, in the semi-infinite equilibrium case, that the bulk critical properties do not depend on the surface terms, nonetheless one expects surface terms to contribute to correlation functions which involve fields on the boundary [@Diehl]. All observables in our problem are given by such correlation functions, since all diagrams originate with the $\np\bphi(0)$ term. Therefore we must check for all relevant [*initial*]{} terms, the $t=0$ analog of the surface terms, which might be generated, as well as those of the bulk. As mentioned above, the only relevant bulk term is that of $\l_1$. The proper framework for determining which terms are relevant is via the rescaled fields (\[rsc\]). Therefore, for an initial term of the type $(\D^{(m,n)}/m!n!)\bphi^m\bpsi^n\vert_{t=0}$ added to (\[S\]) we consider the dimensions of the coupling $[\l_1^{m+n}\D^{(m,n)}]=k^{(n+m)(2-d)+d}$. This power of $\l_1$ also follows from calculating the number of vertices required to attach a $t=0$ vertex of $\D^{(m,n)}$ to a given diagram. These terms are relevant when $$\label{dnm} d<{2(n+m)\over n+m-1}.$$ If $m+n=1$ then the initial term is relevant for all $d$. The case corresponds to the initial density, which has already been demonstrated to be relevant. For the case $n=1$ we first address a symmetry of the theory. When starting with equal initial densities the system is invariant under exchanging $A\exch B$ and $D_A\exch D_B$. Therefore the action must be invariant under the transformation $(\phi,\bphi,\psi,\bpsi,\d)\goto(\phi,\bphi,-\psi,-\bpsi,-\d)$. For what follows we will consider only the case $\d=0$, or , in which case the symmetry forbids the generation of a initial term $\D^{(0,1)}\bpsi$. In section \[uneqdiff\] the case $\d\ne 0$ will be discussed, and it will be demonstrated that again no $n=1$ initial term is generated. For $m+n=2$ symmetry allows only the generation of $\D^{(2,0)}$ and $\D^{(0,2)}$. Below we will address the calculation of these quantities, and demonstrate that . These terms are relevant whenever , as can be seen by equation (\[dnm\]), and therefore must be considered when constructing an effective theory for . In fact, it will be shown that the term $(\D/2)\bpsi^2$ is solely responsible for determining the asymptotic decay of the density. This is an important point. This system is dominated by initial terms, as opposed to the one-species reaction. Therefore techniques which utilize homogeneous source terms and look for a bulk steady state will not work for this problem. Since this initial term dominates the asymptotic behavior of the density, we identify $d_c^*=4$ as a second critical dimension of the system. Higher order initial terms will also be relevant in the range $2<d\le 3$. In fact, as $d\goto 2$ one finds that all initial terms become relevant. While this seems to be an extreme complication, it is in fact possible to calculate exactly the asymptotic density for $2<d\le 4$ and demonstrate that it is independent of such terms. This will be presented in the next section. We now turn to the calculation of the parameter $\D$. .2in .15in The diagrams which must be considered in calculating an effective initial term $(\D/2)\bpsi^2$ are all those in which two $\psi$ lines exit from the left, as shown in Fig. \[surface\](a). The sum of these diagrams gives rise to an effective term $f(t)\bpsi(t)^2$ in the action. If the function $f(t)$ goes to zero for large $t$, and is sharply peaked enough that $\int_0^\infty dtf(t)$ is finite, then a coarse-graining in time gives $f(t)\bpsi(t)^2\to(\D/2)\d(t)\bpsi(0)^2$, where both quantities are understood to be integrated over $t$, and $\D=2\int_0^\infty d\tp f(\tp)$. To calculate this parameter $\D$ we consider first the subset of diagrams given by the tree diagrams, as shown in Fig. \[surface\](b). These diagrams sum to give $f_0(t)=-\l_2/(1+\np\l_1 t)^2$, and so $$\D_0=-2\l_2\int_0^\infty dt{1\over (1+\np\l_1 t)^2}=2\np{\l_2\over\l_1}.$$ Therefore we conclude that this set of diagrams generates an effective initial term $\D_0=2\np\l_2/\l_1$, or, in terms of the parameters in the master equation (\[me\]), $\D_0=n_0$, the initial density of each species. This will be shown to be the leading order term for a small $n_0$ expansion of $\D$. The width of the function $f_0(t)$ is given by $(\np\l_1)^{-1}$, and therefore we expect this coarse-grained picture to be valid for times $t\gg (\np\l_1)^{-1}$. We can group all the diagrams in the full theory (\[S\]) which are of the form specified in Fig. \[surface\](a) in the following way. There is a vertex $\l_2$ which is the leftmost vertex in the diagram. The lines coming into this vertex from the right can either come from mutually distinct or connected diagrams. The tree diagrams are a subset of the former group, and we argue that by letting $\l_1$ go to some bulk effective coupling $\lf$ all diagrams of the former group are included. The connected diagrams can be grouped by the number of times they are connected, and shown in Fig. \[surface\](c) are a set of diagrams which are connected exactly once. Again we argue that by taking $\l_1\goto\lf$ the diagrams of Fig. \[surface\](c) give the entire contribution of the set which are connected exactly once. The sum of these diagrams is evaluated in the appendix, and is found to contribute to $\D$ a term which is higher order in $n_0$ than that given by the tree diagrams. It can be shown in general that the groups with more connections will contribute correspondingly higher order terms, and therefore this classification scheme gives rise to an expansion for $\D$. Of course, an almost identical mechanism will also generate a term $\D^{(2,0)}\bar\phi(0)^2$, and it is straightforward to show on the grounds of symmetry that $\D^{(2,0)}=-\D$. Although this initial term is equally relevant from the renormalization group point of view, nevertheless it does not affect the late time behavior of the density. This is because it acts as a source for late time fluctuations only through the response function (\[resp\]), and this is strongly damped for $t_2\gg t_1$. In contrast the response function of the $\psi$ field is simply the diffusion propagator, which has no such damping. In summary of the discussion above, for $2<d\le 4$ and for large times one can replace the full theory with a simplified action $$\begin{aligned} \label{Sf} S=\int d^dx\biggl[&&\int_0^t dt \Bigl\{\bphi(\dee_t-\nabla^2)\phi+\bpsi(\dee_t-\nabla^2)\psi\nonumber\\ &&-\l_{\rm eff}\bphi(\phi^2-\psi^2)\Bigr\}-n_{\phi}\bphi(0)-{\D\over 2} \bpsi(0)^2\nonumber\\ &&+\hbox{other initial terms}\biggr],\end{aligned}$$ where $\D$ is given by (\[Delta\]). Since the bulk theory, the terms within curly braces, is linear in $\bphi$ and $\bpsi$ these fields can be integrated out to yield the equations of motion $$\label{eom} \pt\phi=\nabla^2\phi-\l_{\rm eff}\phi^2+\l_{\rm eff}\psi^2$$ $$\label{eomii} \pt\psi=\nabla^2\psi.$$ These are equations for classical fields with fluctuations in the initial conditions. They are often taken to be the continuum limit of the master equation (\[me\]), but we stress that only for $d>2$ and large times are these equations valid. In addition, since $\l_{\rm eff}<\l$, it is never correct to say that $\langle ab\rangle\sim\langle a\rangle\langle b\rangle$, but only that they are effectively proportional. Density Calculation for Equal Diffusion Constants ================================================= Effective action: $2<d<4$ {#classden} ------------------------- Starting with the action (\[Sf\]) one can calculate exactly the leading time dependence of the density, as well as correlation functions. We begin with a comment about notation. For this section and the next, where we deal with only the effective field theory, averages over the initial conditions will be denoted by angular brackets. The classical fields $\phi,\psi$ themselves represent bulk averages, or equivalently, averages over reaction and diffusion, of the same fields as written in (\[S\]). Also, the effective coupling is abbreviated to be $\l=\lf$. With this notation, then, the average of equation (\[eom\]) over the translationally invariant initial conditions is $$\label{eomc} {d\over dt}\br\phi=-\l\br{\phi^2}+\l\br{\psi^2},$$ since $\nabla^2\br{\phi}=0$. .2in .15in A diagrammatic expansion for $\br\phi$ is shown in Fig. \[densum\]. Operating on both sides of this expansion with $(\dee_t-\nabla^2)$, the inverse of the Green’s function propagator, gives equation (\[eomc\]). At this point, knowing that $\D$ is relevant for $d<4$, one might attempt to apply the renormalization group to try to find a nontrivial fixed point of order $4-d$. However, no such fixed point exists. This is because there are no corrections to a correlation function $\langle\psi(x_1,t_1)\psi(x_2,t_2)\bar\psi(0)^2\rangle$, so that $\D$ is not renormalized. It therefore flows, for $d<4$, to infinity under renormalization. This would appear to make it very difficult to sum the diagrams in Fig. \[densum\] explicitly. However, it turns out to be possible to solve (\[eomc\]) exactly for late times. There is only one diagram contributing to the value of $\br{\psi^2}$ in equation (\[eomc\]), which is the single $\psi$ loop. Evaluating this loop gives $\br{\psi^2}=\D/(8\pi t)^{d/2}$. It is important to note that this result holds even when all possible higher order initial terms are included. Next, consider a related problem in which $\br{\phi^2}$ in (\[eomc\]) is replaced by $\br\phi^2$, which is equivalent to including only the diagrams in Fig. \[densum\] which are disconnected to the right of the leftmost vertex. This partial sum satisfies a differential equation known as Ricatti’s equation, which, though non-linear, can be solved.[^2] Let $f$ denote the function which satisfies this equation, that is, $$\label{ricatti} {d\over dt}f=-\l f^2+\l{\D\over(8\pi)^{d/2}} t^{-d/2}.$$ It will be shown below that this function $f$ provides a upper bound for the actual density, but first we will discuss the solution of this equation. It is integrable for certain values of $d$, specifically $d=4$ and $d=4\pm 4/(2s+1)$ where $s$ is a non-negative integer. For general values of $d$ a solution can be obtained by transforming the equation via the substitution $f=\dot u/(\l u)$, which gives $$\ddot u={\l^2\D\over (8\pi)^{d/2}}t^{-d/2}u,$$ a linear, second order equation whose solution can be expressed in terms of confluent hypergeometric functions. Therefore the asymptotic behavior of $f$ is rigorously obtained, and is in fact what one naively obtains by assuming $f\sim At^{-\a}$ and inserting it into (\[ricatti\]): $$\label{UBi} f\sim\cases{\D^{1/2}(8\pi t)^{-d/4} & $d<4$\cr A_4^u\>t^{-1} & $d=4$ \cr (\l t)^{-1} & $d>4$. \cr}$$ When $d<4$ the asymptotic behavior comes from balancing the two terms on the right hand side of (\[ricatti\]), whereas for $d>4$ it comes from balancing the $f^2$ and the $\dot f$ terms. For $d=4$ all three terms contribute, and the amplitude is $$\label{UBii} A_4^u={1\over 2\l}+\sqrt{{1\over(2\l)^2}+{\D\over(8\pi)^2}}.$$ The case of $d=4$ will be discussed in more detail in section \[dgefour\]. Notice that the asymptotic behavior of the solution $f$ is independent of the initial conditions. In fact, the initial conditions must be specified at some $t_0>0$, since the equation is singular at $t=0$. A natural choice for this initial time is that given by the coarse-graining time scale of the effective initial conditions, that is $t_0=(\np\l)^{-1}$. Now we show that $f$ provides an upper bound for the actual density $\br\phi$. Our method is to derive an equation for $\chi=f-\br\phi$ and show that asymptotically $\chi\ge 0$. Since $\phi$ is a real field in the effective theory, then $h(t)\equiv\br{\phi^2}-\br\phi^2\ge 0$. Equation (\[eomc\]) can be rewritten $${d\over dt}\br\phi=-\l h(t)-\l\br\phi^2+\l\br{\psi^2},$$ and then substituting $\br\phi=f-\chi$ gives $$\label{UBc} {d\over dt}\chi=\l h+\l(\chi-2f)\chi.$$ Assume that $\chi(t_0)=0$, that is we choose the initial condition for $f$ such that $f(t_0)=\br{\phi(t_0)}$. As mentioned above, the asymptotic value of $f$ is independent of the choice of initial conditions. Since $f$ is known to be positive for all $t>t_0$, then from equation (\[UBc\]) we know that $\dot\chi>0$ whenever $\chi<0$. Now suppose that there exists some $t_1>t_0$ for which $\chi(t_1)<0$. Since $\chi(t)$ is a continuous function, it follows that there must be some intermediate time $t_0<t<t_1$ for which $\chi(t)<0$ and $\dot\chi(t)<0$. This is in contradiction with equation (\[UBc\]), and therefore our assumption that there exists $\chi(t_1)<0$ for $t_1>t_0$ is false. We can also find a lower bound for $\br\phi$ by noting that $\phi({\bf x},t) \ge|\psi({\bf x},t)|$ at all points $({\bf x},t)$. This is equivalent to the statement that $a({\bf x},t), b({\bf x},t)$ are at all points non-negative, when starting from any initial condition in which $a, b$ are everywhere non-negative. While this result is somewhat intuitive, it can be made more rigorous by considering the field equations (\[eom\]), (\[eomii\]) expressed in terms of $a=(\phi+\psi)/\sqrt 2$ and $b=(\phi-\psi)/\sqrt 2$: $$\label{gx} \pt a=\nabla^2a-\sqrt 2\l ab\qquad\qquad \pt b=\nabla^2b-\sqrt 2\l ab.$$ Given that the fields $a,b$ are initially everywhere non-negative, then for the fields to have a negative value at a later time $t_1$ there must be an intermediate time $0<t_0<t_1$ for which both $a(t_0)=0$ and $\dee_t a(t_0)<0$. However, in the case where $a=0$ at a single point in space, then $a>0$ locally around the point, implying that it is a local minimum and $\dee_t a>0$. For a region of $a=0$ equation (\[gx\]) gives $\dee_t a=0$ in the region and $\dee_t a>0$ on the boundary. Therefore the fields cannot pass through zero, and will remain non-negative. Since $\phi\ge|\psi|$ it follows that $\br\phi\ge\br{|\psi|}$. At late times $\psi$ has a normal distribution, independent of the initial distribution, which follows from the fact that $\psi$ obeys the simple diffusion equation (\[eomii\]). Therefore the asymptotic value of $\br{|\psi|}$ can be computed directly. The asymptotic distribution of $\psi$ is given by $$P[\psi(t)]\propto\exp\left\{-{\psi(t)^2\over 2\br{\psi(t)^2}}\right\},$$ from which it follows that $$\label{psi} \br{|\psi(t)|}=\sqrt{{2\over\pi}\br{\psi(t)^2}}={(2\D)^{1/2}\over \pi^{1/2}(8\pi)^{d/4}}t^{-d/4}.$$ Given the upper bound $\langle\phi\rangle\leq f\sim O(t^{-d/4})$ it can be shown that $\br\phi\sim\br{|\psi|}$, that is, that the lower bound gives exactly the density. Since $\br{g^2}\ge\br g^2$ for any real $g$, then $$\br{\phi-|\psi|}^2\le\br{(\phi-|\psi|)^2} =\br{\phi^2}+\br{\psi^2} -2\br{\phi|\psi|}$$ Using again $\phi\ge|\psi|$: $$\label{UBd} \br{\phi-|\psi|}^2\le\br{\phi^2}-\br{\psi^2}= -{1\over\l}\br{\dot\phi} =O(t^{-1-d/4}).$$ Therefore $\br\phi=\br{|\psi|}+O(t^{-1/2-d/8})$, which gives $\br\phi\sim\br{|\psi|}$ for $d<4$. This is actually a statement about segregation in the system, implying that to leading order the density of $a+b$ is the same as $|a-b|$, or equivalently, that the minority species in each region decays faster than the majority. For $2<d<4$ then, we find $$\label{ans} \br a\sim{\D^{1/2}\over\pi^{1/2}(8\pi)^{d/4}}t^{-d/4},$$ as stated in section I, with $\D$ given by (\[Delta\]). Substituting the leading order term in the expansion $\D=n_0+O(n_0^{d/2})$ then gives the result of Toussaint and Wilczek [@TW]. In fact, our method is very similar to theirs, with two exceptions. First, they use a central limit argument to calculate $\D$, whereas we can compute it directly from the full field theory. It is reassuring that the answers agree, to leading order in $n_0$. The other difference is that they calculate $\br{|\psi|}$, and then hypothesize the asymptotic segregation, saying $\br\phi\sim\br{|\psi|}$. Starting from the effective theory (\[Sf\]) we have shown rigorously that these quantities are asymptotically the same. Effective Action: $d\ge 4$ {#dgefour} -------------------------- When $d=4$ the upper and lower bounds for the density from section \[classden\] still hold: $\br{|\psi|}\le\br\phi\le f$. However, it is no longer necessarily true that , since the bound on the corrections, which is of order $O(t^{-1/2-d/8})$, is the same order as the density. The upper bound $\br\phi\leq f\sim A_4^u/t$ is given by (\[UBi\]) and (\[UBii\]). Notice that for small $\l$ or small $\D$ that $A_4^u\goto 1/\l$. Also, when $\l$ is large or $\D$ is large then $A_4^u\goto\D^{1/2}/(8\pi)$. However, in the intermediate region there is a smooth crossover in the upper bound from the $\l$ dependent asymptote to the $\D$ dependent asymptote. The lower bound is given by $\br\phi\ge\br{|\psi|}=A_4^\ell/t$ with $A_4^\ell=\D^{1/2}/8\pi^{3/2}$. For large $\D$, then, the upper and lower bounds differ by a factor of $\sqrt\pi$. The lower bound continues to decrease with $\D$, and therefore is not very useful in the small $\D$ limit. However, since the parameter $\D$ is dimensionless in $d=4$ one can do a perturbative expansion for small $\D$, which results in a better lower bound. It follows from equation (\[eomc\]) that the zeroth order term in this expansion is a constant, and is in fact equal to the small $\D$ limit of the upper bound, $\l^{-1}$. To the next order one has $$\br a={1\over\l t}+{\l\D\over t}+O(\D^2),$$ and it is plausible to conjecture that the amplitude is monotonically increasing with $\D$. The amplitude given by Bramson and Lebowitz [@BLii], has the form $$A_4\propto\cases{\hbox{constant}&$\D<\D_c$\cr\D^{1/2}&$\D>\D_c$,\cr}$$ that is, the amplitude is independent of $\D$ for small $\D$. Their result seems to be at odds with our small $\D$ calculation. However, as discussed in section \[model\], there are differences between our model and the one they study. Since the corrections to the small $n_0$ or $\D$ limit are non-universal, this is a possible explanation of the discrepancy. When $d>4$ then it follows from the power counting of section \[efftheory\] that the $(\D/2)\bpsi^2$ initial term is irrelevant. In this case the density is given asymptotically by $\br a\sim(\l t)^{-1}$. The power law of the density decay is independent of the dimension of space. The amplitude $\l^{-1}$ will depend on the dimension and the microscopic details, but it is independent of initial terms, or equivalently initial conditions. Renormalization for $\lowercase{d}\le 2$ {#renormab} ---------------------------------------- When $d\le 2$ one has to consider the full theory as given by the action (\[S\]), and the subsequent renormalization. Although (\[eomc\]) is still valid formally, since the noise in (\[phin\]) averages to zero, we can no longer apply the upper and lower bounds of the previous section since, in the presence of the imaginary noise term, $\phi$ is no longer real. Much of the contents of this section is directly related to the one-species calculation of Ref. [@Lee], in which more details can be found. The primitively divergent vertex functions were identified by power counting in section \[power\], and were found to be those with two lines coming in and two or fewer lines going out. These primitive divergences are used to define a renormalized coupling, following conventional RG methods [@Amit]. It is found that all the vertices in the action (\[S\]) renormalize identically, with the primitive divergences given by the bubble sums shown in Fig. \[bubbles\]. .2in .15in In this sum all diagrams of a given number of loops come in with the same sign, since replacing a $\phi$ loop with a $\psi$ loop, for example, introduces always two negative signs (see Fig. \[verts\]). At the order of $n$ loops there are $2^n$ diagrams, so these form a geometric sum, with the ratio given by $2$ times the value of a single loop. Denoting this sum by $\l_i(k,t)$ where $i=1,2$ labels the number of outgoing lines, then the Laplace transform, $\l_i(k,s)=\int_0^\infty dt\>e^{-st}\l_i(k,t)$ is given by $$\l_i(k,s)={\l_i\over 1+[4/(8\pi)^{d/2}]\,\l_2\,\Gamma(\eps/2)(s+k^2/2)},$$ where $\eps=2-d$. The renormalized coupling is defined in terms of an arbitrary normalization scale $\k$, which has dimensions of wave number: $g_R=\k^{-\eps} \l_2(k=0,s=\k^2)$. Then the $\beta$ function is $$\beta(g_R)=\k{\dee\over\dee\k}g_R=-\eps g_R+{4\eps\over(8\pi)^{d/2}} \Gamma\left(\eps\over 2\right)g_R^2$$ which gives a fixed point $g_R^*=O(\eps)$. Let the density $n(t)=\br{a(t)}=\br{b(t)}$. Since the density is independent of the normalization scale, then $dn/d\k=0$, which leads to the Callan-Symanzik equation $$\begin{aligned} \biggl[2t\pt-&&d\np{\dee\over\dee\np}-d\D{\dee\over\dee\D} +\beta(g_R){\dee\over\dee g_R}+d\biggr]n(t,g_R,\np,\D)\nonumber\\ &&=0.\end{aligned}$$ The solution is found by the method of characteristics to be $$\label{CSsol} n(t,g_R,\np,\D)=(\k^2t)^{-d/2}n(\k^{-2},\tilde g_R,\tilde\np,\tilde\D),$$ where in the asymptotic limit of large $\k^2t$ the running coupling has the limit $\tilde g_R\rightarrow g_R^*$. However, the running initial couplings go as $\tilde\np=\np(\k^2t)^{d/2}$ and $\tilde\D=\D(\k^2t)^{d/2}$, that is, they flow to a strong coupling limit. The solution (\[CSsol\]) is used to calculate the asymptotic density in the following way: the density is calculated as an expansion in $g_R$ and $\np$, and this expansion is put into the right-hand side of (\[CSsol\]). Then, in the limit of large $\k^2t$, the coupling expansion will yield an $\eps$ expansion, but only if the behavior at large $\tilde\np$ and $\tilde\D$ is controlled. This may be done if the diagrams may be grouped into sums over all powers of $\tilde\np$ and $\tilde\D$, which, when summed, yield a well-defined limit. In the one-species case this grouping was relatively simple. The series may be put in the form of a sum of terms $g_R^n\tilde\np f_n(g_R\tilde\np)$, where $n$ counts the number of loops in a given diagram. The term $n=0$ corresponds to the sum of tree diagrams, given by the solution of the simple rate equation, so that $f_0(g_R\tilde\np)\propto (g_R\tilde\np)^{-1}$ as $\tilde\np\to\infty$. By explicit calculation, it is then possible to show that the $f_n$ for $n>0$ behave in a similar manner. Since $g_R\to g_R^*=O(\eps)$, this lead to the result that $n(t)\sim A/t^{d/2}$ where the amplitude $A$ is in principle calculable to any order in $\eps$. In the present case, the series may similarly be organized as a sum of terms of the form $g_R^n\tilde\np f_n(g_R\tilde\np,g_R^2\tilde\D)$, where now the $n=0$ term is given by the sum of diagrams in Fig. \[densum\]. This is given by the solution of (\[eomc\]), which, by the analysis of the previous section, implies that $f_0(g_R\tilde\np,g_R^2\tilde\D)\propto (g_R^2\tilde\D)^{1/2}/(g_R\tilde\np)$ in this limit. However, unlike the single species case, $n$ does not simply count the loops, since already at $n=0$ there are arbitrarily many $\psi$ loops. In addition, while it is possible to identify those diagrams appearing at $n=1$ for example, it is difficult to see how to express their sum in terms of a suitable generalization of (\[eomc\]), and thereby evaluate it. Assuming that their asymptotic behavior is independent of $n_0$ and thus of $\tilde\np$, there are three conceivable ways in which these higher order terms could affect the result. They either diverge less slowly than ${\tilde\D}^{1/2}$ as $\tilde\D\to\infty$, in which case the $n=0$ result gives the leading behavior, which would then yield the same result as for $d>2$; or they all behave like ${\tilde\D}^{1/2}$, in which case the density behaves as $t^{-d/4}$ but with an amplitude which has a nontrivial expansion in powers of $\eps$; or they diverge more strongly, in which case the density no longer behaves as $t^{-d/4}$ for $d<2$. Since this last possibility is in conflict with numerical experiments and rigorous results (albeit for slightly different models), it is unlikely to be correct. When $d=2$ the running coupling goes to zero as $(\ln t)^{-1}$ for $t\ginf$, rather than to an order $\eps$ fixed point. Therefore the leading order terms for an $\eps$ expansion of the amplitude become the exact asymptotic amplitude, with correction terms which go as $(\ln t)^{-1}$. Therefore, if our conjecture is correct, then density should be given exactly by (\[ans\]) in the large $t$ limit. Density Calculation for Unequal Diffusion Constants {#uneqdiff} =================================================== When the two species of particles no longer have equal diffusion constants, then the vertices which depend on $\d$ must be included in the full theory. Then for $d>2$ an effective theory can be developed, just as before, with the resulting action $$\begin{aligned} \label{Sun} S=\int d^dx\biggl[&&\int_0^t dt\Bigl\{\bphi(\dee_t-\nabla^2)\phi+ \bpsi(\dee_t-\nabla^2)\psi-\bpsi\d\nabla^2\phi\nonumber\\&& -\bphi\d\nabla^2\psi+\l\bphi(\phi^2-\psi^2)\Bigr\}-n_{\phi}\bphi(0) -{\D\over 2}\bpsi(0)^2\nonumber\\&&+\dots\biggr].\end{aligned}$$ The effective theory describes classical fields which evolve via the deterministic equations of motion $$\label{UNa}\pt\phi=\nabla^2\phi+\d\nabla^2\psi-\l\phi^2+ \l\psi^2$$ $$\label{UNb}\pt\psi=\nabla^2\psi+\d\nabla^2\phi,$$ which follows from integrating out the $\bphi,\bpsi$ degrees of freedom in the bulk component of (\[Sun\]). From these equations the density can be calculated exactly by using the same methods as before. First, equation (\[UNa\]) is averaged over the initial conditions to yield equation (\[eomc\]), just as in the $\d=0$ case. The solution to Ricatti’s equation again provides an upper bound $\br\phi\leq f\sim\sqrt{\br{\psi(t)^2}}$, although the value of $\br{\psi(t)^2}$ is changed. It will be shown that $\br{\psi^2}\propto t^{-d/2}$, so the upper bound decays with the same exponent as before. Since the fields are real and $\phi\ge|\psi|$, it then follows that $\br{\phi}\sim\br{|\psi|}$ for $d<4$, as shown in (\[UBd\]). Furthermore, it will be shown that asymptotically $\psi(t)$ has a normal distribution, so the density is given exactly by $\br a=\br\phi/\sqrt 2\sim\sqrt{\br{\psi^2}/\pi}$. Therefore the only change in the asymptotic density from the $\d=0$ case is due to the change in the value of $\br{\psi(t)^2}$. Calculation of $\br{\psi(t)^2}$ ------------------------------- The initial terms in the effective theory are in general changed by the presence of $\d$ in the full theory, and therefore must be computed again. One can show that, as before, no $\D^{(0,1)}\bpsi$ initial term is generated. For any diagram which ends with a single $\psi$ line, the last vertex (first from the left) must be a $\d k^2$ vertex. However, this external line has $k=0$, and so all of these diagrams have no contribution. To leading order $\D=n_0$ is unchanged, as can be seen from the diagrams in Fig. \[surface\]: the leading order contribution to $\D$ comes from diagrams composed of no loops, and so all lines carry wave number $k=0$ and are unaffected by the $\d k^2$ vertex. The correction terms to the small $n_0$ limit of $\D$ will likely be of the same order as before, $O(n_0^{d/2}\l^{d/2})$, but with a different amplitude. This amplitude could be calculated, although it would require a generalization of the response functions discussed below. It will be shown the asymptotic value of $\br{\psi^2}$ depends only on $\D$, and so the other surface terms can be neglected. There are new response functions generated in the bulk theory. With $\d=0$ there was just a bare $\psi$ propagator and a $\phi$ response function. In this theory there are instead four response functions, which connect $\phi,\psi$ to $\bphi,\bpsi$ in each possible way, as shown in Fig. \[undiffrf\]. Each of these response functions, represented by double lines, is an infinite sum over all possible numbers of $\d k^2$ vertices inserted. .2in =3.4in .15in These response functions can be found exactly via coupled integral equations, also shown in Fig. \[undiffrf\]. For our purposes, since just the leading term for small $n_0$ is being calculated, we need only the form of the response functions when the earlier time argument is set to zero. To calculate the higher order terms in the expansion $\D=n_0+\dots$ one needs to derive these response functions with $t_1\neq 0$. Setting $t_2=t$, $t_1=0$ in the equations represented by diagrams (a) gives $$G(k,t)=e^{-k^2t}+\d k^2\int_0^td\tp e^{-k^2(t-\tp)}F(k,\tp)$$ $$F(k,t)=\d k^2\int_0^td\tp e^{k^2(t-\tp)}\left(1+\np\l\tp\over 1+\np\l t \right)^2G(k,\tp),$$ or, in terms of $f,g$ defined by $G(k,t)=e^{-k^2t}g(k,t)$ and $F(k,t)= e^{-k^2t}f(k,t)$ $$\label{abbot} g(k,t)=1+\d k^2\int_0^td\tp f(k,\tp)$$ $$\label{costello} f(k,t)=\d k^2\int_0^td\tp\left(1+\np\l\tp\over 1+\np\l t\right)^2g(k,\tp).$$ Differentiating equations (\[abbot\]) and (\[costello\]) with respect to $t$ gives $$\label{fofg} f(k,t)={1\over\d k^2}\dot g(k,t)$$ $$\pt\left[(1+\np\l t)^2f(k,t)\right]=\d k^2(1+\np\l t)^2g(k,t)$$ Substituting for $f$ into the lower equation and manipulating the expression gives a remarkably simple equation for $g$ $${\dee^2\over\dee t^2}[(1+\np\l t)g]=\d^2k^4[(1+\np\l t)g]$$ which has the general solution $$g(k,t)={1\over 1+\np\l t}\left[A\cosh(\d k^2t)+B\sinh(\d k^2 t) \right].$$ From the integral equation (\[abbot\]) one finds the conditions $g(k,0)=1$, which implies $A=1$, and $g(0,t)=1$, which then implies $B=n_0\l/(\d k^2)$. Therefore the explicit form of $G(k,t)$, and from (\[fofg\]) $F(k,t)$, is calculated: $$\label{Gpt} G(k,t)={e^{-k^2t}\over 1+\np\l t}\left [\cosh(\d k^2t)+{\np\l\over\d k^2}\sinh(\d k^2t)\right],$$ $$\begin{aligned} F(k,t)={e^{-k^2t}\over(1+\np\l t)^2}&&\biggl[\biggl(1+\np\l t- {\np^2\l^2\over\d^2k^4}\biggr)\sinh(\d k^2t)\nonumber\\&& +{\np^2\l^2t\over\d k^2} \cosh(\d k^2t)\biggr].\end{aligned}$$ The other response functions, $H(k,t)$ and $J(k,t)$, defined in diagram Fig. \[undiffrf\], can be found via similar methods. The coupled integral equations shown in Fig. \[undiffrf\](b), written in terms of $h=e^{k^2t}H$ and $j=e^{k^2t}J$, are $$h(k,t)={1\over(1+\np\l t)^2}+\d k^2\int_0^td\tp\left(1+ \np\l\tp\over 1+\np\l t\right)^2j(k,\tp)$$ $$j(k,t)=\d k^2\int_0^t h(k,\tp).$$ Differentiating both equations with respect to $t$ and substituting to eliminate $h$ gives the equation $${\dee^2\over\dee t^2}[(1+\np\l t)j]=\d^2k^4[(1+\np\l t)j]$$ which has the general solution $$\label{jpt} j(k,t)={1\over 1+\np\l t}\left[A\cosh(\d k^2t)+B\sinh(\d k^2 t) \right].$$ The condition that $j(k,0)=0$ implies $A=0$. The general solution of $h$ can be found from (\[jpt\]), and then the condition that $h(k,0)=1$ implies $B=1$. Therefore $H$ and $J$ are given by $$\begin{aligned} H(k,t)={e^{-k^2t}\over(1+\np\l t)^2}&&\biggl[(1+\np\l t)\cosh(\d k^2 t) \nonumber\\&&-{\np\l\over\d k^2}\sinh(\d k^2t)\biggr]\end{aligned}$$ $$\label{Jpt} J(k,t)={e^{-k^2t}\over(1+\np\l t)}\sinh(\d k^2t).$$ .2in =2.5in .15in In section \[classden\] the value of $\br{\psi^2}$ was calculated from the simple loop shown in Fig. \[densum\]. The generalization of this calculation is given by the diagrams shown in Fig. \[undiffloop\], which are composed of the $G(k,t)$ and $J(k,t)$ response functions. The surface couplings $\D^{(0,2)}\neq -\D^{(2,0)}$ beyond the leading small $n_0$ terms, and so the couplings are labeled $\D$ and $\D^{\p}$ respectively. It should be noted that unlike the $\d=0$ case, these are not the only diagrams which contribute to $\br{\psi^2}$. Examples of other diagrams, and arguments for why they are irrelevant, will be given below. First, we compute those of Fig. \[undiffloop\], which give $$\br{\psi(t)^2}=\int{d^dk\over(2\pi)^d}[\D\,G(k,t)^2-\D^{\p} J(k,t)^2].$$ Substituting (\[Gpt\]) and (\[Jpt\]) into the equation above, and rewriting the integral in terms of the variable $u=k^2t$ gives $$\begin{aligned} \br{\psi(t)^2}&=&{t^{-d/2}\over(4\pi)^{d/2}\G(d/2)(1 +\np\l t)^2}\int_0^{\infty}du\>u^{d/2-1}e^{-2u}\nonumber\\ &\mbox{}\times&\biggl[\D\cosh^2(\d u)-\D^{\p}\sinh^2(\d u) +{\D\np\l t\over\d u}\sinh(2\d u)\nonumber\\ &&\mbox{}+\D\left(\np\l t\over\d u\right)^2\sinh(\d u)^2\biggr].\end{aligned}$$ Each term in the square brackets gives a convergent integral for $d>0$. Therefore we can take the large $t$ limit before integrating, and only calculate the leading term in $t$, which is found to be that on the far right in the brackets. Consequently, the value of $\D^{\p}$ is unimportant. Evaluating this integral gives $$\label{psisq} \br{\psi^2}={\D\over(8\pi)^{d/2}}Q(d,\d)\>t^{-d/2}$$ where $$\label{Qd} Q(d,\d)={4\left[(1+\d)^{2-d/2}+(1-\d)^{2-d/2}-2\right]\over\d^2(d-2)(d-4)}.$$ From (\[psisq\]) it follows that $Q\sim\br{\psi^2}_{\d}/ \br{\psi^2}_{0}$, in the small $n_0$ limit. This function $Q$ is non-singular at $\d=0$, and satisfies $Q(d,0)=1$. While $Q$ appears to be divergent at $d=2,4$, it is actually finite everywhere except $d\ge 4$ and $\d=\pm 1$. It is likely that the limits of $t\ginf$ and $\d\goto\pm 1$ do not necessarily commute, and that a separate treatment for the case of an immobile species, at least in this singular case, would be required. For $d<4$ this function has finite values as $\d\goto\pm 1$, but the slope at $\d=\pm 1$ is infinite for $d\ge 2$. .2in .15in While the calculation of $Q(d,\d)$ is only strictly valid for $2<d<4$, it is nonetheless interesting to consider its limits for the integer dimensions from $d=1$ to $d=4$, motivated by section \[renormab\] on $d\le 2$, in which it was conjectured that the “classical” amplitude is also the leading term in an $\eps$ expansion for $d=2-\eps$. From (\[Qd\]) $$\label{Qdii} Q(d,\d)=\cases{ {\ds 4\over\ds 3\d^2}[(1+\d)^{3/2}+(1-\d)^{3/2}-2]&$d=1$\cr\noalign{\medskip} {\ds(1-\d)\ln(1-\d)+(1+\d)\ln(1+\d)\over\ds\d^2}&$d=2$\cr \noalign{\medskip} {\ds 4\over\ds\d^2}[2-\sqrt{1+\d}-\sqrt{1-\d}]&$d=3$\cr \noalign{\medskip} {\ds -\ln(1-\d^2)\over\ds \d^2}&$d=4$\cr}$$ Since the density goes as $\sqrt{\br{\psi^2}}$, the function $\sqrt{Q(d,\d)}$ is plotted in Fig. \[plotQ\] for integer values of $d$. The density amplitude increases monotonically with $|\d|$, but is not changed remarkably for modest values of $\d$. .2in =2.5in .15in There are other diagrams which give contributions to $\br{\psi^2}$, unlike the case. Some of these are shown in Fig. \[extra\](a). All of these diagrams have the similar feature that they contain one of the two sub-diagrams in Fig. \[extra\](b). These sub-diagrams give rise to effective vertices of the form $\l^{\p}\bpsi\nabla^2\psi^2$ and $\l^{\p}\bpsi\nabla^2\phi^2$ in the bulk theory. However, such vertices are irrelevant, which follows from power counting, and so the diagrams which arise from them must be sub-leading in time. Therefore we conclude that asymptotically the value of $\br{\psi^2}$ is given by (\[psisq\]) and (\[Qd\]). Demonstration that $\psi(t)$ has a Normal Distribution ------------------------------------------------------ In order for the calculation of $\br{\psi^2}$ to give the amplitude of the density it is necessary that $\psi(t)$ have a normal distribution. When $\d=0$ this follows directly from the simple diffusion equation satisfied by $\psi$, or equivalently, from central limit arguments. However, $\psi$ evolves via equation (\[UNb\]) for $\d\neq 0$, and so it needs to be shown that it still flows to a normal distribution. What we will show is that the random variable $t^{d/4}\psi$ flows to a static normal distribution, the width of which was calculated above. Consider $\br{\psi^n}$, where $n$ is even. There is one diagram in which $n$ response functions $G(k,t)$ are connected in pairs to $n/2$ initial terms $(\D/2)\bpsi^2$. This diagram contains $n/2$ loops, and is therefore of order $t^{-nd/4}$. It was shown above replacing any of the $G(k,t)$ loops with $J(k,t)$ response functions connected to $(\D^{\p}/2)\bphi^2$ gives a lower order contribution. Similarly, any other diagrams, which would originate from considering higher order surface terms, will involve more than $n/2$ loops, and will therefore decay faster in time. For $n$ odd one finds that there are no diagrams for $\br{\psi^n}$ which decay as slowly as $t^{-nd/4}$. That is, for $n$ odd, $\lim_{t\goto\infty} \br{(t^{d/4}\psi)^n}=0$. Since the distribution of the variable $t^{d/4}\psi$ has only even moments as $t\ginf$, and these moments are just multiples of $\br{(t^{d/4}\psi)^2}$, generated by all possible pair contractions, then the distribution is normal. Correlation Functions for $2<\lowercase{d}<4$ ============================================= When $d>2$, one can use the classical action to calculate the correlation functions. Consider the distribution of the random variable $t^{d/4}\phi({\bf x},t)$ with $2<d<4$. From section \[classden\] we know that as $t\ginf$. Furthermore, from equation (\[UBd\]) it follows that, as $t\ginf$, $\br{(t^{d/4}\phi -t^{d/4}|\psi|)^2}\goto 0$. This suggests that the distributions $P[t^{d/4}\phi]\sim P^{\p}[t^{d/4}|\psi|]$ as $t\ginf$. The latter distribution is known exactly, as $t^{d/4}\psi$ is at late times given by a static normal distribution. It is not correct to say that asymptotically $\phi$ and $|\psi|$ are everywhere equal, since this would imply that there are no regions in which the densities $a$ and $b$ are both non-zero. However, the reaction regions, those in which both densities are non-zero, become negligibly small for large $t$, and the corrections to setting $\phi$ equal to $|\psi|$ in calculating correlation functions will be subleading in time. Stated another way, the leading term in both $\br{\phi_1\phi_2}$ and $\br{|\psi_1||\psi_2|}$ is of order $t^{-d/2}$. To this order the two random variables $\phi$ and $|\psi|$ have identical distributions. This is in contrast to a quantity such as $\phi^2-\psi^2$, which is measuring a subleading term relative to $t^{-d/2}$. We can use the property that $t^{d/4}\phi$ is given by the absolute value of a gaussian random field to calculate correlation functions. This is similar to what is done is the dynamics of phase ordering, where the order parameter field can be mapped to an auxiliary field which is assumed to be a gaussian random field. This analogy will be discussed further below. Since $\phi$ and $|\psi|$ are given by the same distribution, we conclude $\br{\phi_1\phi_2}\sim\br{|\psi_1||\psi_2|}$, where the labels indicate the positions ${\bf x}_1$ and ${\bf x}_2$ at time $t$. The correlation function $\br{|\psi_1||\psi_2|}$ can be calculated exactly by using the fact that, asymptotically, $\psi(t)$ has a normal distribution. The joint probability distribution $P[\psi_1,\psi_2]$ is then also normal, so $$\label{probdist} P[\psi_1,\psi_2]={\sqrt{4\a^2-\b^2}\over2\pi}\exp\left\{ -\a\psi_1^2-\a\psi_2^2-\b\psi_1\psi_2\right\},$$ where we have used translational invariance to set $\br{\psi_1^2}= \br{\psi_2^2}$. The constants $\a$ and $\b$ are determined by the values of $\br{\psi^2}$ and $\br{\psi_1\psi_2}$, which are evaluated from the diagrams. The latter we have only calculated for $\d=0$, or equal diffusion constants, so we consider that case first. For notational convenience we define $\br{\psi^2}\equiv C(t)=\D/(8\pi t)^{d/2}$. The diagram shown in Fig. \[crossterm\](a) is used to calculate the correlation function $\br{\psi({\bf k})\psi({\bf -k})}$, from which one finds $$\br{\psi_1\psi_2}=\int {d^dk\over(2\pi)^d}e^{i{\bf k}\cdot({\bf x}_1- {\bf x}_2)}\br{\psi({\bf k})\psi({\bf -k})}.$$ When $\d=0$ then $\br{\psi({\bf k})\psi({\bf -k})}=\D e^{-2k^2t}$, and $$\label{Deff} \br{\psi_1\psi_2}=C(t)\exp(-r^2/8t)\equiv C(t)f(r^2/t)$$ where $r=|{\bf x}_1-{\bf x}_2|$. In terms of (\[Deff\]) we find for $\a,\b$ $$\a={1\over 2C(1-f^2)}\qquad\qquad\b={f\over C(1-f^2)}.$$ .2in .15in With these values substituted into (\[probdist\]), one can calculate $$\begin{aligned} \br{\phi_1\phi_2}&\sim&\br{|\psi_1||\psi_2|}= \int_{-\infty}^{\infty}d\psi_1\int_{-\infty}^{\infty}d\psi_2\>|\psi_1| |\psi_2|\,P[\psi_1,\psi_2]\nonumber\\ \noalign{\medskip} &=&{2C\over\pi}\left[\sqrt{1-f^2}+f\arctan\left(f\over\sqrt{1-f^2}\right) \right].\end{aligned}$$ This correlation function can be used to find the correlation functions $\br{a_1a_2}$ and $\br{a_1b_2}$. Specifically $$\br{a_1a_2}={1\over 2}\br{\phi_1\phi_2+\psi_1\psi_2},$$ which gives for the connected part $\br{a_1a_2}_c=\br{a_1a_2}-\br a^2$, $$\begin{aligned} &&\br{a_1a_2}_c=\nonumber\\ &&{\D\over\pi(8\pi t)^{d/2}}\left[{\pi\over 2}f-1+\sqrt{1-f^2} +f\arctan\left(f\over\sqrt{1-f^2}\right)\right]\nonumber\\\end{aligned}$$ For large $r$, $f=\exp(-r^2/8t)$ is small, giving $$\label{Lra} \br{a_1a_2}_c\sim{\D\over 2(8\pi t)^{d/2}}e^{-r^2/8t}.$$ Similarly, $\br{a_1b_2}=\br{\phi_1\phi_2-\psi_1\psi_2}/2$, so that $$\begin{aligned} &&\br{a_1b_2}_c=\nonumber\\ &&{\D\over\pi(8\pi t)^{d/2}}\left[-{\pi\over 2}f-1+ \sqrt{1-f^2}+f\arctan\left(f\over\sqrt{1-f^2}\right)\right]\nonumber\\\end{aligned}$$ which for large $r$ goes as $$\label{Lrb} \br{a_1b_2}_c\sim-{\D\over 4(8\pi t)^{d/2}}e^{-r^2/8t}.$$ A plot of these connected correlation functions is shown in Fig. \[corrs\]. The signs $\br{a_1a_2}>0$ and $\br{a_1b_2}<0$ can be understood for short distances to be a consequence of the segregation. Given an $A$ particle at a particular point, there is an increased probability that a nearby particle is also an $A$, and a decreased probability that it is a $B$. .2in .15in For the case $\d\neq 0$ one has $\br{\psi^2}=C(t)Q(d,\d)$, as given by (\[psisq\]). The generalization of $\br{\psi({\bf k})\psi({\bf -k})}$, shown in Fig. \[crossterm\], behaves for small $k$ the same as when $\d=0$. Therefore, for large $r$ one still has $\br{\psi_1\psi_2}=Cf$. When this is put in the expressions for $\br{a_1a_2}$ and $\br{a_1b_2}$ one finds that the large $r$ behavior is given by (\[Lra\]) and (\[Lrb\]) is unaffected by $\d\neq 0$. While these correlation functions and other quantities can be calculated, they ultimately rely on the stronger statement that $\psi$ is a gaussian random field, and that $\phi\sim|\psi|$. The topology of the domains is determined by the random field, with the boundaries between $a$ regions and $b$ regions given by the zeroes of $\psi$. This topology is completely equivalent to an analogous situation in phase ordering. It has been suggested that in the phase ordering of a scalar order parameter an invertible, non-linear mapping from the order parameter field to an auxiliary field results in the latter being a gaussian random field [@Mazenko]. Usually this mapping is chosen to be the solution of a single kink, for example the hyperbolic tangent profile. While this method is no longer believed to be quantitatively correct [@Oono], it does provide a useful picture of the structure of the domains. Again, the zeroes of this gaussian random field determine the boundaries between the equilibrated phase. The difference between these systems lies in how correlation functions are calculated from the random field. In the reaction-diffusion case one is interested in the correlation functions of the field itself, and of the absolute value of the field. Neither of these quantities exhibit remarkable behavior. In the phase ordering one argues that at late times the mapping between the order parameter field and the gaussian field goes to a step function, and therefore order parameter correlations are given by the correlations of the sign of the random field. These sharp boundaries give rise to more interesting features, such as non-analytic terms in the small $r$ limit of the correlation function, or correspondingly power law tails for large wave number in the Fourier transform. Reaction Zones {#srz} ============== It was shown in section \[classden\] that for $d<4$ the particles segregate asymptotically into regions of purely $A$ or $B$ particles. As a result of this segregation there exist interfaces between the two species, and all reactions occur in the interfacial regions. These reaction zones have interesting scaling properties. For example, the width of the interface goes as $t^{\a}$ with the exponent $\a<1/2$. Also, the nearest neighbor distance distribution of the particles in the reaction zone is found to have a characteristic length $\lrz$ that goes as a power of $t$, with an exponent which differs from that of the bulk system, where $\br a^{-1/d}\sim t^{1/4}$. To derive these properties we begin with a related steady-state problem. Consider a system with a source of $A$ particles located at the boundary $x=-L$ which maintains a fixed current $J\hat{\bf x}$, and a similar source $-J\hat{\bf x}$ of $B$ particles positioned at $x=L$. These opposing currents will establish a steady-state profile, in which the average densities will be functions of the transverse coordinate $x$. For a given current $J$ one can choose $L$ to be large enough that the reactions are localized to an interfacial region of width $w\ll L$. In this case, it is found that the densities in the reaction zone, where $|x|\lesssim w$, have universal scaling forms. Also of interest is the reaction rate $R(x)=\lo \br{a(x)b(x)}$, which exhibits scaling, and which is used to define the width of the reaction zone. The power counting of section \[power\] showed that the four-point vertices were irrelevant for $d>2$. Therefore $R(x)\sim\l_{\rm eff}\br a\br b$ in the asymptotic limit—which will be shown to be the small $J$ limit—and the problem reduces to the differential equations of the effective theory: $$\begin{aligned} \label{feq} \pt\br a&=\nabla^2\br a-\l_{\rm eff}\br a\br b\\ \pt\br b&=\nabla^2\br b-\l_{\rm eff}\br a\br b.\end{aligned}$$ From these equations it has been shown that $$\label{mfscale} R\sim J^{4/3}f(xJ^{1/3}),\qquad d>2,$$ implicitly by Gálfi and Rácz [@GR], and later explicitly by Ben-Naim and Redner [@BR]. From (\[mfscale\]) one identifies the width $w\sim J^{-1/3}$, and the characteristic length of the particle distribution within the reaction zone $\lrz\equiv\br{a(x=0)}^{-1/d}\sim J^{-2/3d}$. The latter quantity is derived in Ref. [@LC]. For $d\le 2$ one does not have simply differential equations, and the full field theory must be taken into account. We begin by observing that the current $J_A$ is given by $a^*\dee_x a$ in the notation of section \[power\], and similarly for $J_B$. From dimensional analysis $[J]=k^{d+1}$. We proceed with the renormalization of the theory, as was sketched in section \[renormab\]. A normalization scale $\k$ is introduced, and used to define the renormalized coupling $g_R$. Since physical quantities, such as the width, cannot depend on $\k$, then $$\k{d\over d\k}w=\left[\k{\dee\over\dee\k}+\beta(g_R){\dee\over\dee g_R} \right]w(J,g_R,\d,\k)=0.$$ Note that, since there are no diagrams which can dress the two-point vertices in (\[S\]), $\d$ does not get renormalized, and therefore does not appear in equation above. From dimensional analysis one has $$\left[\k{\dee\over\dee\k}+(d+1)J{\dee\over\dee J}+1\right]w(J,g_R,\d,\k)=0.$$ Combining these equations gives the Callan-Symanzik equation $$\left[(d+1)J{\dee\over\dee J}-\beta(g_R){\dee\over\dee g_R}+1\right]w=0.$$ with the solution $$\label{rzCS} w(J,g_R,\d)=\k J^{-1/(d+1)}w(\k^{d+1},\tilde g_R,\d).$$ In the small $J$ limit then $\tilde g_R\goto g_R^*$, and the right-hand side is given by $$w\sim J^{-1/(d+1)}f(\d,\eps),\qquad d<2.$$ Following the same procedure for any dimensionful quantity results in the scaling behavior being given by dimensional analysis. That is, $\lrz\sim w$, $$\br a,\br b\sim J^{d/(d+1)}F_{a,b}(x\,J^{1/(d+1)}),$$ and $$R(x)\sim J^{(d+2)/(d+1)}G(x\,J^{1/(d+1)}).$$ Note that these results imply that\ $R\sim J^{(2-d)/(d+1)}\br a\br b$. This can be shown explicitly by calculating $R\propto\br{ab}=\l(J)\br a\br b$, where $\l(J)$ is given by the bubble sum shown in Fig. \[bubbles\], with $s=J^{2/(d+1)},k=0$. In the small $J$ limit then $\l(J)\sim J^{(2-d)/(d+1)}$. Since the current $J$ may be thought of as being due to localized sources of $A$ and $B$ particles at $x=\pm L$ respectively, the coupling constant power counting arguments are formally the same as those of Ref. [@Lee] (see Sec. \[renormab\]), in which the sources are localized at $t=0$.[@HC] Thus the various scaling functions above may, in principle, be calculated as an expansion in $g_R^*=O(\eps)$, in which the leading term is given by the solution of the rate equations (\[feq\]). The next order corrections to the reaction profile have been computed in Ref. [@HC], where it was shown that the fluctuation corrections lead to a universal power law tail in this function. For $d=2$ one has $\tilde g_R\sim 1/|\ln J|$ for small $J$, and the leading order result is therefore found by substituting this behavior into the solution of the rate equation (\[feq\]). This leads to the results[^3] $$\begin{aligned} w&\propto& \left(|\ln J|\over J\right)^{1/3},\\ R(x)&\sim&\lambda(J)\br a\br b\nonumber\\ &\sim& J^{4/3}|\ln J|^{-1/3}G(x\,J^{1/3}|\ln J|^{-1/3}),\\ \lrz&\propto& J^{-1/3}|\ln J|^{-1/6}.\end{aligned}$$ As was discussed in Ref. [@LC], the corresponding results for the time-dependent cases of segregated initial conditions or of randomly homogeneous initial conditions is given by substituting $J\propto t^{-1/2}$ or $J\propto t^{-(d+2)/4}$ (with $d=2$), respectively, in the above formulas. These results, for the case of segregated initial conditions, have been obtained recently via heuristic arguments by Krapivsky [@Krapivsky]. The authors are grateful for useful discussions with S. Cornell, M. Droz and M. Howard. This work was supported by a grant from the EPSRC, and by NSF grants DMR 90-07811 and CHE 93-1729. Calculation of the leading correction term for $\D$ {#calculation-of-the-leading-correction-term-for-d .unnumbered} =================================================== In order to calculate the first order correction term to the expansion $\D=\D_0+\dots$ we must first comment on the bulk diagrams which generate $\lf$. The effective coupling can be calculated as an expansion in the bare couplings, via the diagrams shown in Fig. \[leff\]. The loop integrals in this expansion require the cutoff $\L$, and one finds $$\label{newlam} \lf=\l_2-\l_2^2{4\L^{d-2}\over(8\pi)^{d/2}(d-2)}+O(\l_2^3).$$ .2in .15in If the response functions in the loop of Fig. \[surface\](c) were instead just propagators, then this set of diagrams would be included into those of Fig. \[surface\](b) when the substitution $\l_2\goto\lf$ is made via (\[newlam\]). Therefore, the terms which are new and constitute a correction to $\D_0$ are those in Fig. \[surface\](c) with the propagator loop subtracted out. We define the large $t$ limit of these diagrams to be $\D_1$, that is $$\begin{aligned} \D_1=4\l_2^2\np^2\int_0^\infty &&dt_2\int_0^{t_2}dt_1{d^dk\over(2\pi)^d} \biggl[{e^{-2k^2t}(1+\np\l_1t_1)^2\over(1+\np\l_1t_2)^4}\nonumber\\ &&-{\L^{d-2}\over(8\pi)^{d/2}(d-2)(1+\np\l_1 t_2)^2}\biggr].\end{aligned}$$ Performing the wave number integral with the $\L$ cutoff imposed in the same manner as in (\[newlam\]) gives The $t_1$ integral can be evaluated as a Laplace convolution integral, and the cutoff dependent terms cancel. The remaining $t_2$ integral is $$\begin{aligned} \D_1={-8\l_2^2\np^2\over(8\pi)^{d/2}(d-2)}&&\int_0^\infty{dt_2\> t_2^{1-d/2}\over(1+\np\l_1t_2)^4}\nonumber\\ \times&&\biggl[1+{4\np t_2\over 4-d} +{8\np^2t_2^2\over(4-d)(6-d)}\biggr].\end{aligned}$$ This integral can be done exactly, giving $$\D_1={\l_2^2\over\l_1^2}(\np\l_1)^{d/2}{(d+2)(d+4)\over 48(8\pi)^{d/2-1}\sin(\pi d/2)}.$$ In terms of the initial density $n_0$ and the effective coupling then one finds the result (\[Delta\]) for $\D$. Evaluating the diagrams such as those in Fig. \[surface\](c), but containing more loops will then give the higher order terms in this small $n_0$ expansion of $\D$. V. Kuzovkov and E. Kotomin, Rep. Prog. Phys. [**51**]{}, 1479 (1988). A. A. Ovchinnikov, S. F. Timashev, and A. A. Belyy, [*Kinetics of Diffusion Controlled Chemical Processes*]{}, (Nova Science, New York, 1989). A. A. Ovchinnikov and Ya. B. Zeldovich, Chem. Phys. [**28**]{}, 215 (1978). D. Toussaint and F. Wilczek, J. Chem. Phys. [**78**]{}, 2642 (1983). M. Bramson and J. L. Lebowitz, J. Stat. Phys. [**62**]{}, 297 (1991). M. Bramson and J. L. Lebowitz, J. Stat. Phys. [**65**]{}, 941 (1991). K. Kang and S. Redner, Phys. Rev. Lett. [**52**]{}, 955 (1984). S. Cornell, M. Droz, and B. Chopard, Physica A [**188**]{}, 322 (1992). F. Leyvraz, J. Phys. A [**25**]{}, 3205 (1992). S. Cornell and M. Droz, Phys. Rev. Lett. [**70**]{}, 3824 (1993). L. Peliti, J. Physique [**46**]{}, 1469 (1985). L. Peliti, J. Phys. A [**19**]{}, L365 (1986). B. P. Lee, J. Phys. A, [**27**]{}, 2633 (1994). L. Gálfi and Z. Racz, Phys. Rev. A [**38**]{}, 3151 (1988). E. Ben-Naim and S. Redner, J. Phys. A [**25**]{}, L575 (1992). B. P. Lee and J. Cardy, Phys. Rev. E, [**50**]{}, 3287 (1994). M. Doi, J. Phys. A [**9**]{}, 1465 (1976); M. Doi, J. Phys. A [**9**]{}, 1479 (1976). L. S. Schulman, [*Techniques and Applications of Path Integration*]{}, (Wiley, New York, 1981) p. 242 ff. M. Howard, private communication. H. W. Diehl, in [*Phase Transitions and Critical Phenomena*]{}, edited by C. Domb and J. L. Lebowitz (Academic, New York, 1986), Vol. 10. G. N. Watson, [*A Treatise on the Theory of Bessel Functions*]{}, (Cambridge, Cambridge, 1952). see for example D. J. Amit, [*Field Theory, the Renormalization Group, and Critical Phenomena*]{}, (World Scientific, Singapore, 1984). G. Mazenko, Phys. Rev. Lett. [**63**]{}, 1605 (1990); G. Mazenko, Phys. Rev. B [**42**]{}, 4487 (1990). C. Yeung, Y. Oono, and A. Shinozaki, Phys. Rev. E [**49**]{}, 2693 (1994). S. Cornell, M. Droz, and B. Chopard, Phys. Rev. A [**44**]{}, 4826 (1991). M. Howard and J. Cardy, in preparation. P. L. Krapivsky, preprint. [^1]: There is a misprint in the $d=3$ amplitude of Ref. [@TW], Eq. (19c). [^2]: For an interesting presentation of the properties and history of this equation, see [@Watson]. [^3]: These logarithms were incorrectly omitted from Ref. [@LC].

--- abstract: 'The interaction between a polarizable particle and a reflecting wall is examined. A macroscopic approach is adopted in which the averaged force is computed from the Maxwell stress tensor. The particular case of a perfectly reflecting wall and a sphere with a dielectric function given by the Drude model is examined in detail. It is found that the force can be expressed as the sum of a monotonically decaying function of position and of an oscillatory piece. At large separations, the oscillatory piece is the dominant contribution, and is much larger than the Casimir-Polder interaction that arises in the limit that the sphere is a perfect conductor. It is argued that this enhancement of the force can be interpreted in terms of the frequency spectrum of vacuum fluctuations. In the limit of a perfectly conducting sphere, there are cancellations between different parts of the spectrum which no longer occur as completely in the case of a sphere with frequency dependent polarizability. Estimates of the magnitude of the oscillatory component of the force suggest that it may be large enough to be observable.' --- =cmbx10 scaled 2 -36pt 0.64cm 0.64cm quant-ph/9804055\ April 22, 1998\ Revised Oct. 22, 1998 [Casimir Force between a Dielectric Sphere and a Wall:\ 0.2in A Model for Amplification of Vacuum Fluctuations ]{} .4in L.H. Ford[^1]\ .2in Institute of Cosmology\ Department of Physics and Astronomy\ Tufts University\ Medford, Massachusetts 02155\ .3in PACS categories: 12.20.Ds, 77.20.+e, 03.70.+k Introduction {#sec:intro} ============ It was noted some time ago that if one wishes to assign a frequency spectrum to the Casimir force between reflecting planar boundaries, the result is a wildly oscillating function of frequency [@Ford88; @Hacyan]. The integral of this function over all frequencies can only be performed with the aid of a suitable convergence factor. The net Casimir energy is much smaller than the contribution of each individual oscillation peak. The effect of integration over all frequencies is almost, but not quite completely, to cancel the various frequency regions against one another. This leads to the speculation [@Ford93] that one might be able to upset this cancellation in some way, and thereby greatly amplify the magnitude of the Casimir force, and possibly change its sign. In the case of parallel plane boundaries, no natural way to do this has been demonstrated. However, the Casimir-Polder interaction between a polarizable particle and a reflecting plane offers similar possibilities. Casimir and Polder [@CP] originally derived the potential between an atom in its ground state and a perfectly reflecting wall. In the large distance limit, their result takes the particularly simple form[^2] $$V_{CP} \sim - \frac{3\, \alpha_0}{8 \pi\, z^4}\,, \label{eq:CP}$$ where $z$ is the distance to the wall, and $\alpha_0$ is the static polarizability of the atom. This asymptotic potential may be derived from the interaction Hamiltonian $$H_{int} = - \frac{1}{2}\, \alpha_0\, {\bf E}^2 \,, \label{eq:H_int}$$ where ${\bf E}$ is the quantized electromagnetic field operator. If one expands this operator in terms of a complete set of the Maxwell equations in the presence of the boundary, the asymptotic Casimir-Polder may be written as $$\langle H_{int} \rangle = \frac{\alpha_0}{4 \pi \, z^3}\, \int_0^\infty d \omega\, \sigma(\omega) \,,$$ where $$\sigma(\omega) = \Bigl[(2\,\omega^2\,z^2 -1) \sin 2\omega z + 2\,\omega\,z \cos 2\omega z \Bigr] \,.$$ The integrand, $\sigma(\omega)$, is an oscillatory function whose amplitude [*increases*]{} with increasing frequency. Nonetheless, the integral can be performed using a convergence factor ([*e.g.*]{}, insert a factor of $e^{-\beta\, \omega}$ and then let $\beta \rightarrow 0$ after integration). The result is the right hand side of Eq. (\[eq:CP\]). It is clear that massive cancellations have occurred (see Fig. 1), and that the area under an oscillation peak can be much greater in magnitude than the final result. This again raises the possibility of tampering with this delicate cancellation, and dramatically altering the magnitude and sign of the force. \[Figure 1\] The frequency spectrum, $\sigma(\omega)$, for the Casimir-Polder potential. The oscillations almost exactly cancel, leaving a net area under the curve equal to that of the shaded region indicated by the arrow. The purpose of this paper is to explore this question in the context of a specific model. The force between a dielectric sphere and a perfectly conducting plane will be examined. The polarizability of the sphere will be taken to be a function of frequency, thereby introducing the possibility of modifying the contributions of different parts of the spectrum. This or similar problems have been discussed before by several authors. However, it will here be examined from a different viewpoint. The force may be calculated from the Maxwell stress tensor. In Section  \[sec:dipole\], a formula for the force on a small sphere in an arbitrary applied electromagnetic field will be derived in an electric dipole approximation and discussed. In Section  \[sec:interface\], this formula will be applied to the calculation of the force on a dielectric particle near an interface in terms of the Fresnel coefficients of the interface. This result will be applied to the case of a dielectric sphere and a perfectly reflecting boundary in Section  \[sec:perfect\]. It will be shown that the force has a component which is an oscillatory function of position, and that it is possible for the sphere is be in stable equilibrium at a finite distance from the boundary. The results are summarized and discussed in Section \[sec:final\]. Force on a Small Particle ========================= Electric Dipole Approximation {#sec:dipole} ----------------------------- In this section, we will discuss the force which an applied electromagnetic field exerts on a small dielectric sphere. The applied electric field will be taken to be ${\bf E}_a({\bf x},t)$, and the corresponding magnetic field to be ${\bf B}_a({\bf x},t)$. We assume that the induced (scattered) field is that of electric dipole radiation from a time-varying dipole moment ${\bf p}$. Later, ${\bf p}$ will be taken to be linearly related to ${\bf E}_a$, but for now it is unspecified. We further assume that the particle is small compared to the characteristic spatial scale over which ${\bf E}_a({\bf x},t)$ and ${\bf B}_a({\bf x},t)$ vary. The latter assumption is not really independent of the electric dipole approximation: If the size of the sphere is not small then one would in general have to include the contributions of higher multipoles. Just outside the particle, the electric and magnetic fields due to the dipole take the near-zone forms: $$\begin{aligned} {\bf E}_d &\approx& \frac{3 {\bf \hat n} ({\bf \hat n}\cdot {\bf p} - {\bf p}) } {r^3} \nonumber \\ {\bf B}_d &\approx& - \frac{{\bf \hat n} \times {\bf \dot p} }{r^2} \,. \label{eq:dipole}\end{aligned}$$ Here $r$ is the radial distance from the dipole, and ${\bf \hat n}$ is the outward directed unit normal vector. The net force acting upon the particle can be calculated by integrating the Maxwell stress tensor over a spherical surface just outside the particle, $$F^i = \oint d a_j \, T^{ij} \, ,$$ where $$T^{ij} = \frac{1}{4 \pi}\, \Bigl[ E^i\, E^j + B^i\, B^j - \frac{1}{2} \delta^{ij}\,({\bf E}^2 + {\bf B}^2) \Bigr] \,.$$ If we insert the net fields, ${\bf E}_a + {\bf E}_d$ and ${\bf B}_a + {\bf B}_d$, into this expression, there will be three types of terms: those involving only the applied fields, those involving only the dipole fields, and the cross terms. However, the pure dipole terms yield no net contribution. Furthermore, any force due to the pure applied field terms is independent of the polarizability, and hence not of interest. Thus we consider only the cross terms in $T^{ij}$ between the applied and dipole fields: $${\bf F} = \frac{1}{4 \pi}\, \oint d a \,\Bigl[ ({\bf \hat n}\cdot{\bf E}_a){\bf E}_d + ({\bf \hat n}\cdot{\bf E}_d){\bf E}_a + ({\bf \hat n}\cdot{\bf B}_a){\bf B}_d - {\bf \hat n}\bigl({\bf E}_a\cdot{\bf E}_d + {\bf B}_a\cdot{\bf B}_d\bigr) \Bigr] \,. \label{eq:stress}$$ Note that ${\bf \hat n}\cdot{\bf B}_d = 0$. Because the particle is assumed to be small, we may expand ${\bf E}_a$ and ${\bf B}_a$ in a Taylor series around ${\bf x} = {\bf x}_0$, the location of the particle. The leading nonzero contributions to the force come from the zeroth order term in ${\bf B}_a$ and the first order term in ${\bf E}_a$: $$\begin{aligned} {\bf B}_a({\bf x},t) &\approx& {\bf B}_0 \nonumber \\ E_a^i({\bf x},t) &\approx& E^i_0 + r \,{\bf \hat n}\cdot{\bf \nabla}E^i_0 + \cdots \,,\end{aligned}$$ We now insert these expansions and Eq. (\[eq:dipole\]) into Eq. (\[eq:stress\]) and perform the angular integration, using the relation $$\oint d a \, n^i \, n^j = \frac{4\pi r^2}{3} \delta^{ij} \,,$$ to find $$F^i = \frac{2 }{3}\, p^j \,\partial_j E^i_0 + \frac{1}{3}\, p_j \,\partial^i E^j_0 + \frac{2 }{3}\, ({\bf \dot p} \times {\bf B}_0)^i \,. \label{eq:force0}$$ It is of interest to check the static limit of this expression. In this limit, $\dot{\bf p} = 0$ and ${\bf \nabla}\times{\bf E}_0 = 0$. If we use these relations, and set ${\bf p} = \alpha_0 {\bf E}_0$, where $\alpha_0$ is the static polarizability of the particle, the result is $$F^i = \alpha_0 \,p_j \, \partial^i E^j_0 = \frac{1 }{2}\,\alpha_0 \,\partial^i {\bf E}_0^2 \,.$$ This is equivalent to the familiar result that the interaction energy of an induced dipole with a static electric field is $$V = -\frac{1 }{2}\,\alpha_0 \,{\bf E}_0^2 \,.$$ Interaction with a Single Plane Wave {#sec:plane} ------------------------------------ Here we apply the result, Eq. (\[eq:force0\]), to compute the force which a single, linearly polarized plane wave exerts on the particle. The electric and magnetic fields of this wave are given by $$\begin{aligned} {\bf E}_0 &=& Re \Bigl(\hat{\bf \epsilon}\, A e^{i({\bf k}\cdot{\bf x}_0 -\omega t)}\Bigr) = \hat{\bf \epsilon} A \cos({\bf k}\cdot{\bf x}_0 -\omega t) \nonumber \\ {\bf B}_0 &=& \hat{\bf k} \times \hat{\bf \epsilon}\, A \cos({\bf k}\cdot{\bf x}_0 -\omega t)\,,\end{aligned}$$ where $A$ is the amplitude and $\hat{\bf \epsilon}$ the polarization vector. The dipole moment is given by $${\bf p} = Re \bigl(\alpha {\bf E}_0) = \hat{\bf \epsilon} A |\alpha| \cos({\bf k}\cdot{\bf x}_0 -\omega t +\gamma)\,,$$ where $$\alpha = |\alpha| \, e^{i\gamma} = \alpha_1 + i \alpha_2 \,.$$ We are interested in the time-averaged force, measured over time scales long compared to $1/\omega$; so we henceforth understand $F^i$ to be the time average of Eq. (\[eq:force0\]). In the present case, this yields $${\bf F} = \frac{1}{2}\,{\bf k}\, A^2 \,|\alpha| \sin\gamma = \frac{1}{2}\,{\bf k}\, A^2 \,\alpha_2 \,, \label{eq:single_plane}$$ a force proportional to the imaginary part of the polarizability, $\alpha_2$. This result may be given a simple physical interpretation. The rate at which electromagnetic energy is dissipated is given by the usual Joule heating term $$\dot W = \int {\bf J}\cdot {\bf E}\; d^3 x \,,$$ where ${\bf J}$ is the current density, and the integration is taken over the volume of the particle. Because the electric field is approximately constant over this volume, and because one may show [@Jackson] from the continuity equation that $$\int {\bf J}\, d^3 x = \dot{\bf p} \,,$$ we have that the time-averaged power absorbed by the particle is $$\dot W = \frac{1}{2} \omega A^2 |\alpha| \sin\gamma \,.$$ However, each photon carries energy $\omega$ and momentum ${\bf k}$, so the right hand side of Eq. (\[eq:single\_plane\]) is just the rate at which momentum is being absorbed by the particle due to the absorption of photons. There is of course also some momentum being transferred as a result of photon scattering. However, that effect is proportional to $\alpha^2$, and is being neglected here. Force on a Particle near an Interface {#sec:interface} ===================================== In this section, we will derive a formula for the Casimir force on a polarizable particle in the presence of a single plane interface. The interface will be assumed to have arbitrary reflectivity. We will, however, work in an approximation in which evanescent modes are neglected. The quantized electromagnetic field is to be expanded in a complete set of normalized solutions of the Maxwell equations. These solutions fall into three classes: 1. Modes which are in the region above the interface, and which consist of an incident and a reflected part, as illustrated in Fig. 2. 2. Modes which originate on the far side of the interface, and which are outwardly propagating transmitted waves in the region above the interface. 3. Evanescent modes which are propagating inside of the material comprising the interface, but which are exponentially decaying in the region above it. These modes will be left out of the present discussion. \[Figure 2\] The propagating modes above an interface consist of (1) incident, $I$, and reflected, $R$, waves, or (2) transmitted, $T$, waves. Let us focus first on the reflected modes in Class 1. The net electric field is $${\bf E} = {\bf E}_I + {\bf E}_R \,,$$ where the incident wave is $${\bf E}_I = \hat{\bf \epsilon}\, A\, \cos({\bf k}\cdot{\bf x}_0 -\omega t)\,, \label{eq:incident}$$ and the reflected wave is $${\bf E}_R = \hat{\bf \epsilon'}\, A\, R\, \cos({\bf k'}\cdot{\bf x}_0 -\omega t +\delta)\,.$$ The associated magnetic fields are ${\bf B}_I = {\bf \hat k}\times{\bf E}_I$ and ${\bf B}_R = {\bf \hat k'}\times{\bf E}_R$, respectively. Here the complex reflection (Fresnel) coefficient is $${\cal R} = R\, e^{i \delta} \,,$$ where $R$ is the magnitude of the reflection coefficient, and $\delta$ is the phase shift. This mode induces a dipole moment ${\bf p} = Re(\alpha {\bf E})$, where $\alpha = |\alpha|\, e^{i \gamma}$ is again the complex polarizability. The portions of ${\bf p}$ arising from the incident and reflected waves are, respectively, $${\bf p}_I = \hat{\bf \epsilon}\, A\, \cos({\bf k'}\cdot{\bf x}_0 -\omega t +\gamma)\,.$$ and $${\bf p}_R = \hat{\bf \epsilon}\, A\, R\, |\alpha|\, \cos({\bf k}\cdot{\bf x}_0 -\omega t +\delta +\gamma)\,.$$ The force which a particular mode exerts on the polarizable particle is obtained by inserting the above expressions for the fields and dipole moment into Eq. (\[eq:force0\]). The resulting expression should then be summed over all modes. However, it is simpler first to combine it with the corresponding expression arising from the transmitted waves of Class 2. In the region above the interface, the electric field of these modes is of the form $${\bf E}_T = \hat{\bf \epsilon}\, A\, T\, \cos({\bf k}\cdot{\bf x}_0 -\omega t)\,,$$ where $T$ is a transmission coefficient. Here we may think of the interface as being a slab of finite thickness. Below the slab, these modes have the same form as the incident waves above the slab, Eq. (\[eq:incident\]). If the material in the slab is non-absorptive, then the transmission and reflection coefficients satisfy $$T^2 + R^2 = 1 \,. \label{eq:TR}$$ The force due to the modes of Class 1 can be expressed as a sum of three types of terms, those involving only the incident wave, those involving only the reflected wave, and cross terms between the two. (See Fig. 3.) The first two types of contributions are of the form discussed in the previous section for a single plane wave, as are the contributions due to the Class 2 transmitted waves. As a consequence of the relation Eq. (\[eq:TR\]), these three sets of contributions cancel one another, leaving only the incident-reflected-wave cross terms. The resulting force, for a single mode, is $$F^i = \frac{2 }{3}\, (p_I^j\, \partial_j E_R^i + p_R^j\, \partial_j E_I^i) + \frac{1}{3}\, (p_{Ij}\, \partial^i E_R^j + p_{Rj}\, \partial^i E_I^j) + \frac{2 }{3}\,[ (\dot{\bf p}_I \times {\bf B}_R)^i + (\dot{\bf p}_R \times {\bf B}_I)^i ]\,. \label{eq:force1}$$ \[Figure 3\] The force due to an incident wave, $I$, is cancelled by the sum of the forces due to a reflected wave, $R$, and a transmitted wave, $T$. We next insert the explicit forms for the fields and dipole moment and then average the resulting expression over time. The result is $$\begin{aligned} F^i = \frac{1}{6} \,A^2\, R\, \Bigl\{ &\alpha_1& \bigl[ (k^i - k'^i)({\bf \hat\epsilon}\cdot{\bf \hat\epsilon'}) + 2 \hat\epsilon^i ({\bf k}\cdot{\bf \hat\epsilon'}) - 2 \hat\epsilon'^i ({\bf k'}\cdot{\bf \hat\epsilon}) \nonumber \\ &+& 2 \omega\, {\bf \hat\epsilon'}\times({\bf k}\times{\bf \hat\epsilon}) - 2 \omega\, {\bf \hat\epsilon}\times({\bf k'}\times{\bf \hat\epsilon'}) \bigr] \sin\Delta \nonumber \\ + &\alpha_2& \bigl[ (k^i + k'^i)({\bf \hat\epsilon}\cdot{\bf \hat\epsilon'}) + 2 \hat\epsilon^i ({\bf k}\cdot{\bf \hat\epsilon'}) + 2 \hat\epsilon'^i ({\bf k'}\cdot{\bf \hat\epsilon}) \nonumber \\ &+& 2 \omega\, {\bf \hat\epsilon'}\times({\bf k}\times{\bf \hat\epsilon}) +2 \omega\, {\bf \hat\epsilon}\times({\bf k'}\times{\bf \hat\epsilon'}) \bigr] \cos\Delta \Bigr\} \,. \label{eq:force2}\end{aligned}$$ Here $\Delta = ({\bf k'} - {\bf k})\cdot {\bf x}_0 + \delta$ is the phase difference between the incident and reflected waves at the location of the particle. \[Figure 4\] The wavevectors, ${\bf k}$ and ${\bf k'}$, and polarization vectors, $\hat \epsilon$ and $\hat \epsilon'$, for the incident and reflected parts of an S-polarized wave are illustrated. Let us further evaluate this expression. Let the $z$-direction be perpendicular to the interface and let $\theta$ be the angle of incidence. Then $$k'_z = - k_z = \omega \, c \,$$ where $c = \cos \theta$. Furthermore, $$\Delta = 2\, k'_z\, z + \delta = 2\, \omega\, z\, c + \delta \,.$$ We must now specify the polarization state. We adopt a linear polarization basis, using the usual S (${\bf \hat\epsilon}$ perpendicular to the plane of incidence) and P (${\bf \hat\epsilon}$ parallel to the plane of incidence) states. For S-polarization (Fig. 4), we have $${\bf \hat\epsilon'} = {\bf \hat\epsilon}$$ and $${\bf \hat\epsilon}\times({\bf k}\times{\bf \hat\epsilon}) = {\bf k} \,.$$ Only the $z$-component of the force will remain after summation over all modes; so we need only consider that component. For S-polarization, we find $$F_S^z = - A^2\, R_S\, \alpha_1 \, c\, \sin\Delta \,. \label{eq:force_S}$$ For P-polarization (Fig. 5), we have that $${\bf \hat\epsilon}\cdot{\bf \hat\epsilon'} = \cos 2\theta \,,$$ $${\bf \hat\epsilon}\cdot{\bf \hat k'}= {\bf \hat\epsilon'}\cdot{\bf \hat k} = \sin 2\theta \,,$$ $${\bf \hat\epsilon}\times({\bf k'}\times{\bf \hat\epsilon'}) = -{\bf k} \,,$$ $${\bf \hat\epsilon'}\times({\bf k}\times{\bf \hat\epsilon}) = -{\bf k'} \,,$$ and $$\epsilon_z = - \epsilon'_z = \sin\theta \,.$$ With the aid of these relations, Eq. (\[eq:force2\]) can be written for the case of P-polarization as $$F_P^z = A^2\, R_P\, \alpha_1 \, c\,(1 -2c^2)\, \sin\Delta \,. \label{eq:force_P}$$ Note the force produced by the interference of incident and reflected waves depends upon $\alpha_1$, the real part of the polarizability, rather than on the imaginary part as in Eq. (\[eq:single\_plane\]). \[Figure 5\] The wavevectors and polarization vectors for a P-polarized wave are illustrated. The net force is obtained by integration of $F_S^z + F_P^z$ over all modes for which $k_z \leq 0$: $$F = \int d^3 k \, (F_S^z + F_P^z) = 2 \pi \int_0^\infty d \omega \, \omega^2 \int_0^1 d c\, (F_S^z + F_P^z)\,.$$ The modes are correctly normalized if we set $$A^2 = \frac{4\pi\, \omega}{ (2\pi)^3}\,.$$ This leads to our final result for the force in the direction away from the interface: $$F = \frac{1}{\pi} \int_0^\infty d \omega \, \omega^4\, \alpha_1(\omega)\, \int_0^1 d c\, c\, \Bigl[ -R_S\, \sin(2\, \omega\,z\, c + \delta_S) + R_P\, (1-2c^2)\, \sin(2\, \omega\,z\, c + \delta_P) \Bigr] \,. \label{eq:force3}$$ It is of interest to note that this result may also be derived from an effective interaction Hamiltonian of the form of Eq. (\[eq:H\_int\]), except with the static polarizability $\alpha_0$ replaced by the real part of the dynamic polarizability, $\alpha_1(\omega)$. The interaction potential is given by first order perturbation theory [@TS; @milonni] to be $$\begin{aligned} V &=& \langle H_{int} \rangle = \frac{1}{2 \pi} \int_0^\infty d \omega \, \omega^3\, \alpha_1(\omega)\, \nonumber \\ &\times& \int_0^1 d c\, \Bigl[ -R_S\, \cos(2\, \omega\,z\, c + \delta_S) + R_P\, (1-2c^2)\, \cos(2\, \omega\,z\, c + \delta_P) \Bigr] \,, \label{eq:pot}\end{aligned}$$ so that $$F = - {\bf \nabla} V \,.$$ The Force between a Dielectric Sphere and a Perfectly Conducting Plane {#sec:perfect} ====================================================================== Let us consider the the limit of Eq. (\[eq:force3\]) in which the interface is a perfect conductor. In this limit, we have $$R_S = R_P = 1\, , \label{eq:Rperfect}$$ and $$\delta_S = \delta_P = \pi \,. \label{eq:deltaperfect}$$ This leads to $$\begin{aligned} F &=& \frac{2}{\pi} \int_0^\infty d \omega \, \omega^4\, \alpha_1(\omega)\, \int_0^1 d c\, c^3\, \sin(2\, \omega\,z\, c) \nonumber \\ &=& - \frac{1}{4\pi z^4} \int_0^\infty d \omega \, \alpha_1(\omega)\, \nonumber \\ &\times& \Bigl[ 3\sin 2\omega z -6 z\,\omega\, \cos 2\omega z -6 z^2\,\omega^2\,\sin 2\omega z + 4z^3\,\omega^3\,\cos 2\omega z \Bigr] \,. \label{eq:force4}\end{aligned}$$ Note that there are no evanescent modes in this case; so the previous approximation of ignoring such modes is not needed here. Now consider a sphere of radius $a$ composed of a uniform material with dielectric function $\varepsilon(\omega)$. The complex polarizability is given by $$\alpha(\omega) = a^3\, \frac{\varepsilon(\omega) - 1}{\varepsilon(\omega) +2} \,. \label{eq:alpha}$$ We will take the dielectric function to be that of the Drude model, $$\varepsilon(\omega) = 1 - \frac{\omega_p^2}{\omega(\omega + i\gamma)} \,, \label{eq:epsilon}$$ where $\omega_p$ is the plasma frequency and $\gamma$ is the damping parameter. From Eqs. (\[eq:alpha\]) and (\[eq:epsilon\]), we find that the real part of the polarizability is given by $$\alpha_1 = a^3\, \omega_p^2\, \frac{\omega_p^2 -3 \omega^2}{(3 \omega^2 -\omega_p^2)^2 + 9 \omega^2 \gamma^2} \,. \label{eq:alpha_1}$$ Note that although $\alpha(\omega)$ has poles only in the lower half-$\omega$ plane, its real part, $\alpha_1(\omega)$, has poles in both the upper and lower half planes. If we insert Eq. (\[eq:alpha\_1\]) into Eq. (\[eq:force4\]), we must evaluate the following set of integrals: $$I_1 = \int_0^\infty d \omega \, \alpha_1(\omega)\, \sin(2\, \omega\, z) = Im \int_0^\infty d \omega \, \alpha_1(\omega)\, e^{2\,i\, \omega\,z} \,, \label{eq:I1}$$ $$I_2 = \frac{1}{2} \frac{d I_1}{d z} = \int_0^\infty d \omega \, \alpha_1(\omega)\, \omega\, \cos 2\omega z \,, \label{eq:I2}$$ $$I_3 = -\frac{1}{2} \frac{d I_2}{d z} = \int_0^\infty d \omega \, \alpha_1(\omega)\, \omega^2\, \sin 2\omega z \,, \label{eq:I3}$$ and $$I_4 = \frac{1}{2} \frac{d I_3}{d z} = \int_0^\infty d \omega \, \alpha_1(\omega)\, \omega^3\, \cos 2\omega z \,, \label{eq:I4}$$ In terms of these integrals, the force between the sphere and the plate is $$F = - \frac{1}{4\pi z^4}\, \Bigl( 3\, I_1 -6\,z\,I_2 -6\,z^2\,I_3 + 4 \,z^3\,I_4 \Bigl) \,.$$ The second integral in Eq. (\[eq:I1\]) may be evaluated by rotating the contour of integration to the positive imaginary axis (Fig. 6). However, in this process we will also acquire a contribution from the residue of the pole of $\alpha_1(\omega)$ at $\omega = \Omega + \frac{1}{2}\, i\, \gamma$, where $$\Omega = \frac{1}{6} \sqrt{12 \omega_p^2 - 9 \gamma^2} \,.$$ The result may be written as $$I_1 = J_1 + P_1\,.$$ Here integrating over imaginary frequency yields $$J_1 = \int_0^\infty d \xi \, \alpha_1(i\xi)\, e^{-2\, \xi\, z} = a^3\, \omega_p^2\, \int_0^\infty d \xi \, \frac{3 \xi^2 +\omega_p^2}{(3 \xi^2 +\omega_p^2)^2 - 9 \xi^2 \gamma^2} \, e^{-2 z \xi} \,,$$ and the residue of the pole is $$P_1 = - \frac{\pi\,a^3\, \omega_p^2}{6\, \Omega}\, e^{-\gamma z} \, \cos 2\Omega z \,.$$ \[Figure 6\] The contours of integration for integrals of the form of Eq. (\[eq:I1\]) are illustrated. The integral on real $\omega$ can be expressed as a sum of an integral on imaginary $\omega$, plus a contribution, $C$, coming from the pole at $\omega=\Omega + \frac{1}{2} i \gamma$. These results may be combined to obtain our final expression for the force between the sphere and the plate, which may be written as $$F = J + P \,,$$ where $J$ is the net contribution from integrals along the imaginary axis, and $P$ is that from the pole at $\omega = \Omega + \frac{1}{2}\, i\, \gamma$. The explicit forms of these two contributions are $$J = - \frac{a^3\, \omega_p^2}{4 \pi \, z^4}\, \int_0^\infty d \xi \, \frac{(3 \xi^2 +\omega_p^2)(4 z^3 \xi^3 +6z^2 \xi^2 +6z\xi +3)} {(3 \xi^2 +\omega_p^2)^2 - 9 \xi^2 \gamma^2}\; e^{-2 z \xi} \,, \label{eq:J}$$ and $$\begin{aligned} P &=& - \frac{a^3\, \omega_p^2}{48 \,\Omega\, z^4}\, e^{-\gamma z} \, \Bigl[ 2\Omega \,z\,(4\,\Omega^2\, z^2 - 3\gamma^2\, z^2 -6\gamma z -6) \sin 2\Omega z \nonumber \\ &+& (12\gamma \,\Omega^2\, z^3 -\gamma^3 \,z^3 +12\Omega^2\, z^2 -3\gamma^2\, z^2 -6\gamma \,z -6) \cos 2\Omega z \Bigr] \,. \label{eq:P}\end{aligned}$$ (Here and at other points in this paper, the calculations were performed with the aid of the symbolic algebra program MACSYMA.) In the case that $\gamma =0$, the integral for $J$ may be evaluated in terms of sine and cosine integral functions. In the limit of small separations, one finds for this case that $$J \sim a^3\, \omega_p\, \left( -\frac{\sqrt{3}}{8 z^4} + \frac{\omega_p}{6\, \pi\, z^3} +O(z^{-1}) \right)$$ and that $$P \sim a^3\, \omega_p\, \left( \frac{\sqrt{3}}{8 z^4} + O(z^{0}) \right)\,.$$ Thus the leading terms cancel, and we find a repulsive force in this limit: $$F \sim \frac{a^3\, \omega_p^2}{6\, \pi\, z^3} +O(z^{-1})\,, \qquad a \ll z \ll \omega_p^{-1} \,.$$ It is of particular interest that $P$ contributes an oscillatory term to the force. In the large separation limit, $z \gg 1/\omega_p$, we have that $$J \sim - \frac{3\, a^3}{2 \pi\, z^5} \,. \label{eq:CPforce}$$ This is just the attractive force due to the asymptotic Casimir-Polder potential, Eq.  (\[eq:CP\]), where $\alpha_0 = a^3$ is the static polarizability of the sphere. The oscillatory term becomes, in the large distance limit, $$P \sim - \frac{\Omega\, \omega_p^2\, a^3}{12\, z} \; e^{-\gamma\, z} \; \Bigl( 2\Omega\, \sin 2\Omega z + 3\gamma\, \cos 2\Omega z \Bigr)\,. \label{eq:large_z}$$ Although this term is exponentially decaying, it is possible for it still to be significant in the asymptotic region if, as is typically the case, $\gamma \ll \omega_p$. In this case, the oscillatory term $P$ will dominate the Casimir-Polder term, $J$, and lead to a series of stable equilibrium points at finite distance from the boundary, separated from one another by a distance of approximately $\ell = \pi/\Omega$. A plot of the force at various separations is given in Fig. 7. \[Figure 7 \] The force $F$ between a sphere and a perfectly reflecting wall is illustrated in the case where $\gamma = 0.005 \omega_p$, with $F$ in units of $\omega_p^5 a^3$ and $z$ in units of $\omega_p^{-1}$. The stable equilibrium points are the zeros of $F$ where the slope is negative. Here $F >0$ corresponds to repulsion. The dotted line is the contribution of $J$, the imaginary frequency integral Eq. (\[eq:J\]), and the dashed line is that of $P$, the pole contribution Eq. (\[eq:P\]). One might imagine trying to levitate the spheres in the Earth’s gravitational field by this means. This will occur if $F_{max} \geq F_g$, where $F_{max}$ is $F$ evaluated at a peak value, and $F_g$ is the force of gravity. The ratio of these two forces may be expressed as $$\frac{F_{max}}{F_g} \approx 27\, \left(\frac{\omega_p}{1 eV}\right)^4\, \left(\frac{1 \mu m}{z}\right)\, \left(\frac{1 g/cm^3}{\rho}\right) \, e^{-5\,(\gamma/ 1eV)\,(z/ 1 \mu m)} \,,$$ where $\rho$ is the mass density of the sphere. We have assumed that $\gamma \ll \omega_p$, so $\Omega \approx \sqrt{3}\,\omega_p/3$. Let $z = z_c$ be the distance at which this ratio of forces is unity, and hence the maximum distance above the interface at which levitation can occur. In Table 1, values of $z_c$ for various alkali metals are given, along with appropriate input parameters. ---- -------- ------------ ---------- -------- ------- $\rho$ $\omega_p$ $\gamma$ $\ell$ $z_c$ Li 0.53 6.6 0.031 0.16 49 Na 0.97 5.6 0.028 0.19 46 K 0.86 3.8 0.021 0.28 47 ---- -------- ------------ ---------- -------- ------- : Parameters for some alkali metals. The Drude model parameters, $\omega_p$ and $\gamma$, taken from Ref. [@Palik] , are in eV. The maximum levitation height, $z_c$, and the separation between equilibrium points, $\ell$, are in $\mu m$. The maximum elevation $z_c$ at which a sphere could levitate is in the range of $46\, -\, 49 \mu m$. This is rather larger than the distance at which Casimir forces are usually expected to have a noticeable effect. Recall that all of the discussion in this paper is at zero temperature. Thermal effects at finite temperature can mask this vacuum energy effect. For example, for a sodium sphere of radius $a = 50 nm$ near the maximum levitation height, the difference in potential energy between successive equilibrium points corresponds to a temperature of approximately $0.1 K$, and would be observable only at low temperatures. On the other hand, the corresponding energy difference near the minimum levitation height is about $2000 K$. Thus the first several equilibrium points should be observable at room temperature. The use of a perfectly reflecting wall should be a reasonable approximation so long as the plasma frequency of the material in the wall is large compared to that in the sphere. Thus, a wall composed of aluminum ($\omega_p = 14.8 eV$) [@Palik2] is a good reflector at frequencies of the order of the plasma frequencies of the alkali metals. Discussion {#sec:final} ========== In the previous sections, we have seen that a polarizable sphere with a dispersive polarizability in the vicinity of a perfectly reflecting boundary can experience a Casimir force which is much larger than would be experienced by a perfectly conducting sphere at the same separation. This can be understood in terms of the oscillatory frequency spectrum of vacuum energy effects. Cancellations between different parts of the spectrum which occur in the perfectly conducting limit seem to be upset by the dispersive properties of the sphere’s material. A perfectly reflecting sphere would have a frequency independent polarizability of $\alpha =\alpha_0 = a^3$, and the force exerted by the wall would be given by Eq. (\[eq:CPforce\]) at all separations. In addition to its amplification, the force now becomes an approximately oscillatory function of position, leading to the possibility of trapping the sphere in stable equilibrium. Note that this type of oscillatory force does not arise in the case of a pair of half-spaces of dielectric material separated by a gap. If the material in the half-spaces is a homogeneous dielectric, whose dielectric function satisfies the Kramers-Kronig relations, then the Lifshitz theory [@Lif] predicts a force of attraction which is always less than that in the case of two perfectly conducting planes. Apparently, the effect of the infinite spatial volume of the half-spaces is to average over the spatial oscillations. A similar result was found recently by Lambrecht [*et al.*]{} [@LJR] for the case of mirrors for a scalar field in one spatial dimension. It is of interest to compare the macroscopic sphere problem discussed in this paper with the problem of an atom near a perfect mirror. The case where the atom is in the ground state was discussed in the original Casimir-Polder paper [@CP], where a monotonically decreasing expression was obtained which reduces to Eq. (\[eq:CP\]) in the large $z$ limit. This result is of the same form as the contribution $J$ to the net force found in Section \[sec:interface\] coming from the integration over imaginary frequencies. Various authors [@TS; @milonni; @SDM] have treated the problem of a polarizable particle near an interface. However, these authors were primarily interested in the case where the polarizable particle is an atom in its ground state, and hence included only the imaginary frequency contribution. The case of an atom in an excited state was treated by Barton [@Barton] and other authors [@CPS; @WS; @HS], who found that the potential now has an oscillatory component. Furthermore, this oscillatory term at large distances has a form similar to Eq. (\[eq:large\_z\]), with the magnitude of the oscillatory part decreasing as $1/z$. Thus at large separations, the net potential is dominated by this oscillatory term. In the case of the atom in an excited state, the oscillating potential can be given a classical interpretation: The atom behaves like a radiating antenna in the presence of a mirror. Such an antenna will experience an oscillatory backreaction force whose sign depends upon whether the reflected wave interferes constructively or destructively with the original radiated wave. The oscillatory force found in the present problem does not seem to have such an interpretation, because the dielectric sphere is not radiating. Nonetheless, it is plausible that there should be some parallels between an atom in an excited state and a macroscopic system such as the sphere with a continuum of quantum states just above the ground state. The oscillatory force can be understood in this case as arising from a position dependence of the cancellation of the different parts of the frequency spectrum. One can see from Fig. 1 that a particle whose polarizability is nonzero only in a narrow band of frequency will experience an oscillatory force. (See Ref. [@Ford93] for further discussion of this point.) The delicate cancellation is perhaps one reason that it is difficult to predict the sign of a Casimir force in advance of an explicit calculation. Finally, let us recall the assumptions which were employed in the analysis of this paper. The general formula for the force, Eq. (\[eq:force3\]), was derived in Sect. \[sec:interface\] assuming the scattered wave is dipole and that there are no evanescent modes. The dipole approximation should be valid so long as the size of the particle is small compared to the wavelength of any modes which contribute significantly to Eq. (\[eq:force3\]). The assumption of no evanescent modes places some restrictions on the material of the interface. In particular, a perfectly conducting interface will have no evanescent modes. More generally, in frequency ranges in which the real part of the index of refracion is less than unity, there will be no such modes. This will be the case for all frequencies if the interface is composed of a metal for which the collisionless Drude model (Eq. (\[eq:epsilon\]) with $\gamma =0$) is a good approximation. In Sect. \[sec:perfect\], we made some further approximations. These included the assumption that the particle is a small sphere whose dielectric function has the form given by the Drude model, Eq. (\[eq:epsilon\]). Here the dipole approximation is expected to be valid when $a \ll \omega_p^{-1}$. A final approximation was made in assuming that the interface is perfectly conducting. This is expected to be valid when the interface is composed of a metal whose plasma frequency is large compared to that of the sphere. Then the dominant contributions to Eq. (\[eq:force3\]), those for which $\alpha_1 \not= 0$, come from modes for which Eqs. (\[eq:Rperfect\]) and (\[eq:deltaperfect\]) are approximately valid. The extension of the results of this paper to the case where the interface is an imperfect reflector is currently under investigation. .8cm [**Acknowledgments:**]{} I would like to thank G. Barton, T. Jacobson, P.W. Milonni, V. Sopova, and L. Spruch for useful conversations. This work was supported in part by the National Science Foundation (Grant No. PHY-9507351). [–]{} L.H. Ford, Phys. Rev. D [**38**]{}, 528 (1988). S. Hacyan, R. Jáuregui, F. Soto, and C. Villarreal, J. Phys. A [**23**]{}, 2401 (1990). L.H. Ford, Phys. Rev. A [**48**]{}, 2962 (1993). H.B.G. Casimir and D. Polder, Phys. Rev. [**73**]{}, 360 (1948). J.D. Jackson, [*Classical Electrodynamics*]{}, 3rd ed. (Wiley, New York, 1975) p. 410. Y. Tikochinsky and L. Spruch, Phys. Rev. A [**48**]{}, 4223 (1993). P.W. Milonni,[*The Quantum Vacuum*]{}, (Academic Press, San Diego, 1994) Sect. 8.4. D.W. Lynch and W.R. Hunter, in [*Handbook of Optical Constants of Solids II*]{}, E.D. Palik, ed. (Academic Press, San Diego, 1998) pp. 350, 359, and 369. E.D. Palik, in [*Handbook of Optical Constants of Solids II*]{} (Ref. [@Palik]), p 318. E.M. Lifshitz, Zh. Eksp. Teor. Fiz. [**29**]{}, 94 (1954) \[Sov. Phys. JETP [**2**]{}, 73 (1956)\]. A. Lambrecht, M.-T. Jaekel, and S. Reynaud, Phys. Lett. A [**225**]{}, 188 (1997), quant-ph/9801055. J. Schwinger, L.L. DeRaad, and K.A. Milton, Ann. Phys. (N.Y.) [**115**]{}, 1 (1978). G. Barton, J. Phys. B [**16**]{}, 2134 (1974). R.R. Chance, A. Prock, and R. Silbey, Phys. Rev. B [**12**]{}, 1448, (1975). J.M. Wylie and J.E. Sipe, Phys. Rev. A [**30**]{}, 1185 (1984); [**32**]{}, 2030 (1985). E.A. Hinds and V. Sanoghar, Phys. Rev. A [**43**]{}, 398 (1991). [^1]: email: ford@cosmos2.phy.tufts.edu [^2]: Gaussian units with $c=\hbar=1$ will be used in this paper.

--- author: - 'Dutra-Ferreira, L. [^1]' - 'Pasquini, L.' - 'Smiljanic, R.' - 'Porto de Mello, G. F.,' - 'Steffen, M.' bibliography: - 'ldferreira.bib' date: 'Received; accepted ' subtitle: A benchmark test using the Hyades title: 'Consistent metallicity scale for cool dwarfs and giants [^2]' --- [In several instances chemical abundances of main-sequence and giant stars are used simultaneously under the assumption that they share the same abundance scale. This assumption, if wrong, might have important implications in different astrophysical contexts.]{} [It is therefore crucial to understand whether the metallicity or abundance differences among dwarfs and giants are real or are produced by systematic errors in the analysis. We aim to ascertain a methodology capable of producing a consistent metallicity scale for giants and dwarfs.]{} [To achieve that, we analyzed giants and dwarfs in the Hyades open cluster, under the assumption that they share the same chemical composition. All the stars in this cluster have archival high-resolution spectroscopic data obtained with HARPS and UVES. In addition, the giants have interferometric measurements of the angular diameters. We analyzed the sample with two methods. The first method constrains the atmospheric parameters independently from spectroscopic method. For that we present a novel calibration of microturbulence based on 3D model atmospheres. The second method is the classical spectroscopic analysis based on lines. We also tested two different line lists in an attempt to minimize possible non-LTE effects and to optimize the treatment of the giants.]{} [We show that it is possible to obtain a consistent metallicity scale between dwarfs and giants. The preferred method should constrain the three parameters $T_{\rm eff}$, $\log~g$, and $\xi$ independent of spectroscopy. A careful selection of  lines is also important. In particular, the lines should not be chosen based on the Sun or other dwarfs, but specifically to be free of blends in the spectra of giants. When attention is paid to the line list, the classical spectroscopic method can also produce consistent results. In our test, the metallicities derived with the well-constrained set of stellar parameters are consistent independent of the line list used. Therefore, for this cluster we favor the metallicity of +0.18$\pm$0.03 dex obtained with this method. The classical spectroscopic analysis, using the line list optimized for the giants, provides a metallicity of +0.14$\pm$0.03 dex, in agreement with previous works.]{} Introduction ============ The determination of chemical abundances in stars by spectral synthesis or curve of growth is a rather well-established technique, largely available to most astronomers. A more detailed view of stellar abundances has become possible since the advent of 8-10m class telescopes coupled with high-efficiency spectrographs, which produced high quality spectra for many stars. As a consequence, uncertainties in abundance analyses are mainly dominated by systematic errors rather than by observational limitations regarding spectral resolution and/or signal-to-noise ratio (S/N). In this context, we should recall that the determination of stellar abundances requires a number of assumptions, some of which might not be valid for stars in different evolutionary stages, such as dwarfs and giants. Although cool dwarfs and giants are often analyzed following a single methodology, the differences in their photospheric properties might introduce distinct systematic effects on the final atmospheric parameters derived during the analysis. An example is the large experiment with multiple analysis pipelines conducted within the Gaia-ESO Spectroscopic Survey. In this survey, it was clearly demonstrated that the multiple analysis methodologies perform differently in distinct regions of the parameter space [see @Smiljanic14]. It is thus important to understand if differences on metallicities and abundances between giants and dwarfs are real or produced by systematic errors in the analysis. The interpretation of some open questions in astronomy depends on these kinds of comparisons. For instance, metallicity plays a role in one of the most accepted scenarios in planet formation theories, which is the core accretion scenario. [see, e.g., @Pollack96]. Concerning the gas-giant planet formation, many works have shown that main-sequence FGK-type stars hosting giant planets ($>$ 1.0 $\rm M_{J}$) are usually metal-rich when compared with their counterparts without giant planets [@Gonzalez97; @Santos04; @FisVal05a]. However, a giant planet versus metallicity correlation is not well established for evolved stars. [@Pasquini07] showed that giant stars with planets do not tend to have high metallicities. A similar result was found by other works [@Schuler05; @Takeda08; @Ghezzi10; @Mortier13], but not by [@HekkerMelen07; @Reffert15] who also argued in favor of a planet metallicity correlation in giants. The situation for giants is more complex because of many factors as, for example, the limited sample size of giants hosting planets, inhomogeneities in the planetary properties, and the higher masses of giant stars in comparison to dwarfs. Indeed, there is a stellar mass vs. planet vs. metallicity correlation when we consider the giant planet occurrence frequency in evolved stars. Some authors have reported that there is a trend toward higher metallicities for stars with masses above 1.5$M_{\odot}$ [@Johnson10; @Maldonado13]. Understanding the metallicity distribution of stars with planets would provide an important constraint for planet formation theories. The comparison between the metallicity scale of giants and dwarfs is also important in studies of Galactic chemical evolution. For instance, the metallicity distribution (MD) of the Galactic bulge provides clues about how similar bulge stars are to thin and thick disk stars. As a consequence, it aids the determination of the bulge age. Initial studies of the bulge were focused on giant stars because these are intrinsically brighter objects. Later, the microlensing technique was used to observe dwarfs and subgiants in the Galactic bulge. This revealed discrepancies between the MD of bulge giants and dwarfs [@Cohen08; @Bensby10; @Bensby11]. More recent studies, however, show a better agreement between the two MDs [@Bensby13; @Anders14], although the authors recognize some evidence of a bias toward the high-metallicity tail of the giants’ distribution . Indeed, it is difficult to analyze giants in the high-metallicity regime since their spectra are severely affected by blending and molecules features because of their cool atmospheres. A better picture of the MD of giants requires a full knowledge of the source of this bias in the metal-rich regime. Methodological limitations may affect the analysis of giants and dwarfs differently, e.g., how realistic are the adopted atmospheric models, continuum normalization, and atomic and molecular constants used. The use of the same line list can be a challenge since the intensity of the spectral lines is different in these objects. Moreover, departures from the local thermodynamical equilibrium (LTE) are particularly important for low-gravity and low-metallicity stars [@Asplund05]. Spectroscopic gravities derived by the ionization equilibrium may be unreliable because of overionization. High-temperature dwarfs ($T_{\rm{eff}}$ $>$ 6000 K) may also suffer from non-LTE effects [@Mashon10]. Additionally, the *gf* values need to be very accurate for giants since a differential analysis with respect to the Sun does not cancel out uncertainties in these constants. Open clusters are the optimal sites to evaluate the limitations of abundance analyses when applied to giants and dwarfs. They are made of stars with basically the same distance, and in general, it is reasonable to assume that all stars share the same overall chemical composition, except for the elements affected by mixing in giants [see, e.g., @Takeda08; @Smiljanic09; @Verne13]. Stars in an open cluster also share the same age, outlining a common isochrone curve in the HR diagram, making it easier to constrain stellar parameters in these environments. Few studies so far have attempted the simultaneous analysis of giants and dwarfs in open clusters, aiming to explore possible discrepancies in the abundance patterns between these two classes of objects. [@Pasquini04] investigated dwarfs and giants of the intermediate-age cluster IC 4651, and found, in general, excellent agreement between stars of different evolutionary status. [@Pace2010] investigated the metallicity of five dwarfs and three giants in two open clusters and found differences in the metallicity of up to 0.10 dex for one of the clusters. They also pointed out enhancements of sodium, aluminium, and silicon for the giants. A similar study was performed by @San2009 [@San2012], who investigated the abundance pattern of several open clusters performing a simultaneous and homogeneous spectroscopic analysis. These authors noticed that the discrepancy on the metallicity scale of giants and dwarfs belonging to the same cluster may depend on the line list used. The explanation about the source of these differences is still a matter of investigation. In this work, we chose the Hyades open cluster to perform a simultaneous and homogeneous study of giant and dwarf stars. Our aim is to define an analysis method that can deliver metallicities in a consistent scale for both types of stars. Once this method is tested and established, it can be applied to different astrophysical problems, such as the comparison of MDs of planet host dwarfs and giants. We choose the Hyades, as atmospheric parameters can be constrained by other methods than the classical spectroscopic analysis (see below), and because this cluster is young enough that it should be free of atomic diffusion effects. The paper is organized as follows. Section \[sec:hyades\] reviews the main properties of the cluster, while Section \[sec:data\] presents a description of the data. Section \[sec:analysis\] is dedicated to the analysis and Section \[sec:discussion\] to the discussion of the results. In Section \[sec:conclusion\] we draw our final conclusions. Hyades: The benchmark test {#sec:hyades} ========================== The Hyades is a relatively young cluster, with an estimated age of $\sim$ 625 $\pm$ 50 $\rm{Myr}$ [@P98], and is the closest open cluster to the Sun [$\sim$ 46.5 pc, @vanLeeuwen09]. Spectroscopic studies of FGK-type dwarfs in the Hyades find a metallicity of about +0.13 dex [@Cayrel85; @Boesgaard90; @Paulson03; @Schuler06a]. Regarding the giants, the metallicity values range from +0.10 up to +0.20 dex, where this scatter is usually attributed to the star HIP 20455, a spectroscopic binary [@Schuler06a; @CarPan11]. The slightly over-solar metallicity provides a safe regime to test the classical spectroscopic analysis. At this regime, departures from LTE are not expected to be significant for . Also, close to the solar metallicity, the mean temperature stratification is close to the radiative equilibrium expectation, and therefore, the difference between 1D and 3D model atmospheres is expected to be relatively small [@Asplund05]. A relevant advantage of the stars in the Hyades is that their atmospheric parameters can be very well constrained. From the standpoint of the giants, interferometric measurements of the angular diameters are available, which enables the ** determination of absolute effective temperatures [@Boyajian09]. Furthermore, precise Hipparcos parallaxes [@vanLeeuwen07] are available. Thus, a reliable determination of the surface gravities is possible. Finally, a large amount of spectroscopic data is available for the numerous dwarfs and all four giants of this cluster. All these make the Hyades the optimal benchmark to test the limitations of the classical spectroscopic analysis method. The Hyades is also a target of planet searches using the radial velocity technique [@Cochran02; @Paulson04]. So far, one of the four giants in the Hyades was reported to host a giant planet. The clump giant HIP 20889 has a long period planet ($\sim$ 594 days) with $\sim$ 7.6 $M_{J}$ [@Sato07]. More recently, [@Quinn14] reported the discovery of the first hot Jupiter orbiting a K dwarf HD 285507 in this cluster. Our sample was selected as follows. The dwarfs were selected primarily from the list of [@vanBueren52] and then cross-checked with the reliable sample of cluster members defined by [@P98]. Three additional dwarfs were selected exclusively from [@P98] to complement the cooler end of our sample (these are the stars without the vB number in Table \[tab1\]). Among them, one cool dwarf (HIP 13976) appears slightly away from the cluster’s main sequence. However, this object was classified as a reliable member of the cluster by [@P98] and has radial velocity and distance fully compatible with the cluster distribution. Moreover, this star is present in many analyses of the cluster [e.g., @Paulson03; @Yong04; @Schuler06a]. Several studies investigated binaries in the Hyades and our stars were chosen to avoid binary systems. Among the giants, we excluded HIP 20885, as it is a spectroscopic binary (SB1) with a blue companion. [@Torres97] estimated that the secondary star contributes about 3% of the flux of the primary. An accurate abundance analysis should take this contribution into account. Finally, whenever possible, we chose stars that have been studied in previous works for comparison purposes [in particular, @Paulson03; @Schuler06a]. Table \[tab1\] presents the basic data of the 14 dwarfs and the three giants selected for our sample. Their position in the CMD is shown in Fig. \[fig1\]. In this figure, the magnitude $V$ and the $(B-V)$ color are from [@P98]. We highlight that our sample encompasses stars in a large range of effective temperatures (4700 K $\le$ $T_\mathrm{eff}$ $\le$ 6200 K). This facilitates the investigation of possible systematic effects from the analysis as a function of this parameter. ------- ----- ------------ ------- ------------ ------------- ---------------------------- ---------------------------------- ---------- ---------- HIP vB Spec. Type V RA(J2000) Dec(J2000) ($\pi$ $\pm$ $\sigma_\pi$) ($V_{r}$ $\pm$ $\sigma_{V_{r}}$) $M_\sun$ S/N (mas) ($\mathrm{km~s^{-1}}$) @ 609 nm (a) (b) (c) (d) (e) 20205 28 K0III 3.66 04:19:47.6 +15:37:39.5 21.17$\pm$1.17 +39.28$\pm$0.11 2.48 400 20455 41 K0IV 3.77 04:22:56.1 +17:32:33.0 21.29$\pm$0.93 +39.65$\pm$0.08 2.48 380 20889 70 K0III 3.52 04:28:36.9 +19:10:49.5 21.04$\pm$0.82 +39.37$\pm$0.06 2.48 440 13976 ... K2.5V 7.95 03:00:02.8 +07:44:59.1 42.66$\pm$1.22 +28.35$\pm$0.18 0.83 220 16529 4 G5D 8.88 03:32:50.1 +23:41:31.9 22.78$\pm$1.26 +32.72$\pm$0.17 0.87 210 18946 ... K5D 10.13 04:03:39.0 +19:27:18.0 23.07$\pm$2.12 +36.93$\pm$0.26 0.75 150 19098 ... K2D 9.29 04:05:39.7 +17:56:15.7 19.81$\pm$1.39 +37.61$\pm$0.05 0.88 160 19148 10 G0V 7.85 04:06:16.1 +15:41:53.2 21.41$\pm$1.47 +38.04$\pm$0.17 1.08 310 19781 17 G5V 8.46 04:14:25.6 +14:37:30.1 21.91$\pm$1.27 +39.24$\pm$0.06 0.97 290 19793 15 G3V 8.09 04:14:32.3 +23:34:29.8 21.69$\pm$1.14 +38.21$\pm$0.23 1.01 320 19934 21 G5D 9.15 04:16:33.5 +21:54:26.9 19.48$\pm$1.17 +38.46$\pm$0.19 0.92 230 20130 26 G9V 8.63 04:18:57.9 +19:54:24.1 23.53$\pm$1.25 +39.58$\pm$0.06 0.93 330 20146 27 G8V 8.46 04:19:08.0 +17:31:29.1 21.24$\pm$1.32 +38.80$\pm$0.08 0.94 300 20899 73 G2V 7.85 04:28:48.3 +17:17:07.7 21.09$\pm$1.08 +39.37$\pm$0.06 1.06 470 21112 88 F9V 7.78 04:31:29.3 +13:54:12.5 19.46$\pm$1.02 +40.98$\pm$0.31 1.13 320 22422 118 F8D 7.74 04:49:32.1 +15:53:19.5 19.68$\pm$0.96 +42.04$\pm$0.14 1.10 320 22566 143 F8D 7.90 04:51:23.2 +15:26:00.5 17.14$\pm$1.00 +42.92$\pm$0.19 1.17 250 ------- ----- ------------ ------- ------------ ------------- ---------------------------- ---------------------------------- ---------- ---------- ![Color magnitude diagram (CMD) of the Hyades. The large black triangles correspond to the sample selected for our analysis. The small gray dots are the other cluster members. The magnitude $V$ and the $(B-V)$ color are from [@P98] and all stars shown fulfill the membership criterion of this work.[]{data-label="fig1"}](fig1.eps){width="45.00000%"} Observations and data reduction {#sec:data} =============================== All spectra used in this work were downloaded from the European Southern Observatory (ESO) science archive facility[^3]. The giants were observed during ESO program 088.C-0513 with the HARPS [High Accuracy Radial velocity Planet Searcher, @Mayor03] high-resolution spectrograph (R = 110000), situated at the 3.6 m telescope in La Silla, Chile. The spectral range covers from 3800 to 6900 Å, with a small gap between 5300–5330 Å because of the arrangement of the CCD mosaic. The average S/N @6109 Å is about 400. All the spectra were reprocessed by the last version of the HARPS pipeline (Data Reduction Software version 3.5). We only carried out the Doppler correction and the continuum normalization with standard IRAF[^4] routines. The dwarfs were observed during ESO program 70D-0356 with UVES [Ultraviolet and Visual Echelle Spectrograph, @Dekker00] at the 8.2m Kueyen telescope of the VLT (Very Large Telescope) with and spectral resolution of R $\sim$ 60000. For consistency, we used spectra acquired with the same instrument configuration for all the observations: slit width of 0.8$\arcsec$ and central wavelength at 580 nm in the red arm. The wavelength coverage is from 4780–6800 Åwith a gap between 5750–5830 Å because of thee arrangement of the CCD mosaic. The S/N varies from 150 to 400. Data reduction was carried out with the ESO Reflex environment [@Freudling13] version 2.3, with the exception of Doppler correction, which was performed with IRAF routines. We adopt two spectra as solar proxies. For comparison with the giants, we use the reflected spectrum of the Jupiter’s moon Ganymede[^5], which was obtained with the same HARPS configuration as the spectra of the giants. For the comparison with the dwarfs, we use the UVES solar spectrum[^6] , which was obtained with the moonlight illuminating the slit. Both spectra have, on average, a S/N $\geq$ 300. Analysis {#sec:analysis} ======== The atmospheric parameters (effective temperature, $T_{\rm eff}$, surface gravity, $\log~g$, and microturbulence, $\xi$) of the sample stars were determined using two different methods. In this way, we can compare the final metallicity scale obtained under different assumptions. The first method, hereafter M1 and described in Section \[sec:m1\], takes constraints into account that do not depend on the classical spectroscopic method, but that are still fine-tuned using and lines. The second method, hereafter M2 and described in Section \[sec:m2\], is the classical spectroscopic analysis where the parameters are determined using the and ionization and excitation equilibria. In addition, in the implementation of each method, we make use of two different line lists. We, therefore, derived four sets of atmospheric parameters for each star. Before describing the two methods, we present the two line lists and discuss the measurement of equivalent widths (EWs). Line lists selection -------------------- The two line lists that we adopted were assembled with two goals. First, we aim to minimize non-LTE effects as it will differentially affect stars of different gravities and temperatures. Therefore, the first list contain a set of lines for which non-LTE effects would be minimized or neglected, at least in the metallicity regime of the Hyades. [@Mashon11] evaluated the non-LTE line formation of the two ions of iron in cool reference stars, some of them, with metallicity comparable to the Hyades. In that study, the authors concluded that non-LTE effects are virtually negligible for lines (i.e., they affect the abundances by less than 0.01 dex) and are very small for lines, for stars of metallicity slightly higher than the Sun. We selected a total of 42 and 15 , from their line list, among weak to moderately strong transitions, which were well isolated and as free as possible from blending features. We excluded 14 lines of the original list for which the measured EW in the Sun was in the saturated regime of the curve of growth. [@Mashon11] claimed that the accuracy of their iron abundances might be affected by the uncertainties in the *gf* values that were used. Their *gf* values were obtained from experimental measurements collected from different papers. We improved some of the *gf* values in this line list with more recent determinations that were kindly provided by Maria Bergemann (private communication). Table \[tab:mash\] lists the selected lines, their atomic data and, in addition, the equivalent widths and individual abundances obtained for the solar reflected spectrum of Ganymede using M1. This line list is hereafter referred to as MASH. The second list was chosen to be suitable for the analysis of giants. Often, line lists assembled for the analysis of the Sun may not be optimized for giants because of the more pronounced spectral transitions present in these stars. We have used a line list with transitions carefully chosen to avoid blends in giants and with accurate *gf* values determination provided through the courtesy of Dr. Martin Asplund (private communication). This list includes a total of 34 and 7 transitions and hereafter is referred to as ASPL. Table \[tab:aspl\] lists the lines, as well their atomic data, and the equivalent widths and individual abundances for the solar reflected spectrum of Ganymede using M1. There are 15 lines in common between the MASH and ASPL lists. Nevertheless, the abundances in the Sun are sometimes distinct because of the different atomic data adopted in each list. This difference is greater than 0.05 dex for about 20$\%$ of the lines, while for the remaining it is about $\sim$ 0.02 dex on average. Equivalent width measurements {#sec:ews} ----------------------------- We used the code ARES [Automatic Routine for line Equivalent widths in stellar Spectra; @Sousa07] to perform automatic measurements of the EWs of the and lines. This code applies a Gaussian fit to the profile of the absorption lines, taking a local continuum into account, which is determined on the basis of a set of input parameters provided by the user. We tested different combinations of input parameters for both UVES and HARPS spectra, and chose those which visually produced the best fits to the line profile. The best fits were determined by a visual check and by comparing the EWs computed with ARES and the EWs computed with IRAF for a given set of stars. These were the same stars selected for the EWs comparison between IRAF and ARES (see below). With this approach, the input fitting parameters for ARES were optimized. The spectral resolution we use is such that, in general, the instrumental profile dominates the observed profile, and therefore, a Gaussian fit can well reproduce the observed profile of the lines. Additionally, we removed all EWs $<$ 5 mÅ from our analysis to avoid lines severely affected by noise or uncertainties related to the continuum fit, therefore, with larger relative errors in the EW measurement. We also removed EWs $>$ 120 mÅ, to avoid the flat part of the curve of growth, in which a Gaussian fit may not adequately reproduce the observed line profiles. The quality of the automatic measurements is comparable to the manual method obtained with the task *splot* of IRAF for the majority of the HARPS and UVES spectra as shown by [@Sousa07]. Notwithstanding, we repeated this comparison with the line lists used in our work to verify if the manual EWs obtained with *splot* are comparable with those computed using ARES for both giants and dwarfs. {height="8cm" width="6cm"} {height="8cm" width="6cm"} {height="8cm" width="6cm"} Figure \[fig:ews\] shows the results of such comparison. In addition to the solar spectrum, we select as examples the hot dwarf HIP 22566, where the continuum placement is more challenging because of the enhanced rotation, and the giant HIP 20205, where we have fewer good lines available for the analysis. There is in general a good agreement between the automatic and the manual measurements. The mean difference between the measurements is 0.717 $\pm$ 1.505 mÅ for Ganymede, 1.907 $\pm$ 1.933 mÅ for HIP 20205, and 1.074 $\pm$ 2.413 mÅ for HIP 22566. As seen in Fig. \[fig:ews\], the slope and the intercept values of the linear regressions between the two sets of EWs are close to one and zero, respectively. Although the difference between the two sets of EWs is small, it is not negligible, but the comparison indicates the same trend for all stars tested, i.e, the EWs obtained with *splot* are systematically higher than those obtained with ARES. To quantify the effect of the EW measurement differences on the final metallicities, we computed the abundances for the stars shown in Fig \[fig:ews\] using both the automatic and the manual measurements. The mean difference of the metallicities is $\sim$ 0.03 dex for the Sun, and $\sim$ 0.04 dex for both HIP 22205 and HIP 22566. Based on the small differential variation found in the derived abundances, we therefore decided to measure all the EWs using ARES. We visually inspected the fits for each line, and removed from our analysis those fits that were judged of poor quality. Bad line fits were mainly attributed to poor estimates of the local continuum regions or lines that were significantly affected by noise. Method M1: Well-constrained parameters independent of spectroscopy {#sec:m1} ------------------------------------------------------------------ The first method used to constrain the atmospheric parameters makes use of input values that have been determined independent of the classical spectroscopic method. With this new approach, we expect to be able to investigate which are the possible systematics that can arise from the classical spectroscopic analysis. For the red giants, we adopted the direct determinations of $T_\mathrm{eff}$ by [@Boyajian09]. Angular diameters were obtained with long-baseline optical interferometry and transformed into linear radii using the Hipparcos parallaxes. The direct estimations of $T_\mathrm{eff}$ are calculated combining these radii and the bolometric flux of the star [see @BohmVit81 for the definition of direct measurements of $T_\mathrm{eff}$ for FGK stars]. The bolometric fluxes were determined using the bolometric corrections from [@AllendePrieto99] and assuming $M_{\rm BOL,\odot}\,=\,$4.74. Effective temperatures derived by this method can reach an accuracy of 1$\%$ and are the state of the art in $T_\mathrm{eff}$ determinations. For the dwarfs, as interferometric angular diameters are not available, we obtained $T_\mathrm{eff}$ from photometric calibrations derived with the InfraRed Flux Method (IRFM). Although the IRFM provides semidirect estimations of $T_\mathrm{eff}$, the temperature scale is almost model independent. We use the $JHK$s magnitudes from 2MASS [@2MASS] and the $(B-V)$ color from the Hipparcos catalog [@vanLeeuwen07]. We adopt the calibrations from [@Casagrande10] to derive photometric temperatures using the colors $(V-J)$, $(V-H)$, $(V-K_{s}),$ and $(B-V)$. The adopted photometric $T_\mathrm{eff}$ was calculated with an average of the four different temperatures estimates weighted by the errors of the calibrations. Effective temperatures derived by this method reach and accuracy of a few percent [@Casagrande14]. Since the $T_\mathrm{eff}$ of the giants were derived from interferometric measurements and those of the dwarfs were derived from IRFM, it is interesting to evaluate how well the two methods agree for the giants. The IRFM $T_\mathrm{eff}$ for the giants are shown in Table \[teffphot\]. [@daSilva06] showed that 2MASS colors are unsuitable to determine $T_\mathrm{eff}$ of bright stars. For this reason, we used only the $(B-V)$ color combined with the calibrations of [@GonzalezHer09], [@Ramirez05] and [@Alonso96], which are more appropriated for evolved stars. Interferometric effective temperatures ($T_\mathrm{eff}$(int)) obtained from [@Boyajian09] are also shown for comparison in Table \[teffphot\]. The average IRFM $T_\mathrm{eff}$ are in a excellent agreement with the interferometric values within less than 50 K, as shown in the last column of Table \[teffphot\]. A difference between individual IRFM and interferometric $T_\mathrm{eff}$ as large as 100 K is found only for the star HIP 20889 when using the calibration of [@GonzalezHer09], which gives values systematically lower than the others. HIP $T_\mathrm{eff}$(GH09) $T_\mathrm{eff}$(AL96) $T_\mathrm{eff}$(RM05) $<T_\mathrm{eff}>$ $T_\mathrm{eff}$(int) $\Delta$$T_\mathrm{eff}$ ------- ------------------------ ------------------------ ------------------------ -------------------- ----------------------- -------------------------- -- -- -- 20205 4782 4874 4892 4849 4844 5 20455 4778 4870 4888 4845 4826 19 20889 4720 4810 4820 4783 4827 -44 Surface gravities were determined according to the following equation: $$\rm{log}(g_{*}/g_{\odot})=\rm{log}(M_{*}/M_{\odot}) +4~\rm{log}(T_{\rm{eff*}}/T_{\rm{eff\odot}})-\rm{logg}(L_{*}/L{\odot}).$$ For the dwarfs, the masses were computed with the theoretical evolutionary tracks of [@Girardi2000] and a Bayesian estimation method, which takes the error of each quantity for the mass determination into account [see @PARAM for details]. Luminosities were taken from [@Bruijne01] and the $T_\mathrm{eff}$ are those computed as mentioned above. For the giants we adopted a value of mass of 2.48 $M_\sun$; the mass of a clump giant in a @Girardi02 isochrone of 625 Myr and \[Fe/H\] = +0.13 [see @Smiljanic12 for details]. One free parameter that can not be constrained by our observational knowledge of the cluster is the microturbulence velocity ($\xi$). For M1, we fixed $\xi$ adopting predictions obtained with 3D atmospheric models. The details on how these values were obtained are described in Section \[sec:micro\]. To calculate the metallicities, we used 1D-LTE plane-parallel atmospheric models computed using the Linux version of the ATLAS9 code [@Sbordone04; @Sbordone05] originally developed by Kurucz [see, e.g., @Kurucz93] and adopting the ODFNEW opacity distribution from [@CastelliKurucz03]. The mixing length parameter adopted was 1.25 and no overshooting was considered for both giants and dwarfs. Chemical abundances of and were derived using WIDTH package [@Kurucz93], under some minor optimizations to facilitate handling the input data. Our main innovation in method M1 is that the spectroscopy independent parameters described above were used as input values and subjected to a further controlled fine-tuning. The best set of parameters for each star was determined as follows. First, we allowed the stellar parameters to vary within a range of conservative errors to find the best agreement between the abundances of the and lines. The ranges were $\pm$50 K for $T_\mathrm{eff}$, with steps of 10 $\mathrm{K}$; $\pm$0.10 $\mathrm{dex}$ for $\log~g$, with steps of 0.05 $\mathrm{dex}$; and $\pm$0.10 $\mathrm{km~s^{-1}}$ for $\xi$, with steps of 0.05 $\mathrm{km~s^{-1}}$. These ranges were chosen because they represent typical error values of the classical abundance analysis. We tested variations using smaller steps, but no major improvement on the final solution was found. Next, after reaching a solution, we applied a 2$\sigma$-clipping to remove lines classified as outliers with respect to the average abundances of the previous iteration. The final set of atmospheric parameters is given by looking for the best possible agreement between the abundances of and in the 3D-plane $T_\mathrm{eff}$ – $\log~g$ – $\xi$. The stellar parameters and metallicities derived with this approach, for giants and dwarfs, with both MASH and ASPL lists, are shown in Table \[tab:paramm1\]. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------ ---------- ------- ------------------------------ ------------------- --------------------------- -------------------- ------------------ ---------- ------- ------------------------------ ------------------- --------------------------- -------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- HIP $T_\mathrm{eff}$ $\log~g$ $\xi$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ $T_\mathrm{eff}$ $\log~g$ $\xi$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ (r)[2-8]{}(l)[9-15]{} 20205 4874 2.61 1.30 7.62$\pm$0.10 22 7.62$\pm$0.09 11 4884 2.61 1.30 7.59$\pm$0.06 20 7.59$\pm$0.06 7 20455 4876 2.59 1.30 7.59$\pm$0.10 22 7.62$\pm$0.09 11 4876 2.59 1.35 7.56$\pm$0.07 20 7.56$\pm$0.06 7 20889 4817 2.65 1.35 7.70$\pm$0.09 18 7.70$\pm$0.10 10 4827 2.60 1.35 7.63$\pm$0.08 18 7.63$\pm$0.06 6 *[average]{} & & & & ***7.64$\pm$0.06** && ***7.65$\pm$0.05** & & & & & ***7.59$\pm$0.04** & & ***7.59$\pm$0.04**\ HIP & $T_\mathrm{eff}$ & $\log~g$ & $\xi$ & log$\epsilon_{\mathrm{FeI}}$ & $\mathrm{N(FeI)}$ & log$\epsilon_{\rm{FeII}}$ & $\mathrm{N(FeII)}$ & $T_\mathrm{eff}$ & $\log~g$ & $\xi$ & log$\epsilon_{\mathrm{FeI}}$ & $\mathrm{N(FeI)}$ & log$\epsilon_{\rm{FeII}}$ & $\mathrm{N(FeII)}$\ (r)[2-8]{}(l)[9-15]{} 18946 & 4661 & 4.54 & 0.75 & 7.63$\pm$0.09 & 21 & 7.63$\pm$0.11 & 6 & 4691 & 4.64 & 0.70 & 7.56$\pm$0.09 & 17 & 7.56$\pm$0.21 & 5\ 13976 & 5023 & 4.51 & 0.85 & 7.69$\pm$0.07 & 22 & 7.69$\pm$0.11 & 8 & 5013 & 4.61 & 0.80 & 7.64$\pm$0.08 & 22 & 7.64$\pm$0.08 & 5\ 19098 & 5178 & 4.59 & 0.85 & 7.62$\pm$0.09 & 22 & 7.62$\pm$0.06 & 9 & 5138 & 4.54 & 0.80 & 7.63$\pm$0.09 & 24 & 7.63$\pm$0.06 & 7\ 16529 & 5237 & 4.51 & 0.80 & 7.64$\pm$0.10 & 22 & 7.64$\pm$0.08 & 10 & 5207 & 4.51 & 0.85 & 7.62$\pm$0.08 & 22 & 7.62$\pm$0.05 & 6\ 19934 & 5361 & 4.57 & 0.90 & 7.64$\pm$0.06 & 23 & 7.64$\pm$0.06 & 9 & 5341 & 4.57 & 0.85 & 7.62$\pm$0.08 & 27 & 7.62$\pm$0.04 & 6\ 20130 & 5531 & 4.45 & 0.95 & 7.66$\pm$0.09 & 28 & 7.66$\pm$0.07 & 10 & 5511 & 4.55 & 0.90 & 7.62$\pm$0.06 & 26 & 7.62$\pm$0.07 & 7\ 20146 & 5563 & 4.33 & 1.00 & 7.67$\pm$0.08 & 26 & 7.67$\pm$0.08 & 10 & 5553 & 4.43 & 0.95 & 7.62$\pm$0.07 & 28 & 7.62$\pm$0.04 & 6\ 19781 & 5641 & 4.35 & 0.85 & 7.69$\pm$0.09 & 30 & 7.69$\pm$0.08 & 10 & 5621 & 4.40 & 0.90 & 7.60$\pm$0.06 & 27 & 7.60$\pm$0.05 & 7\ 19793 & 5831 & 4.31 & 1.00 & 7.73$\pm$0.08 & 29 & 7.73$\pm$0.09 & 10 & 5781 & 4.41 & 0.95 & 7.64$\pm$0.08 & 28 & 7.64$\pm$0.05 & 7\ 20899 & 5916 & 4.31 & 0.90 & 7.66$\pm$0.06 & 26 & 7.66$\pm$0.08 & 10 & 5886 & 4.31 & 0.95 & 7.61$\pm$0.08 & 27 & 7.61$\pm$0.04 & 8\ 19148 & 6021 & 4.37 & 1.00 & 7.61$\pm$0.06 & 28 & 7.61$\pm$0.09 & 10 & 5961 & 4.37 & 0.85 & 7.62$\pm$0.08 & 28 & 7.62$\pm$0.04 & 7\ 22422 & 6074 & 4.34 & 1.05 & 7.66$\pm$0.09 & 32 & 7.66$\pm$0.08 & 10 & 6004 & 4.34 & 1.00 & 7.63$\pm$0.06 & 30 & 7.63$\pm$0.01 & 6\ 21112 & 6161 & 4.36 & 1.10 & 7.60$\pm$0.09 & 34 & 7.63$\pm$0.04 & 8 & 6161 & 4.26 & 1.05 & 7.60$\pm$0.06 & 29 & 7.60$\pm$0.03 & 7\ 22566 & 6251 & 4.30 & 1.10 & 7.69$\pm$0.08 & 27 & 7.69$\pm$0.09 & 9 & 6211 & 4.30 & 1.10 & 7.66$\pm$0.08 & 26 & 7.66$\pm$0.07 & 6\ *[average]{} & & & & ***7.66$\pm$0.04** & & ***7.66$\pm$0.03** & & & & & ***7.62$\pm$0.02** & & ***7.62$\pm$0.02**\ ********** ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------ ---------- ------- ------------------------------ ------------------- --------------------------- -------------------- ------------------ ---------- ------- ------------------------------ ------------------- --------------------------- -------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- The main aspect of M1 is that the final parameters shown in Table \[tab:paramm1\] can vary within a very narrow range, constrained by independent methods. Thus, the errors for the stellar parameters in M1 are the uncertainties in the input parameters. An error of about $\sim$50 K corresponds to the error of the IRFM calibrations and is comparable with the interferometric errors presented in the work of [@Boyajian09 see their Table 4]. For $\log~g$, we adopted an error of 0.10 dex because we have very small errors in the parallaxes, thus the main source of error in the gravities comes from our mass determination. An error of 20$\%$ in the masses changes the surface gravities by $\sim$ 0.10 dex. Since the evolutionary status of the Hyades is well known, we consider the gravity error of 0.1 dex as conservative and most likely the highest source of uncertainty is in the mass-loss estimate for the giants. For $\xi$ we adopted an error of 0.10 $\mathrm{km~s^{-1}}$, which corresponds to the uncertainty estimated with the 3D microturbulence calibration (see Section \[sec:micro\]). The dispersion for the and abundances shown in Table \[tab:paramm1\] correspond to the standard deviation of each abundance distribution. We also evaluate how the uncertainty in the physical parameters of the stars affect the retrieved abundances when using M1. To this end, we assumed the parameters presented in Table \[tab:paramm1\] for the giant HIP 20205 and the dwarf HIP 19148, and then vary $T_\mathrm{eff}$, $\log~g$ and $\xi$ separately for both ASPL and MASH lists. The results are given in Table \[tab:errorM1\]. We emphasize that the variations in and follow the same direction for both, the giant and main-sequence stars. \[!htbp\] [cccccccccc]{}\ El. & $T_\mathrm{eff}$+50 K & $T_\mathrm{eff}$-50 K & $\log~g$+0.10 dex & $\log~g$-0.10 dex & $\xi$+0.10 $\mathrm{km~s^{-1}}$ & $\xi$-0.10 $\mathrm{km~s^{-1}}$\ \ & 0.03 & -0.01 & 0.01 & 0.00 & -0.04 & 0.06\ & -0.04 & 0.03 & 0.05 & -0.05 & -0.02 & 0.01\ \ & 0.04 & -0.04 & -0.01 & 0.01 & -0.02 & 0.06\ & -0.01 & 0.01 & 0.03 & -0.03 & -0.02 & 0.01\ \ \ & 0.03 & -0.03 & 0.00 & -0.01 & -0.05 & 0.04\ & -0.04 & 0.04 & 0.05 & -0.05 & -0.04 & 0.04\ \ & 0.04 & -0.03 & -0.01 & 0.01 & -0.02 & 0.02\ & -0.01 & 0.02 & 0.04 & -0.03 & -0.02 & 0.02\ We obtain a good agreement between the metallicity of giants and dwarfs with this method. Moreover, since we are able to retrieve stellar parameters with a good agreement between and abundances, we do not see evidence for significant LTE departures either among the giants or the dwarfs according to this method. In Table \[tab:paramm1\], we find the metallicities obtained using MASH list are slightly higher than those using the ASPL list. The systematic offset of about $\sim$ 0.04-0.06 dex is likely related to the different selection of lines, but can also be considered part of the internal errors of the analysis, as an accuracy better than 0.05 dex can hardly ever be achieved without a line-by-line differential analysis. We further discuss this behavior in Section \[sec:discussion\]. ### Probing microturbulence velocities with 3D models {#sec:micro} Aside from the different techniques one can adopt to determine the microturbulence velocity, it is important to recall that this parameter arises from a limitation of the classical 1D model atmospheres into fully describing all the velocity fields present in the stellar photosphere. As a consequence, the inclusion of an extra velocity field is required to describe the broadening observed in the lines placed in the partly saturated regime of the curve of growth [@StruveElvey34; @vanParadijs72]. In practice, it is essential to use this parameter to obtain the same abundance for lines with small and large EWs. Thus, the optimal value of $\xi$ is obtained by imposing the absence of a correlation between the abundances and the EWs of a set of lines. This analysis, of course, depends on a good statistics of weak and moderately strong lines. This is usually a challenge in the simultaneous analysis of giant and dwarf stars. One alternative to overcome the limitations cited above is to use 3D atmospheric models to predict $\xi$. Three-dimensional models treat convection in a physically consistent way, without the need of defining free parameters like $\xi$. This approach has been adopted by [@Steffen09; @Steffen13], and it is based on the comparison between lines computed with spectral synthesis using 3D and 1D models. Given a sample of spectral lines, the EWs computed from the 3D model with a fixed abundance are taken to represent the observation. For each line, the 1D abundance is obtained by matching the “observed” 3D EW with the synthetic line profiles derived from the 1D models (using exactly the same atomic line parameters as in the 3D synthesis). The best value of $\xi$ for use with 1D models is taken to be that which eliminates the correlation between line strength and derived 1D abundance. In fact, the best $\xi$ depends somewhat on the choice of the line list. Ideally, the lines should be insensitive to temperature fluctuations and have similar properties as the lines to be used for the abundance determinations. High-excitation lines and lines are an obvious choice. We used 3D hydrodynamic models taken from the CIFIST grid [@Ludwig09][^7] computed with the $\rm{CO^{5}BOLD}$ code[^8] [@Freytag12]. The metallicity of all selected models is solar since this grid does not have models available for stars more metal-rich than the Sun yet. Indeed, the differences between temperature stratification of 3D and 1D models are expected to be small at slightly super solar metallicity. At this regime, the line opacities act toward heating the optically thin layers and, as a result, the mean temperature of the layer is close to radiative equilibrium. We do not expect the small metallicity differences to have any noticeable effect on the derived microturbulence velocities. The required 1D hydrostatic reference models were computed with LHD code using the same stellar parameters and opacity scheme as the 3D models. Table \[tab:micro\] shows the stellar parameters of the computed models used for two giants, four subgiants, and six dwarfs, which were selected to be representative of the Hyades HR diagram. For giants and dwarfs, the models were selected to cover as well as possible the range of temperatures and gravities of our sample (e.g., Table \[tab:paramm1\]). For the subgiants, the models cover the cool part of the temperature range expected for subgiants in the Hyades, but have slightly higher $\log~g$ ($\sim$ 3.50) with respect to what is expected from the cluster theoretical isochrone ($\log~g \sim$ 3.0). Models for subgiants with $\log~g$ smaller than 3.50 were not available in the grid. For computing the $\xi$ values, we used 30 plus 7 lines from the ASPL list and applied a method similar to method 3a of [@Steffen13]. The selected lines are identified in Table\[tab:aspl\]. Here again, lines with EW $>$ 120mÅ were rejected to make this analysis compatible with the clipping criteria that were adopted in both M1 and M2. Briefly, for each line, we computed the equivalent width using the 3D model, $W_\mathrm{3D}$. Then, we computed for the very same lines a 2D curve of growth from the adopted 1D reference model, $W_\mathrm{1D}$($\Delta$log$\epsilon$, $\xi$), where $\Delta$log$\epsilon$ is the abundance difference with respect to the original abundance used in the 3D spectrum synthesis. This grid allows us to find, by interpolation for given $\xi$ value, $\Delta$log$\epsilon_i$ for each line $i$ from the condition $W_{\mathrm{3D}_{i}}=W_{\mathrm{1D}_{i}}$. In other words, this abundance correction, $\Delta$log$\epsilon_i$, is the difference of the abundance computed in the 1D model by fitting the equivalent width of the 3D line, and the true abundance used in the 3D spectrum synthesis. We computed $\Delta$log$\epsilon_i$ for a grid of microturbulence values ranging from 0 to 2 $\mathrm{km~s^{-1}}$ with intervals of 0.1 $\mathrm{km~s^{-1}}$. The only exception was the giant with $T_\mathrm{eff}$ = 5000 K, where the grid was ranging from 0 to 3 $\mathrm{km~s^{-1}}$ in steps of 0.15 $\mathrm{km~s^{-1}}$. For each $\xi$ value, we plot $\Delta$log$\epsilon_i$ as a function of $W_\mathrm{3D}$ (see Fig. \[fig:xi\], left side, for an example). This graph illustrates the classical concept of defining the microturbulence. We then determine the slope of the linear regression from each of the plots described above. These slopes are plotted against the corresponding microturbulence values, as shown in Fig. \[fig:xi\] (right side). The best value of $\xi$ for each star is taken to be that where the slope is zero in this curve. By following the procedure described here, we estimated the microturbulence values for each of the model stars with parameters as listed in Table \[tab:micro\]. ---------------------------------------- ---------------------------------------- {width="50.00000%"} {width="50.00000%"} ---------------------------------------- ---------------------------------------- \[htbp\] Star $T_\mathrm{eff}$ K $\log g$ dex $ \left \langle \xi \right \rangle$ ---------- -------------------- -------------- ------------------------------------- Giant 4477 2.5 0.89 Giant 4968 2.5 1.40 Subgiant 4582 3.2 0.84 Subgiant 4923 3.5 0.90 Subgiant 5432 3.5 1.13 Subgiant 5884 3.5 1.27 Dwarf 4509 4.5 0.71 Dwarf 4982 4.5 0.85 Dwarf 5488 4.5 0.87 Dwarf 5865 4.5 0.95 Dwarf 6233 4.5 1.02 Dwarf 6456 4.5 1.10 : Microturbulence values computed from 3D models in correspondence with 1D models.[]{data-label="tab:micro"} We used the data listed in Table \[tab:micro\] to establish an empirical relation of the microturbulence as a function of effective temperature and surface gravity. We tested different functional forms for the calibration and adopted that which gave the best statistical response. We searched for a calibration that shows no significant trend in the residuals distribution, but we have also assessed the quality of the fit by inspecting its correlation coefficient ($R$), the standard deviation $\sigma$ of the fit, and the $p-$values for each term of the calibration. The best fit was found for the following equation: $$\begin{split} \xi \,\,(\rm{km\,s^{-1}}) = 0.998 + 3.16\times10^{-4}\,X - 0.253\,Y\\ -2.86 \times 10^{-4}\,X\,Y + 0.165\,Y^2\, , \label{eq:micro} \end{split}$$ where $X \equiv T_{\rm eff} - 5500$ \[K\] and $Y \equiv \log g - 4.0$. The rms scatter of the residuals of this relation is 0.05 $\mathrm{km~s^{-1}}$. However, we assume as the total uncertainty of the calibration a more conservative value of 0.10 $\mathrm{km~s^{-1}}$ because of the small number of stars used in the fit. Although other calibrations in the literature [e.g., @E93; @F98; @Bruntt12] have considered a larger number of stars, their estimates of microturbulence are based on spectroscopic analyses with different line lists, while Eq. (\[eq:micro\]) presents a relation that reflects predictions from 3D models with a single line list, albeit for a limited set of stellar parameters. The results for the microturbulence values on Table \[tab:micro\] are consistent with those presented in Table 3 of [@Steffen13], although these authors adopted a different selection of lines ( lines with a lower excitation potential greater than 2eV only) and a slightly different method (their method 3b). In this case, the microturbulence is given by the value that minimizes the scatter of the abundance corrections $\Delta$log$\epsilon_i$. [@Steffen13] prefer method 3b over 3a, arguing that 3a is more susceptible to details of the spectral line sample. Nevertheless, we decided to follow 3a because it is closer to the usual analysis applied to determine $\xi$ in the literature, and thus facilitates the comparison with values derived in a classical spectroscopic analysis. We remark that our values display a trend for increasing $\xi$ toward higher temperatures and (perhaps) lower gravities. However, the limited number of stars tested here is not sufficient to provide more than a rough idea of the microturbulence behavior across larger areas of the HR diagram. More details about these trends can be found in [@Steffen13]. Method M2: The classical spectroscopic analysis {#sec:m2} ----------------------------------------------- The second method used to constrain the atmospheric parameters is the classical spectroscopic analysis based on the ionization and excitation equilibria of the and lines. We want to investigate whether a consistent metallicity scale between giants and dwarfs can be recovered through this method. Effective temperatures are calculated, forcing the line abundances to be independent of the excitation potential, i.e., forcing the excitation equilibrium. The microturbulence velocity is determined, forcing the abundances to be independent of the EWs. Surface gravities are calculated, forcing the lines of and to produce the same abundance, fulfilling the ionization equilibrium. As a consequence, the metallicity (\[Fe/H\]) is obtained as a spin-off of this procedure. All these criteria must converge in a fully consistent way and the final solution should be independent of the initial input parameters and iteration path. For this method, we use the same models as in Section \[sec:m1\], i.e, 1D-LTE plane-parallel models (ATLAS9+ODFNEW). We also used the package WIDTH9 for the abundances computation, under some modifications to facilitate the handling of the input/output files. Extra IDL[^9] routines were written to optimize the calculation through the spectroscopic requirements mentioned before. First, the program computes the atmospheric model correspondent to the values of $T_\mathrm{eff}$, $\log~g$, \[Fe/H\], and $\xi$ given as initial guess. The code optimizes one parameter at a time, always checking if the optimization of the previous parameter is still valid. If the previous optimization is still satisfied the code moves forward to the next parameter. If not, the code returns to the previous parameter and recalculates the optimization. The final set of spectroscopic parameters is obtained once all parameters are optimized consistently. An example of the convergence of M2 is shown in Fig. \[fig:m2\] for the star HIP 20205, and for both ASPL and MASH lists. ---------------------------------------- ---------------------------------------- {width="50.00000%"} {width="50.00000%"} ---------------------------------------- ---------------------------------------- The internal errors of the M2 set of spectroscopic stellar parameters were obtained as follows: for the $T_\mathrm{eff}$, we changed the slope of the linear regression in the \[Fe/H\] versus $\chi$ diagram by its own 1$\sigma$ error. The error in temperature is the difference between this new temperature and the previous best value. Similarly, for $\xi$, we applied the same idea, changing the slope of the linear regression in the \[Fe/H\] versus $\log(EW$/$\lambda)$ diagram. The $\log~g$ error is estimated by changing this parameter until we obtain a difference between the or abundances that is equal to the larger of their dispersions. For the metallicity, we adopted the standard deviation of the abundance distribution. Table \[tab:m2\] shows the stellar parameters derived using M2 for ASPL and MASH lists. The metallicities derived by this method agree between giants and dwarfs, within the presented uncertainties. However, the internal metallicity uncertainties of M2 are relatively larger than for M1, when compared within the same line list. On average, the internal uncertainties of M2 are about $\sim$ 0.10 dex when using the MASH list and $\sim$ 0.06 with the ASPL list. The difference in the metallicity scale obtained for both lists are also larger when using M2. These differences are more significant for the giants and for the hottest dwarf of the sample. This result reflects how the line lists were assembled. The MASH line list was selected using the Sun as a proxy and has a larger number of transitions. This reduces the number of well-isolated lines, free from neighboring features, in the spectra of the giants. On the other hand, the ASPL list was chosen to have only isolated and uncontaminated transitions in the spectra of giants, which consequently reduced the number of lines used in the analysis. We discuss the quality of the line lists again in Section \[sec:discussion\]. On average, the difference between the atmospheric parameters obtained with the MASH and ASPL lists for M2 is +68$\pm$92 K for $T_\mathrm{eff}$, +0.12$\pm$0.24 for $\log~g$, $-$0.13$\pm$0.21 for $\xi$ and +0.09$\pm$0.08 for \[Fe/H\]. These values are not significantly different from typical errors of spectroscopic analyses and from typical comparisons between multiple analyses methods [see, e.g., @Hinkel14; @Smiljanic14]. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------ --------------- --------------- ------------------------------ ------------------- --------------------------- -------------------- ------------------ --------------- --------------- ------------------------------ ------------------- --------------------------- -------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- HIP $T_\mathrm{eff}$ $\log~g$ $\xi$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ $T_\mathrm{eff}$ $\log~g$ $\xi$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ (r)[2-8]{}(l)[9-15]{} 20205 4914$\pm$109 2.88$\pm$0.07 1.34$\pm$0.16 7.62$\pm$0.11 25 7.62$\pm$0.07 10 4875$\pm$23 2.71$\pm$0.06 1.43$\pm$0.06 7.53$\pm$0.07 21 7.52$\pm$0.06 7 20455 5010$\pm$162 3.03$\pm$0.09 1.15$\pm$0.21 7.72$\pm$0.13 24 7.72$\pm$0.11 14 4816$\pm$61 2.55$\pm$0.07 1.35$\pm$0.05 7.50$\pm$0.06 19 7.51$\pm$0.07 7 20889 4955$\pm$229 3.36$\pm$0.11 1.08$\pm$0.23 7.86$\pm$0.16 25 7.86$\pm$0.12 12 4833$\pm$38 2.74$\pm$0.08 1.41$\pm$0.06 7.59$\pm$0.09 19 7.59$\pm$0.07 6 *[average]{} & & & & ***7.73$\pm$0.12** && ***7.73$\pm$0.12** & & & & & ***7.54$\pm$0.05** & & ***7.54$\pm$0.04**\ HIP & $T_\mathrm{eff}$ & $\log~g$ & $\xi$ & log$\epsilon_{\mathrm{FeI}}$ & $\mathrm{N(FeI)}$ & log$\epsilon_{\rm{FeII}}$ & $\mathrm{N(FeII)}$ & $T_\mathrm{eff}$ & $\log~g$ & $\xi$ & log$\epsilon_{\mathrm{FeI}}$ & $\mathrm{N(FeI)}$ & log$\epsilon_{\rm{FeII}}$ & $\mathrm{N(FeII)}$\ (r)[2-8]{}(l)[9-15]{} 18946 & 4815$\pm$141 & 4.56$\pm$0.17 & 0.91$\pm$0.10 & 7.50$\pm$0.08 & 20 & 7.49$\pm$0.25 & 10 & 4813$\pm$104 & 4.78$\pm$0.27 & 0.54$\pm$0.20 & 7.57$\pm$0.11 & 18 & 7.57$\pm$0.34 & 7\ 13976 & 4900$\pm$87 & 4.31$\pm$0.09 & 0.70$\pm$0.19 & 7.60$\pm$0.07 & 20 & 7.59$\pm$0.12 & 9 & 4915$\pm$42 & 4.44$\pm$0.08 & 0.63$\pm$0.17 & 7.58$\pm$0.06 & 20 & 7.58$\pm$0.08 & 5\ 19098 & 5120$\pm$75 & 4.54$\pm$0.06 & 1.00$\pm$0.19 & 7.53$\pm$0.07 & 21 & 7.53$\pm$0.06 & 9 & 5025$\pm$62 & 4.35$\pm$0.06 & 0.93$\pm$0.13 & 7.51$\pm$0.08 & 25 & 7.53$\pm$0.07 & 7\ 16529 & 5090$\pm$86 & 4.22$\pm$0.18 & 0.83$\pm$0.13 & 7.55$\pm$0.10 & 23 & 7.55$\pm$0.11 & 11 & 5100$\pm$48 & 4.30$\pm$0.06 & 0.83$\pm$0.10 & 7.53$\pm$0.05 & 20 & 7.53$\pm$0.05 & 6\ 19934 & 5330$\pm$59 & 4.57$\pm$0.07 & 0.81$\pm$0.14 & 7.62$\pm$0.05 & 21 & 7.60$\pm$0.08 & 10 & 5200$\pm$45 & 4.35$\pm$0.08 & 0.81$\pm$0.11 & 7.52$\pm$0.06 & 25 & 7.52$\pm$0.08 & 7\ 20130 & 5590$\pm$104 & 4.73$\pm$0.12 & 0.98$\pm$0.17 & 7.62$\pm$0.02 & 17 & 7.61$\pm$0.04 & 8 & 5505$\pm$40 & 4.56$\pm$0.06 & 1.01$\pm$0.07 & 7.55$\pm$0.05 & 25 & 7.54$\pm$0.05 & 6\ 20146 & 5500$\pm$92 & 4.34$\pm$0.10 & 0.58$\pm$0.12 & 7.68$\pm$0.11 & 27 & 7.68$\pm$0.12 & 11 & 5540$\pm$40 & 4.48$\pm$0.05 & 0.96$\pm$0.08 & 7.57$\pm$0.06 & 27 & 7.58$\pm$0.04 & 6\ 19781 & 5695$\pm$53 & 4.63$\pm$0.08 & 0.78$\pm$0.12 & 7.66$\pm$0.03 & 17 & 7.66$\pm$0.06 & 8 & 5625$\pm$41 & 4.46$\pm$0.05 & 0.96$\pm$0.07 & 7.56$\pm$0.05 & 27 & 7.57$\pm$0.03 & 6\ 19793 & 5790$\pm$88 & 4.47$\pm$0.05 & 0.80$\pm$0.20 & 7.66$\pm$0.08 & 28 & 7.65$\pm$0.04 & 7 & 5710$\pm$34 & 4.32$\pm$0.07 & 1.13$\pm$0.07 & 7.54$\pm$0.05 & 26 & 7.53$\pm$0.05 & 7\ 20899 & 5855$\pm$105 & 4.39$\pm$0.12 & 0.85$\pm$0.17 & 7.62$\pm$0.07 & 28 & 7.62$\pm$0.05 & 8 & 5885$\pm$47 & 4.43$\pm$0.05 & 1.09$\pm$0.07 & 7.58$\pm$0.05 & 26 & 7.58$\pm$0.01 & 5\ 19148 & 5985$\pm$91 & 4.46$\pm$0.07 & 0.75$\pm$0.15 & 7.62$\pm$0.09 & 30 & 7.63$\pm$0.05 & 9 & 5970$\pm$50 & 4.43$\pm$0.05 & 1.09$\pm$0.07 & 7.55$\pm$0.05 & 26 & 7.55$\pm$0.02 & 6\ 22422 & 6029$\pm$104 & 4.30$\pm$0.08 & 0.57$\pm$0.28 & 7.73$\pm$0.10 & 32 & 7.73$\pm$0.10 & 11 & 5975$\pm$53 & 4.40$\pm$0.06 & 0.94$\pm$0.09 & 7.62$\pm$0.06 & 30 & 7.62$\pm$0.02 & 6\ 21112 & 6135$\pm$113 & 4.30$\pm$0.07 & 0.83$\pm$0.24 & 7.66$\pm$0.13 & 36 & 7.66$\pm$0.07 & 11 & 6095$\pm$44 & 4.31$\pm$0.04 & 1.16$\pm$0.09 & 7.55$\pm$0.04 & 24 & 7.55$\pm$0.03 & 7\ 22566 & 6343$\pm$152 & 4.56$\pm$0.11 & 1.24$\pm$0.14 & 7.69$\pm$0.14 & 33 & 7.69$\pm$0.13 & 11 & 6010$\pm$59 & 4.07$\pm$0.08 & 1.16$\pm$0.09 & 7.55$\pm$0.05 & 24 & 7.56$\pm$0.08 & 6\ *[average]{} & & & & ***7.62$\pm$0.06** & & ***7.62$\pm$0.07** & & & & & ***7.56$\pm$0.03** & & ***7.56$\pm$0.03**\ ********** ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------ --------------- --------------- ------------------------------ ------------------- --------------------------- -------------------- ------------------ --------------- --------------- ------------------------------ ------------------- --------------------------- -------------------- -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- Solar abundances and differential analysis ------------------------------------------ We also computed the solar metallicity using our solar proxy spectra. This is useful to understand the behavior of the metallicity scale derived with methods M1 and M2. As mentioned in Section \[sec:data\], for the comparison with the giant stars we used the solar reflected spectrum of Ganymede, hereafter Sun HARPS. For the comparison with the dwarf stars, we used the solar spectrum observed with UVES, hereafter Sun UVES. We decided to have two solar proxies to avoid any inconsistencies that may arise from the use of two different spectrographs and, therefore, different spectral resolution, instrumental profiles, or possible scattered light influence that affects a particular spectrograph. Solar abundances were derived applying the same two methods presented before and for both MASH and ASPL lists. For M1, in particular, we fixed the solar atmospheric parameters to 5777/4.44/0.90 instead of ranging them within its expected errors. For the Sun, we prefer to keep fixed these parameters since its errors are too small to produce a noticeable difference in the metallicity determination. A 2$\sigma$ clipping of the lines was also applied. For M2, the solar parameters were computed exactly as describe in Section \[sec:m2\]. Table \[tab:sun\] shows the abundances of and for our solar proxies, according to M1 and M2, using the ASPL and MASH lists. The solar atmospheric parameters found through M2 are very similar to the canonical values adopted as fixed in M1. The mean differences between M1 and M2 for the solar parameters are 41 K for $T_\mathrm{eff}$, with minimum and maximum values of 16 K and 52 K; 0.01 dex for $\log~g$, with minimum and maximum values of 0.01 dex and 0.02 dex; and 0.09 $\mathrm{km~s^{-1}}$ for $\xi$, with minimum and maximum values of 0.01 $\mathrm{km~s^{-1}}$ and 0.18 $\mathrm{km~s^{-1}}$. Finally, we used the solar values of Table \[tab:sun\] as a reference and derived the abundances of and with respect to the Sun for the Hyades stars. The results of this differential analysis are shown in Table \[tab:m1m2diff\], for M1 and M2 and for both ASPL and MASH lists. Also given are the average metallicities obtained for giants and dwarfs according to each methodology. The internal dispersions presented in the table were obtained by the squared sum of the internal uncertainty relative to the abundance of the star and the Sun. \[t:4.7\] M1 ------------------------------- ------------------------------ ------------------- --------------------------- -------------------- ------------------------------ ------------------- --------------------------- -------------------- -- log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ (r)[2-5]{}(l)[6-9]{} Sun UVES 7.46$\pm$0.10 36 7.45$\pm$0.05 14 7.45$\pm$0.06 31 7.42$\pm$0.02 5 Sun HARPS 7.45$\pm$0.07 37 7.44$\pm$0.06 14 7.45$\pm$0.05 34 7.43$\pm$0.02 7 M2 log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ log$\epsilon_{\mathrm{FeI}}$ $\mathrm{N(FeI)}$ log$\epsilon_{\rm{FeII}}$ $\mathrm{N(FeII)}$ (r)[2-5]{}(l)[6-9]{} Sun UVES 7.47$\pm$0.05 22 7.48$\pm$0.08 15 7.41$\pm$0.03 25 7.41$\pm$0.05 6 Sun HARPS 7.50$\pm$0.05 26 7.49$\pm$0.07 14 7.42$\pm$0.05 33 7.42$\pm$0.02 5 [@c c c c c c c c c c@]{}& &\ & & & &\ HIP & \[FeI/H\] & \[FeII/H\] & \[FeI/H\] & \[FeII/H\] & \[FeI/H\] & \[FeII/H\] & \[FeI/H\] & \[FeII/H\]\ (r)[2-3]{}(l)[4-5]{} (l)[6-7]{}(l)[8-9]{} 20205 & 0.17$\pm$0.12 & 0.18$\pm$0.11 & 0.14$\pm$0.08 & 0.16$\pm$0.06 & 0.12$\pm$0.12 & 0.13$\pm$0.10 & 0.11$\pm$0.09 & 0.10$\pm$0.06\ 20455 & 0.14$\pm$0.12 & 0.18$\pm$0.11 & 0.11$\pm$0.09 & 0.13$\pm$0.06 & 0.22$\pm$0.14 & 0.23$\pm$0.13 & 0.08$\pm$0.09 & 0.09$\pm$0.07\ 20889 & 0.25$\pm$0.11 & 0.26$\pm$0.12 & 0.18$\pm$0.09 & 0.20$\pm$0.06 & 0.36$\pm$0.17 & 0.37$\pm$0.14 & 0.17$\pm$0.10 & 0.17$\pm$0.07\ & [**0.19$\pm$0.07**]{} & [**0.21$\pm$0.05**]{} & [**0.14$\pm$0.03**]{} & [**0.16$\pm$0.03**]{} & [**0.23$\pm$0.12**]{} & [**0.24$\pm$0.12**]{} & [**0.12$\pm$0.04**]{} & [**0.12$\pm$0.04**]{}\ 18946 & 0.17$\pm$0.13 & 0.18$\pm$0.15 & 0.11$\pm$0.11 & 0.14$\pm$0.21 & 0.03$\pm$0.09 & 0.01$\pm$0.26 & 0.16$\pm$0.11 & 0.16$\pm$0.34\ 13976 & 0.23$\pm$0.12 & 0.18$\pm$0.12 & 0.19$\pm$0.10 & 0.22$\pm$0.08 & 0.13$\pm$0.09 & 0.11$\pm$0.14 & 0.17$\pm$0.07 & 0.17$\pm$0.09\ 19098 & 0.16$\pm$0.13 & 0.17$\pm$0.08 & 0.18$\pm$0.11 & 0.21$\pm$0.06 & 0.06$\pm$0.09 & 0.05$\pm$0.10 & 0.10$\pm$0.08 & 0.12$\pm$0.09\ 16529 & 0.18$\pm$0.14 & 0.19$\pm$0.09 & 0.17$\pm$0.10 & 0.20$\pm$0.05 & 0.08$\pm$0.11 & 0.07$\pm$0.14 & 0.12$\pm$0.06 & 0.12$\pm$0.07\ 19934 & 0.18$\pm$0.12 & 0.19$\pm$0.08 & 0.17$\pm$0.08 & 0.17$\pm$0.08 & 0.15$\pm$0.07 & 0.12$\pm$0.11 & 0.11$\pm$0.07 & 0.11$\pm$0.09\ 20130 & 0.20$\pm$0.13 & 0.21$\pm$0.09 & 0.17$\pm$0.08 & 0.17$\pm$0.07 & 0.15$\pm$0.05 & 0.13$\pm$0.09 & 0.14$\pm$0.06 & 0.13$\pm$0.07\ 20146 & 0.23$\pm$0.12 & 0.24$\pm$0.09 & 0.17$\pm$0.09 & 0.17$\pm$0.04 & 0.21$\pm$0.12 & 0.20$\pm$0.14 & 0.16$\pm$0.07 & 0.17$\pm$0.06\ 19781 & 0.23$\pm$0.13 & 0.24$\pm$0.09 & 0.15$\pm$0.08 & 0.18$\pm$0.05 & 0.19$\pm$0.06 & 0.18$\pm$0.10 & 0.15$\pm$0.06 & 0.16$\pm$0.06\ 19793 & 0.27$\pm$0.13 & 0.28$\pm$0.10 & 0.19$\pm$0.10 & 0.22$\pm$0.05 & 0.19$\pm$0.09 & 0.17$\pm$0.09 & 0.13$\pm$0.06 & 0.12$\pm$0.07\ 20899 & 0.20$\pm$0.12 & 0.21$\pm$0.09 & 0.16$\pm$0.10 & 0.19$\pm$0.04 & 0.15$\pm$0.09 & 0.14$\pm$0.09 & 0.17$\pm$0.06 & 0.17$\pm$0.05\ 19148 & 0.15$\pm$0.12 & 0.16$\pm$0.10 & 0.17$\pm$0.09 & 0.20$\pm$0.04 & 0.14$\pm$0.10 & 0.15$\pm$0.09 & 0.14$\pm$0.06 & 0.14$\pm$0.05\ 22422 & 0.20$\pm$0.13 & 0.21$\pm$0.09 & 0.18$\pm$0.08 & 0.21$\pm$0.02 & 0.26$\pm$0.11 & 0.25$\pm$0.13 & 0.21$\pm$0.06 & 0.21$\pm$0.09\ 21112 & 0.14$\pm$0.13 & 0.18$\pm$0.06 & 0.15$\pm$0.08 & 0.18$\pm$0.04 & 0.19$\pm$0.14 & 0.18$\pm$0.11 & 0.14$\pm$0.07 & 0.14$\pm$0.05\ 22566 & 0.23$\pm$0.13 & 0.24$\pm$0.10 & 0.22$\pm$0.10 & 0.25$\pm$0.07 & 0.22$\pm$0.15 & 0.21$\pm$0.15 & 0.14$\pm$0.06 & 0.15$\pm$0.09\ & [**0.20$\pm$0.04**]{} & [**0.21$\pm$0.03**]{} & [**0.17$\pm$0.02** ]{} & [**0.19$\pm$0.03**]{} & [**0.15$\pm$0.06**]{} & [**0.14$\pm$0.06**]{} & [**0.15$\pm$0.03** ]{} & [**0.15$\pm$0.03**]{}\ Results and discussion {#sec:discussion} ====================== The quality of the EWs ---------------------- One step in the spectral analysis that has a substantial influence on the final parameters and abundance is the measurement of the EWs. In this work, EWs were measured with the automatic code ARES. This code been used on many abundance analyses in the literature [e.g., @Adibekyan12; @Tabernero12]. Such kind of automatic codes are a fast and systematic way to compute EWs that minimize the subjectivity of manual measurements using, for example, IRAF. Nevertheless, it has been shown in the analysis of the Gaia-ESO Survey spectra that different groups measuring EWs in the same spectra, using the same code, can still find considerably different values [see the discussion in Section 6 of @Smiljanic14]. This is the case mainly for two reasons: i) there are still some crucial free parameters that need to be adjusted for the optimal measurement of EWs by automatic codes, in particular, for continuum fitting; and ii) quality control of the measured EWs is still important, as the codes just measure all possible EWs without recognizing potential local problems (e.g., unrecognized blends). Regarding the continuum placement, in general, there is no consensus on the best way to define the continuum, which can be done adopting either a local or a global solution. It is not our aim in this paper to advocate in favor of choosing a global or local continuum normalization. We just stress that regardless of the type of normalization that was chosen, care is needed to evaluate whether the continuum solution was of good quality or whether it is affecting the measured EWs in a negative way. As ARES adopts a local continuum normalization, lines in crowded regions of the spectrum can have their EWs underestimated because of a low local continuum solution. This effect can become important in the spectra of giants, where many lines are stronger than in the spectra of dwarfs. Of course, if the line list is composed solely of well-isolated spectral features, the local and the global continuum solution should give similar results. However, this is not always the case, especially when analyzing stars in a wide range of $T_{\rm eff}$. For this reason, as stated before, we were careful to visually inspect and exclude measurements that could be underestimated because of continuum misplacement. This was particularly important in the MASH list, where the lines were basically chosen in the solar spectrum and, consequently, this list has spectral features that can be blended in the spectra of giants. We find that the EWs obtained with IRAF are systematically higher than the EWs obtained with ARES. We attribute this behavior mainly to the different continuum normalization. The larger discrepancies were found for hot dwarfs and for the giants. If we compare the EWs of all transitions, without excluding those with suspicious normalization, the differences between IRAF and ARES reach up to 13.6 mÅ for the hot star HIP 22566 and up to 7.4 mÅ for the giant HIP 20205. The use of all these measurements would certainly lead to an extra source of uncertainties. After the visual inspection and exclusion, the difference in EWs causes an effect on the abundances that is smaller than 0.05 dex (see Section \[sec:ews\]). This is similar to typical errors found in abundance analyses that are not line-by-line differential. We also noticed some difficulties of measuring EWs with ARES in hot stars where rotational broadening becomes to be important. After several tests with different ARES input options, we decided to limit our sample to those stars where we were confident that the EWs were well measured with this automatic tool. Therefore, we do not include stars hotter than $\sim$6300 K. Finally, we also tested the influence of using spectra obtained with two different instruments on the EWs scale. To this end, we used a spectrum of the giant HIP 20455 observed with UVES, under the same configuration set as the dwarfs. We computed metallicities using M1 for the two sets of EWs of this star. The mean differences were 0.03 dex for the ASPL list and 0.05 dex for the MASH list. Although the difference is slightly higher for the MASH list, it is still within the internal uncertainties of this method. Consequently, if there is any systematic error in the metallicity introduced by the use of different spectrographs, this is comparable to the internal errors of our analysis. Therefore, we concluded that the use of spectra from both UVES and HARPS instruments should not compromise the metallicity scale between giants and dwarfs. Comparison between the line lists --------------------------------- The use of different line lists has already been reported as one of the possible sources of the discrepancies between the metallicity scale of giants and dwarfs in open clusters [@San2009; @San2012]. However, in those works only the classical spectroscopic analysis was tested. Here, we are able to compared the spectroscopic method (M2) with a set of parameters that is independent (method M1). Thus, we can test the performance of the different line lists in these different approaches. When considering the analysis performed with the MASH list, the metallicity scale of giants and dwarfs agree better under M1. The difference between the average metallicity of giants and dwarfs is $-$0.02 dex with M1 and increases to 0.11 dex with M2. This increased difference seems to be driven by the giants. The uncertainties of the atmospheric parameters are larger for the giants than for the dwarfs (Table \[tab:m2\]). Moreover, the metallicity and the surface gravity are substantially higher for two of the giants when M2 is applied with this list. On the other hand, with the ASPL list there is a better agreement between giants and dwarfs in both methods. The difference between the average metallicity of giants and dwarfs is $-$0.03 dex for M1 and $-$0.02 dex for M2. We attribute the increased scatter in the metallicity of the giants seen when using the MASH list under M2 to the fact the lines were chosen from the solar spectrum. For this reason, they are likely more affected by blending features that were weak in the Sun, but are significant in the spectrum of a giant star. This result indicates that the use of fewer transitions, but that are carefully chosen to avoid contaminating blends, is fundamental for an accurate determination of atmospheric parameters and of the metallicity scale through the spectroscopic method. We also remark that the improved *gf* values that we adopted for the MASH list did not reduce significantly the scatter of the metallicities. Indeed, our values of metallicity dispersion are around 0.08-0.10 dex. These are similar to the ones reported in [@Mashon11], which are about $\sim$ 0.09-0.11 dex. Perhaps the reduced number of lines in our analysis is not enough to reduce the internal statistical errors even improving the *gf* values sources. For the list ASPL, we found an average dispersion on the metallicity of around 0.07 dex, which is smaller but still closer to the above mentioned values. We found a systematic difference of 0.04-0.06 dex between the abundances derived using MASH and ASPL lists, and higher for the giants when M2+MASH is applied. The metallicities derived with the MASH list are higher than the ones derived with ASPL list but also the total average temperature scale is hotter for the MASH list by about 20 K when using M1 and by about 70 K when using M2 (see Figure \[fig:linelists\]). These variations in $T_{\rm eff}$ correspond to an increase in abundances of approx. 0.02-0.06 dex. In addition, the MASH list contains lines which are likely more affected by blending, thus we would expect that the cool giants show much larger abundances for the MASH list than for the ASPL list, which is the case. Thus, we attribute these systematic differences to the higher retrieved $T_{\rm eff}$ and to the presence of blended lines in the MASH list. More details about the comparison between M1 and M2 are presented in the next section. --------------------------------------- -- {width="50.00000%"} --------------------------------------- -- As mentioned at the beginning of this subsection, many works have demonstrated how abundances (and abundance differences or trends) depend on the adopted line list. From our analysis clearly emerges that a proper line list maximises the robustness of the results. A similar conclusion has been reached by other authors who recently proposed to improve existing line list by customising it to the stars observed, using an empirical approach. This method has been applied to cool dwarfs [@Tsantaki13] and to evolved stars [@Adibekyan15], reaching substantial improvements with respect to what obtained with a not optimised line list. In terms of metallicity dispersion and compatibility of the metallicity scale between giants and dwarfs, the ASPL list shows a good performance for both methods tested in this work. This is likely related to the selection of well isolated line transitions which are equally appropriate for the analysis of giants and dwarfs. We, therefore, recommend the use of such list for this kind of analysis. Comparison between the methods ------------------------------ In this Section, we compare the performance of methods M1 and M2 to investigate possible systematic effects between them. Figure \[fig:methods\] shows the difference between the atmospheric parameters obtained with M1 and M2 as a function of the M1 parameters. For each parameter, we performed a linear regression to check for significant trends. These trends would appear in case of considerable differences between the atmospheric parameters derived with each method. In case of a significant trend, the ratio between the slope and its own uncertainty should be less than 0.5, i.e., $x$ $>$ 2$\sigma_{x}$. ---------------------------------------- ---------------------------------------- {width="50.00000%"} {width="50.00000%"} ---------------------------------------- ---------------------------------------- According to Fig. \[fig:methods\], only $\log~g$ appears significantly different between M1 and M2, and only when using the MASH list (see the middle plot on the left panel of Fig. \[fig:methods\]). On average, the spectroscopy surface gravities are larger when using M2 for both line lists, about 0.10 dex, but in the case of MASH the $\log~g$ values are significantly overestimated. This is probably caused by overestimated EWs, due to blending contaminants, which result in larger effective temperatures and surface gravities. The $\xi$, on the other hand, is higher on average for M1 than for M2. For the ASPL list, the plots shown in Fig. \[fig:methods\] do not indicate any significant difference between the parameters derived with M1 and M2. However, the M1 $T_{\rm eff}$ values are, on average, $\sim$ 47 K higher than the M2 ones. Although this is a small difference, it is probably responsible for the larger metallicities found with M1 in comparison with M2 (in absolute values, 0.05 dex for the giants and 0.06 dex for the dwarfs). The metallicities derived with both line lists show a better agreement when M1 is applied. This behavior can be more easily noticed by inspecting Figure \[fig:linelists\]. The differences between the \[Fe/H\] scale derived with the two line lists with M1 are, on average, about $\sim$0.05 dex, and constrained within the range of $\pm$0.1 dex (dotted lines) of this Figure. For M2, these differences are on average higher about $\sim$0.10 dex, and present a larger scatter especially for the giants and for the hottest dwarf HIP 22566. Figure \[fig:linelists\] shows that a good constraint in the atmospheric parameters of a given set of star can produce a consistent metallicity scale for giants and dwarfs even when using line lists selected with different approaches, not optimized for the analysis of giants. While the agreement between M1 and M2 is, on average, good, there are a few individual stars for which the differences are significant. The $T_\mathrm{eff}$ differences are about 100-150 K for the cooler stars ($T_\mathrm{eff}$ &lt; 5300 K), with both line lists. Such discrepancies for the cooler stars do not compromise the overall agreement between the two $T_\mathrm{eff}$ scales, but this results are in agreement with [@Ramirez07] who argue that differences between the spectroscopic and the photometric temperature scales are more significant for stars with $T_\mathrm{eff}$ $<$ 5000 K. HIP 18946 and HIP 22566 are the stars that present the highest discrepancies between the effective temperatures of M1 and M2, reaching up to $\sim$ 150 K for the MASH list. Finally, we tested the compatibility of the giants temperature scale obtained with the interferometric angular diameters and the IRFM. This last test was done in order too see if, in the absence of angular diameters for the giants, the IRFM $T_\mathrm{eff}$ scale would give similar results. As mentioned in Sect. \[sec:m1\], the agreement between the IRFM and the interferometric $T_\mathrm{eff}$s is excellent. The average absolute difference between them is -6$\pm$33 K. This results indicates that, for similar giants, the IRFM $T_\mathrm{eff}$ could also be used as input values of temperature in M1. In general, the metallicities derived using M1 are more consistent between giants and dwarfs for the two line lists tested here. This results seems to be a consequence of the well constrained set of atmospheric parameters. The metallicities of giants and dwarfs are only consistent when derived using M2 with the ASPL list. When using M2 associated with the MASH list we found systematic differences especially in the giants. Our results confirm that, for the Hyades, the line list is a primary source of systematic differences on the metallicity scale of giants and dwarfs. The Hyades metallicity scale ---------------------------- In this Section we discuss the metallicity scale that we obtained for the Hyades. In general, the two methods tested favor the ASPL over the MASH list, as this last one seems to introduce systematic effects on the results, in particular for M2. Thus, we consider the results obtained with the ASPL list to be of higher quality. There is a good agreement between the metallicity of giants and dwarfs when using this list for both methods. According to M1 the difference between the average values of the iron abundances of giants and dwarfs is 0.03 dex. When M2 is considered, the difference between the metallicity of the giants and the dwarfs is 0.02 dex. These values are smaller than the internal uncertainties of our methods and the typical errors found in classical abundance analysis ($\sim$ 0.05-0.10 dex). Therefore, we consider all the stars together, giants and dwarfs, analyzed with ASPL list to revisit the metallicity scale of the Hyades. We combined the average abundances of for giants and dwarfs of each list to obtain the metallicity of the cluster. We choose abundances because they are safeer abundance indicators than lines, since lines may suffer more of non-LTE effects. We found that the average metallicity of the Hyades is 0.18$\pm$0.03 dex or 0.14$\pm$0.03 dex, according to M1 and M2, respectively, using the ASPL list. These values are compatible with each other within 2$\sigma$. For the MASH list, under the M1 analysis, we found a metallicity for the Hyades of 0.21$\pm$0.04 dex. At this point, we reinforce two main aspects of our work. When investigating giants and dwarfs simultaneously using the classical spectroscopic analysis, it is recommended to use a line list especially suitable for the spectra of giants. In case the line list was selected based on the spectrum of a dwarf star, as the Sun, we recommend that it is more appropriate to use a method in the lines of M1. This minimizes the systematic effects on the metallicity scale retrieved by the analysis. Comparison with other works --------------------------- We present here a comparison of our results with selected previous analyses of Hyades stars: [@Paulson03], [@Schuler06a] and [@CarPan11]. These works were chosen because they have a larger number of stars in common with our analysis. As our final recommendation we select the results obtained with M1 using the ASPL list. [@Paulson03] analyzed a sample of 90 stars of the Hyades, of which 13 are present in our study. Those authors performed a differential spectroscopic analysis to derive the atmospheric parameters of the stars. For the sample in common, the average difference in $T_{\rm eff}$ between [@Paulson03] and this work is 58$\pm$59 K. The metallicity reported by [@Paulson03] is 0.13$\pm$0.01 dex, which is within a 2$\sigma$ range of the values found with M1 (0.18$\pm$0.03 dex), but it is in excellent agreement with our metallicity derived according the same spectroscopic method (M2) of 0.14$\pm$0.03 dex. [@Schuler06a] analyzed the giants and dwarfs of the Hyades using a method similar to M1. The $T_{\rm eff}$ was obtained with the IRFM by . Surface gravities and $\xi$ were determined fitting synthetic spectra (1D+LTE) to the observed spectra. For the giants, the agreement between this work and our results is remarkably good for $\log~g$, $\xi$ and \[Fe/H\], and reasonable for $T_\mathrm{eff}$. The mean differences are 76$\pm$12 K for $T_\mathrm{eff}$, 0.03$\pm$0.06 dex for $\log~g$, 0.06$\pm$0.05 $\mathrm{km~s^{-1}}$ for $\xi$ and 0.02$\pm$0.01 dex for \[Fe/H\]. Among the eight dwarfs analyzed by this work, two were also included in our analysis. For these objects the agreement between the stellar parameters is good for $T_\mathrm{eff}$ and $\log~g$, but the differences increase up to 0.40 $\mathrm{km~s^{-1}}$ for $\xi$. The overall metallicity for the Hyades recommended by these authors is 0.13 dex, which is also compatible within a 2$\sigma$ range of our value retrieved with M1 and ASPL list (0.18$\pm$0.03 dex). Another analysis of the giants of the Hyades was performed by [@CarPan11]. These authors determined the atmospheric parameters via spectroscopy and photometry. The agreement between the set of parameters presented in [@CarPan11] and our work is reasonably good for $T_\mathrm{eff}$ and good for \[Fe/H\], with average values of about $-$79$\pm$53 K and $-$0.03$\pm$0.03 dex, respectively. However, the surface gravities reported by these authors are compatible with those that we derived using M2, the classical spectroscopic analysis. The differences between the surface gravities is likely responsible for the marginal agreement between the metallicity scales: 0.11$\pm$0.01 dex derived by [@CarPan11] and our value of 0.18$\pm$0.03 dex. Although the value of metallicity reported by these authors is in excellent agreement with our value of 0.12$\pm$0.04 dex, which is the determination of the metallicity for the giants also derived using the spectroscopy method (M2 and ASPL list). Conclusions {#sec:conclusion} =========== We investigated the metallicity scale obtained in a simultaneous analysis of dwarfs and giants. Inconsistencies between the metallicities of these types of stars have been reported before in the literature. Understanding whether these inconsistencies are a real or systematic effect of the analysis methods is important for advancing in a number of different areas of astrophysics. Of particular interest is the comparison of metallicities between stars of different evolutionary phases hosting planets. As a test case, we chose a sample of giants and dwarfs in the Hyades open cluster. All stars in an open cluster are expected to share the same initial chemical composition. They are therefore optimal to test the consistency of analysis methods. We computed metallicities using two different methods and, for each method, using two different line lists. One line list was assembled based on the solar spectrum and with the aim of minimizing non-LTE effects in the determination of the atmospheric parameters (MASH line list). The other was assembled with the specific purpose of analyzing giant stars, and therefore only includes lines that are relatively free of blends in the spectra of these objects (ASPL line list). Analysis method M1 was based on atmospheric parameters, which are independent of spectroscopy. Effective temperatures were computed using interferometric angular diameters for the giants, and the IRFM for the dwarfs. Surface gravities were computed with the help of theoretical evolutionary tracks. For use with M1, we presented a new calibration of microturbulence based on 3D hydrodynamical models. In this way, one can also use an estimate of $\xi$ that is independent of spectroscopic measurements. Method M2 was the classical spectroscopic method using and lines to constrain the stellar parameters. We investigate the main steps in the analysis that may affect the metallicity scale of dwarfs and giants of the Hyades. Our careful evaluation of EWs has shown that differences in the continuum placement can have an effect of about $\sim$ 0.03 dex in the precision of the final metallicities. An effect of similar magnitude (0.05 dex) can be introduced using spectra obtained with different instruments. From the stand point of the line lists, we found larger systematic differences between the metallicity scale for giants and dwarfs using MASH list, where the line selection criteria are not optimized for evolved stars. Additionally, we found a systematic difference of 0.04-0.06 dex between the two lists used in this work. Lastly, the difference on the metallicity scale for the giants and dwarfs using the two different methods (M1 and M2) is of about 0.04-0.06 dex, disregarding the results of the giants obtained with M2+MASH list. One interesting aspect is that perhaps the use of the solar spectrum as a differential reference does not cancel out all these effects. Considering all these points, we conclude that the limit of precision of an abundance analysis that is not line-by-line strictly differential cannot be better than 0.03 dex. We show that with a careful determination of atmospheric parameters and a well -selected line list, it is possible to simultaneously analyze giants and dwarfs and obtain consistency between their metallicities. When attention to the line list is considered, M1 and M2 produce results with a good agreement. The metallicity scale of the Hyades obtained with our preferred line list (ASPL) is 0.18$\pm$0.03 dex for M1 and 0.14$\pm$0.03 dex for M2, and these values are consistent among each other. At the very least, it is clear that assembling a line list well suited for the analysis of giants is mandatory to obtain consistency between the metallicity scales. When using the spectroscopy technique, the metallicity scale of giant and dwarfs using ASPL list agree with previous spectroscopic analyses of the Hyades. However, M1 does show a more efficient capability to recover the metallicity scale of giants and dwarfs regardless of the line list used. The results obtained with M1 for both MASH and ASPL lists are consistent among each other, indicating that a set of very well-constrained atmospheric parameters might compensate possible systematics from the features of the line list. Therefore, we believe that M1 results are more robust. In the view of our results, we favor as our final metallicity for the stars in Hyades open cluster the value of 0.18$\pm$0.03 dex (based on M1 and the ASPL line list). For similar studies of giants and dwarfs, in a more general approach, we suggest the use of M1 whenever possible. A simultaneous spectroscopic analysis of giants an dwarfs, as done in M2, is only recommended under the use of a line list optimized for giant stars. L. D. F. acknowledges support through the ESO Studentship Programme, CAPES Programme and CNPq fellowship Programme (166880/2014-0). L. P. acknowledges the Visiting Researcher program of CNPq Brazilian Agency, at the Fed. Univ. of Rio Grande do Norte, Brazil. Support for C.C. is provided by the Ministry for the Economy, Development and Tourism’s Programa Iniciativa Científica do Milênio through grant IC20009, awarded to the Millennium Institute of Astrophysics (MAS). G.F.P.M. acknowledges financial support by CNPq grant 476909/2006-6, FAPERJ grant APQ1/26/170.687/2004 and ESO, for financial support for a trip to Garching, which enabled participation in this work. We thank Dr. Martin Asplund and Dr. Maria Bergemann for their contribution with the line lists used in this work. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France, of NASA’s Astrophysics Data System, of the WEBDA database, operated at the Department of Theoretical Physics and Astrophysics of the Masaryk University, and of data products from the Two Micron All Sky Survey, which is a joint project of the University of Massachusetts and the Infrared Processing and Analysis Center/California Institute of Technology, funded by the National Aeronautics and Space Administration and the National Science Foundation. Specie $\lambda$ (Å) $\chi_{\rm{low}}$ (eV) log *gf* EW (mÅ) log$\epsilon_{\rm{Fe}}$ Clipped -------- --------------- ------------------------ ---------- --------- ------------------------- --------- -- 4445.47 0.09 -5.410 40.9 7.470 4574.72 2.28 -2.970 54.9 7.489 4726.14 3.00 -3.250 18.8 7.588 4808.15 3.25 -2.790 26.9 7.600 4994.13 0.91 -3.002 104.2 7.275 5197.94 4.30 -1.540 36.9 7.525 5198.72 2.22 -2.113 97.2 7.396 5216.27 1.61 -2.082 117.7 7.156 $\ast$ 5217.40 3.21 -1.116 111.3 7.282 5236.20 4.19 -1.497 32.4 7.298 5247.05 0.09 -4.975 66.5 7.530 5250.21 0.12 -4.918 64.6 7.462 5285.13 4.43 -1.540 27.9 7.449 5295.31 4.42 -1.590 29.5 7.518 5379.58 3.69 -1.514 61.5 7.505 5397.62 3.63 -2.528 28.0 7.702 $\ast$ 5491.83 4.19 -2.188 13.4 7.445 5517.06 4.21 -2.370 17.5 7.791 $\ast$ 5522.45 4.21 -1.450 42.9 7.463 5607.66 4.15 -3.437 14.6 8.699 $\ast$ 5638.26 4.22 -0.770 76.1 7.394 $\ast$ 5662.52 4.18 -0.573 92.8 7.402 5679.02 4.65 -0.820 59.0 7.501 5778.45 2.59 -3.430 22.1 7.409 5807.78 3.29 -3.410 8.5 7.563 5852.22 4.55 -1.230 39.3 7.470 5855.08 4.61 -1.478 23.8 7.441 5858.78 4.22 -2.160 12.2 7.383 5916.25 2.45 -2.914 54.1 7.471 5930.18 4.65 -0.230 86.8 7.335 6065.48 2.61 -1.470 116.7 7.285 6082.71 2.22 -3.571 33.7 7.456 6105.13 4.55 -2.532 12.8 8.084 $\ast$ 6151.62 2.18 -3.312 50.0 7.494 6200.31 2.61 -2.405 73.4 7.497 6213.43 2.22 -2.481 82.6 7.386 6229.23 2.85 -2.805 37.6 7.379 $\ast$ 6252.56 2.40 -1.727 119.8 7.437 6421.35 2.28 -2.020 103.7 7.326 6481.88 2.28 -2.985 62.4 7.497 6518.37 2.83 -2.373 59.0 7.361 6608.03 2.28 -3.930 17.6 7.428 4491.40 2.84 -2.710 77.8 7.468 4508.29 2.84 -2.440 84.8 7.313 4582.83 2.83 -3.180 57.0 7.455 4620.52 2.82 -3.210 50.3 7.304 5197.58 3.22 -2.220 81.4 7.315 5264.81 3.22 -3.130 44.3 7.434 5284.11 2.88 -3.195 59.9 7.525 5414.07 3.21 -3.580 25.6 7.414 5425.26 3.20 -3.220 41.5 7.429 5991.38 3.15 -3.647 29.5 7.506 6239.95 3.89 -3.573 13.8 7.635 $\ast$ 6247.56 3.89 -2.435 52.1 7.512 6369.46 2.89 -4.110 18.6 7.407 6432.68 2.89 -3.570 41.3 7.442 6456.38 3.90 -2.185 61.5 7.456 Specie $\lambda$ (Å) $\chi_{\rm{low}}$ (eV) log *gf* EW (mÅ) log$\epsilon_{\rm{Fe}}$ 3D–$\xi$ list -------- --------------- ------------------------ ---------- --------- ------------------------- --------------- 5044.21 2.851 -2.058 73.9 7.396 5225.52 0.110 -4.789 70.9 7.482 $\dag$ 5247.05 0.087 -4.946 66.5 7.501 $\dag$ 5250.21 0.121 -4.938 64.6 7.481 5651.47 4.473 -1.750 18.6 7.444 5661.35 4.284 -1.756 22.0 7.369 $\dag$ 5679.02 4.652 -0.750 59.0 7.431 $\dag$ 5701.54 2.559 -2.216 83.7 7.504 $\dag$ 5705.46 4.301 -1.355 38.6 7.373 $\dag$ 5793.91 4.220 -1.619 33.3 7.435 $\dag$ 5809.22 3.883 -1.710 50.5 7.516 $\dag$ 5855.08 4.608 -1.478 23.8 7.441 $\dag$ 5956.69 0.859 -4.605 52.3 7.532 $\dag$ 6027.05 4.076 -1.090 63.1 7.397 $\dag$ 6065.48 2.609 -1.530 116.7 7.345 $\dag$ 6093.64 4.607 -1.300 30.8 7.395 $\dag$ 6096.67 3.984 -1.810 37.5 7.455 $\dag$ 6151.62 2.176 -3.299 50.0 7.481 $\dag$ 6165.36 4.143 -1.460 44.9 7.454 $\dag$ 6173.33 2.223 -2.880 67.7 7.484 $\dag$ 6200.31 2.609 -2.437 73.5 7.532 $\dag$ 6213.43 2.223 -2.520 82.6 7.425 6240.65 2.223 -3.233 48.5 7.427 $\dag$ 6252.56 2.404 -1.687 119.8 7.397 $\dag$ 6265.13 2.176 -2.550 84.9 7.448 $\dag$ 6270.23 2.858 -2.540 52.2 7.430 $\dag$ 6430.85 2.176 -2.006 110.0 7.318 $\dag$ 6498.94 0.958 -4.699 45.8 7.543 $\dag$ 6703.57 2.759 -3.023 37.7 7.488 $\dag$ 6705.10 4.607 -0.980 46.5 7.446 $\dag$ 6713.75 4.795 -1.400 21.1 7.438 $\dag$ 6726.67 4.607 -1.030 46.7 7.499 $\dag$ 6750.15 2.424 -2.621 74.3 7.504 $\dag$ 6810.26 4.607 -0.986 48.9 7.413 $\dag$ 5197.58 3.231 -2.220 81.4 7.415 $\dag$ 5234.62 3.221 -2.180 82.8 7.388 $\dag$ 5264.81 3.230 -3.130 44.3 7.461 $\dag$ 5414.07 3.221 -3.580 25.6 7.414 $\dag$ 5425.26 3.200 -3.220 41.5 7.453 $\dag$ 6369.46 2.891 -4.110 18.6 7.407 $\dag$ 6432.68 2.891 -3.570 41.3 7.442 $\dag$ [^1]: [^2]: Based on data obtained from the ESO Science Archive Facility. The observations were made with ESO Telescopes at the La Silla and Paranal Observatories under programmes ID 070.D-0356, 088.C-0513 and 070.D-0421. [^3]: <http://archive.eso.org/eso/eso_archive_adp.html> [^4]: *Image Reduction and Analysis Facility* - IRAF provided by *Association of Universities for Research in Astronomy*- AURA, EUA. [^5]: <http://www.eso.org/sci/facilities/lasilla/instruments/harps/inst/monitoring/sun.html> [^6]: <http://www.eso.org/observing/dfo/quality/UVES/pipeline/solar_spectrum.html> [^7]: An extended grid with respect to 2009. [^8]: COnservative COde for the COmputation of COmpressible COnvection in a BOx of L Dimensions, L=2,3. <http://www.astro.uu.se/~bf/co5bold_main.html> [^9]: [IDL]{} (Interactive Data Language) is a registered trademark of ITT Visual Information Solutions.

--- abstract: 'Connections between integration along hypersufaces, Radon transforms, and neural networks are exploited to highlight an integral geometric mathematical interpretation of neural networks. By analyzing the properties of neural networks as operators on probability distributions for observed data, we show that the distribution of outputs for any node in a neural network can be interpreted as a nonlinear projection along hypersurfaces defined by level surfaces over the input data space. We utilize these descriptions to provide new interpretation for phenomena such as nonlinearity, pooling, activation functions, and adversarial examples in neural network-based learning problems.' author: - | \  \ \ [^1] bibliography: - 'IEEEabrv.bib' - 'NN\_Generalized\_Radon.bib' title: 'Neural Networks, Hypersurfaces, and Radon Transforms' --- Introduction {#introduction .unnumbered} ============ Artificial Neural Networks (NN) have long been used as a mathematical modeling method and have recently found numerous applications in science and technology including computer vision, signal processing and machine learning [@lecun2015deep] to name a few. Although NN-based methods are recognized as powerful techniques, much remains to be explored about neural networks as a mathematical operator (one notable exception is the function approximation results in [@baldi1989neural; @cybenko1989approximation]). As a consequence, numerous doubts often accompany NN practitioners such as: how does depth add nonlinearity in a NN? What is the effect of different activation functions? What are the effects of pooling?, and many others. This didactic note is meant to highlight an alternative interpretation of NN-based techniques and their use in supervised learning problems. By investigating the connections of machine learning classification methods with projections along hyperplanes and hypersurfaces, we highlight the links between different NN architectures and the integration geometry of linear and nonlinear Radon transforms. We then use these concepts to highlight different properties of neural networks, which may help shed light on the questions highlighted above, as well as potentially provide a path for novel studies and developments. For brevity and to reduce pre-requisites, the derivations presented fall short of rigorous mathematical proofs. The Python code to reproduce all of the figures used here is available at <https://github.com/rohdelab/radon-neural-network>. Statistical regression and classification {#statistical-regression-and-classification .unnumbered} ========================================= Let $X$ be a compact domain of a manifold in Euclidean space (the space corresponding to input digital data) and let $h: X \rightarrow Y$, with $Y \in \mathbb{R}^{K}$ represent a map (oracle) which ascertains outputs (e.g. labels) to input data (random variable) $x \in X$. In learning problems, $y \in \mathbb{R}^{K}$ is usually a vector for which the value of the $k^{\mbox{th}}$ element represents the probability that the sample $x$ belongs to the $k^{\mbox{th}}$ class, although other regression problems can also be formulated with the same approach. {width="\linewidth"} Omitting here a measure theoretic formulation (see [@cucker2002mathematical] for a more complete development) let $p_X, p_Y$, and $p_{X,Y} \in L_1$ (space of absolutely integrable functions) define the probability density functions (PDFs) for random variables $X$, $Y$, and $(X,Y)$, respectively. Now utilizing a technique often used in the theoretical physics community [@gillespie1983theorem], known as random variable transformation (RVT), we can write the PDF of the output $p_Y$ as a function of $p_X$ via: $$p_Y(y) = \int_{X}p_X(x) \delta(y-h(x))dx, \label{eq:RVT}$$ where $\delta$ is the standard Dirac distribution. See the supplementary material for a derivation. The same transformation of random variables technique can be used to derive $$p_{f_\theta}(z) = \int_{X}p_X(x) \delta(z-f_\theta(x))dx \label{eq:p_F}$$ and $$p_{f_\theta,Y}(z,y) = \int_{X}p_X(x) \delta(y-h(x))\delta(z-f_\theta(x))dx.$$ The goal in a regression task is to estimate $f_\theta$ so that it accurately ‘predicts’ the dependent variable $y$ for each input $x$. In other words, we wish to find $f_{\theta} \sim h$ over the distribution of the input space. To that end “goodness of fit" measures are used to fit a model $f_{\theta}$ to given labeled data (supervised learning). One popular model is to find $\theta$ that minimizes the discrepancy between $y_n$ and $f_\theta(x_n)$ according to a dissimilarity measure $\mathcal{L}$: $$\min_\theta \sum_{n=1}^{N} \mathcal{L}(y_n,f_\theta(x_n)), $$ which can be interpreted in relation to random variables $Y = h(X)$ and $f_\theta$ and their respective distributions. For instance, the cross entropy minimization strategy $\frac{-1}{N}\sum_{k=1}^{N} y_k\cdot\log(f_\theta(x_k))$ can be viewed as an estimate of $\mathbb{E}_{x\sim p_X}\left( h(x)\cdot\log(f_\theta(x))\right)$, which is equivalent to minimizing the KL-divergence between $p_Y$ and $p_{f_\theta}$. Next, we consider the formulations for the standard Radon transform, and its generalized version and demonstrate a connection between this transformation and the statistical learning concepts reviewed above. Radon transform ---------------- The standard Radon transform, $\mathcal{R}$, maps distribution $p_X$ to the infinite set of its integrals over the hyperplanes of $\mathbb{R}^d$ and is defined as, $$\begin{aligned} \mathcal{R} p_X(t,\theta):=\int_{X} p_X(x)\delta(t-x\cdot\theta)dx, \label{eq:radon}\end{aligned}$$ where $\delta$ is the one-dimensional Dirac delta function. For $\forall\theta\in \mathbb{S}^{d-1}$ where $\mathbb{S}^{d-1}$ is the unit sphere in $\mathbb{R}^{d}$, and $\forall t \in \mathbb{R}$. Each hyperplane can be written as: $$H(t,\theta)=\{x\in \mathbb{R}^d |x\cdot\theta=t\} \label{eq:hyperplanes}$$ which alternatively could be thought as the level set of the function $g(x,\theta)=x\cdot\theta=t$. For a fixed $\theta$, the integrals over all hyperplanes orthogonal to $\theta$ define a continuous function, $\mathcal{R}p_X(\cdot,\theta):\mathbb{R}\rightarrow\mathbb{R}$, that is a projection/slice of $p_X$. We note that the Radon transform is more broadly defined as a linear operator $\mathcal{R}: L_1(\mathbb{R}^d)\rightarrow L_1(\mathbb{R}\times \mathbb{S}^{d-1})$, where $L_1(X):=\{ I:X \rightarrow \mathbb{R} | \int_{X} |I(x)|dx \leq \infty\}$. Figure \[fig:radon\_slices\] provides a visual representation of the Radon transform, the integration hyper-planes $H(t,\theta)$ (i.e., lines for $d=2$), and the slices $\mathcal{R}p_X(\cdot,\theta)$. The Radon transform is an invertible linear transformation (i.e. linear bijection). The inverse of the Radon transform denoted by $\mathcal{R}^{-1}$ is defined as: $$\begin{aligned} p_X(x)&=&\mathcal{R}^{-1}(\mathcal{R}p_X(t,\theta))\nonumber\\&=& \int_{\mathbb{S}^{d-1}} (\mathcal{R}p_X(\cdot,\theta)*\eta(\cdot))\circ (x\cdot\theta)d\theta\end{aligned}$$ where $\eta(.)$ is a one-dimensional high-pass filter with corresponding Fourier transform $\mathcal{F}\eta(\omega)\approx c|\omega|^{d-1}$ (it appears due to the Fourier slice theorem, see the supplementary material) and ‘$*$’ denotes the convolution operation. The above definition of the inverse Radon transform is also known as the filtered back-projection method, which is extensively used in image reconstruction in the biomedical imaging community. Intuitively each one-dimensional projection/slice, $\mathcal{R}p_X(\cdot,\theta)$, is first filtered via a high-pass filter and then smeared back into $X$ along $H(\cdot,\theta)$ to approximate $p_X$. The filtered summation of all smeared approximations then reconstructs $p_X$. Note that in practice acquiring infinite number of projections is not feasible therefore the integration in the filtered back-projection formulation is replaced with a finite summation over projections. ### Radon transform of empirical PDFs In most machine learning applications one does not have direct access to the actual distribution of the data but to its samples, $x_n\sim p_X$. In such scenarios the empirical distribution of the data is used as an approximation for $p_X$: $$\begin{aligned} p_X(x)\approx \hat{p}_X(x)= \frac{1}{N}\sum_{n=1}^N \delta(x-x_n) \label{eq:edf}\end{aligned}$$ where $\delta$ is the Dirac delta function in $\mathbb{R}^d$. Then, it is straightforward to show that the Radon transform of $\hat{p}_X$ is: $$\begin{aligned} \mathcal{R}\hat{p}_X(t,\theta)= \frac{1}{N}\sum_{n=1}^N \delta(t-x_n\cdot\theta) \label{eq:rtEmp}\end{aligned}$$ See supplementary material for detailed derivations of Equations . Given the high-dimensional nature of estimating density $p_X$ in $\mathbb{R}^d$ one requires large number of samples. The projections/slices of $p_X$, $\mathcal{R}p_X(\cdot,\theta)$, however, are one dimensional and therefore it may not be critical to have large number of samples to estimate these one-dimensional densities. ![The linear classifier slices the distribution of the data $p_X$ at an optimal $\theta$, for which the data is best discriminated. Therefore, one can think of the distribution of the output of the classifier as a slice of the Radon transform of the distribution $p_X$.[]{data-label="fig:linear_classifier"}](linear_classifier.png){width="\columnwidth"} ### Linear classification and the Radon transform Now, let us consider the supervised learning of a linear binary classifier. Given the data samples $\{x_n\sim p_X\}_{n=1}^N$ and their corresponding labels $\{y_n\in\{0,1\}\}_{n=1}^N$, the task is to learn a linear function of the input samples, $f_\theta(x)=\theta\cdot x$ for $\theta\in\mathbb{S}^{d-1}$ such that, $$\theta\cdot x \underset{y=0}{\overset{y=1}{\gtrless}} b.$$ Many methods exist to obtain the optimal $\theta$, e.g., Support Vector Machines or Logistic Regression. While the projection $f_\theta(x)=\theta\cdot x$ is applied to each sample, we can consider $f_\theta(\cdot)$ as an operator and inquire about the distribution $p_{f_\theta}$. Here $p_{f_\theta}(z)$ is the density of $z=f_\theta(x)$ when $x\sim p_X$. One can clearly see that $p_{f_\theta}$ corresponds to a slice of the input distribution $p_X$ with respect to $\theta\in\mathbb{S}^{d-1}$, hence there is a natural relationship between the Radon transform and linear classification. Figure \[fig:linear\_classifier\] depicts this phenomenon. Generalized Radon transform ---------------------------- Generalized Radon transform (GRT) extends the original idea of the classic Radon transform introduced by J. Radon [@radon1917uber] from integration over hyperplanes of $\mathbb{R}^d$ to integration over hypersurfaces [@ehrenpreis2003universality; @kuchment2006generalized] (i.e. $(d-1)$-dimensional manifolds). GRT has various applications including Thermoacoustic Tomography (TAT), where the hypersurfaces are spheres, and Electrical Impedance Tomography (EIT), where integration over hyperbolic surfaces appear. To formally define the GRT, we introduce a function $g$ defined on $X \times (\mathbb{R}^n \backslash \{ 0 \})$ with $X \subset \mathbb{R}^d$. We say that $g$ is a *defining function* when it satisfies the four conditions below: 1. $g(x,\theta)$ is a real-valued $C^\infty$ function on $X \times (\mathbb{R}^n \backslash \{ 0 \})$ 2. $g(x, \theta)$ is homogeneous of degree one in $\theta$, *i.e.* $$\forall \lambda \in \mathbb{R},\; g(x, \lambda \theta) = \lambda g(x, \theta)$$ 3. $g$ is non-degenerate in the sense that $d_x g(x, \theta) \neq 0$ in $X \times \mathbb{R}^n \backslash \{ 0 \}$ 4. The mixed Hessian of $g$ is strictly positive, *i.e.* $$\text{det}\left( \frac{\partial^2 g}{\partial x^i \partial \theta^j} \right) > 0$$ For a given defining function, $g$, the generalized Radon transform is a linear operator $\mathcal{G}:L^1(X)\rightarrow L^1(X \times \Omega_\theta)$, where $\Omega_\theta\subseteq (\mathbb{R}^n \backslash \{0\})$ and is defined as: $$\begin{aligned} \mathcal{G} p_X(t,\theta):=\int_{X} p_X(x)\delta(t-g(x,\theta))dx \label{eq:grt}\end{aligned}$$ From a geometrical perspective and for a fixed $t$, $\mathcal{G}p_X(t,\theta)$ is the integral of $p_X$ along the hypersurface $H(t,\theta)=\{x\in X|g(x,\theta)=t\}$. Note that the classic Radon transform is a special case of the generalized Radon transform where $g(x,\theta)=x\cdot\theta$. The investigation of the sufficient and necessary conditions for showing the injectivity of GRTs is a long-standing topic [@ehrenpreis2003universality; @kuchment2006generalized]. The conditions 1-4 for a defining function, $g$, enumerated in this section, are necessary conditions for injectivity but not sufficient. Though the topic related to inversion of the GRT is beyond the scope of this article, an inversion approach is given in [@uhlmann2003inside]. Here, we list a few examples of known defining functions that lead to injective GRTs. The circular defining function, $g(x,\theta) = \|x-r*\theta\|_2$ with $r\in\mathbb{R}^+$ and $\Omega_\theta = \mathbb{S}^{d-1}$ was shown to provide an injective GRT [@kuchment2006generalized]. More interestingly, homogeneous polynomials with an odd degree also yield an injective GRT [@ehrenpreis2003universality], *i.e.* $g(x,\theta) = \sum_{|\alpha| = m} \theta_\alpha x^\alpha$, where we use the multi-index notation $\alpha = (\alpha_1, \dots, \alpha_{d_\alpha}) \in \mathbb{N}^{d_\alpha}$, $|\alpha| = \sum_{i=1}^{d_\alpha} \alpha_i$, and $x^\alpha = \prod_{i=1}^{d_\alpha} x_i^{\alpha_i}$. Here, the summation iterates over all possible multi-indices $\alpha$, such that $|\alpha| = m$, where $m$ denotes the degree of the polynomial and $\theta_\alpha \in \mathbb{R}$. The parameter set for homogeneous polynomials is then set to $\Omega_\theta=\mathbb{S}^{d_\alpha-1}$. We can observe that choosing $m=1$ reduces to the linear case $ g(x,\theta)=x\cdot\theta$, since the set of the multi-indices with $|\alpha|=1$ becomes $\{ (\alpha_1, \dots, \alpha_d); \alpha_i = 1 \text{ for a single } i\in \llbracket 1, d \rrbracket, \text{ and } \alpha_j = 0, \quad \forall j \neq i\}$ and contains $d$ elements. {width="\linewidth"} neural networks and the generalized Radon transform {#neural-networks-and-the-generalized-radon-transform .unnumbered} ==================================================== To illustrate the relationship between deep neural networks and the generalized Radon transform we start by describing the link between perceptrons and the standard Radon transform. Single perceptron ----------------- Let $z=\sigma(\theta\cdot x)$, with $\|\theta\|=1$, define a perceptron for input data $x\sim p_X$, where we dissolved the bias, $b$, into $\theta$. Treating $z\sim p_Z$ as a random variable and using RVT, it is straightforward to show that $p_{Z}$ is isomorphic to a single slice of $p_X$, $\mathcal{R}p_X(t,\theta)$, when $\sigma$ is invertible (see supplementary material for a proof). The isomorphic relationship provides a fresh perspective on perceptrons, stating that the distribution of the perceptron’s output, $f_\theta(x)$, is equivalent to integration of the original data distribution, $p_X$, along hyperplanes $H(t,\theta)=\{x|x\cdot\theta=\sigma^{-1}(t)\}$ (see Equation ). In addition, one can show that the distribution of the output of a perceptron is equal to the generalized Radon transform with $g(x,\theta)=f_\theta(x)$, $$\begin{aligned} p_{f_\theta}(z)=\mathcal{G}p_X(z,\theta)=\int_X p_X(x)\delta(z-\sigma(x\cdot\theta))dx.\end{aligned}$$ An important and distinctive point here is that here we are interested in the distribution of the output of a perceptron, $\mathcal{G}p_X(z,\theta)$, and its relationship to the original distribution of the data, $p_X$, as opposed to the individual responses of the perceptron, $z_n=g(x_n,\theta)$. Columns (a) and (b) in Figure \[fig:perceptron\] demonstrate the level sets (or level curves since $d=2$) and the line integrals for $g(x,\theta)=x\cdot\theta$ and $g(x,\theta)=\sigma(x\cdot\theta)$, where $\theta\in\mathbb{S}^1$. Note that samples that lay on the same level set will be mapped to a fixed projection (a constant value $z$). In other words, the samples that lay on the same level sets of $g(x,\theta)$ are indistinguishable in the range of the perceptron. Next we discuss the case of having multiple perceptrons. Multilayer (Deep) neural networks --------------------------------- To obtain a hierarchical (multilayer) model, the concept of a perceptron can be applied recursively. As before, let $\Theta^1$ and $\Theta^2$ correspond to two matrices whose rows contain a set of projection vectors (different $\theta$’s in the preceding section): e.g. $\Theta^{1} = [\theta_{1}^{(1)^T}, \theta_{2}^{(1)T}, \cdots]$ where $\theta_{1}^{(1)^T}$ is the transpose of projection vector corresponding to the first node/perceptron in layer 1. A two layer NN model can be written as $\sigma(\Theta^2\sigma(\Theta^1) x)$. Expanding the idea further, we then may define a general formula for a $K$-layer NN as $$g(x,\theta)= \sigma(\theta_1^K\cdot \sigma(\Theta^{K-1}\sigma(\Theta^{K-2}(...\sigma(\Theta^1x))))) \label{eq:mlp}$$ Note that $\theta_1^K$ above refers to a column vector which collapses the output of the neural network to one node and that $\Theta^k=[\theta_1^k,...,\theta_{L_k}^k]$ where $L_k$ is the number of neurons in the $k$’th layer of a deep neural network. Now, let $\sigma$ be a Lipschitz continuous nonlinear activation function. Its self-composition is therefore also Lipschitz continuous. For invertible activation functions $\sigma$, and for $\Theta^{k}$ square and invertible, the gradient of a multi-layer perceptron in equation does not vanish in any compact subset of $\mathbb{R}^d$ and therefore the level sets are well-behaved. Therefore, from the definition in we have that the distribution over the output node $p_Y(y)$ could be considered as a slice of the generalized Radon transform of $p_X$ evaluated at $\theta$: $p_Y(y) = \mathcal{G}p_X(t,\theta)$ with $$\mathcal{G}p_X(t,\theta) = \int_{X}p_X(x)\delta(t-\sigma(\theta_1^K\cdot \sigma(\Theta^{K-1}...\sigma(\Theta^1x)))) dx \notag$$ Figure \[fig:perceptron\] columns (c) and (d) demonstrate the level sets and the line integrals of $p_X(x)$ using a multi-layer perceptron as $g_\theta$. Column (c) is initialized randomly and column (d) shows $g_\theta$ after the network parameters are trained in a supervised classification setting to discriminate the modes of the half-moon distribution. It can be seen that after training, the level sets, $H(t,\theta)$, only traverse a single mode of the half-moon distribution, which indicates that the samples from different modes are not projected onto the same point (i.e. the distribution is not integrated across different modes). It also readily becomes apparent the facility with which neural networks have to generate highly nonlinear functions, even with relatively few parameters (below we compare these to other polynomials). We note that with just one layer, NN’s can form nonlinear decision boundaries, as the superposition of surfaces formed by $\sigma(\theta_1^1 \cdot x)+\sigma(\theta_2^1 \cdot x) +\cdots$ can add curvature to the resulting surface. Note also that generally speaking, the integration streamlines (hypersurfaces for higher dimensional data) have the ability to become more curved (nonlinear) as the number of layers increases. ![Level curves of nodes introduced by different activation functions. Parameters $\theta_1 \in \mathbb{R}^2, \Theta^1 \in \mathbb{R}^{50\times 2}, \theta_1^2\in \mathbb{R}^{50}$ are randomly initialized (with the same seed) for the first three column demonstrations. Parameters for the last column $\Theta^1 \in \mathbb{R}^{50\times 2}, \Theta^2 \in \mathbb{R}^{100\times 50}, \theta^3_1 \in \mathbb{R}^{100}$ are optimized by minimizing a misclassification loss.[]{data-label="fig:activations"}](activation_curves.pdf){width="0.9\columnwidth"} Activation functions {#activation-functions .unnumbered} -------------------- It has been noted recently that NN’s (e.g. convolutional neural networks) can at times work better when $\sigma$ is chosen to be the rectified linear unit (ReLU) function, as opposed the sigmoid option [@nair2010rectified; @glorot2011deep; @krizhevsky2012imagenet]. The experience has encouraged others to try different activation functions such as the ‘leaky’-ReLU [@maas2013rectifier]. While theory describing which type of activation function should be used with which type of learning problem is yet to emerge, the interpretation of NN’s as nonlinear projections can help highlight the differences between activation function types. Specifically, Figure \[fig:activations\] can help visualize the effects of different activation functions on the integration geometry over the input data space $X$. First note that the ReLU is a non-invertible map, given that negative values all map to zero. This will cause the surface generated by a perceptron constructed with ReLU to have a region over $X$ which is flat, whereby all points in that region are integrated and mapped to the same value (zero) in the output space. This ability may provide ReLU neural networks with the flexibility to construct adaptable characteristic function-type models for different regions in the data space. Although, the outcome of the optimization procedure will dictate whether such regions would emerge in the final model. Finally, note that both ReLU and the leaky-ReLU activation functions contain non-differentiable points, which are also imparted on the surface function (hence the sharp ‘kinks’ that appear over iso-surfaces lines). {width="\linewidth"} Pooling {#pooling .unnumbered} ------- Pooling (e.g. average or maximum sampling) operations are typically used in large neural networks, especially the CNN kind. The reasons are often practical, as subsampling can be used as a way to control and reduce the dimension (number of parameters) of the model. Another often stated reason is that pooling can also add a certain amount of invariance (e.g. translation) to the NN model. In the case of average pooling, it is clear that the operation can also be written as a linear operator $\Theta^k$ in equation where the pooling operation can be performed by replacing a particular row of $\Theta^k$ by the desired linear combination between two rows of $\Theta^k$, for example. ‘Max’-pooling on the other hand selects the maximum surface value (perceptron response), over all surfaces (each generated by different perceptrons) in a given layer. Figure \[fig:max\_pooling\] shows a graphical description of the concept, though it should be noted that as defined above, the integration lines are not being added, rather the underlying ‘level’ surfaces. Adversarial examples {#adversarial-examples .unnumbered} -------------------- It has often been noted that highly flexible nonlinear learning systems such as CNN’s can be ‘brittle’ in the sense that a seemingly small perturbation of the input data can have cause the learning system to produce confident, but erroneous, outputs. Such perturbed examples are often termed as adversarial examples. Figure \[fig:adversarial\] utilizes the integral geometric perspective described above to provide a visualization of how neural networks (as well as other classification systems) can be fooled by small perturbations. To find the minimum displacement that could cause misclassificaiton, using the blue point as the starting point $x_0$, we perform gradient ascent $ {x}_{n+1}= {x}_{n}+\gamma \nabla g(x_n, \theta)$, until we reach the other side of the decision boundary (which is indicated by the orange point). We limit the magnitude of the displacement small enough so that the two points belong to the same distribution. However, once integrated along the isosurfaces corresponding to the NN drawn in the figure, due to the uneven curvature of the corresponding surface, the two points are projected onto opposite ends of the output node, thus fooling the classifier to make a confident, but erroneous, prediction. {width="0.9\columnwidth"} \[fig:spherical\] {width="0.8\columnwidth"} \[fig:adversarial\] Summary and future directions {#summary-and-future-directions .unnumbered} ============================= In this note we explored links between Radon transforms and Artificial Neural Networks to study the properties of the later as an operator on the probability density function $p_X$ associated with a given learning problem. More specifically, it can be shown that the probability density function associated with any output node of a neural network (or a single perceptron within a hierarchical NN) can be interpreted as a particular hyperplane integration operation over the distribution $p_X$. This interpretation has natural connections with the $N$-dimensional generalized Radon transforms, which similarly proposes that a high-dimensional PDF can be represented by integrals defined over linear or nonlinear hyperplanes in $X$. The projections can be linear (in the case of simple linear logistic regression) or nonlinear in which case the projection are computed over nonlinear hyperplanes. The analogy has limitations, however, given that depending on the number of nodes in each layer, as well as the choice of activation function, the conditions for the generalized Radon transforms in [@uhlmann2003inside] (i.e. invertibility, homogeneity, etc) may not be satisfied with specific neural network architectures. Despite these limitations, the analogy is useful to provide a mechanistic understanding of NN operators, and it may also be useful as a path to study the effect of different neural network architectures and related concepts (number of layers, number of nodes in each layer, choice of activation function, recurrency, etc.) as well as to provide ideas for alternative models. For example, other types of projections may be considered within a learning problem. Figure \[fig:spherical\] compares linear projections, circular projections, a homogeneous polynomial of degree 5, and an ANN of depth 1, all trained to minimize the logistic regression cost function. While it is clear that linear and circular projections don’t have enough ‘flexibility’ to solve the separation problem, a polynomial degree of degree 5 seems to emulate the behavior of an ANN of depth 1. It is possible that in the future, the point of view provided by analyzing the nonlinear projections associated with different NN’s can provide inspiration for alternative models. Supplementary material ====================== Inverse of Radon transform -------------------------- To define the inverse of the Radon transform we start by the Fourier slice theorem. Let $\mathcal{F}_d$ be the d-dimensional Fourier transform, then the one dimensional Fourier transform of a projection/slice is: $$\begin{aligned} \mathcal{F}_1(\mathcal{R}I(\cdot,\theta))(\omega)&=&\int_\mathbb{R}\mathcal{R}I(t,\theta)e^{-i\omega t}dt\\ &=&\int_\mathbb{R}\int_{\mathbb{R}^d} I(x)e^{-i\omega t}\delta(t-x\cdot \theta)dxdt\\ &=&\int_{\mathbb{R}^d} I(x)e^{-i(\omega\theta)\cdot x}dx=\mathcal{F}_dI(\omega\theta)\end{aligned}$$ which indicates that the one-dimensional Fourier transform of each projection/slice is equal to a slice of the d-dimensional Fourier transform in a spherical coordinate. Taking the inverse d-dimensional Fourier transform of $\mathcal{F}_dI(\omega\theta)$ in the Cartesian coordinate, $u\in\mathbb{R}^d$, would lead to the reconstruction of $I$. $$\begin{aligned} I(x)&=&\mathcal{F}_d^{-1}(\mathcal{F}_dI(u))\\ &=& \int_\mathbb{R}\int_{\mathbb{S}^{d-1}} \mathcal{F}I(\omega\theta)e^{i\omega\theta\cdot x}|\omega|^{d-1} c(\theta)d\theta dt\end{aligned}$$ where, $$c(\theta)=\sin^{d-2}(\theta_1)\sin^{d-3}(\theta_2)...\sin(\theta_{d-2})$$ where $\theta=[\theta_1,...,\theta_d-1]$, and $c(\theta)$ is often approximated as a small angle-independent constant, $c$. The RVT theorem --------------- Here we show the derivations for Equation . Recall that $h(x)=y$ is the true map (oracle) from the data samples to their corresponding labels. Let $g:\mathbb{R}\rightarrow \mathbb{R}$ be a real function, then we can write $g(y)=g(h(x))$. By definition, the average of the quantity on the left with respect to $y$ should be equal to the average of the quantity on the right with respect to $x$, and we can write: $$\begin{aligned} \int_\mathbb{R} g(y)p_Y(y)dy&=& \int_X g(h(x))p_X(x)dx \nonumber \\ &=& \int_X\int_\mathbb{R} g(y)\delta(y-h(x))p_X(x)dydx \nonumber\\ \label{eq:rvt_proof}\end{aligned}$$ Now let $g(y)=\delta(y-y')$ and for the left hand side of Equation we have: $$\begin{aligned} \int_\mathbb{R} \delta(y-y')p_Y(y)dy&=& p_Y(y')\end{aligned}$$ and for the right hand side of Equation we have: $$\begin{aligned} \int_X\int_\mathbb{R} &\delta(y-y')\delta(y-h(x))p_X(x)dydx =& \\ &\int_X p_X(x)\delta(y'-h(x))dx&\end{aligned}$$ which yields: $$\begin{aligned} p_Y(y') = \int_X p_X(x)\delta(y'-h(x))dx,\end{aligned}$$ and concludes our proof the derivation. For a more complete analysis, see [@gillespie1983theorem]. Isomorphic relationship between a perceptron and a standard Radon slice ------------------------------------------------------------------------ For the perceptron, $f_\theta(x)=\sigma(x\cdot\theta)$, where $\|\theta\|=1$ and $\sigma:\mathbb{R}\rightarrow U=(0,1)$, the distribution of the output could be obtained from: $$\begin{aligned} p_{f_\theta}(z)=\int_X p_X(x)\delta(z-\sigma(x\cdot\theta))dx\end{aligned}$$ on the other hand, the Radon slice of $p_X$ is obtained from $$\begin{aligned} \mathcal{R}p_X(t,\theta)=\int_X p_X(x)\delta(t-x\cdot\theta)dx\end{aligned}$$ We first show that having $\mathcal{R}p_X(\cdot,\theta)$ one can recover $p_{f_\theta}$. Let $z=\sigma(t)$, where $t\sim \mathcal{R}p_X(\cdot,\theta)$ therefore using RVT the distribution of $z$ is equal to: $$\begin{aligned} p_Z(z)&=&\int_\mathbb{R} \mathcal{R}p_X(t,\theta)\delta(z-\sigma(t))dt\\ &=&\int_\mathbb{R}\int_X p_X(x)\delta(t-x\cdot\theta)\delta(z-\sigma(t))dxdt\\ &=&\int_X p_X(x)\delta(z-\sigma(x\cdot\theta))dx\\ &=& p_{f_\theta}(z)\end{aligned}$$ Now we show the reverse arguement. For invertible $\sigma$, let $t=\sigma^{-1}(z)$ where $z\sim p_{f_\theta}$, then we can obtain the distribution of $t$ from: $$\begin{aligned} p_T(t)&=&\int_U p_{f_\theta}(z)\delta(t-\sigma^{-1}(z))dz\\ &=&\int_U\int_X p_X(x)\delta(z-\sigma(x\cdot\theta))\delta(t-\sigma^{-1}(z))dxdz\\ &=&\int_X p_X(x)\delta(t-\sigma^{-1}(\sigma(x\cdot\theta)))dx\\ &=&\int_X p_X(x)\delta(t-x\cdot\theta)dx\\ &=& \mathcal{R}p_X(t,\theta)\end{aligned}$$ therefore the two distributions, $\mathcal{R}p_X(\cdot,\theta)$ and $p_{f_\theta}$, are isomorphic. [^1]: This work was supported in part by NIH award GM130825.

--- abstract: 'We demonstrate a method for efficient coupling of guided light from a single mode optical fiber to nanophotonic devices. Our approach makes use of single-sided conical tapered optical fibers that are evanescently coupled over the last $\sim10\,\mu $m to a nanophotonic waveguide. By means of adiabatic mode transfer using a properly chosen taper, single-mode fiber-waveguide coupling efficiencies as high as $97(1)\%$ are achieved. Efficient coupling is obtained for a wide range of device geometries which are either singly-clamped on a chip or attached to the fiber, demonstrating a promising approach for integrated nanophotonic circuits, quantum optical and nanoscale sensing applications.' author: - 'T. G. Tiecke$^{1,2}$' - 'K. P. Nayak$^{1,3}$' - 'J. D. Thompson$^{1}$' - 'T. Peyronel$^{1}$' - 'N. P. de Leon$^{1,4}$' - 'V. Vuletić$^{2}$' - 'M. D. Lukin$^{1}$' title: 'Efficient fiber-optical interface for nanophotonic devices' --- [^1] [^2] [^3] Introduction ============ The field of nanophotonics [@joannopoulos08] opened new avenues for applications such as nanophotonic integrated circuits[@chen11; @miller10], sensing [@pohl84; @lewis84; @tan92; @yan12] and scalable quantum information processing [@kimble08; @ladd10; @vanmeter10]. Moreover, sub-wavelength confinement of optical fields enabled strong light-matter interaction at the single quantum level [@yoshie04; @tiecke14]. A major challenge in the field is to efficiently integrate the nanophotonic devices with conventional optical fiber networks. This challenge is due to a large mismatch between the size of the fundamental mode of the optical fiber and that of the optical modes of nanophotonic devices. This mismatch has to be bridged in order to achieve efficient coupling. Highly efficient coupling is crucial for applications such as quantum repeaters [@briegel98] or quantum networks [@kimble08] since the performance of these systems, in the limit of many nodes, deteriorates near-exponentially with photon loss between individual nodes. Additionally, highly efficient coupling enables distribution of non-classical states of light which are extremely fragile to photon loss. A wide range of coupling techniques are currently being explored, including grating coupling [@chen11] and end-firing from macroscopic fibers [@cohen13] where coupling efficiencies up to 70-80% to on-chip waveguides have been achieved. More recently, on-chip photonic waveguides have been coupled to the waist of a biconical fiber taper [@groblacher13] with an efficiency as high as 95%. In this Letter, we demonstrate a novel method to efficiently couple a single mode fiber to a dielectric nanophotonic waveguide using a conical tapered fiber tip. The coupling is based on an adiabatic transfer of the fundamental mode of the optical fiber to the fundamental mode of the nanophotonic waveguide. Our method can be applied to general dielectric one-dimensional waveguides. In contrast to biconical tapered fibers [@groblacher13], our devices are single-sided, thereby offering alternative geometries and mechanical support for nanophotonic devices [@thompson13a; @tiecke14] and opening the door for new applications. Adiabatic coupling ================== Adiabatic mode transformation is widely used to obtain efficient power transfer through nonuniform optical waveguides [@loveBook]. The key idea is to change the waveguide cross-section slowly along the propagation direction of the light such that all the optical power remains in a single eigenmode of the composite waveguide, while the coupling to other modes is suppressed. More specifically, two eigenmodes $\nu$ and $\mu$ with respective effective indices $n_\nu$ and $n_\mu$ define a characteristic beating length scale between the modes given by $z_b=\lambda/(n_\nu-n_\mu)$, where $\lambda$ is the wavelength in vacuum. In order to achieve adiabatic transfer the characteristic length scale $z_t$ over which the waveguide changes has to be large compared to $z_b$ [@love91]. While the exact coupling strength depends on the details of the spatial mode profiles, we design our devices according to this intuitive length-scale argument and use numerical simulations to verify the design. Figure 1a shows a schematic of a typical device. A fiber is tapered down to a conical tip and an inversely tapered silicon nitride ($\mathrm{Si_3N_4}$) waveguide is attached over $7.5\,\mu$m. In what follows, we separate the fiber tips in two regions (see Fig.1a): the adiabatic fiber-waveguide coupler (I) and the tapered fiber (II), separated by plane P at the start of the waveguide. In both regions we design our devices according to the adiabaticity criterion by ensuring $z_t>z_b$. In region I the optical modes of the fiber and waveguide are coupled through their evanescent fields and form a set of hybridized supermodes. $z_b$ is determined by the difference between the effective indices of the fundamental supermode and the higher order supermode with the closest propagation constant (in this case radiation mode) while $z_t$ is limited by the length of the coupling region. In region II the fiber is tapered down from a standard single mode fiber, adiabatically transforming the core-guided HE11 mode to a cladding-guided HE11 mode. We ensure that the local taper angle $\Omega(z)\equiv\rho(z)/z_t \ll \rho(z) /z_b$, where $\rho(z)$ is the fiber radius at position $z$ along the taper. In region II $z_b$ is determined by the coupling of the fundamental fiber mode to the nearest higher-order mode. We follow the conventional design for biconical adiabatic fiber tapers where recently transmission efficiencies as high as $99.95\,\%$ [@hoffman14] have been achieved (see also [@love91; @stiebeiner10] for details). For a tapered nanophotonic waveguide which vanishes at plane P (such as that shown in figure 1) the cross-section of refractive index profile changes smoothly along the complete coupler. For such waveguides the criteria given above are sufficient to design the coupler. However, for a waveguide with a non-vanishing cross-section at P (such as the rectangular waveguide shown in figure 1d), the refractive index cross-section changes discontinuously. We approximate the power transfer through P by projecting the fundamental fiber mode on the fundamental supermode at P. In order to achieve efficient power transfer, we design the coupler such that this projection is close to unity. In this case the effect of the waveguide is only a small perturbation of the fundamental fiber mode at plane P, thus enabling us to design couplers based on simple eigenmode calculations. ![**Adiabatic transfer between fiber and waveguide modes.** **a.** Schematic of the fiber-waveguide coupling. The fiber (right) has a conical shape and is attached to a tapered $\mathrm{Si_3N_4}$ rectangular waveguide (left) and we consider modes polarized along $\hat x$. **b.** Effective index $n_\mathrm{eff}$ of the fiber and waveguide modes for an opening angle of the fiber (waveguide) of $5^\circ$ ($4^\circ$). The blue dotted (dashed) lines are the separate fiber (waveguide) modes and the blue solid line corresponds to the fundamental supermode of the combined structure. The red line shows the power in the fundamental supermode obtained from an FDTD simulation of the coupler (see text). **c.** Cross sections of $|E|^2$ obtained from the FDTD simulation at various points along the coupler. **d.** The fraction of the power in the fundamental supermode of the combined structure as a function of the waveguide width $dx$, obtained from a mode decomposition (solid line). The transmission through a tapered coupler (see *inset*) obtained with an FDTD simulation (circles) agrees well with the estimated transmission obtained from the mode decomposition. The two data points for $dx\leq 200\,\mathrm{nm}$ (open circles) are calculated using a shallower fiber angle ($2^\circ$) to ensure $z_t>z_b$. The dotted line shows the same geometry except that the fiber and waveguide are in contact on the $xz$-plane instead of the $yz$-plane. The fiber-waveguide cross-sections used for this simulation are shown in the *inset*, $\rho=450\,\mathrm{nm}$. \[fig:fig1\]](figure1_v6.pdf){width="1.0\columnwidth"} ![**Characterization of adiabatic tapers.** **a.** The fiber angle as a function of the local fiber diameter along the taper axis $z$. The dashed line and shaded area indicate the adiabaticity criterion $z_t>z_b$ as discussed in the text. Fiber tapers which have a profile below the dotted line are expected to be adiabatic. For a diameter smaller than $1.1\,\mu \mathrm{m}$ the HE12 mode is cut off. The taper profiles for 4 tapers (blue (A), red (B), purple (C) and green (D)) are shown. **b.** Far field mode profiles. Tapers A, B and C show Gaussian profiles, while taper D has clear contributions from higher order modes. For tapers C and D cuts through the center of the profiles are shown together with a Gaussian fit. **c.** The transmission versus pulling time of a taper similar to A-C, the dashed line indicates $99\%$ transmission. The sudden drop in transmission at $\simeq 87\,\mathrm{s}$ arises from the fast pull by the electromagnetic coil. **d.** The taper profile of taper C (blue) and of a biconical taper (dashed) using the same pulling parameters but without pulsing the electromagnet to create the tip. \[fig:fig2\]](figure2_v3.pdf){width="1.0\columnwidth"} We next verify these design criteria using Finite Difference Time Domain (FDTD) simulations. Figures 1b and 1c show simulations for the device presented in figure 1a. We consider a conical fiber with an opening angle $d(2\rho)/dz=5^\circ$ and a $dy=175\,\mathrm{nm}$ thick $\mathrm{Si_3N_4}$ waveguide with a taper angle $d(dx)/dz=4^\circ$ to a width of $dx=500\,\mathrm{nm}$. We focus on TE-polarized ($\hat x$) modes, but we have verified that the same reasoning can be applied for TM-polarized modes. Figure 1b shows the effective indices of the fiber mode, the waveguide mode and of the hybridized mode of the combined structure (supermode). For this geometry, the latter has an effective index of $n_{\mathrm{eff}}>1.2$ over the entire length of the coupler. The combined structure supports only one other mode, which, however, has orthogonal polarization and therefore does not couple to the fundamental supermode. The relevant beat length is therefore set by the fundamental supermode and the free space modes ($n_0 = 1$), and corresponds to $z_b\simeq 4\,\mu \mathrm{m}$. We chose the length of the coupler ($z_t \simeq 7\,\mu \mathrm{m}$) to be longer than $z_b$. In order to verify adiabaticity we perform a FDTD simulation (see figure 1b, c), in which we excite the fiber taper at $z=11\,\mu \mathrm{m}$ with the fundamental HE11 mode polarized along $\hat x$ and propagating along $-\hat z$. Along the coupler we decompose the optical fields in the basis of local eigenmodes of the combined fiber-waveguide structure and find that essentially all the optical power ($>99\,\%$) is in the fundamental mode across the complete fiber-waveguide coupler, thereby confirming the adiabaticity of the coupler. In the case of a rectangular waveguide (Figure 1d), we model the sudden onset of the waveguide by decomposing the fundamental fiber mode in the basis of supermodes of the combined structure. This decomposition is performed using the fields of the eigenmodes of the fiber and the combined structure which we obtain using the MIT Photonic Bands (MPB) mode-solver [@johnson01]. To verify that the projection indeed describes the power transfer accurately we compare the mode decomposition results with FDTD simulations. Figure 1d shows the power in the fundamental supermode of the combined structure as obtained from the mode decomposition and from an FDTD simulation for a rectangular $\mathrm{Si_3N_4}$ waveguide with varying width $dx$ and a fiber with radius $\rho=450\,\mathrm{nm}$. We find that the loss of transmission through the coupler obtained from the FDTD simulation is well described by the mode decomposition. For the simulated conditions the losses can be made small for waveguide dimensions below 200 nm. Design and fabrication ====================== We next discuss the optimization and characterization of the tapered fiber tips in region II. For our experiments we use single mode fiber (Thorlabs 780HP), with a $4.4\,\mu \mathrm{m}$ ($125\,\mu \mathrm{m}$) core (cladding) diameter and we optimize our design for a wavelength of $\lambda=780\,\mathrm{nm}$. Figure 2a shows the critical angle ($\Omega_c(z)=\rho(z) /z_b(z)$) for this fiber. In a cylindrically symmetric geometry, modes with different angular momentum do not couple, therefore, the coupling occurs between the HE11 and HE12 modes. At large fiber-diameters ($d> 50\,\mu \mathrm{m}$) the adiabaticity criterion is determined by coupling of the HE11 core guided mode and the cladding guided modes, while for $d<50\,\mu \mathrm{m}$ the adiabaticity criterion is determined by the coupling of the HE11 and HE12 cladding guided modes. We fabricate fiber tapers using a conventional heat-and-pull setup [@mazur03; @ward14] where the fiber is heated using an isobutane torch ($140\,\mathrm{mL/min}$ flow), with an effective flame length of $L=4.3\,\mathrm{mm}$. The flame is continuously brushed back and forth to heat the fiber over a variable length, which is adjusted during the pulling to obtain the desired fiber profile (see Refs. [@mazur03; @bambrilla10]). This results in a $24\,\mathrm{mm}$ long biconical fiber taper with a minimum diameter of $\sim\,30\mu \mathrm{m}$. At this stage we apply a fast pull to one of the stages holding the fiber, which quickly ($\sim 10\,\mathrm{ms}$) pulls the fiber out of the flame thereby creating a $14\,\mathrm{mm}$ long fiber taper with a conical tip (see figure 2d). The fast pull is generated by an electromagnet, composed of a hard-drive head with its arm connected to one of the two fiber clamps. The clamp itself is mounted on a linear ball-bearing translation stage and a current pulse through the electromagnet results in a constant acceleration of the fiber, creating a smooth fiber tip which is well described by a parabolic shape at larger fiber diameters and a constant opening angle over the last tens of microns. We find that the acceleration changes linearly with the applied current over a range of $17\,\mathrm{m/s^2}$ to $46\,\mathrm{m/s^2}$ and we typically use an acceleration of $33\,\mathrm{m/s^2}$. By optimizing the heat-and-pull-parameters we realize the requirements of the taper angle for large diameters, while the electromagnet current and fiber-diameter at which the pulse is applied controls the fiber taper angle at smaller diameter. We note that the resulting parabolic shape of the fiber taper conveniently has the same scaling ($\Omega\sim1/\rho$) as the adiabaticity criterion at the relevant range of fiber diameters ($2 - 50\,\mu \mathrm{m}$, see figure 2a). Additionally, since our fiber tips have sub-wavelength dimensions only over $\sim 10\,\mu \mathrm{m}$, the requirements on the cleanliness of the flame and the fabrication environment are less stringent as compared to those for creating efficient biconical tapered fibers [@hoffman14; @ward14]. Characterization ================ We characterize our devices with several measurements. First, we measure the taper profiles to ensure the local angle is smaller than the critical angle set by the adiabaticity condition. Figure 2 shows three fibers (A, B, C) which are made under the same conditions, while fiber D is made using different pulling parameters for the purpose of illustrating the performance of a sub-optimal fiber taper. In figure 2a we show the fiber profiles for each fiber, which are measured using optical and scanning electron microscopy (SEM). Fibers A, B and C show nearly identical profiles which satisfy the adiabaticity criterion, indicating that our fabrication method yields reproducible fiber tapers. Fiber D has a somewhat steeper angle for fiber diameters around $\sim 3-30\,\mu \mathrm{m}$ and is therefore expected to be non-adiabatic. Second, in figure 2b we show the far-field profiles of the fiber mode imaged directly on a CMOS camera. The optical modes of fibers A, B, C are all nearly Gaussian, indicating that at the end of the taper, most of the power is in the fundamental HE11 mode. Fiber D shows clearly a multimode structure, in agreement with our expectation that this taper does not match the adiabaticity criterion. To quantify the single mode character of the profiles we calculate the coefficient of determination ($R^2$) of the Gaussian fits, resulting in $R^2=0.98,0.99,0.99,0.87$ for tapers A, B, C and D, respectively. We find this method of measuring the far field profiles to be a very fast, sensitive and reliable for verifying the single mode character of our fiber tips. Third, in figure 2c we show the total transmission during the time of the pulling-process. We observe $>99\%$ transmission during the complete pulling process; consistent with the full biconical fiber taper being single-mode before the fast pull occurs. After confirming that our fiber tips are single mode, we measure the coupling efficiency ($\eta_c$) to a tapered $\mathrm{Si_3N_4}$ waveguide (see Refs. [@thompson13a; @tiecke14] for details of the device fabrication). The waveguide we use has a photonic crystal cavity, which, away from the cavity resonance, reflects all the incident light. We measure the reflected power $P_\mathrm{r}$, normalized to the power $P_\mathrm{in}$ in the fiber before the taper is pulled (see Fig. 3a). The normalized reflection is given by $P_\mathrm{r}/P_\mathrm{in}=\eta_c^2 \eta_m \eta_{bs} \eta_{FC}$, where $\eta_{bs}$ and $\eta_{FC}$ are the fiber beamsplitter and FC-FC coupling efficiencies and $\eta_m$ is the Bragg mirror reflectivity (see SI). We obtain a coupling efficiency of $\eta_c=0.97(1)$, where the error bars reflect drifts of the input power and calibrations over the course of our measurements. In comparison we achieve coupling efficiencies of $\eta_c\simeq 0.5 - 0.6$ for fiber tips with parameters such as fiber D. Figure 3d shows measurements of the coupling efficiency for waveguide angles of $2^\circ \leq \alpha \leq 8^\circ$ and for rectangular waveguides with a width ranging from: $100 <dx< 250\,\mathrm{nm}$, all waveguides have a $dy=175\,\mathrm{nm}$ thickness. For these measurements the waveguides are attached to the chip on one side only (see figure 3b). We observe that the coupling efficiency for most of these devices is $\geq 95\,\%$. The coupling efficiencies for $dx\leq 150\,\mathrm{nm}$ are slightly lower, consistent with a not fully adiabatic coupler ($z_b>z_t$) since for decreasing waveguide width $z_b$ increases while in our measurement we keep $z_t$ constant. ![**Coupling to photonic crystal waveguide cavities.** **a.** Setup to measure fiber-waveguide coupling efficiency. A tunable probe laser is coupled weakly to the fiber connecting to the device using a 99:1 fiber beamsplitter. The polarization at the waveguide is adjusted by means of a fiber polarization controller and the light is in and out coupled of the fiber network using fiber collimators (FC) **b.** SEM image of an array of singly-clamped photonic crystal waveguide cavities used for on-chip measurements. **c.** SEM image of a photonic crystal cavity attached to the fiber tip, *inset* shows a zoom of the fiber-waveguide coupler. **d.** Schematic of the various waveguide geometries. **e.** Coupling efficiencies for a range of waveguides; the devices are either a tapered waveguide with an opening angle $\alpha$ or rectangular waveguides with a varying width $dx$ and $5\,\mu \mathrm{m}$ long before adiabatically expanding to the photonic crystal cavity. All waveguides are $175\,\mathrm{nm}$ thick and attached to the chip as in panel **b**. []{data-label="fig:fig3"}](figure3_6.pdf){width="1.0\columnwidth"} We also detach cavities from the chip and attach them to the fiber such that they are solely connected to the fiber tip (see SI). A typical device attached in free space is shown in figure 3c for which we measure a coupling efficiency of $\eta_c=0.96(1)$. We find that our alignment procedure (see SI) allows to optimize the coupling efficiency in a reliable and reproducible manner, however, since we perform the alignment under an optical microscope we do not have exact knowledge of the fiber-waveguide interface. From our simulations we find that for a range of configurations the coupling efficiency is close to unity and consistent with the design criteria defined above. Outlook ======= We have presented a method for highly efficient fiber coupling to nanophotonic waveguides. Our measurements indicate coupling efficiencies as high as $97(1)\%$ for a range of devices. These results open the door for a range of unique applications in quantum optics and nano photonics. In particular, in combination with our recent results demonstrating strong coupling of a single atom to photonic crystals [@thompson13a; @tiecke14], efficient coupling to fibers can enable the creation of highly non-classical Schrödinger cat states of light [@wang05] and realization of efficient protocols for scalable quantum networks [@kimble08]. Moreover, the flexible geometries as well as the fiber-based mechanical support for nanophotonic devices, allowed by this approach open the door for new applications in nanoscale biosensing [@tan92; @yan12; @shambat13]. Funding Information {#funding-information .unnumbered} =================== Financial support was provided by the NSF, the Center for Ultracold Atoms, the Natural Sciences and Engineering Research Council of Canada, the Air Force Office of Scientific Research Multidisciplinary University Research Initiative and the Packard Foundation. KPN acknowledges support from Strategic Innovation Program of Japan Science and Technology Agency (JST). JDT acknowledges support from the Fannie and John Hertz Foundation and the NSF Graduate Research Fellowship Program. This work was performed in part at the Center for Nanoscale Systems (CNS), a member of the National Nanotechnology Infrastructure Network (NNIN), which is supported by the National Science Foundation under NSF award no. ECS-0335765. CNS is part of Harvard University. [10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} , , & ** (, ). , & . ** ****, (). <http://dx.doi.org/10.1038/asiamat.2010.194>. ** ****, (). <http://dx.doi.org/10.1038/nphoton.2009.240>. , & . ** ****, (). <http://scitation.aip.org/content/aip/journal/apl/44/7/10.1063/1.94865>. , , & . ** ****, (). <http://www.sciencedirect.com/science/article/pii/0304399184902018>. , , , & . ** ****, (). <http://www.sciencemag.org/content/258/5083/778.abstract>. *et al.* . ** ****, (). <http://dx.doi.org/10.1038/nnano.2011.226>. . ** ****, (). <http://dx.doi.org/10.1038/nature07127>. *et al.* . ** ****, (). <http://dx.doi.org/10.1038/nature08812>. , , & . ** ****, (). *et al.* . ** ****, (). <http://dx.doi.org/10.1038/nature03119>. *et al.* . ** ****, (). <http://dx.doi.org/10.1038/nature13188>. , , & . ** ****, (). <http://link.aps.org/doi/10.1103/PhysRevLett.81.5932>. , & . ** ****, (). <http://www.opticsexpress.org/abstract.cfm?URI=oe-21-9-11227>. , , , & . ** **** (). <http://scitation.aip.org/content/aip/journal/apl/103/18/10.1063/1.4826924>. *et al.* . ** ****, (). <http://www.sciencemag.org/content/340/6137/1202.abstract>. & ** (, ). *et al.* . ** ****, (). *et al.* . ** (). , & . ** ****, (). <http://www.opticsexpress.org/abstract.cfm?URI=oe-18-22-22677>. & . ** ****, (). <http://www.opticsexpress.org/abstract.cfm?URI=OPEX-8-3-173>. *et al.* . ** ****, (). <http://dx.doi.org/10.1038/nature02193>. , , & . ** (). . ** ****, (). <http://stacks.iop.org/2040-8986/12/i=4/a=043001>. & . ** ****, (). <http://link.aps.org/doi/10.1103/PhysRevA.72.022320>. *et al.* . ** ****, (). Supplementary Information {#supplementary-information .unnumbered} ========================= For all measurements we assume perfect reflection from the Bragg mirror ($\eta_m\equiv 1$) and we correct for the independently calibrated beamsplitter ratio ($\eta_{bs}=0.99$) and the FC-FC fiber coupling (with a typical value of $\eta_{FC}=0.89$). We note that our results for $\eta_c$ are conservative since we observe a small amount of scattering from the entrance Bragg mirror ($\eta_m<1$) and did not account for propagation losses through the $\simeq 5\,\mathrm{m}$ of fiber. We estimate these combined losses to be on the percent level, however, our current method has not sufficient accuracy to determine $\eta_c$ to a higher precision. Aside from the high coupling efficiency another important property of the coupler are its possible reflections. If light is reflected from the fiber-waveguide interface rather than the Bragg mirror this could affect measurements performed in reflection, such as the fiber-wg coupling efficiency characterization performed here, or quantum optics experiments. In order to characterize spurious reflections from our coupler we perform a measurement using a critically-coupled double-sided cavity. In the absence of cavity losses the reflection vanishes on resonance (enabling us to measure spurious reflections from the coupler) and the cavity is fully reflective off resonance (enabling us to measure the fiber-wg coupling efficiency as described above and in the main text). We choose a low quality factor ($Q\simeq 2000$) in order to minimize the effect of cavity losses and detach the cavity from the chip to avoid additional reflections from the chip. For this specific device we achieve a slightly sub-optimal coupling efficiency of $\eta_c=0.87$ and for TE-polarized light (see figure S1) we observe an on-resonance reflection of $2.0(4)\times10^{-3}$. Additionally, we do not observe any reflection ($0.1(3)\times10^{-3}$) of the fiber tip when no cavity is attached. Finally, in an independent measurement using TM-polarized light we measure a reflectivity of $4(1)\times10^{-3}$. These values set an upper limit of the reflection from the coupler since we have assumed a perfectly symmetric cavity, no cavity losses, perfect polarized probe light and no reflection of TM-polarized light by the Bragg mirror. Here, we have assigned the polarizations based on maximum and minimum reflection from the Bragg mirror which is designed to have a band gap for TE-polarized light at 780 nm. In order to align the relative fiber-waveguide position we mount the fiber under an angle of $\sim 10-20^\circ$ and align the fiber with respect to the waveguides under an optical microscope using a three axis translation stage with micrometer and piezo control. When the fiber is brought into contact with the waveguide, they stick together, allowing stable alignment to be maintained over long periods of time. They can be released from each other by pulling them apart with the translation stage, apparently without damage. While we have not investigated the mechanism of the sticking, we note that “stiction” is very common and has been extensively studied in the context of micro- and nano-mechanical systems. We obtain the optimal fiber-waveguide coupling by adjusting the fiber position, optimizing for the reflected power. Typically, the optimal coupling is achieved by slightly lifting the fiber off the chip which we attribute to the waveguide bending and aligning with the angle-mounted fiber taper. The procedure to detach cavities from the chip and attach them to the fiber taper is as follows: we break off and pick up a waveguide using a tungsten tip and transfer it to the fiber tip in free space using three axis translation mounts for both the fiber and tungsten tip. We move the waveguide using the tungsten tip while optimizing for optimal reflection. ![**Spurious reflections from coupler.** Reflection spectrum of a critically damped cavity (*Inset* shows full spectrum) for TE-polarized light (red) and for the bare fiber taper without a cavity attached (green).[]{data-label="fig:fig3"}](figureSI1_1.pdf){width="1.0\columnwidth"} [^1]: These authors contributed equally to this work [^2]: These authors contributed equally to this work [^3]: These authors contributed equally to this work

--- abstract: 'Online variants of the Expectation Maximization (EM) algorithm have recently been proposed to perform parameter inference with large data sets or data streams, in independent latent models and in hidden Markov models. Nevertheless, the convergence properties of these algorithms remain an open problem at least in the hidden Markov case. This contribution deals with a new online EM algorithm which updates the parameter at some deterministic times. Some convergence results have been derived even in general latent models such as hidden Markov models. These properties rely on the assumption that some intermediate quantities are available in closed form or can be approximated by Monte Carlo methods when the Monte Carlo error vanishes rapidly enough. In this paper, we propose an algorithm which approximates these quantities using Sequential Monte Carlo methods. The convergence of this algorithm and of an averaged version is established and their performance is illustrated through Monte Carlo experiments.' author: - 'Sylvain Le Corff[^1] [^2] and Gersende Fort' bibliography: - './onlineblock.bib' title: 'Convergence of a Particle-based Approximation of the Block Online Expectation Maximization Algorithm' --- ------------------------------------------------------------------------ This extended version of the paper “Convergence of a Particle-based Approximation of the Block Online Expectation Maximization Algorithm“, by S. Le Corff and G. Fort, provides detailed proofs which have been omitted in the submitted paper since they are very close to existing results. These additional proofs are postponed to Appendix \[app:extended:version\]. ------------------------------------------------------------------------ Introduction {#sec:intro} ============ The Expectation Maximization (EM) algorithm is a well-known iterative algorithm to solve maximum likelihood estimation in incomplete data models, see  [@dempster:laird:rubin:1977]. Each iteration is decomposed into two steps: in the E-step the conditional expectation of the complete log-likelihood (log of the joint distribution of the hidden states and the observations) given the observations is computed; and the M-step updates the parameter estimate. The EM algorithm is mostly practicable if the model belongs to the curved exponential family, see [@mclachlan:krishnan:1997 Section $1.5$] and [@cappe:moulines:ryden:2005 Section $10.1$], so that we assume below that our model belongs to this family. Under mild regularity conditions, this algorithm is known to converge to the stationary points of the log-likelihood of the observations, see [@wu:1983]. However, the original EM algorithm cannot be used to perform online estimation or when the inference task relies on large data sets. Each iteration requires the whole data set and each piece of data needs to be stored and scanned to produce a new parameter estimate. [*Online*]{} variants of the EM algorithm were first proposed for independent and identically distributed (i.i.d.) observations: [@cappe:moulines:2009] proposed to replace the original E-step by a stochastic approximation using the new observation. Solutions have also been proposed in hidden Markov models (HMM):  [@cappe:2011] provides an algorithm for finite state-space HMM which relies on recursive computations of the filtering distributions combined with a stochastic approximation step. Note that, since the state-space is finite, deterministic approximations of these distributions are available. This algorithm has been extended to the case of general state-space models, the approximations of the filtering distributions being handled with Sequential Monte Carlo (SMC) algorithms, see [@cappe:2009], [@delmoral:doucet:singh:2010a] and [@lecorff:fort:moulines:2011]. Unfortunately, it is quite challenging to address the asymptotic behavior of these algorithms (in the HMM case) since the recursive computation of the filtering distributions relies on approximations which are really difficult to control. In [@lecorff:fort:2011], another online variant of the EM algorithm in HMM is proposed, called the Block Online EM (BOEM) algorithm. In this case, the data stream is decomposed into blocks of increasing sizes. Within each block, the parameter estimate is kept fixed and the update occurs at the end of the block. This update is based on a single scan of the observations, so that it is not required to store any block of observations. [@lecorff:fort:2011] provides results on the convergence and on the convergence rates of the BOEM algorithms. These analyses are established when the E-step (computed on each block) is available in closed form and when it can be approximated using Monte Carlo methods, under an assumption on the ${\mathrm{L}}_{p}$-error of the Monte Carlo approximation. In this paper, we consider the case when the E-step of the BOEM algorithm is computed with SMC approximations: the filtering distributions are approximated using a set of random weighted particles, see [@cappe:moulines:ryden:2005] and [@delmoral:2004]. The Monte Carlo approximation is based on an online variant of the Forward Filtering Backward Smoothing algorithm (FFBS) proposed in [@cappe:2011] and [@delmoral:doucet:singh:2010a]. This method is appealing for two reasons: first, it can be implemented forwards in time i.e. within a block, each observation is scanned once and never stored and the approximation computed on each block does not require a backward step - this is crucial in our online estimation framework. Secondly, recent work on SMC approximations provides ${\mathrm{L}}_{p}$-mean control of the Monte Carlo error, see e.g. [@dubarry:lecorff:2011] and [@delmoral:doucet:singh:2010b]. This control, combined with the results in [@lecorff:fort:2011], sparks off the convergence results and the convergence rates provided in this contribution. The paper is organized as follows: our new algorithm called the [*Particle Block Online EM*]{} algorithm (P-BOEM) is derived in Section \[sec:BOEM:description\] together with an [*averaged*]{} version. Section \[sec:MCexperiments\] is devoted to practical applications: the P-BOEM algorithm is used to perform parameter inference in stochastic volatility models and in the more challenging framework of the Simultaneous Localization And Mapping problem (SLAM). The convergence properties and the convergence rates of the P-BOEM algorithms are given in Section \[sec:convergence\]. The Particle Block Online EM algorithms {#sec:BOEM:description} ======================================= In Section \[subsec:BOEM:description:notation\], we fix notation that will be used throughout this paper. We then derive our online algorithms in Sections \[subsec:BOEM:description:PBOEM\] and \[subsec:BOEM:description:PBOEM:aver\]. We finally detail, in Section \[subsec:BOEM:description:SMC\], the SMC procedure that makes our algorithm a true online algorithm. Notations and Model assumptions {#subsec:BOEM:description:notation} ------------------------------- A hidden Markov model on $\Xset\times \Yset$ is defined by an initial distribution $\chi$ on $(\Xset,\sigmaX)$ and two families of transition kernels. In this paper, the transition kernels are parametrized by ${\theta}\in{\Theta}$, where ${\Theta}\subseteq {\mathbb{R}}^{d_{\theta}}$ is a compact set. In the sequel, the initial distribution $\chi$ on $(\Xset,\sigmaX)$ is assumed to be known and fixed. The parameter is estimated online in the maximum likelihood sense using a sequence of observations ${\mathbf{Y}}$. Online maximum likelihood parameter inference algorithms were proposed either with a gradient approach or an EM approach. In the case of finite state-spaces HMM, [@legland:mevel:1997] proposed a recursive maximum likelihood procedure. The asymptotic properties of this algorithm have recently been addressed in [@tadic:2010]. This algorithm has been adapted to general state-spaces HMM with SMC methods (see [@doucet:poyiadjis:singh:2009]). The main drawback of gradient methods is the necessity to scale the gradient components. As an alternative to performing online inference in HMM, online EM based algorithms have been proposed for finite state-spaces (see [@cappe:2011]) or general state-spaces HMM (see [@cappe:2009], [@delmoral:doucet:singh:2010a] and [@lecorff:fort:2011]). [@delmoral:doucet:singh:2010a] proposed a SMC method giving encouraging experimental results. Nevertheless, it relies on a combination of stochastic approximations and SMC computations so that its analysis is quite challenging. In [@lecorff:fort:2011], the convergence of an online EM based algorithm is established. This algorithm requires either the exact computation of intermediate quantities (available explicitly only in finite state-spaces HMM or in linear Gaussian models) or the use of Monte Carlo methods to approximate these quantities. We propose to apply this algorithm to general models where these quantities are replaced by SMC approximations. We prove that the Monte Carlo error is controlled in such a way that the convergence properties of [@lecorff:fort:2011] hold for the P-BOEM algorithms. We now detail the model assumptions. Consider a family of transition kernels $\{m_{\theta}(x,x') {\mathrm{d}}\lambda(x')\}_{{\theta}\in{\Theta}}$ on $\Xset\times\sigmaX$, where $\Xset$ is a general state-space equipped with a countably generated $\sigma$-field $\sigmaX$, and $\lambda$ is a finite measure on $(\Xset,\sigmaX)$. Let $\{g_{\theta}(x,y) {\mathrm{d}}\nu(y)\}_{{\theta}\in{\Theta}}$ be a family of transition kernels on $\Xset \times \sigmaY$, where $\Yset$ is a general space endowed with a countably generated $\sigma$-field $\sigmaY$ and $\nu$ is a measure on $(\Yset,\sigmaY)$. Let ${\mathbf{Y}}=\{{\mathbf{Y}}_t\}_{t\in\Zset}$ be the observation process defined on $\left(\Omega,{{\mathbb{P}}},\mathcal{F}\right)$ and taking values in $\Yset^\Zset$. The batch EM algorithm is an offline maximum likelihood procedure which iteratively produces parameter estimates using the complete data log-likelihood (log of the joint distribution of the observations and the states) and a fixed set of observations, see [@dempster:laird:rubin:1977]. In the HMM context presented above, given $T$ observations ${\mathbf{Y}}_{1:T}$, the missing data $x_{0:T}$ and a parameter ${\theta}$, the complete data log-likelihood may be written as (up to the initial distribution $\chi$ which is assumed to be known) $$\label{eq:complete:loglikeli} \ell_{{\theta}}(x_{0:T},{\mathbf{Y}}_{1:T}){\ensuremath{\stackrel{\mathrm{def}}{=}}}\sum_{t=1}^{T}\left\{\log m_{{\theta}}(x_{t-1},x_{t})+\log g_{{\theta}}(x_{t},{\mathbf{Y}}_{t})\right\}{\;},$$ where we use $x_{r:t}$ as a shorthand notation for the sequence $(x_r,\dots, x_t)$, $r \leq t$. Each iteration of the batch EM algorithm is decomposed into two steps. The E-step computes, for all ${\theta}\in{\Theta},$ an expectation of the complete data log-likelihood under the conditional probability of the hidden states given the observations and the current parameter estimate $\hat {\theta}$. In the HMM context, due to the additive form of the complete data log-likelihood (\[eq:complete:loglikeli\]), the E-step is decomposed into $T$ expectations under the conditional probabilities ${\Phi_{\hat {\theta},t,T}^{\chi,0}}(\cdot,{\mathbf{y}})$ where $$\label{eq:define-Phi} {\Phi_{{\theta},s,t}^{\chi,r}}(h,{\mathbf{y}}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{\int \chi({\mathrm{d}}x_r) \{ \prod_{i=r}^{t-1} m_{\theta}(x_i, x_{i+1})g_{{\theta}}(x_{i+1},{\mathbf{y}}_{i+1}) \} \, h(x_{s-1},x_{s},{\mathbf{y}}_{s}) \, {\mathrm{d}}\lambda(x_{r+1:t})}{\int \chi({\mathrm{d}}x_r) \{\prod_{i=r}^{t-1} m_{\theta}(x_i, x_{i+1})g_{{\theta}}(x_{i+1},{\mathbf{y}}_{i+1}) \}\, {\mathrm{d}}\lambda(x_{r+1:t})}{\;},$$ for any bounded function $h$, any ${\theta}\in{\Theta}$, any $r <s \leq t$ and any sequence ${\mathbf{y}}\in \Yset^\Zset$. Then, given the current value of the parameter $\hat{\theta}$, the E-step amounts to computing the quantity $$\label{eq:QEM} Q_T({\theta},\hat{\theta}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{T} \sum_{t=1}^{T}{\Phi_{\hat{\theta},t,T}^{\chi,0}}\left( \log m_{{\theta}}+\log g_{{\theta}}, {\mathbf{Y}}\right){\;},$$ for any ${\theta}\in {\Theta}$. The M-step sets the new parameter estimate as a maximum of this expectation over ${\theta}$. The computation of ${\theta}\mapsto Q_{T}({\theta},\hat{\theta})$ for any ${\theta}\in {\Theta}$ is usually intractable except in the case of complete data likelihood belonging to the curved exponential family, see [@mclachlan:krishnan:1997 Section $1.5$] and [@cappe:moulines:ryden:2005 Section $10.1$]. Therefore, in the sequel, the following assumption is assumed to hold: \[assum:exp\] (a) \[assum:exp:decomp\]There exist continuous functions $\phi : {\Theta}\to {\mathbb{R}}$, $\psi : {\Theta}\to {\mathbb{R}}^d$ and $S: \Xset \times \Xset \times \Yset \to {\mathbb{R}}^d$ s.t. $$\label{eq:exponential:family} \log m_{\theta}(x,x') + \log g_{\theta}(x',y) = \phi({\theta}) + {\left\langle S(x,x,',y), \psi({\theta}) \right\rangle} {\;},$$ where ${\left\langle \cdot, \cdot \right\rangle}$ denotes the scalar product on ${\mathbb{R}}^d$. (b) \[assum:exp:convex\] There exists an open subset ${\mathcal{S}}$ of ${\mathbb{R}}^d$ that contains the convex hull of $S(\Xset \times \Xset \times \Yset)$. (c) \[assum:exp:max\]There exists a continuous function $\bar {\theta}: {\mathcal{S}}\to {\Theta}$ s.t. for any $s \in {\mathcal{S}}$, $$\bar {\theta}(s) = \mathrm{argmax}_{{\theta}\in {\Theta}} \; \left\{ \phi({\theta}) + {\left\langle s, \psi({\theta}) \right\rangle} \right\} {\;}.$$ Under A\[assum:exp\], the quantity $Q_{T}({\theta},\hat{\theta})$ defined by becomes $$\label{eq:QEM-exp} Q_{T}({\theta},\hat{\theta}) = \phi({\theta}) + {\left\langle \frac{1}{T}\sum_{t=1}^{T}{\Phi_{\hat{\theta},t,T}^{\chi,0}}\left(S, {\mathbf{Y}}\right), \psi({\theta}) \right\rangle}{\;},$$ so that the definition of the function ${\theta}\mapsto Q_{T}({\theta},\hat{\theta})$ requires the computation of an expectation $\frac{1}{T}\sum_{t=1}^{T}{\Phi_{\hat{\theta},t,T}^{\chi,0}}\left(S, {\mathbf{Y}}\right)$ independently of ${\theta}$. The M-step of the batch EM iteration amounts to computing $$\bar{\theta}\left(\frac{1}{T}\sum_{t=1}^{T}{\Phi_{\hat{\theta},t,T}^{\chi,0}}\left(S, {\mathbf{Y}}\right)\right){\;}.$$ This batch EM algorithm is designed for a fixed set of observations. A natural extension of this algorithm to the online context is to define a sequence of parameter estimates by $${\theta}_{t+1} = \mathrm{argmax}_{{\theta}} \ Q_{t+1}({\theta}, {\theta}_t) {\;}.$$ Unfortunately, the computation of $Q_{t+1}({\theta}, {\theta}_t)$ requires the whole set of observations to be stored and scanned for each estimation. For large data sets the computation cost of the E-step makes it intractable in this case. To overcome this difficulty, several online variants of the batch EM algorithm have been proposed, based on a recursive approximation of the function ${\theta}\mapsto Q_{t+1}(\cdot,{\theta}_t)$ (see [@cappe:2009], [@delmoral:doucet:singh:2010a] and [@lecorff:fort:2011]). In this paper, we focus on the Block Online EM (BOEM) algorithm, see [@lecorff:fort:2011]. Particle Block Online EM (P-BOEM) {#subsec:BOEM:description:PBOEM} --------------------------------- The BOEM algorithm, introduced in [@lecorff:fort:2011], is an online variant of the EM algorithm. The observations are processed sequentially per block and the parameter estimate is updated at the end of each block. Let $\{\tau_k\}_{k \geq 1}$ be a sequence of positive integers denoting the length of the blocks and set $$\label{eq:timeupdate} T_n{\ensuremath{\stackrel{\mathrm{def}}{=}}}\sum_{k=1}^n \tau_k\quad\mbox{and}\quad T_0 {\ensuremath{\stackrel{\mathrm{def}}{=}}}0{\;};$$ $\{T_k\}_{k \geq 1}$ are the deterministic times at which the parameter updates occur. Define, for all integers $\tau>0$ and $T\geq 0$ and all ${\theta}\in{\Theta}$, $$\label{eq:rewrite:barS} \bar S_{\tau}^{\chi,T}({\theta}, {\mathbf{Y}}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{\tau} \sum_{t=T+1}^{T+\tau}{\Phi_{{\theta},t,T+\tau}^{\chi,T}}\left( S, {\mathbf{Y}}\right){\;}.$$ The quantity $\bar S_{\tau}^{\chi,T}({\theta}, {\mathbf{Y}})$ corresponds to the intermediate quantity in with the observations ${\mathbf{Y}}_{T+1:T+\tau}$. The BOEM algorithm iteratively defines a sequence of parameter estimates $\{{\theta}_n\}_{n \geq 0}$ as follows: given the current parameter estimate ${\theta}_{n}$, (i) \[BOEM:stat\] compute the quantity $\bar S_{\tau_{n+1}}^{\chi,T_{n}}({\theta}_{n}, {\mathbf{Y}})$, (ii) \[BOEM:param\]compute a candidate for the new value of the parameter: ${\theta}_{n+1} = \bar{\theta}\left(\bar S_{\tau_{n+1}}^{\chi,T_{n}}({\theta}_{n}, {\mathbf{Y}}) \right)$, To make the exposition easier, we assume that the initial distribution $\chi$ is the same on each block. The dependence of $\bar S_{\tau}^{\chi,T}({\theta}, {\mathbf{Y}})$ on $\chi$ is thus dropped from the notation for better clarity. The quantity $\bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})$ is available in closed form only in the case of linear Gaussian models and HMM with finite state-spaces. In HMM with general state-spaces $\bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})$ cannot be computed explicitly and we propose to compute an approximation of $\bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})$ using SMC algorithms thus yielding the [*Particle-BOEM*]{} (P-BOEM) algorithm. Different methods can be used to compute these approximations (see e.g. [@delmoral:doucet:singh:2010b], [@delmoral:doucet:singh:2010a] and [@douc:garivier:moulines:olsson:2010]). We will discuss in Section \[subsec:BOEM:description:SMC\] below some SMC approximations that use the data sequentially. Denote by $\widetilde S_{n}({\theta}, {\mathbf{Y}})$ the SMC approximation of $\bar S_{\tau_{n+1}}^{T_{n}}({\theta}, {\mathbf{Y}})$ computed with $N_{n+1}$ particles. The P-BOEM algorithm iteratively defines a sequence of parameter estimates $\{{\theta}_n\}_{n \geq 0}$ as follows: given the current parameter estimate ${\theta}_{n}$, (i) \[PBOEM:stat\] compute the quantity $\widetilde S_{n}({\theta}_n, {\mathbf{Y}})$, (ii) \[PBOEM:param\]compute a candidate for the new value of the parameter: $${\theta}_{n+1} = \bar{\theta}\left(\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) \right){\;}. $$ We give in Algorithm \[alg:PBOEM\] lines $1$ to $9$ an algorithmic description of the P-BOEM algorithm. Note that the idea of processing the observations by blocks is proposed in [@mclachlan:ng:2003] to fit a normal mixture model. The incremental EM algorithm discussed in [@mclachlan:ng:2003] is an alternative to the batch EM algorithm for very large data sets. Contrary to our framework, in the algorithm proposed by [@mclachlan:ng:2003], the number of observations is fixed and the same observations are scanned several times. Averaged Particle Block Online EM {#subsec:BOEM:description:PBOEM:aver} --------------------------------- Following the same lines as in [@lecorff:fort:2011], we propose to replace the P-BOEM sequence $\{{\theta}_n \}_{n \geq 0}$ by an [*averaged*]{} sequence. This new sequence can be computed recursively, simultaneously with the P-BOEM sequence, and does not require additional storage of the data. The proposed averaged P-BOEM algorithm is defined as follows (see also lines $5$ and $6$ of Algorithm \[alg:PBOEM\]): the step of the P-BOEM algorithm presented above is followed by (iv) \[ABOEM:stat\] compute the quantity $$\label{eq:PBonem:averaged} \widetilde\Sigma_{n+1} = \frac{T_{n}}{T_{n+1}}\widetilde\Sigma_{n} + \frac{\tau_{n+1}}{T_{n+1}}\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}){\;},$$ (v) \[ABOEM:param\] define $$\label{eq:abonem:recursion} \widetilde {\theta}_{n+1} {\ensuremath{\stackrel{\mathrm{def}}{=}}}\bar{\theta}\left(\widetilde\Sigma_{n+1}\right){\;}.$$ We set $\widetilde\Sigma_{0} =0$ so that $$\label{eq:tildeSigma:bis} \widetilde\Sigma_{n} = \frac{1}{T_n} \sum_{j=1}^n \tau_j \, \widetilde S_{j-1}({\theta}_{j-1}, {\mathbf{Y}}) {\;};$$ we will prove in Section \[sec:averaging\] that the rate of convergence of the averaged sequence $\{ \widetilde {\theta}_n\}_{n \geq 0}$, computed from the averaged statistics $\{\widetilde\Sigma_{n}\}_{n \geq 0}$, is better than the non-averaged one. We will also observe this property in Section \[sec:MCexperiments\] by comparing the variability of the P-BOEM and the averaged P-BOEM sequences in numerical applications. $\theta_{0}$, $\{\tau_{n}\}_{n\geq 1}$, $\{N_{n}\}_{n\geq 1}$, $\{{\mathbf{Y}}_{t}\}_{t\geq 0}{\;}.$ $\{{\theta}_{n}\}_{n\geq 0}$ and $\{\widetilde{\theta}_{n}\}_{n\geq 0}{\;}.$ Set $\widetilde\Sigma_{0} = 0$. Compute sequentially $\widetilde S_{i}({\theta}_{i}, {\mathbf{Y}}){\;}.$ Set ${\theta}_{i+1} = \bar{\theta}\left(\widetilde S_{i}({\theta}_{i}, {\mathbf{Y}})\right){\;}.$ Set $$\widetilde\Sigma_{i+1} = \frac{T_{i}}{T_{i+1}}\widetilde\Sigma_{i} + \frac{\tau_{i+1}}{T_{i+1}}\widetilde S_{i}({\theta}_{i}, {\mathbf{Y}}){\;}.$$ Set $\widetilde{\theta}_{i+1} = \bar{\theta}\left(\widetilde \Sigma_{i+1}\right){\;}.$ The SMC approximation step {#subsec:BOEM:description:SMC} -------------------------- As the P-BOEM algorithm is an online algorithm, the SMC algorithm should use the data sequentially: no backward pass is allowed to browse all the data at the end of the block. Hence, the approximation is computed recursively within each block, each observation being used once and never stored. These SMC algorithms will be referred to as [*forward only SMC*]{}. We detail below a forward only SMC algorithm for the computation of $\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}})$ which has been proposed by [@cappe:2011] (see also [@delmoral:doucet:singh:2010a]). For notational convenience, the dependence on $n$ is omitted. For block $n$, the algorithm below has to be applied with $(\tau,N) \leftarrow (\tau_{n+1}, N_{n+1})$, $Y_{1:\tau} \leftarrow Y_{T_{n}+1,T_{n}+\tau_{n+1}}$ and $\theta \leftarrow \theta_{n}$. The key property is to observe that $$\label{eq:SMCapprox:tool1} \bar S_{\tau}^{0}({\theta}, {\mathbf{Y}}) = \phi_\tau^{\theta}(R_{{\theta},\tau})$$ where $\phi_{t}^{\theta}$ is the filtering distribution at time $t$, and the functions $R_{t,{\theta}}:\Xset \to {\mathcal{S}}$, $1 \leq t \leq \tau$, satisfy the equations $$\label{eq:SMCapprox:tool2} R_{t,{\theta}}(x) = \frac{1}{t} \mathrm{B}_{t}^{\theta}\left(x, S(\cdot,x,Y_t)\right) + \frac{t-1}{t} \mathrm{B}_{t}^{\theta}\left(x, R_{t-1,{\theta}}\right) {\;},$$ where $\mathrm{B}_{t}^{\theta}$ denotes the backward smoothing kernel at time $t$ $$\label{eq:SMCapprox:tool3} \mathrm{B}_{t}^{\theta}(x, {\mathrm{d}}x') = \frac{m_{\theta}(x',x) }{\int m_{\theta}(u,x) \phi_{t-1}^{\theta}({\mathrm{d}}u) }\phi_{t-1}^{\theta}({\mathrm{d}}x') {\;}.$$ By convention, $R_{0,{\theta}}(x) = 0$ and $\phi_{0}^{\theta}= \chi$. A proof of the equalities (\[eq:SMCapprox:tool1\]) to (\[eq:SMCapprox:tool3\]) can be found in [@cappe:2011] and [@delmoral:doucet:singh:2010a]. Therefore, a careful reading of Eqs (\[eq:SMCapprox:tool1\]) to (\[eq:SMCapprox:tool3\]) shows that, for an iterative particle approximation of $\bar S_{\tau}^{0}({\theta}, {\mathbf{Y}})$, it is sufficient to update from time $t-1$ to $t$ (i) $N$ weighted samples $\left\{\left({\xi_{t}^{\ell}},{\omega_{t}^{\ell}}\right); \ell\in\{1,\dots,N\}\right\}$ used to approximate the filtering distribution $\phi_{t}^{\theta}$. (ii) the intermediate quantities $\{R_{t,{\theta}}^{\ell}\}_{\ell=1}^{N}$, approximating the function $R_{t,{\theta}}$ at point $x = {\xi_{t}^{\ell}}$, $\ell \in \{1,\cdots, N\}$. We describe below such an algorithm. An algorithmic description is also provided in Appendix \[app:alg\], Algorithm \[alg:FSMC\]. Given [*instrumental*]{} Markov transition kernels $\{{\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}(x,x'), t \leq \tau \}$ on $\Xset\times\mathcal{X}$ and adjustment multipliers $\{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}, t\leq \tau \}$, the procedure goes as follows: (i) [*line $1$ in Algorithm \[alg:FSMC\]:*]{} sample independently $N$ particles $\{{\xi_{0}^{\ell}}\}_{\ell=1}^{N}$ with the same distribution $\chi$. (ii) [*line $6$ in Algorithm \[alg:FSMC\]:*]{} at each time step $t\in\{1,\dots,\tau\}$, pairs $\{ (J_t^{\ell}, {\xi_{t}^{\ell}}) \}_{\ell = 1}^{N}$ of indices and particles are sampled independently (conditionally to $Y_{1:t}$, ${\theta}$ and $\{ (J_{t-1}^{\ell}, {\xi_{t-1}^{\ell}}) \}_{\ell = 1}^{N}$) from the instrumental distribution: $$\label{eq:instrumental-distribution-filtering} {\ensuremath{\pi_{t}}}(i, {\mathrm{d}}x) \propto {\omega_{t-1}^{i}} {\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{i}})}}{\mathrm{erreur}}}}} {\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}({\xi_{t-1}^{i}},x)\lambda({\mathrm{d}}x) {\;},$$ on the product space $\{1, \dots, N\} \times \Xset$. For any $t\in\{1,\dots,\tau\}$ and any $\ell\in\{1,\dots,N\}$, $J_t^{\ell}$ denotes the index of the selected particle at time $t-1$ used to produce ${\xi_{t}^{\ell}}$. (iii) [*line $7$ in Algorithm \[alg:FSMC\]:*]{} once the new particles $\{{\xi_{t}^{\ell}}\}_{\ell = 1}^{N}$ have been sampled, their importance weights $\{ {\omega_{t}^{\ell}}\}_{\ell = 1}^{N}$ are computed. (iv) [*lines $8$ in Algorithm \[alg:FSMC\]:*]{} update the intermediate quantities $\{R_{t,{\theta}}^{\ell}\}_{\ell=1}^{N}$. If, for all $x\in\Xset$, ${\ifthenelse{\equal{}{}}{\ifthenelse{\equal{x}{}}{\upsilon_{t}}{\upsilon_{t}(x)}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{x}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}(x)}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{x}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}(x)}}{\mathrm{erreur}}}}}=1$ and if the kernels ${\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}$ are chosen such that ${\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}} = m_{{\theta}}$, lines $6$-$7$ in Algorithm \[alg:FSMC\] are known as the [*Bootstrap filter*]{}. Other choices of ${\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}{}$ and ${\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}$ can be made, see e.g. [@cappe:moulines:ryden:2005]. Applications to Bayesian inverse problems in Hidden Markov Models {#sec:MCexperiments} ================================================================= Stochastic volatility model --------------------------- Consider the following stochastic volatility model: $$X_{t+1} = \phi X_t + \sigma U_{t}{\;}, \qquad \qquad Y_t = \beta {\mathrm{e}}^{\frac{X_t}{2}} V_t{\;},$$ where $X_0\sim\mathcal{N}\left(0, (1-\phi^2)^{-1} \sigma^2\right)$ and $\{U_t\}_{t\geq 0}$ and $\{V_t\}_{t\geq 0}$ are two sequences of i.i.d. standard Gaussian r.v., independent from $X_0$. We illustrate the convergence of the P-BOEM algorithms and discuss the choice of some design parameters such as the pair ($\tau_n, N_n$). Data are sampled using $\phi = 0.95$, $\sigma^{2} = 0.1$ and $\beta^{2} = 0.6$; we estimate ${\theta}= (\phi, \sigma^2, \beta^2)$ by applying the P-BOEM algorithm and its averaged version. All runs are started from $\phi = 0.1$, $\sigma^{2} = 0.6$ and $\beta^{2} = 2$. Figure \[fig:boxplotSVM\] displays the estimation of the three parameters as a function of the number of observations, over $50$ independent Monte Carlo runs. The block-size sequence is of the form $\tau_{n}\propto n^{1.2}$. For the SMC step, we choose $N_n = 0.25\cdot \tau_n$; particles are sampled as described in Algorithm \[alg:FSMC\] (see Appendix \[app:alg\]) with the bootstrap filter. For each parameter, Figure \[fig:boxplotSVM\] displays the empirical median (bold line) and upper and lower quartiles (dotted line). The averaging procedure is started after $1500$ observations. Both algorithms converge to the true values of the parameters and, once the averaging procedure is started, the variance of the estimation decreases (estimation of $\phi$ and $\beta^{2}$). The estimation of $\sigma^{2}$ shows that, if the averaging procedure is started with too few observations, the estimation can be slowed down. \ \ \ We now discuss the role of the pairs $(\tau_n, N_n)$. Roughly speaking (see section \[sec:convergence\] for a rigorous decomposition), $\tau$ controls the rate of convergence of $\bar S_{\tau}^{T}({\theta}, {\mathbf{Y}})$ to $\lim_{\tau \to \infty} \bar S_{\tau}^{T}({\theta}, {\mathbf{Y}}) $; and $N$ controls the error between $\bar S_{\tau}^{T}({\theta}, {\mathbf{Y}})$ and its SMC approximation. We will show in Section \[sec:convergence\] that $\lim_n \tau_n = \lim_n N_n = +\infty$ are part of some sufficient conditions for the P-BOEM algorithms to converge. We thus choose increasing sequences $\{\tau_n, N_n \}_{n \geq 1}$. The role of $\tau_n$ has been illustrated in [@lecorff:fort:2011 Section $3$]. Hence, in this illustration, we fix $\tau_n$ and discuss the role of $N_n$. Figure \[fig:varSVM\] compares the algorithms when applied with $\tau_n \propto n^{1.1}$ and $N_{n} = \sqrt{\tau_{n}}$ or $N_{n}=\tau_{n}$. The empirical variance (over $50$ independent Monte Carlo runs) of the estimation of $\beta^{2}$ is displayed, as a function of the number of blocks. First, Figure \[fig:varSVM\] illustrates the variance decrease provided by the averaged procedure, whatever the block size sequence. Moreover, increasing the number of particles per block improves the variance of the estimation given by the P-BOEM algorithm while the impact on the variance of the averaged estimation is less important. On average, the variance is reduced by a factor of $3.0$ for the P-BOEM algorithm and by a factor of $1.8$ for its averaged version when the number of particles goes from $N_{n}=\sqrt{\tau_{n}}$ to $N_{n}=\tau_{n}$. These practical considerations illustrate the theoretical results derived in Section \[sec:averaging\]. \ Finally, we discuss the role of the initial distribution $\chi$. In all the applications above, we have the same distribution $\chi\equiv\mathcal{N}\left(0, (1-\phi^2)^{-1} \sigma^2\right)$ at the beginning of each block. We could choose a different distribution $\chi_n$ for each block such as, e.g., the filtering distribution at the end of the previous block. We have observed that this particular choice of $\chi_{n}$ leads to the same behavior for both algorithms. To end this section, the P-BOEM algorithm is compared to the Online EM algorithm outlined in [@cappe:2011] and [@delmoral:doucet:singh:2010a]. These algorithms rely on a combination of stochastic approximation and SMC methods. According to classical results on stochastic approximation, it is expected that the rate of convergence of the Online EM algorithm behaves like $\gamma_n^{1/2}$, where $\{\gamma_n\}_{n\geq 0}$ is the so called step-size sequence. Hence, $\gamma_{n}$ in the Online EM algorithm is chosen such that $\gamma_n \propto n^{-0.55}$ and the block-size sequence in the P-BOEM algorithm such that $\tau_n \propto n^{1.2}$. The number of particles used in the Online EM algorithm is fixed and chosen so that the computational costs of both algorithms are similar. Provided that $N_{n}\propto \tau_{n}$ in the P-BOEM algorithm, this leads to a choice of $70$ particles for the Online EM algorithm. We report in Figure \[fig:boxplotSVM\_BOEM-OEM\], the estimation of $\phi$ and $\sigma^{2}$ for a Polyak-Ruppert averaged Online EM algorithm (see [@polyak:1990]) and the averaged P-BOEM algorithm as a function of the number of observations. The averaging procedure is started after about $1500$ observations. As noted in [@lecorff:fort:2011 Section $3$] for a constant sequence $\{N_{n}\}_{n\geq 0}$ this figure shows that both algorithms behave similarly. For the estimation of $\phi$ and $\beta^{2}$, the variance is smaller for the P-BOEM algorithm and the convergence is faster for the P-BOEM algorithm in the case of $\beta^{2}$. Conclusions are different for the estimation of $\sigma^{2}$: the variance is smaller for the P-BOEM algorithm but the Online EM algorithm converges a bit faster. The main advantage of the P-BOEM algorithm is that it relies on approximations which can be controlled in such a way that we are able to show that the limiting points of the P-BOEM algorithms are the stationary points of the limiting normalized log-likelihood of the observations. \ \ Simultaneous Localization And Mapping ------------------------------------- The Simultaneous Localization And Mapping (SLAM) problem arises when a mobile device wants to build a map of an unknown environment and, at the same time, has to estimate its position in this map. The common statistical approach for the SLAM problem is to introduce a state-space model. Many solutions have been proposed depending on the assumptions made on the transition and observation models, and on the map (see e.g. [@burgard:fox:thrun:2006], [@martinezcantin:2008] and [@montemerlo:2003]). In [@martinezcantin:2008] and [@lecorff:fort:moulines:2011], it is proposed to see the SLAM as an inference problem in HMM: the localization of the robot is the hidden state with Markovian dynamic, and the map is seen as an unknown parameter. Therefore, the mapping problem is answered by solving the inference task, and the localization problem is answered by approximating the conditional distribution of the hidden states given the observations. In this application, we consider a statistical model for a landmark-based SLAM problem for a bicycle manoeuvring on a plane surface. Let $x_t {\ensuremath{\stackrel{\mathrm{def}}{=}}}\{x_{t,i}\}_{i=1}^{3}$ be the robot position, where $x_{t,1}$ and $x_{t,2}$ are the robot’s cartesian coordinates and $x_{t,3}$ its orientation. At each time step, deterministic controls are sent to the robot so that it explores a given part of the environment. Controls are denoted by $( v_t,\psi_t)$ where $\psi_t$ stands for the robot’s heading direction and $v_t$ its velocity. The robot position at time $t$, given its previous position at time $t-1$ and the noisy controls $(\hat{v}_{t},\hat{\psi}_{t})$, can be written as $$\label{eq-transitionmodel} x_t = f(x_{t-1},\hat{v}_{t},\hat{\psi}_{t}){\;},$$ where $(\hat{v}_{t},\hat{\psi}_{t})$ is a $2$-dimensional Gaussian distribution with mean $( v_t,\psi_t)$ and known covariance matrix $Q$. In this contribution we use the kinematic model of the front wheel of a bicycle (see e.g. [@bailey:2006]) where the function $f$ in is given by $$f(x_{t-1},\hat{v}_{t},\hat{\psi}_{t}) = x_{t-1} + \begin{pmatrix} \hat{v}_{t}d_{t}\cos(x_{t-1,3} + \hat{\psi}_{t})\\\ \hat{v}_{t}d_{t}\sin(x_{t-1,3} + \hat{\psi}_{t})\\ \hat{v}_{t}d_{t}B^{-1}\sin(\hat{\psi}_{t})\end{pmatrix}{\;},$$ where $d_{t}$ is the time period between two successive positions and $B$ is the robot wheelbase. The $2$-dimensional environment is represented by a set of landmarks $\theta {\ensuremath{\stackrel{\mathrm{def}}{=}}}\{\theta_{j}\}_{ 1 \leq j \leq q}$, $\theta_{j} \in \mathbb{C}$ being the position of the $j-$th landmark. The total number of landmarks $q$ and the association between observations and landmarks are assumed to be known. At time $t$, the robot observes the distance and the angular position of all landmarks in its neighborhood; let $c_t \subseteq \{1, \cdots, q\}$ be the set of observed landmarks at time $t$. It is assumed that the observations $\{y_{t,i}\}_{i \in c_t}$ are independent and satisfy $$y_{t,i} = h(x_t,\theta_{i})+\delta_{t,i}{\;},$$ where $h$ is defined by $$h(x,\boldsymbol{\kappa}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\begin{pmatrix}\sqrt{(\boldsymbol{\kappa}_1-x_1)^2 + (\boldsymbol{\kappa}_2-x_2)^2 }\\\arctan{\frac{\boldsymbol{\kappa}_2-x_2}{\boldsymbol{\kappa}_1-x_1}}-x_3\end{pmatrix}{\;},$$ and the noise vectors $\{\delta_{t,i}\}_{t,i}$ are i.i.d Gaussian $\mathcal{N}\left(0,R\right)$. $R$ is assumed to be known. The model presented in this Section does not take into account all the issues arising in the SLAM problem (such as the association process which is assumed to be known and the known covariance matrices). The aim is to prove that the BOEM algorithm and its averaged version have satisfying behavior even in the challenging framework described above. The observation and motion models are highly nonlinear and we show that the BOEM algorithm remains stable in this experiment. Several solutions have been proposed to solve the association problem (see e.g. [@burgard:fox:thrun:2006] for a solution based on the likelihood of the observations) and could be adapted to our case. We want to estimate ${\theta}= \{{\theta}_j\}_{j =1}^q$ by applying the P-BOEM algorithms. In this paper, we use simulated data. $q=15$ landmarks are drawn in a square of size $45m\mathrm{x}45m$. The robot path is sampled with a given set of controls. Using the [*true*]{} positions of all landmarks in the map and the true path of the robot (see the dots and the bold line on Figure \[fig:BOEMSLAM\]), observations are sampled by setting: $R = \begin{pmatrix}\sigma_r^2 & \rho\\ \rho & \sigma_b^2\end{pmatrix}{\;},$ where $\sigma_r = 0.5\mbox{m}$, $\sigma_b = \frac{\pi}{60}\mbox{rad}$ and $\rho = 0.01$. We choose $Q = \mbox{diag}(\sigma_v^2, \sigma_{\phi}^2)$ where $\sigma_v = 0.5\mbox{m.s}^{-1}$, $\sigma_{\psi} = \frac{\pi}{60}\mbox{rad}$ and $B=1.5\mathrm{m}$. In this model, the transition denoted by $m_{{\theta}}$ does not depend on the map ${\theta}$ (see ) and the marginal likelihood $g_{{\theta}}$ is such that the complete data likelihood does not belong to the curved exponential family: $$\label{eq:expression:logg} \sum_{i\in c_{t}} \ln g_{{\theta}}(x_{t},y_{t,i})\propto \sum_{i\in c_{t}} \left[y_{t,i}-h(x_{t},{\theta}_{i})\right]^{\star}R^{-1}\left[y_{t,i}-h(x_{t},{\theta}_{i})\right] {\;}.$$ Hence, in order to apply Algorithm \[alg:PBOEM\], at the beginning of each block, $g_{{\theta}}$ is approximated by a function depending on the current parameter estimate so that the resulting approximated model belongs to the curved exponential family (see [@lecorff:fort:moulines:2011]). As can be seen from (\[eq:expression:logg\]), approximating the function $\boldsymbol{\kappa}\mapsto h(x,\boldsymbol{\kappa})$ by its first-order Taylor expansion at ${\theta}_{i}$ leads to a quadratic approximation of $g_{{\theta}}$. This approach is commonly used in the SLAM framework to use the properties of linear Gaussian models (see e.g. [@burgard:fox:thrun:2006]). As the landmarks are not observed all the time, we choose a slowly increasing sequence $\{\tau_{n}\propto n^{1.1}\}_{n\geq 1}$ so that the number of updates is not too small (in this experiment, we have $60$ updates for a total number of observations of $2000$). As the total number of observations is not so large (the largest block is of length $60$), the number of particles is chosen to be constant on each block: for all $n\geq 1$, $N_{n} = 50$. For the SMC step, we apply Algorithm \[alg:FSMC\] with the [*bootstrap filter*]{}. For each run the estimated path (equal to the weighted mean of the particles) and the estimated map at the end of the loop ($T=2000$) are stored. Figure \[fig:BOEMSLAM\] represents the mean estimated path and the mean map over $50$ independent Monte Carlo runs. It highlights the good performance of the P-BOEM algorithm in a more complex framework. ![True trajectory (bold line) and true landmark positions (balls) with the estimated path (dotted line) and the landmarks’ estimated positions (stars) at the end of the run ($T=2000$).[]{data-label="fig:BOEMSLAM"}](./BOEM_SLAM.pdf){width="60.00000%"} We also compare our algorithm to the [*marginal SLAM*]{} algorithm proposed by [@martinezcantin:2008]. In this algorithm, the map is also modeled as a parameter to learn in a HMM model; SMC methods are used to estimate the map in the maximum likelihood sense. The Marginal SLAM algorithm is a gradient-based approach for solving the recursive maximum likelihood procedure. Note that, in the case of i.i.d. observations, [@titterington:1984] proposed to update the parameter estimate each time a new observation is available using a stochastic gradient approach. Figure \[fig:BOEM\_Marginal\_SLAM\] illustrates the estimation of the position of each landmark. The P-BOEM algorithm is applied using the same parameters as above and the [*marginal SLAM*]{} algorithm uses a sequence of step-size $\{\gamma_{n}\propto n^{-0.6}\}_{n\geq 1}$. We use the averaged version of the P-BOEM algorithm and a Polyak-Ruppert based averaging procedure for the [*marginal SLAM*]{} algorithm (see [@polyak:1990]). For each landmark the last estimation (at the end of the loop) of the position is stored for each of the $50$ independent Monte Carlo runs. Figure \[fig:BOEM\_Marginal\_SLAM\] displays the distance between the estimated position and the true position for each landmark. In this experiment, the P-BOEM based SLAM algorithm outperforms the [*marginal SLAM*]{} algorithm. ![Distance between the final estimation and the true position for each of the $15$ landmarks with the averaged [*marginal SLAM*]{} algorithm (left) and the averaged P-BOEM algorithm (right).[]{data-label="fig:BOEM_Marginal_SLAM"}](./BOEM_Marginal_SLAM.pdf){width="60.00000%"} Convergence of the Particle Block Online EM algorithms {#sec:convergence} ====================================================== In this section, we analyze the limiting points of the P-BOEM algorithm. We prove in Theorem \[th:Pbonem:conv\] that the P-BOEM algorithm has the same limit points as a so-called [*limiting EM*]{} algorithm, which defines a sequence $\{{\theta}_n\}_{n \geq 0}$ by ${\theta}_{n+1} = \bar {\theta}\left[\bar S({\theta}_n)\right]$ where $\bar S({\theta})$ is the a.s. limit $\lim_{\tau \to +\infty} \bar S_{\tau}^{T}({\theta}, {\mathbf{Y}})$ (defined by (\[eq:rewrite:barS\])). As discussed in [@lecorff:fort:2011 Section 4.3.], the set of limit points of the limiting EM algorithm is the set of stationary points of the contrast function $\ell({\theta})$, defined as the a.s. limit of the normalized log-likelihood of the observations, when $T\to +\infty$. The convergence result below on the P-BOEM algorithm requires two sets of assumptions: conditions A\[assum:strong\] to A\[assum:block-size\] are the same as in [@lecorff:fort:2011] and imply the convergence of the BOEM algorithm; assumptions A\[assum:dub:lec\] and A\[assum:moment:w\] are introduced to control the Monte Carlo error. Assumptions ----------- Consider the following assumptions \[assum:strong\] There exist $\sigma_{-}$ and $\sigma_{+}$ s.t. for any $\left(x,x^{\prime}\right)\in\Xset^2$ and any ${\theta}\in {\Theta}$, $0 <\sigma_{-} \leq m_{\theta}(x,x^{\prime})\leq \sigma_{+}$. Set $\rho {\ensuremath{\stackrel{\mathrm{def}}{=}}}1 - (\sigma_-/\sigma_+){\;}.$ Define, for all $y\in\Yset$, $$\label{eq:b+:b-} b_-(y) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\inf_{{\theta}\in {\Theta}} \int g_{\theta}(x,y) \lambda({\mathrm{d}}x) \quad\mbox{and}\quad b_+(y) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\sup_{{\theta}\in {\Theta}} \int g_{\theta}(x,y) \lambda({\mathrm{d}}x) {\;}.$$ For any sequence of r.v. $Z{\ensuremath{\stackrel{\mathrm{def}}{=}}}\{Z_{t}\}_{t\in\Zset}$ on $(\Omega,\widetilde{{\mathbb{P}}},\mathcal{F})$, let $$\label{eq:Zfield} \mathcal{F}_{k}^{Z} {\ensuremath{\stackrel{\mathrm{def}}{=}}}\sigma\left(\{Z_{u}\}_{u\leq k}\right)\quad \mbox{and}\quad \mathcal{G}_{k}^{Z} {\ensuremath{\stackrel{\mathrm{def}}{=}}}\sigma\left(\{Z_{u}\}_{u\geq k}\right)$$ be $\sigma$-fields associated to $Z$. We also define the mixing coefficients by, see [@davidson:1994], $$\label{eq:def:mixing} {\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{Z}(n) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\underset{u\in\Zset}{\sup}\,{\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{}(\mathcal{G}_{u+n}^{Z},\mathcal{F}_{u}^{Z}){\;}, \forall\; n\geq0{\;},$$ where for any $\sigma$-algebras $\mathcal{F}$ and $\mathcal{G}$, $$\label{eq:def:mixing:2} {\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{}(\mathcal{G},\mathcal{F}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\underset{B\in\mathcal{G}}{\sup}\,|\widetilde{{\mathbb{P}}}(B| \mathcal{F}) - \widetilde{{\mathbb{P}}}(B)|{\;}.$$ For $p>0$ and $Z$ a ${\mathbb{R}}^{d}$-valued random variable measurable w.r.t. the $\sigma$-algebra $\mathcal{F}$, set $${\ensuremath{\left\| Z \right\|_{p}}}{\ensuremath{\stackrel{\mathrm{def}}{=}}}\left({ \ifthenelse{\equal{}{}}{\mathbb{E}^{}_{}\left[ |Z|^p\right]}{\mathbb{E}^{}_{}\left[ |Z|^p\middle | \right]} }\right)^{1/p} {\;}.$$ \[assum:moment:sup\] -($p$) ${\ensuremath{\left\| \sup_{x,x' \in \Xset^2} \, |S(x,x',{\mathbf{Y}}_{0})| \right\|_{p}}}<+\infty{\;}.$ \[assum:obs\] (a) \[assum:obs:erg\] ${\mathbf{Y}}$ is a stationary sequence such that there exist $C\in[0,1)$ and ${\ifthenelse{\equal{}{a}}{\alpha}{\beta} }\in(0,1)$ satisfying, for any $n\geq 0$, ${\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{{\mathbf{Y}}}(n) \leq C{\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{n}$, where ${\ifthenelse{\equal{}{a}}{\alpha}{\beta} }^{{\mathbf{Y}}}$ is defined in . (b) \[assum:obs:b+:b-\] ${ \ifthenelse{\equal{}{}}{\mathbb{E}^{}_{}\left[|\log b_-({\mathbf{Y}}_0)| +|\log b_+({\mathbf{Y}}_0)| \right]}{\mathbb{E}^{}_{}\left[|\log b_-({\mathbf{Y}}_0)| +|\log b_+({\mathbf{Y}}_0)| \middle | \right]} }<+\infty$. \[assum:block-size\] There exist $c>0$ and $a>1$ such that for all $n\ge 1$, $\tau_{n} = \lfloor cn^{a}\rfloor$. Assumptions A\[assum:strong\] to A\[assum:block-size\] are the same as in [@lecorff:fort:2011]. A\[assum:strong\], referred to as the [ *strong mixing condition*]{}, is used to prove the uniform forgetting property of the initial condition of the filter, see e.g. [@delmoral:guionnet:1998] and [@delmoral:ledoux:miclo:2003]. This assumption is easy to check in finite state-space HMM or when the state-space is compact when the Markov kernel $\boldsymbol{m_{{\theta}}}$ is sufficiently regular. As noted in [@lecorff:fort:2011], it can fail to hold in quite general situations. Nevertheless, the exponential forgetting property needed to ensure the convergence results could be checked under weaker assumptions (see [@douc:fort:moulines:priouret:2009] for a Doeblin assumption). However, it would imply quite technical supplementary results out of the scope of this paper. Examples of observation sequences satisfying A\[assum:obs\] include, for example, stationary $\psi$-irreducible and positive recurrent Markov chains which are geometrically ergodic (see e.g. [@meyn:tweedie:1993] for Markov chains theory). We need to control the ${\mathrm{L}}_{p}$-mean error on each block between $\bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})$ and its SMC approximation. This control is discussed in Section \[sec:theory:SMC\] below when the SMC approximation is computed as described in Section \[subsec:BOEM:description:SMC\]. ${\mathrm{L}}_{p}$-error of the SMC approximation {#sec:theory:SMC} ------------------------------------------------- For each block $n$, denote by $\{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t,n}}{\upsilon_{t,n}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t,n}}{\tilde{\upsilon}_{t,n}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t,n}}{\upsilon^\star_{t,n}()}}{\mathrm{erreur}}}}}\}_{t\leq \tau_{n+1}}$ and $\{{\ifthenelse{\equal{}{}}{q_{t,n}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t,n}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t,n}}{\mathrm{erreur}}}}}\}_{t\leq \tau_{n+1}}$ respectively the adjustment multipliers and the instrumental kernels in the SMC propagation step (see ). For all $y\in\Yset$, define $${\omega_{+}^{}}(y)=\underset{{\theta}\in{\Theta}}{\sup}\,\underset{\substack{(x,x^{'})\in\Xset\times\Xset \\ t\geq 0, n\geq 0}}{\sup} \dfrac{m_{{\theta}}(x,x^{\prime})g_{{\theta}}(x',y)}{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{x}{}}{\upsilon_{t,n}}{\upsilon_{t,n}(x)}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{x}{}}{\tilde{\upsilon}_{t,n}}{\tilde{\upsilon}_{t,n}(x)}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{x}{}}{\upsilon^\star_{t,n}}{\upsilon^\star_{t,n}(x)}}{\mathrm{erreur}}}}}{\ifthenelse{\equal{}{}}{q_{t,n}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t,n}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t,n}}{\mathrm{erreur}}}}}(x,x^{\prime})} {\;}.$$ Consider the following assumptions. \[assum:dub:lec\] $|{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_{\infty}{\ensuremath{\stackrel{\mathrm{def}}{=}}}\sup_{t,n}|{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t,n}}{\upsilon_{t,n}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t,n}}{\tilde{\upsilon}_{t,n}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t,n}}{\upsilon^\star_{t,n}()}}{\mathrm{erreur}}}}}|_{\infty} < \infty$. \[assum:moment:w\] -($p$) ${\ensuremath{\left\| \frac{{\omega_{+}^{}}({\mathbf{Y}}_{0})}{b_{-}({\mathbf{Y}}_{0})} \right\|_{p}}}<+\infty{\;}.$ In the case of the [*Bootstrap filter*]{}, A\[assum:dub:lec\] holds (since $v_{t,n} =1$) and ${\omega_{+}^{}}(y)= \underset{{\theta}\in{\Theta}}{\sup}\,\underset{\substack{x\in\Xset}}{\sup} \,g_{{\theta}}(x,y)$. \[prop:check:lp:control\] Let $S : \Xset^{2}\times\Yset\longrightarrow {\mathbb{R}}^{d}$ be a measurable function s.t. A\[assum:moment:sup\]-($\bar p$) holds for some $\bar p >2$. Assume A\[assum:strong\], A\[assum:obs\], A\[assum:dub:lec\]. Define $\Delta p {\ensuremath{\stackrel{\mathrm{def}}{=}}}2\bar p p/(\bar p- p)$ and assume A\[assum:moment:w\]-($\Delta p$) holds for some $p\in(2,\bar p)$. Then, there exists a constant $C$ s.t. for all $n\ge 0$, $${\left\| \widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) - \bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}}) \right\|_{p}}\leq C\left(\frac{1}{N_{n+1}}+\frac{1}{\tau_{n+1}^{1/2}N_{n+1}^{1/2}}\right){\;},$$ where $\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}})$ is computed with the algorithm described in Section \[subsec:BOEM:description:SMC\]. Asymptotic behavior of the Particle Block Online EM algorithms {#sec:PBONEM} -------------------------------------------------------------- Following [@lecorff:fort:2011], we address the convergence of the P-BOEM algorithm as the convergence of a perturbed version of the [*limiting EM*]{} recursion. The following result, which is proved in [@lecorff:fort:2011 Theorem 4.1.], shows that when $\tau$ is large, the BOEM statistic $\bar S^{T}_\tau({\theta}, {\mathbf{Y}})$ is an approximation of a deterministic quantity ${\bar{\mathrm{S}}}({\theta})$; the [*limiting EM*]{} algorithm is the iterative procedure defined by ${\theta}_{n+1} = {\mathrm{R}}({\theta}_n)$ where $$\label{eq:limem} {\mathrm{R}}({\theta}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\bar{\theta}({\bar{\mathrm{S}}}({\theta})) {\;}, \forall {\theta}\in{\Theta}{\;};$$ the mapping $\bar {\theta}$ is given by A\[assum:exp\]. \[th:LGN\] Let $S : \Xset^{2}\times\Yset\longrightarrow {\mathbb{R}}^{d}$ be a measurable function s.t. A\[assum:moment:sup\]-($1$) holds. Assume A\[assum:strong\] and A\[assum:obs\](\[assum:obs:erg\]). For any ${\theta}\in{\Theta}$, there exists a ${{\mathbb{P}}}$-integrable r.v. ${ \ifthenelse{\equal{{\mathbf{Y}}}{}}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\right]}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\middle | {\mathbf{Y}}\right]} }$ s.t. for any $T>0$, $$\label{th:LGN:ergodic} \bar S_{\tau}^{T}({\theta}, {\mathbf{Y}}) \underset{\tau\rightarrow +\infty}{\longrightarrow} {\bar{\mathrm{S}}}({\theta}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}{ \ifthenelse{\equal{}{}}{\mathbb{E}^{}_{}\left[{ \ifthenelse{\equal{{\mathbf{Y}}}{}}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\right]}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\middle | {\mathbf{Y}}\right]} }\right]}{\mathbb{E}^{}_{}\left[{ \ifthenelse{\equal{{\mathbf{Y}}}{}}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\right]}{\mathbb{E}^{}_{{\theta}}\left[S(X_{-1},X_{0},{\mathbf{Y}}_{0})\middle | {\mathbf{Y}}\right]} }\middle | \right]} }{\;},\quad {{{\mathbb{P}}}-\mathrm{a.s.}}$$ Moreover, ${\theta}\mapsto {\bar{\mathrm{S}}}({\theta})$ is continuous on ${\Theta}$. The asymptotic behavior of the limiting EM algorithm is addressed in [@lecorff:fort:2011 Section $4.2$]: the main ingredient is that the map ${\mathrm{R}}$ admits a positive and continuous Lyapunov function ${\mathrm{W}}$ w.r.t. the set $$\label{eq:staset} {\mathcal{L}}{\ensuremath{\stackrel{\mathrm{def}}{=}}}\{{\theta}\in{\Theta}; {\mathrm{R}}({\theta})={\theta}\}{\;},$$ i.e. *(i)* ${\mathrm{W}}\circ{\mathrm{R}}({\theta}) \geq {\mathrm{W}}({\theta})$ for any ${\theta}\in {\Theta}$ and, *(ii)* for any compact subset $\mathcal{K}$ of ${\Theta}\setminus {\mathcal{L}}$, $\inf_{{\theta}\in \mathcal{K}} {\mathrm{W}}\circ {\mathrm{R}}({\theta}) - {\mathrm{W}}({\theta})>0$. This Lyapunov function is equal to $\exp(\ell({\theta}))$, where the contrast function $\ell({\theta})$ is the (deterministic) limit of the normalized log-likelihood of the observations when $T\to +\infty$ (see [@lecorff:fort:2011-supp Theorem $4.9$]). Theorem \[th:Pbonem:conv\] establishes the convergence of the P-BOEM algorithm to the set ${\mathcal{L}}$ defined by . The proof of Theorem \[th:Pbonem:conv\] is an application of [@lecorff:fort:2011 Theorem $4.4$]. An additional assumption on the number of particles per block is required to check [@lecorff:fort:2011 A$6$] (note indeed that A\[assum:particle\] below and Proposition \[prop:check:lp:control\] imply the condition in [@lecorff:fort:2011] about the ${\mathrm{L}}_{p}$-control of the error). \[assum:particle\] There exist $c>0$ and $d\ge (a+1)/2a$ (where $a$ is given by A\[assum:block-size\]) such that, for all $n\ge 1$, $N_{n} = \lfloor c \tau_{n}^{d}\rfloor$. \[th:Pbonem:conv\] Assume A\[assum:exp\]-\[assum:strong\], A\[assum:moment:sup\]-($\bar p$), A\[assum:obs\]-\[assum:dub:lec\] and A\[assum:particle\] for some $\bar p>2$. Define $\Delta p {\ensuremath{\stackrel{\mathrm{def}}{=}}}2\bar p p/(\bar p- p)$ and assume A\[assum:moment:w\]-($\Delta p$) holds for some $p\in(2,\bar p)$. Assume in addition that ${\mathrm{W}}({\mathcal{L}})$ has an empty interior. Then, there exists $w_{\star}$ s.t. $\{{\mathrm{W}}({\theta}_{n})\}_{n\geq 0}$ converges almost surely to $w_{\star}$ and $\{{\theta}_{n}\}_{n\geq 0}$ converges to $\{{\theta}\in{\mathcal{L}};{\mathrm{W}}({\theta})=w_{\star}\}$. The assumption on ${\mathrm{W}}({\mathcal{L}})$ made in Theorem \[th:Pbonem:conv\] is in common use to prove the convergence of EM based procedures or stochastic approximation algorithms. It is used in [@wu:1983] to find the limit points of the classical EM algorithm. See also [@delyon:lavielle:moulines:1999] and [@fort:moulines:2003] for the stability of the Monte Carlo EM algorithm and of a stochastic approximation of the EM algorithm. If ${\mathrm{W}}$ is sufficiently regular, Sard’s theorem states that ${\mathrm{W}}({\mathcal{L}})$ has Lebesgue measure $0$ and hence has an empty interior. Under the assumptions of Theorem \[th:Pbonem:conv\], it can be proved that, along any converging P-BOEM sequence $\{{\theta}_{n}\}_{n\geq 0}$ to ${\theta}_{\star}$ in ${\mathcal{L}}$, the averaged P-BOEM statistics $\{ \widetilde{\Sigma}_{n}\}_n$ defined by (\[eq:PBonem:averaged\]) (see also (\[eq:tildeSigma:bis\])) converge to ${\bar{\mathrm{S}}}({\theta}_{\star})$, see Proposition \[prop:conv:averaged\]. Since $\bar{\theta}$ is continuous, the averaged P-BOEM sequence $\{\widetilde{{\theta}}_{n}\}_{n\geq 0}$ converges to $\bar{\theta}({\bar{\mathrm{S}}}({\theta}_{\star}))={\mathrm{R}}({\theta}_{\star})$. Since ${\theta}_\star \in {\mathcal{L}}$, ${\mathrm{R}}({\theta}_{\star})={\theta}_{\star}$, showing that the averaged P-BOEM algorithm has the same limit points as the P-BOEM algorithm. Rate of convergence of the Particle Block Online EM algorithms {#sec:averaging} -------------------------------------------------------------- In this section, we consider a converging P-BOEM sequence $\{{\theta}_n \}_{n \geq 0}$ with limiting point ${\theta}_\star \in {\mathcal{L}}$. It can be shown, as in [@lecorff:fort:2011-supp Proposition $3.1$], that the convergence of the sequence $\{{\theta}_n \}_{n \geq 0}$ is equivalent to the convergence of the sufficient statistics $\{ \widetilde S_{n}({\theta}_{n}, {\mathbf{Y}})\}_{n\geq 0}$: along any P-BOEM sequence converging to ${\theta}_{\star}$, this sequence of sufficient statistics converges to $s_{\star} = {\bar{\mathrm{S}}}({\theta}_{\star})$. Let ${\mathrm{G}}: {\mathcal{S}}\to {\mathcal{S}}$ be the limiting EM map defined on the space of sufficient statistics by $$\label{eq:mapping:limitingEM} {\mathrm{G}}(s) {\ensuremath{\stackrel{\mathrm{def}}{=}}}{\bar{\mathrm{S}}}(\bar{\theta}(s)) {\;}, \qquad \forall s\in{\mathcal{S}}{\;}.$$ To that goal consider the following assumption. \[assum:stable:fixpoint\] (a) ${\bar{\mathrm{S}}}$ and $\bar{\theta}$ are twice continuously differentiable on ${\Theta}$ and ${\mathcal{S}}$. (b) $\mathrm{sp}(\nabla_{s}{\mathrm{G}}(s_{\star}))\in (0,1)$ where $\mathrm{sp}$ denotes the spectral radius. We will use the following notation: for any sequence of random variables $\{Z_{n}\}_{n\geq 0}$, write $Z_{n} = O_{{\mathrm{L}}_{p}}(1)$ if $\limsup_{n}{\ensuremath{\left\| Z_{n} \right\|_{p}}}<\infty$; and $Z_{n} = O_{\mathrm{a.s}}(1)$ if $\sup_{n}|Z_{n}|<+\infty$ ${{{\mathbb{P}}}-\mathrm{a.s.}}$ \[th:rate:sto\] Assume A\[assum:exp\]-\[assum:strong\], A\[assum:moment:sup\]-($\bar p$), A\[assum:obs\]-\[assum:dub:lec\] and A\[assum:particle\]-\[assum:stable:fixpoint\] for some $\bar p>2$. Define $\Delta p {\ensuremath{\stackrel{\mathrm{def}}{=}}}2\bar p p/(\bar p- p)$ and assume A\[assum:moment:w\]-($\Delta p$) holds for some $p\in(2,\bar p)$. Then, $$\label{eq:rate:nonaverage:sto} \left[{\theta}_{n} - {\theta}_{\star}\right]{\ensuremath{\mathbf{1}}}_{\lim_{k}{\theta}_{k} ={\theta}_{\star}} = O_{{\mathrm{L}}_{p/2}}\left(\frac{1}{\tau_{n}^{1/2}}\right) O_{\mathrm{a.s}}\left(1\right){\;}.$$ On the other hand, for the averaged sequence, $$\label{eq:rate:average:sto} \left[\widetilde{\theta}_{n} - {\theta}_{\star}\right]{\ensuremath{\mathbf{1}}}_{\lim_{k}{\theta}_{k}={\theta}_{\star}} = O_{{\mathrm{L}}_{p/2}}\left(\frac{1}{T_n^{1/2}}\right)O_{\mathrm{a.s}}\left(1\right) {\;}.$$ The proof of Theorem \[th:rate:sto\] is obtained by checking the assumptions of [@lecorff:fort:2011 Theorem $5.1$ and Theorem $5.2$]. Eq. (\[eq:rate:nonaverage:sto\]) shows that the error ${\theta}_{n} - {\theta}_{\star}$ has a ${\mathrm{L}}_{p/2}$-norm decreasing as $\tau_{n}^{-1/2}$. This result is obtained by assuming $N_n \sim \tau_n^{d}$, with $d\ge (a+1)/2a$, which implies that the SMC error and the BOEM error are balanced. Unfortunately, such a rate is obtained after a total number of observations $T_n$; therefore, as discussed in [@lecorff:fort:2011], it is quite sub-optimal. Eq (\[eq:rate:average:sto\]) shows that the rate of convergence equal to the square root of the total number of observations up to block $n$, can be reached by using the averaged P-BOEM algorithm: the ${\mathrm{L}}_{p/2}$-norm of the error $\widetilde {\theta}_{n}-{\theta}_{\star}$ has a rate of convergence proportional to $T_{n}^{-1/2}$. Here again, note that since $N_{n}$ is chosen as in A\[assum:particle\] the SMC error and the BOEM error are balanced. Proofs {#sec:proofs:sto} ====== For a function $h$, define ${\mathrm{osc}}(h) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\sup_{z,z'} | h(z)-h(z')|$. Proof of Proposition \[prop:check:lp:control\] ---------------------------------------------- For any $t\in\{0,\dots, \tau_{n+1}\}$, define the $\sigma$-algebra $\mathcal{F}_{n,t}^{N_{n+1}}$ by $$\label{eq:Filtrations:Fnt} \mathcal{F}_{n,t}^{N_{n+1}}{\ensuremath{\stackrel{\mathrm{def}}{=}}}\sigma\left\{{\theta}_n,{\mathbf{Y}}_{T_n+1:T_n+t+1},\left({\xi_{s}^{\ell}},{\omega_{s}^{\ell}}\right); \ell\in\{1,\dots,N_{n+1}\}; 0\leq s\leq t\right\}{\;}.$$ We use $S_{s}(x,x')$ as a shorthand notation for $S(x,x',{\mathbf{Y}}_s)$. Under A\[assum:strong\] and A\[assum:dub:lec\], Propositions B.$5$.,  B.$8$. and B.$9$. in Appendix B of [@lecorff:fort:2011b] can be applied so that $$\label{eq:diff:sto} \mathbb{E}_{}\left[\left\vert \widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) - \bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})\right\vert^{p}\right]\leq C\left(I_{1,n}+I_{2,n}\right){\;},$$ where $$\begin{aligned} I_{1,n}&{\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{\tau_{n+1}^{\frac{p}{2}+1}N_{n+1}^{\frac{p}{2}}} \times\sum_{t=0}^{\tau_{n+1}}\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S_{s+T_n}\}\right|^{p}\right]{\;},\\ I_{2,n}&{\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{\tau_{n+1}N_{n+1}^{p}}\,\\ &\hspace{0.9cm}\times\sum_{t=0}^{\tau_{n+1}}\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\right|^{2p}\mathbb{E}_{}\left[\left|\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S_{s+T_n}\}\right|^{\bar p}\middle|\mathcal{F}_{n,t-1}^{N_{n+1}}\right]^{p/\bar p}\right]{\;}.\end{aligned}$$ By the Hölder inequality applied with $\alpha {\ensuremath{\stackrel{\mathrm{def}}{=}}}\bar p /p\geq 1$ and $\beta^{-1} {\ensuremath{\stackrel{\mathrm{def}}{=}}}1 - \alpha^{-1}$, $$I_{1,n}\leq \frac{1}{\tau_{n+1}^{\frac{p}{2}+1}N_{n+1}^{\frac{p}{2}}}\sum_{t=1}^{\tau_{n+1}}{\left\| \sum_{s=1}^{\tau_{n+1}}\rho^{\vert t-s\vert}{\mathrm{osc}}\{S_{s+T_{n}}\} \right\|_{\bar p}}^{p}\times{\left\| \frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_{n}})}{b_{-}({\mathbf{Y}}_{t+T_{n}})} \right\|_{2\bar p p /(\bar p -p)}}^{2p} {\;}.$$ By A\[assum:strong\], A\[assum:moment:sup\]-($\bar p$), A\[assum:obs\] and A\[assum:moment:w\]-($\Delta p$), we have $$I_{1,n}\leq\frac{C}{\tau_{n+1}^{\frac{p}{2}}N_{n+1}^{\frac{p}{2}}}{\;}.$$ Using similar arguments for $I_{2,n}$ yields $I_{2,n}\leq C \, N_{n+1}^{-p}$, which concludes the proof. ${\mathrm{L}}_{p}$-controls ----------------------------- \[prop:lpcontrol:sto\] Let $S : \Xset^{2}\times\Yset\longrightarrow {\mathbb{R}}^{d}$ be a measurable function s.t. A\[assum:moment:sup\]-($\bar p$) holds for some $\bar p >2$. Assume A\[assum:strong\], A\[assum:obs\], A\[assum:dub:lec\] and A\[assum:moment:w\]-($\Delta p$) for some $p\in(2,\bar p)$, where $\Delta p {\ensuremath{\stackrel{\mathrm{def}}{=}}}2\bar p p/(\bar p- p)$. There exists a constant $C$ s.t. for any $n\geq 1$, $${\ensuremath{\left\| \widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) - {\bar{\mathrm{S}}}({\theta}_n) \right\|_{p}}} {}\leq C\left(\frac{1}{\sqrt{\tau_{n+1}}}+\frac{1}{N_{n+1}}\right) {\;}.$$ Under A\[assum:strong\], A\[assum:moment:sup\]-($\bar p$) and A\[assum:obs\], by [@lecorff:fort:2011 Theorem $4.1$], there exists a constant $C$ s.t. $${\ensuremath{\left\| \bar S_{\tau_{n+1}}^{T_{n}}({\theta}_n, {\mathbf{Y}})-{\bar{\mathrm{S}}}({\theta}_n) \right\|_{p}}} \leq \frac{C}{\sqrt{\tau_{n+1}}} {\;}.$$ Moreover, under A\[assum:dub:lec\] and A\[assum:moment:w\]-($\Delta p$), by Proposition \[prop:check:lp:control\], we have $${\left\| \widetilde S_{n}({\theta}_n, {\mathbf{Y}}) - \bar S_{\tau_{n+1}}^{T_n}({\theta}_n, {\mathbf{Y}}) \right\|_{p}}\leq C\left(\frac{1}{N_{n+1}}+\frac{1}{\tau^{1/2}_{n+1} N_{n+1}^{1/2}}\right){\;},$$ which concludes the proof. \[prop:conv:averaged\] Let $S : \Xset^{2}\times\Yset\longrightarrow {\mathbb{R}}^{d}$ be a measurable function s.t. A\[assum:moment:sup\]-($\bar p$) holds for some $ \bar p >2$. Assume A\[assum:strong\], A\[assum:obs\]-\[assum:block-size\], A\[assum:dub:lec\]-\[assum:particle\] and A\[assum:moment:w\]-($\Delta p$) for some $p\in(2,\bar p)$, where $\Delta p {\ensuremath{\stackrel{\mathrm{def}}{=}}}2\bar p p/(\bar p- p)$. Let $\{{\theta}_n\}_n$ be the P-BOEM sequence. For any ${\theta}_{\star}\in{\Theta}$, on the set $\{\lim_{n}{\theta}_{n}={\theta}_{\star}\}$, $$\widetilde \Sigma_{n} \longrightarrow {\bar{\mathrm{S}}}({\theta}_{\star}){\;},\quad {{{\mathbb{P}}}-\mathrm{a.s.}}{\;},$$ where ${\bar{\mathrm{S}}}$ is defined in and $\widetilde \Sigma_{n}$ in . By , $\widetilde\Sigma_{n}$ can be written as $$\label{eq:Sigmatilde:decomp} \widetilde\Sigma_{n} = \frac{1}{T_n} \sum_{j=1}^n \tau_j \, \left[\widetilde S_{j-1}({\theta}_{j-1}, {\mathbf{Y}})-{\bar{\mathrm{S}}}({\theta}_{j-1})\right] + \frac{1}{T_n} \sum_{j=1}^n \tau_j \, {\bar{\mathrm{S}}}({\theta}_{j-1}){\;}.$$ By Theorem \[th:LGN\], ${\bar{\mathrm{S}}}$ is continuous so, by the Cesaro Lemma, the second term in the right-hand side of converges to ${\bar{\mathrm{S}}}({\theta}_{\star})$ ${{\mathbb{P}}}$-a.s., on the set $\{\lim_{n}{\theta}_{n}={\theta}_{\star}\}$. By Proposition \[prop:lpcontrol:sto\], there exists a constant $C$ such that for any $n$, $${\ensuremath{\left\| \widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) - {\bar{\mathrm{S}}}({\theta}_n) \right\|_{p}}} \leq C\left(\frac{1}{\sqrt{\tau_{n+1}}}+\frac{1}{N_{n+1}}\right) {\;}.$$ Hence, by A\[assum:block-size\], A\[assum:particle\] and the Borel-Cantelli Lemma, $$\left|\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}}) - {\bar{\mathrm{S}}}({\theta}_n)\right|\longrightarrow 0{\;},\quad {{{\mathbb{P}}}-\mathrm{a.s.}}$$ The proof is concluded by applying the Cesaro Lemma. Detailed SMC algorithm {#app:alg} ====================== In this section, we give a detailed description of the SMC algorithm used to compute sequentially the quantities $\widetilde S_{n}({\theta}_{n}, {\mathbf{Y}})$, $n\geq 0$. This is the algorithm proposed by  [@cappe:2011] and [@delmoral:doucet:singh:2010a]. At each time step, the weighted samples are produced using sequential importance sampling and sampling importance resampling steps. In Algorithm \[alg:FSMC\], the instrumental proposition kernel used to select and propagate the particles is ${\ensuremath{\pi_{t}}}$ (see and [@douc:garivier:moulines:olsson:2010; @doucet:defreitas:gordon:2001; @liu:2001] for further details on this SMC step). It is readily seen from the description below that the observations ${\mathbf{Y}}_t$ are processed sequentially. ${\theta}_n$, $\tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}{\;}.$ $\widetilde S_{n}({\theta}_n, {\mathbf{Y}}){\;}.$ Sample $\{\xi_{0}^{\ell}\}_{\ell = 1}^{N}$ i.i.d. with distribution $\chi{\;}.$ Set $\omega_{0}^{\ell} = 1/N$ for all $\ell\in\{1,\dots,N\}{\;}.$ Set $R_{0,{\theta}_n}^{\ell} = 0$ for all $\ell\in\{1,\dots,N\}{\;}.$ Conditionally to $({\theta}_n, Y_{T_n+1:T_n+t}, \{J_{t-1}^{\ell}, {\xi_{t-1}^{\ell}} \}_{\ell=1}^{N})$, sample independently $(J_t^{\ell}, {\xi_{t}^{\ell}})\sim {\ensuremath{\pi_{t}}}(i, {\mathrm{d}}x){\;},$ where ${\ensuremath{\pi_{t}}}(i, {\mathrm{d}}x) \propto {\omega_{t-1}^{i}} {\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{i}})}}{\mathrm{erreur}}}}} {\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}({\xi_{t-1}^{i}},x)\lambda({\mathrm{d}}x) {\;}.$ Set $${\omega_{t}^{\ell}} = \frac{{\ensuremath{m}}_{{\theta}_n}({\xi_{t-1}^{J_t^{\ell}}},{\xi_{t}^{\ell}}) g_{{\theta}_n}({\xi_{t}^{\ell}},{\mathbf{Y}}_{T_n+t})}{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{J_t^{\ell}}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{J_t^{\ell}}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{J_t^{\ell}}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{J_t^{\ell}}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{J_t^{\ell}}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{J_t^{\ell}}})}}{\mathrm{erreur}}}}} {\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}({\xi_{t-1}^{J_t^{\ell}}},{\xi_{t}^{\ell}})} {\;}.$$ Set $$R_{t,{\theta}_n}^{\ell} = \frac{1}{t}\sum_{j=1}^{N}{\omega_{t-1}^{j}}m_{{\theta}_n}({\xi_{t-1}^{j}},{\xi_{t}^{\ell}})\frac{ S({\xi_{t-1}^{j}},{\xi_{t}^{\ell}},{\mathbf{Y}}_{T_n+t})+ (t-1) R_{t-1,{\theta}_n}^{j}}{\sum_{k=1}^{N}{\omega_{t-1}^{k}}{\ensuremath{m}}_{{\theta}_n}({\xi_{t-1}^{k}},{\xi_{t}^{\ell}})} {\;}.$$ Set $$\widetilde S_{n}({\theta}_n, {\mathbf{Y}}) = \sum_{\ell=1}^{N}{\omega_{\tau_{n+1}}^{\ell}} R_{\tau_{n+1},{\theta}_n}^{\ell}{\;}.$$ ${\mathrm{L}}_{p}$-controls of SMC approximations {#app:extended:version} ================================================= In this section, we give further details on the ${\mathrm{L}}_{p}$ control on each block (see  (\[eq:diff:sto\])): $$\mathbb{E}_{}\left[\left\vert \widetilde S_{\tau_{n+1}}^{N,T_{n}}({\theta}_{n}, {\mathbf{Y}}) - \bar S_{\tau_{n+1}}^{T_{n}}({\theta}_{n}, {\mathbf{Y}})\right\vert^{p}\right]{\;},$$ $\bar S_{\tau}^{T}$ is defined by (we recall that, $\chi$ being fixed, it is dropped from the notations) and $\widetilde S_\tau^{N,T}$ is the SMC approximation of $\bar S_{\tau}^{T}$ based on $N$ particles computed as described in Section \[subsec:BOEM:description:SMC\]. The following results are technical lemmas taken from [@douc:garivier:moulines:olsson:2010] (stated here for a better clarity) or extensions of the ${\mathrm{L}}_{p}$ controls derived in [@dubarry:lecorff:2011]. Hereafter, “time $t$” corresponds to time $t$ in the block $n$. Therefore, even if it is not explicit in the notations (in order to make them simpler), the following quantities depend upon the observations ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. Denote by ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{s}}} } } } } }^{{\theta}}$ the filtering distribution at time $s$, and let $${\mathrm{B}_{{ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{t}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{t}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{t}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{t}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{t}}} } } } } }^{{\theta}}}}^{{\theta}}(x, {\mathrm{d}}x') {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{m_{\theta}(x',x)}{\int m_{\theta}(u,x) { \ifthenelse{\equal{}{}}{\ensuremath{\phi_{t}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{t}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{t}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{t}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{t}}} } } } } }^{\theta}({\mathrm{d}}u)}{ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{t}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{t}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{t}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{t}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{t}}} } } } } }^{\theta}({\mathrm{d}}x')$$ be the backward kernel smoothing kernel at time $t+1$. For all $0 \leq s \leq \tau-1$ and for all bounded measurable function $h$ on $\Xset^{\tau-s+1}$, define recursively ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s:\tau|\tau}}}} } } }^{\theta}[h]$ backward in time, according to $$\label{eq:smoothing:backw_decomposition_recursion} { \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s:\tau|\tau}}}} } } }^{{\theta}}[h] = \idotsint {\mathrm{B}_{{ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{s}}} } } } } }^{{\theta}}}}^{{\theta}}(x_{s+1}, {\mathrm{d}}x_s) \, { \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s+1:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s+1:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s+1:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s+1:\tau|\tau}}}} } } }^{{\theta}}({\mathrm{d}}{\ifthenelse{\equal{}{}}{\ensuremath{{x}_{s+1:\tau}}}{\ensuremath{x^}_{s+1:\tau}} }) \, h({\ifthenelse{\equal{}{}}{\ensuremath{{x}_{s:\tau}}}{\ensuremath{x^}_{s:\tau}} }) {\;},$$ starting from ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{\tau:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{\tau:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{\tau:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{\tau:\tau|\tau}}}} } } }^{{\theta}} ={ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{\tau}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{\tau}}} } } } } }^{{\theta}}$. By convention, ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{0}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{0}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{0}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{0}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{0}}} } } } } }^{\theta}= \chi$. For $t \geq 1$, let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}})\right\}_{\ell = 1}^{N}$ be the weighted samples obtained as described in Section \[subsec:BOEM:description:SMC\] (see also Algorithm \[alg:FSMC\] in Appendix \[app:alg\]); it approximates the filtering distribution ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{t}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{t}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{t}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{t}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{t}}} } } } } }^{\theta}$. Denote by ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{t}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{t}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{t}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{t}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{t}}} } } } } }^{N,{\theta}}$ this approximation. For $0 \leq s \leq \tau-1$, an approximation of the backward kernel can be obtained $${\mathrm{B}_{{ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{s}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{s}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{s}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{s}}} {\ifthenelse{\equal{}{aux}}{\ensuremath{\phi^{N,\mathrm{a}}_{s}}} } } } } }^{N,{\theta}}}}(x, h) = \sum_{i = 1}^{N} \frac{{\omega_{s}^{i}} {\ensuremath{m}}_{{\theta}}({\xi_{s}^{i}}, x)}{\sum_{\ell = 1}^{N} {\omega_{s}^{\ell}} {\ensuremath{m}}_{{\theta}}({\xi_{s}^{\ell}},x)} h\left( {\xi_{s}^{i}} \right) {\;};$$ and inserting this expression into gives the following particle approximation of the fixed-interval smoothing distribution ${ \ifthenelse{\equal{}{}}{\ensuremath{\phi_{0:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{0:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{0:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{0:\tau|\tau}}}} } } }^{{\theta}}[h]$ $$\label{eq:forward-filtering-backward-smoothing} { \ifthenelse{\equal{}{}}{\ensuremath{\phi_{0:\tau|\tau}}}{\ifthenelse{\equal{}{hat}}{\ensuremath{\phi^{N}_{0:\tau|\tau}}} {\ifthenelse{\equal{}{tilde}}{\ensuremath{\tilde{\phi}^{N}_{0:\tau|\tau}}} {\ifthenelse{\equal{}{tar}}{\ensuremath{\phi^{N,\mathrm{t}}_{0:\tau|\tau}}}} } } }^{N,{\theta}}[h] = \sum_{i_0 = 1}^{N} \dots \sum_{i_{\tau} = 1}^{N} \left(\prod_{u=1}^{\tau} \frac{{\omega_{u-1}^{i_{u-1}}} {\ensuremath{m}}_{{\theta}}({\xi_{u-1}^{i_{u-1}}},{\xi_{u}^{i_u}})}{\sum_{\ell=1}^{N} {\omega_{u-1}^{\ell}} {\ensuremath{m}}_{{\theta}}({\xi_{u-1}^{\ell}},{\xi_{u}^{i_u}})}\right) \times \frac{{\omega_{\tau}^{i_{\tau}}}}{{\ifthenelse{\equal{n}{}}{\ensuremath{\Omega_{\tau}}}{\ensuremath{\Omega_{\tau}^{n}}}}} h\left({\xi_{0}^{i_0}}, \dots, {\xi_{\tau}^{i_{\tau}}}\right) {\;},$$ with ${\ifthenelse{\equal{N}{}}{\ensuremath{\Omega_{\tau}}}{\ensuremath{\Omega_{\tau}^{N}}}}{\ensuremath{\stackrel{\mathrm{def}}{=}}}\sum_{\ell=1}^{N}{\omega_{\tau}^{\ell}}$. \[lem:DeltaStoDeltaPhi\] Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. Then, $$\begin{gathered} \label{eq:rewrite:smerr} \left[\widetilde S_{\tau_{n+1}}^{N,T_n}({\theta}_n, {\mathbf{Y}})-\bar S_{\tau_{n+1}}^{T_n}({\theta}_n, {\mathbf{Y}})\right] \\= \frac{1}{\tau_{n+1}} \left(\phi^{N,{\theta}_n}_{0:\tau_{n+1}|\tau_{n+1}}\left[{\mathsf{S}_{\tau_{n+1}}}\right]-\phi_{0:\tau_{n+1}|\tau_{n+1}}^{{\theta}_n}\left[{\mathsf{S}_{\tau_{n+1}}}\right] \right) {\;},\end{gathered}$$ where $${\mathsf{S}_{\tau}}(x_{0:\tau}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\sum_{s=1}^{\tau}S(x_{s-1},x_{s},{\mathbf{Y}}_{s+T}){\;}. \label{eq:addfuncbis}$$ For all $t\in\{0,\dots,\tau\}$ and all bounded measurable function $h$ on $\Xset^{\tau+1}$, define the kernel ${\mathrm{L}}_{t,\tau}:\Xset^{t+1}\times \sigmaX^{\otimes \tau+1}\rightarrow [0,1]$ by $$\label{eq:ldroit} {\mathrm{L}}_{t,\tau}^{{\theta}}h(x_{0:t}){\ensuremath{\stackrel{\mathrm{def}}{=}}}\int \prod_{u=t+1}^{\tau}{\ensuremath{m}}_{{\theta}}(x_{u-1},x_u)g_{{\theta}}(x_u,{\mathbf{Y}}_{u+T})h(x_{0:\tau})\lambda({\mathrm{d}}x_{t+1:\tau}){\;};$$ by convention, ${\mathrm{L}}_{\tau,\tau}^{{\theta}}h =h$. Let $\mathcal{L}_{t,\tau}^{N,{\theta}}$ and $\mathcal{L}_{t,\tau}^{{\theta}}$ be two kernels on $\Xset \times \sigmaX^{\otimes (\tau+1)}$ defined for all $x_t\in\Xset$ by $$\begin{aligned} \mathcal{L}_{t,\tau}^{{\theta}}h(x_t) &{\ensuremath{\stackrel{\mathrm{def}}{=}}}\int{\mathrm{B}}_{\phi_{t-1}^{{\theta}}}^{{\theta}}(x_t,{\mathrm{d}}x_{t-1})\cdots{\mathrm{B}}_{\phi_{0}^{{\theta}}}^{{\theta}}(x_1,{\mathrm{d}}x_{0}){\mathrm{L}}_{t,\tau}^{{\theta}}h(x_{0:t}) \label{eq:defL} \\\mathcal{L}_{t,\tau}^{N,{\theta}}h(x_t) &{\ensuremath{\stackrel{\mathrm{def}}{=}}}\int{\mathrm{B}}_{\phi_{t-1}^{N,{\theta}}}^{{\theta}}(x_t,{\mathrm{d}}x_{t-1})\cdots{\mathrm{B}}_{\phi_{0}^{N,{\theta}}}^{{\theta}}(x_1,{\mathrm{d}}x_{0}){\mathrm{L}}_{t,\tau}^{{\theta}}h(x_{0:t})\label{eq:defLN}{\;}.\end{aligned}$$ Note that $$\label{eq:lcal} \mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x_t) =\int m_{\theta}(x_t,x') g_{\theta}(x',{\mathbf{Y}}_{T+t+1}) \ \mathcal{L}_{t+1,\tau}^{{\theta}} {\ensuremath{\mathbf{1}}}(x') \lambda({\mathrm{d}}x') {\;}.$$ Lemma \[lem:rew:err\], Proposition \[prop:DeltaStoC-D\], Lemma \[lem:esperancecond\] and \[Lem:Upperbounds\] can be found in [@douc:garivier:moulines:olsson:2010]. \[lem:rew:err\] Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. Then, $$\label{Eq:Err} \phi^{N,{\theta}_n}_{0:\tau_{n+1}|\tau_{n+1}}\left[h\right]-\phi_{0:\tau_{n+1}|\tau_{n+1}}^{{\theta}_n}\left[h\right] = \sum_{t=0}^{\tau_{n+1}}\frac{\sum_{\ell=1}^{N}{\omega_{t}^{\ell}} \ G_{t,\tau_{n+1}}^{N,{\theta}_n}h({\xi_{t}^{\ell}})}{\sum_{\ell=1}^{N}{\omega_{t}^{\ell}} \ \mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}({\xi_{t}^{\ell}})}{\;},$$ with $G_{t,\tau}^{N,{\theta}}$ is a kernel on $\Xset \times \sigmaX^{\otimes (\tau+1)}$ defined, for all $x\in\Xset$ and all bounded and measurable function $h$ on $\Xset^{\tau+1}$, by $$\label{eq:defG} G_{t,\tau}^{N,{\theta}}h(x){\ensuremath{\stackrel{\mathrm{def}}{=}}}\mathcal{L}_{t,\tau}^{N,{\theta}}h(x) - \frac{\phi_{t-1}^{N,{\theta}}[\mathcal{L}_{t-1,\tau}^{N,{\theta}}h]}{\phi_{t-1}^{N,{\theta}}[\mathcal{L}_{t-1,\tau}^{N,{\theta}}{\ensuremath{\mathbf{1}}}]}\mathcal{L}_{t,\tau}^{N,{\theta}}{\ensuremath{\mathbf{1}}}(x){\;}.$$ By definition of ${\mathrm{L}}_{t,\tau}^{{\theta}}$, $$\phi_{0:\tau|\tau}^{{\theta}}[h] = \frac{\phi_{0:t|t}^{{\theta}}\left[{\mathrm{L}}_{t,\tau}^{{\theta}}h\right]}{\phi_{0:t|t}^{{\theta}}\left[{\mathrm{L}}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right]}{\;}.$$ We write $$\phi^{N,{\theta}}_{0:\tau|\tau}\left[h\right]-\phi_{0:\tau|\tau}^{{\theta}}\left[h\right] = \sum_{t=0}^{\tau} \left\{ \frac{\phi_{0:t|t}^{N,{\theta}}\left[{\mathrm{L}}_{t,\tau}^{{\theta}}h\right]}{\phi_{0:t|t}^{N,{\theta}}\left[{\mathrm{L}}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right]}-\frac{\phi_{0:t-1|t-1}^{N,{\theta}}\left[{\mathrm{L}}_{t-1,\tau}^{{\theta}}h\right]}{\phi_{0:t-1|t-1}^{N,{\theta}}\left[{\mathrm{L}}_{t-1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right]} \right \} \nonumber{\;},$$ where we used the convention $$\frac{\phi_{0:-1|-1}^{N,{\theta}}\left[{\mathrm{L}}_{-1,\tau}^{{\theta}}h\right]}{\phi_{0:-1|-1}^{N,{\theta}}\left[{\mathrm{L}}_{-1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right]} = \frac{\chi\left[{\mathrm{L}}_{0,\tau}^{\tau}h\right]}{\chi\left[{\mathrm{L}}_{0,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right]} = \phi_{0:\tau|\tau}^{{\theta}}[h]{\;}.$$ We have for all $0\leq t\leq \tau$, $$\phi_{0:t|t}^{N,{\theta}}\left[{\mathrm{L}}_{t,\tau}h\right] =\int\phi_{t}^{N,{\theta}}({\mathrm{d}}x_{t}) \, \prod_{j=0}^{t-1} {\mathrm{B}}_{\phi_{j}^{N,{\theta}}}(x_{j+1},{\mathrm{d}}x_{j}) \ {\mathrm{L}}_{t,\tau}^{{\theta}}h(x_{0:t}) = \phi_{t}^{N,{\theta}}[\mathcal{L}_{t,\tau}^{N,{\theta}}h]{\;}.$$ Therefore, for all $1\leq t \leq \tau$, $$\frac{\phi_{0:t|t}^{N,{\theta}}[{\mathrm{L}}_{t,\tau}^{{\theta}}h]}{\phi_{0:t|t}^{N,{\theta}}[{\mathrm{L}}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}]}-\frac{\phi_{0:t-1|t-1}^{N,{\theta}}[{\mathrm{L}}_{t-1,\tau}^{{\theta}}h]}{\phi_{0:t-1|t-1}^{N,{\theta}}[{\mathrm{L}}_{t-1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}]} =\frac{\phi_{t}^{N,{\theta}}[\mathcal{L}_{t,\tau}^{N,{\theta}}h]}{\phi_{t}^{N,{\theta}}[\mathcal{L}_{t,\tau}^{N,{\theta}}{\ensuremath{\mathbf{1}}}]} - \frac{\phi_{t-1}^{N,{\theta}}[\mathcal{L}_{t-1,\tau}^{N,{\theta}}h]}{\phi_{t-1}^{N,{\theta}}[\mathcal{L}_{t-1,\tau}^{N,{\theta}}{\ensuremath{\mathbf{1}}}]} =\frac{\phi_{t}^{N,{\theta}}[G_{t,\tau}^{N,{\theta}}h ]}{\phi_{t}^{N,{\theta}}[\mathcal{L}_{t,\tau}^{N,{\theta}}{\ensuremath{\mathbf{1}}}]}\ {\;}.$$ \[prop:DeltaStoC-D\] Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with input variables ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. Then, $$\begin{gathered} \left[\widetilde S_{\tau_{n+1}}^{N,T_n}({\theta}_n, {\mathbf{Y}})-\bar S_{\tau_{n+1}}^{T_n}({\theta}_n, {\mathbf{Y}})\right] = \frac{1}{\tau_{n+1}}\sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\\ + \frac{1}{\tau_{n+1}}\sum_{t=0}^{\tau_{n+1}}C_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}}){\;},\end{gathered}$$ where ${\mathsf{S}_{\tau}}$ is given by (\[eq:addfuncbis\]) and $$\begin{aligned} \label{eq:defD} & D_{t,\tau_{n+1}}^{N,{\theta}_n}(h) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{\phi_{t-1}^{N,{\theta}_n}[{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}]}{\phi_{t-1}^{N,{\theta}_n}\left[\frac{\mathcal{L}_{t-1,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right]}N^{-1}\sum_{\ell=1}^{N}{\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}h({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}{\;}; \\ & C_{t,\tau_{n+1}}^{N}(h) {\ensuremath{\stackrel{\mathrm{def}}{=}}}\left[\frac{1}{N^{-1}\sum_{i=1}^{N}{\omega_{t}^{i}} \frac{\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}({\xi_{t}^{i}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}}-\frac{\phi_{t-1}^{N,{\theta}_n}[{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}]}{\phi_{t-1}^{N,{\theta}_n}\left[\frac{\mathcal{L}_{t-1,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right]}\right] \label{eq:defC}\\ &\hspace{7cm}\times N^{-1}\sum_{\ell=1}^{N}{\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}h({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}} {\;}.\nonumber\end{aligned}$$ can be rewritten as follows: $$\label{eq:rewrite:smerrCD} \phi^{N,{\theta}_n}_{0:\tau_{n+1}|\tau_{n+1}}\left[h\right]-\phi_{0:\tau_{n+1}|\tau_{n+1}}^{{\theta}_n}\left[h\right] = \sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}(h) + \sum_{t=0}^{\tau_{n+1}}C_{t,\tau_{n+1}}^{N,{\theta}_n}(h){\;}.$$ The proof is concluded by Lemma \[lem:DeltaStoDeltaPhi\]. For any $t\in\{0,\dots, \tau_{n+1}\}$, we recall the definition of $\mathcal{F}_{n,t}^{N}$ given by $$\mathcal{F}_{n,t}^{N}= \sigma\left\{{\theta}_n,{\mathbf{Y}}_{T_n+1:T_n+t+1},\left({\xi_{s}^{\ell}},{\omega_{s}^{\ell}}\right); \ell\in\{1,\dots,N\}; 0\leq s\leq t\right\}{\;},$$ where $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}})\right\}_{\ell = 1}^{N}$ are the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with input variables ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. \[lem:esperancecond\] Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. Then, for any $1 \leq t \leq \tau_{n+1}$ and any $1 \leq \ell \leq N$, $$\label{eq:whExpect} {\mathbb{E}_{}\left[{\omega_{t}^{\ell}}h({\xi_{t}^{\ell}}) \middle| \mathcal{F}_{n,t-1}^{N}\right]} =\frac{\phi_{t-1}^{N,{\theta}_n}\left[ \int {\ensuremath{m}}_{{\theta}_n}(\cdot,x) g_{{\theta}_n}(x, {\mathbf{Y}}_{T_n+t}) \ h(x) \ \lambda({\mathrm{d}}x)\right]}{\phi_{t-1}^{N,{\theta}_n}[{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}]}{\;}.$$ By definition of the weighted particles, $$\begin{aligned} &{\mathbb{E}_{}\left[{\omega_{t}^{\ell}}h({\xi_{t}^{\ell}}) \middle| \mathcal{F}_{n,t-1}^{N}\right]} \nonumber\\ &= {\mathbb{E}_{}\left[\frac{{\ensuremath{m}}_{{\theta}_n}({\xi_{t-1}^{I_{t}^{1}}},{\xi_{t}^{1}}) g_{{\theta}_n}({\xi_{t}^{1}},{\mathbf{Y}}_{t+T_n})}{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{I_{t}^{1}}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{I_{t}^{1}}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{I_{t}^{1}}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{I_{t}^{1}}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{I_{t}^{1}}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{I_{t}^{1}}})}}{\mathrm{erreur}}}}} {\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}({\xi_{t-1}^{I_{t}^{1}}},{\xi_{t}^{1}})}h({\xi_{t}^{1}}) \middle| \mathcal{F}_{n,t-1}^{N}\right]} \nonumber\\ &= \left(\sum_{i=1}^{N}{\omega_{t-1}^{i}} {\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}{({\xi_{t-1}^{i}})}\right)^{-1}\sum_{i=1}^{N}\int{\omega_{t-1}^{i}} {\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{i}})}}{\mathrm{erreur}}}}}{\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}({\xi_{t-1}^{i}},x)\\ &\hspace{5.5cm}\times\frac{{\ensuremath{m}}_{{\theta}_n}({\xi_{t-1}^{i}},x) g_{{\theta}_n}(x,{\mathbf{Y}}_{t+T_n})}{{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon_{t}}{\upsilon_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}({\xi_{t-1}^{i}})}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{{\xi_{t-1}^{i}}}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}({\xi_{t-1}^{i}})}}{\mathrm{erreur}}}}} {\ifthenelse{\equal{}{}}{q_{t}} {\ifthenelse{\equal{}{fully}}{q^{\star}_{t}} {\ifthenelse{\equal{}{smooth}}{\tilde{r}_{t}}{\mathrm{erreur}}}}}{}({\xi_{t-1}^{i}},x)}h(x)\lambda({\mathrm{d}}x) \nonumber\\ &= \left(\sum_{i=1}^{N}{\omega_{t-1}^{i}} {\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}{({\xi_{t-1}^{i}})}\right)^{-1}\sum_{i=1}^{N}\int{\omega_{t-1}^{i}}{\ensuremath{m}}_{{\theta}_n}({\xi_{t-1}^{i}},x) g_{{\theta}_n}(x,{\mathbf{Y}}_{t+T_n})h(x)\lambda({\mathrm{d}}x) {\;}.\end{aligned}$$ \[Lem:Upperbounds\] Assume A\[assum:strong\] and A\[assum:dub:lec\]. Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with input variables ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. (i) \[Lem:Upperbounds:ind\] For any $t\in\{0,\dots, \tau_{n+1}\}$ and any measurable function $h$ on $\Xset^{\tau_{n+1}+1}$, the random variables $\displaystyle \left\{ {\omega_{t}^{\ell}} \ G^{N,{\theta}_n}_{t,\tau_{n+1}}h({\xi_{t}^{\ell}}) \ |\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty^{-1}\right\}_{\ell=1}^{N}$ are: 1. \[Lem:Upperbounds:ind:claim1\] conditionally independent and identically distributed given $\mathcal{F}_{n,t-1}^{N}$, 2. \[Lem:Upperbounds:ind:claim2\]centered conditionally to $\mathcal{F}_{n,t-1}^{N}$. (ii) \[Lem:Upperbounds:boundup\] For any $t\in\{0,\dots, \tau_{n+1}\}$: $$\label{eq:bound-G} \left| \dfrac{G^{N,{\theta}_n}_{t,\tau_{n+1}}\mathsf{S}_{\tau_{n+1}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty}\right|\leq \sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot,{\mathbf{Y}}_{s+T_n})\}{\;},$$ where $\mathsf{S}_{\tau}$ is defined by . (iii) \[Lem:Upperbounds:boundlow\] For all $x\in\Xset$ and any $t\in\{0,\dots, \tau_{n+1}\}$, $$\displaystyle \dfrac{\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}(x)}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty} \geq \dfrac{\sigma_-}{\sigma_+} {\;}, \qquad \displaystyle \dfrac{\mathcal{L}_{t-1,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}(x)}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty} \geq \dfrac{\sigma_-^{2}}{\sigma_+}b_{-}({\mathbf{Y}}_{t+T_n}) {\;}.$$ The proof of is given by [[@douc:garivier:moulines:olsson:2010 Lemma 3]]{}. [*Proof of* ]{}. Let $\Pi_{s-1:s,\tau}$ be the operator which associates to any bounded and measurable function $h$ on $\Xset\times\Xset$ the function $\Pi_{s-1:s,\tau}h$ given, for any $(x_0,\dots,x_{\tau}) \in \Xset^{\tau+1}$, by $$\Pi_{s-1:s,\tau}h(x_{0:\tau}) {\ensuremath{\stackrel{\mathrm{def}}{=}}}h(x_{s-1:s}){\;}.$$ Using this notation, we may write ${\mathsf{S}_{\tau}} = \sum_{s=1}^{\tau} \Pi_{s-1:s,\tau}S(\cdot,\cdot,{\mathbf{Y}}_{s+T})$ and $G^{N,{\theta}}_{t,\tau}{\mathsf{S}_{\tau}} = \sum_{s=1}^{\tau}G^{N,{\theta}}_{t,\tau}\Pi_{s-1:s,\tau}S(\cdot,\cdot,{\mathbf{Y}}_{s+T})$. Following the same lines as in [[@douc:garivier:moulines:olsson:2010 Lemma 10]]{}, $$\begin{aligned} |G^{N,{\theta}}_{t,\tau}\Pi_{s-1:s,\tau}S(\cdot,\cdot,{\mathbf{Y}}_{s+T})|_\infty &\leq \rho^{s-1-t} {\mathrm{osc}}(S(\cdot,\cdot,{\mathbf{Y}}_{s+T})) |\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}|_\infty \quad \mbox{if}\quad t\leq s-1{\;},\\ |G^{N,{\theta}}_{t,\tau}\Pi_{s-1:s,\tau}S(\cdot,\cdot,{\mathbf{Y}}_{s+T})|_\infty &\leq \rho^{t-s} {\mathrm{osc}}(S(\cdot,\cdot,{\mathbf{Y}}_{s+T})) |\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}|_\infty \qquad \mbox{if}\quad t\geq s{\;}.\end{aligned}$$ Consequently, $$\begin{gathered} \left|G^{N,{\theta}}_{t,\tau}{\mathsf{S}_{\tau}}\right|_\infty \leq \sum_{s=1}^{\tau} |G_{t,\tau}^{N,{\theta}}\Pi_{s-1:s,\tau}S(\cdot,\cdot,{\mathbf{Y}}_{s+T})|_\infty\\ \leq \left(\sum_{s=1}^{\tau}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot,{\mathbf{Y}}_{s+T})\}\right) |\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}|_\infty{\;},\end{gathered}$$ which shows . [*Proof of* ]{}. By the definition , for all $x\in\Xset$ and all $t\in\{1,\dots,\tau\}$, $$\begin{gathered} \mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x) = \int m_{{\theta}}(x,x_{t+1})g_{{\theta}}(x_{t+1},{\mathbf{Y}}_{t+T+1}) \ \\ \times\prod_{u=t+2}^{\tau}{\ensuremath{m}}_{{\theta}}(x_{u-1},{\mathrm{d}}x_u)g_{{\theta}}(x_{u},{\mathbf{Y}}_{u+T})\lambda({\mathrm{d}}x_{t+1:\tau}){\;}.\end{gathered}$$ Hence, by A\[assum:strong\], $$\begin{aligned} \left|\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}\right|_{\infty}&\leq \sigma_+\int g_{{\theta}}(x_{t+1},{\mathbf{Y}}_{t+T+1})\mathcal{L}_{t+1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x_{t+1})\lambda({\mathrm{d}}x_{t+1})\\ \mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x)&\geq \sigma_- \int g_{{\theta}}(x_{t+1},{\mathbf{Y}}_{t+T+1})\mathcal{L}_{t+1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x_{t+1})\lambda({\mathrm{d}}x_{t+1}){\;},\end{aligned}$$ which concludes the proof of the first statement. By (\[eq:lcal\]), A\[assum:strong\] and , $$\dfrac{\mathcal{L}_{t-1,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x)}{|\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}|_\infty} =\int {\ensuremath{m}}_{{\theta}}(x,x^{\prime})g_{{\theta}}(x^{\prime},{\mathbf{Y}}_{t+T}) \dfrac{\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}(x^{\prime})}{|\mathcal{L}_{t,\tau}^{{\theta}}{\ensuremath{\mathbf{1}}}|_\infty}\lambda({\mathrm{d}}x') \geq \dfrac{\sigma_-^{2}}{\sigma_+}b_{-}({\mathbf{Y}}_{t+T}){\;}.$$ The proofs of Propositions \[Prop:NormD\] and \[Prop:NormC\] follow the same lines as [@dubarry:lecorff:2011 Propositions $1$-$2$]. The upper bounds given here provide an explicit dependence on the observations. \[Prop:NormD\] Assume A\[assum:strong\] and A\[assum:dub:lec\]. Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with input variables ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. For all $p > 1$, there exists a constant $C$ such that $$\begin{gathered} \label{Eq:NormD} \mathbb{E}_{}\left[\left|\sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\right|^{p}\right]\\ \leq C \frac{\tau_{n+1}^{(\frac{p}{2}-1) \vee 0}}{N^{p-(\frac{p}{2}\vee 1)}} \sum_{t=0}^{\tau_{n+1}}\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot, {\mathbf{Y}}_{s+T_n})\}\right|^{p}\right]{\;}.\end{gathered}$$ where $D_{t,\tau}^{N,{\theta}}$ is defined in . By Lemma \[Lem:Upperbounds\], $$\frac{\phi_{t-1}^{N,{\theta}_n}[{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}]}{\phi_{t-1}^{N,{\theta}_n}\left[\frac{\mathcal{L}_{t-1,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right]} \leq \frac{\sigma_{+} |{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_{\infty}}{\sigma_{-}^{2}b_{-}({\mathbf{Y}}_{t+T_n})}{\;}.$$ By Lemma \[Lem:Upperbounds\](\[Lem:Upperbounds:ind\]) and since ${\theta}_{n}$ is $\mathcal{F}_{n,t}^{N}$-measurable for all $t\in\{0,\dots,\tau_{n+1}\}$, $\left\{D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}}),\mathcal{F}_{n,t}^{N}\right\}_{0\leq t\leq \tau_{n+1}}$ is a martingale difference. Since $p> 1$, Burkholder’s inequality (see [@hall:heyde:1980 Theorem 2.10, page 23]) states the existence of a constant $C$ depending only on $p$ such that: $$\mathbb{E}_{}\left[\left|\sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\right|^{p}\right]\leq C\mathbb{E}\left[\left|\sum_{t=0}^{\tau_{n+1}} \left|D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}}) \right|^2\right|^{p/2}\right]{\;}.$$ Hence, $$\begin{gathered} \mathbb{E}_{}\left[\left|\sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\right|^{p}\right] \leq C \left(\frac{\sigma_{+} |{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_{\infty}}{\sigma_{-}^{2}}\right)^p\\ \times\mathbb{E}_{}\left[\left|\sum_{t=0}^{\tau_{n+1}}\left|N^{-1}\sum_{\ell=1}^{N}\frac{{\omega_{t}^{\ell}}}{b_{-}({\mathbf{Y}}_{t+T_n})} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^2\right|^{p/2}\right]{\;},\end{gathered}$$ which implies, using the convexity inequality $(\sum_{k=1}^\tau a_k)^{p/2} \leq \tau^{(p/2-1) \vee 0} \sum_{k=1}^\tau a_k^{p/2}$, $$\begin{gathered} \mathbb{E}_{}\left[\left|\sum_{t=0}^{\tau_{n+1}}D_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\right|^{p}\right]\leq C\left(\frac{\sigma_{+} |{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_{\infty}}{\sigma_{-}^{2}}\right)^p\\ \times \frac{\left(\tau_{n+1}+1\right)^{(\frac{p}{2}-1) \vee 0}}{N^{p}} \sum_{t=0}^{\tau_{n+1}} \mathbb{E}_{}\left[\left|\frac{1}{b_{-}({\mathbf{Y}}_{t+T_n})} \sum_{\ell=1}^{N} {\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\right]{\;}.\end{gathered}$$ Since ${\mathbf{Y}}_{t+T_n}$ and ${\theta}_{n}$ are $\mathcal{F}_{n,t-1}^{N}$-measurable, $$\begin{gathered} \mathbb{E}_{}\left[\left|\frac{1}{b_{-}({\mathbf{Y}}_{t+T_n})} \sum_{\ell=1}^{N} {\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\right]\\ = \mathbb{E}_{}\left[\mathbb{E}_{}\left[\left| \sum_{\ell=1}^{N} {\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]\frac{1}{b_{-}({\mathbf{Y}}_{t+T_n})^{p}}\right]\end{gathered}$$ By Lemma \[Lem:Upperbounds\], using again the Burkholder and convexity inequalities, there exists $C$ s.t. $$\begin{aligned} \mathbb{E}_{}\left[\left| \sum_{\ell=1}^{N} {\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\middle|\mathcal{F}_{t-1,n}^{N}\right]&\leq CN^{(\frac{p}{2}-1)\vee 0}\mathbb{E}_{}\left[\sum_{\ell=1}^{N} \left|{\omega_{t}^{\ell}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]\\ &\leq CN^{\frac{p}{2}\vee 1}\mathbb{E}_{}\left[\left|{\omega_{t}^{1}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{1}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]{\;}.\end{aligned}$$ The proof is concluded by . \[Prop:NormC\] Assume A\[assum:strong\] and A\[assum:dub:lec\]. Let $\left\{ ({\xi_{t}^{\ell}}, {\omega_{t}^{\ell}}), 1 \leq \ell \leq N, 0 \leq t \leq \tau_{n+1} \right\}$ be the weighted samples obtained by Algorithm \[alg:FSMC\] in Appendix \[app:alg\], with input variables ${\theta}_n, \tau_{n+1}$, $N$, ${\mathbf{Y}}_{T_n+1:T_n+\tau_{n+1}}$. For all $\bar p > 1$ and all $p\in (1,\bar p)$, there exists a constant $C$ s.t. for any $t\in\{0,\dots,\tau_{n+1}\}$, $$\begin{gathered} \label{Eq:NormC} \mathbb{E}_{}\left[\left|C_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})\right|^{p}\right]\leq CN^{(\frac{p}{2}\vee \frac{1}{\alpha})+(\frac{p}{2}\vee \frac{1}{\beta})-2p}\\ \times\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\right|^{2p}\mathbb{E}_{}\left[\left|\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot, {\mathbf{Y}}_{s+T_n})\}\right|^{\bar p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{p/\bar p}\right]{\;},\end{gathered}$$ where $C_{t,\tau}^{N,{\theta}}$ is defined in and $\alpha{\ensuremath{\stackrel{\mathrm{def}}{=}}}\bar p/ p$ and $\beta^{-1} = 1 - \alpha^{-1}$. Lemma \[lem:esperancecond\] applied with the function $h = \mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}$ and (\[eq:lcal\]) yield for any $1 \leq \ell \leq N$ $${\mathbb{E}_{}\left[{\omega_{t}^{\ell}} \mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}({\xi_{t}^{\ell}}) \middle| \mathcal{F}_{n,t-1}^{N}\right]} = \frac{\phi_{t-1}^{N,{\theta}_n}\left[\mathcal{L}_{t-1,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}\right]}{\phi_{t-1}^{N,{\theta}_n}[{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{t}}{\upsilon_{t}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{t}}{\tilde{\upsilon}_{t}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{t}}{\upsilon^\star_{t}()}}{\mathrm{erreur}}}}}]}{\;}.$$ Therefore, by definition of $C_{t,\tau}^{N,{\theta}}$ (see ), $C_{t,\tau_{n+1}}^{N,{\theta}_n}({\mathsf{S}_{\tau_{n+1}}})$ is equal to $$\begin{gathered} \frac{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \ B_{n,t}^N {\;}= \left( {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N\right) \cdots \\ \times \frac{\Omega_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \left( \frac{B_{n,t}^N}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} } + \frac{B_{n,t}^N}{\Omega_{n,t}^N{\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]}} \left( {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]}- \Omega_{n,t}^N\right)\right) \\ = B_{n,t}^N \left( {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N\right) \frac{\Omega_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \frac{1}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} } \\ + \left( {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N \right) \left( {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]}- \Omega_{n,t}^N\right)\frac{B_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \frac{1}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} } \ \end{gathered}$$ with $$\begin{aligned} A_{n,t}^N & {\ensuremath{\stackrel{\mathrm{def}}{=}}}N^{-1}\sum_{\ell=1}^{N}{\omega_{t}^{\ell}} \frac{\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}} {\;}, \\ B_{n,t}^N & {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{N} \sum_{\ell=1}^N {\omega_{t}^{\ell}} \frac{G^{N,{\theta}_n}_{t,\tau_{n+1}}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty} {\;}, \\ \Omega_{n,t}^N & {\ensuremath{\stackrel{\mathrm{def}}{=}}}\frac{1}{N} \sum_{\ell=1}^N {\omega_{t}^{\ell}} {\;}.\end{aligned}$$ This can be rewritten, $$C_{t,\tau}^{N,{\theta}} = C_{1} + C_{2}{\;},$$ with $$C_{1} = B_{n,t}^N \left( {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N\right) \frac{\Omega_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \frac{1}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} }$$ and $$\begin{gathered} C_{2} = \left( {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - A_{n,t}^N \right) \left( {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]}- \Omega_{n,t}^N\right)\\ \times\frac{B_{n,t}^N}{ {\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N} \frac{1}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} } {\;}.\end{gathered}$$ By Lemmas \[lem:esperancecond\] and Lemmas \[Lem:Upperbounds\], and A\[assum:dub:lec\], $$\frac{1}{ {\mathbb{E}_{}\left[ \Omega_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} } \leq \frac{ \sigma_-|{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_\infty}{b_{-}({\mathbf{Y}}_{t+T_n})} {\;}; \qquad \frac{\Omega_{t,n}^N}{{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N } \leq \left(\frac{\sigma_+}{\sigma_-}\right)^2 \frac{|{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_\infty}{\sigma_- b_{-}({\mathbf{Y}}_{t+T_n})};$$ and by Lemma \[Lem:Upperbounds\] $$\frac{B_{n,t}^N}{{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ A_{n,t}^N } \leq \left(\frac{\sigma_+}{\sigma_-}\right)^2 \frac{|{\ifthenelse{\equal{}{}}{\ifthenelse{\equal{}{}}{\upsilon_{}}{\upsilon_{}()}} {\ifthenelse{\equal{}{smooth}}{\ifthenelse{\equal{}{}}{\tilde{\upsilon}_{}}{\tilde{\upsilon}_{}()}} {\ifthenelse{\equal{}{fully}}{\ifthenelse{\equal{}{}}{\upsilon^\star_{}}{\upsilon^\star_{}()}}{\mathrm{erreur}}}}}|_\infty}{\sigma_- b_{-}({\mathbf{Y}}_{t+T_n})} \left( \sum_{s=1}^{\tau_{n+1}} \rho^{|t-s|} {\mathrm{osc}}(S(\cdot,\cdot,{\mathbf{Y}}_{s+T_n}) \right) {\;}.$$ Therefore, there exists a constant $C$ s.t. $$\begin{aligned} \mathbb{E}_{}\left[\left|C_{1}\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]\leq C\left|\frac{1}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\right|^{2p}\mathbb{E}_{}\left[\left|B_{n,t}^{N}\right|^{p}\left|{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ - A_{n,t}^N\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]\end{aligned}$$ Applying the Holder inequality with $\alpha {\ensuremath{\stackrel{\mathrm{def}}{=}}}\bar p/p\geq 1$ and $\beta^{-1} {\ensuremath{\stackrel{\mathrm{def}}{=}}}1-\alpha^{-1}$ yields $$\begin{gathered} \mathbb{E}_{}\left[\left|B_{n,t}^{N}\right|^{p}\left|{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ - A_{n,t}^N\right|^{p}\middle|\mathcal{F}_{n,t-1}^{N}\right]\\ \leq \mathbb{E}_{}\left[\left|B_{n,t}^{N}\right|^{\alpha p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\alpha}\mathbb{E}_{}\left[\left|{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ - A_{n,t}^N\right|^{\beta p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\beta}{\;}.\end{gathered}$$ By Proposition \[Prop:NormD\], $$\begin{aligned} \mathbb{E}_{}\left[\left|B_{n,t}^{N}\right|^{\alpha p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\alpha}\\ &\hspace{-2cm}\leq CN^{(\frac{p}{2}\vee \frac{1}{\alpha})-p}{\mathbb{E}_{}\left[\left|{\omega_{t}^{1}} \frac{G_{t,\tau_{n+1}}^{N,{\theta}_n}{\mathsf{S}_{\tau_{n+1}}}({\xi_{t}^{1}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_{\infty}}\right|^{\alpha p}\middle | \mathcal{F}_{n,t-1}^{N}\right]}^{1/\alpha}\\ &\hspace{-2cm}\leq CN^{(\frac{p}{2}\vee \frac{1}{\alpha})-p}{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})^{p}\mathbb{E}_{}\left[\left|\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot, {\mathbf{Y}}_{s+T_n})\}\right|^{\alpha p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\alpha}{\;},\end{aligned}$$ Given $\mathcal{F}_{n,t-1}^{N}$, the random variables $\left\{\mathbb{E}_{}\left[\left.{\omega_{t}^{\ell}} \frac{\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}({\xi_{t}^{1}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty}\right|\mathcal{F}_{n,t-1}^{N}\right]- {\omega_{t}^{\ell}} \frac{\mathcal{L}_{t,\tau_{n+1}}{\ensuremath{\mathbf{1}}}({\xi_{t}^{\ell}})}{|\mathcal{L}_{t,\tau_{n+1}}^{{\theta}_n}{\ensuremath{\mathbf{1}}}|_\infty}\right\}_{\ell=1}^{N}$ are conditionally independent, centered and bounded by Lemma \[Lem:Upperbounds\]. Following the same steps as in the proof of Proposition \[Prop:NormD\], there exists a constant $C$ such that $$\mathbb{E}_{}\left[\left|{\mathbb{E}_{}\left[ A_{n,t}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} \ -A_{n,t}^N\right|^{\beta p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\beta}\leq CN^{(\frac{p}{2}\vee \frac{1}{\beta})-p}{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})^{p}{\;}.$$ Hence, $$\begin{gathered} \mathbb{E}_{}\left[\left|C_{1}\right|^{p}\right]\leq CN^{(\frac{p}{2}\vee \frac{1}{\alpha})+(\frac{p}{2}\vee \frac{1}{\beta})-2p}\\ \times\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\right|^{2p}\mathbb{E}_{}\left[\left|\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot, {\mathbf{Y}}_{s+T_n})\}\right|^{\bar p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{p/\bar p}\right]{\;}.\end{gathered}$$ Similarly, using $$\mathbb{E}_{}\left[\left| {\mathbb{E}_{}\left[ \Omega_{t,n}^N \middle| \mathcal{F}_{n,t-1}^{N}\right]} - \Omega_{t,n}^N\right|^{\alpha p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{1/\alpha}\leq CN^{(\frac{p}{2}\vee \frac{1}{\alpha})-p}{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})^{p}{\;},$$ yields $$\begin{gathered} \mathbb{E}_{}\left[\left|C_{2}\right|^{p}\right]\leq CN^{-p}\\ \times\mathbb{E}_{}\left[\left|\frac{{\omega_{+}^{}}({\mathbf{Y}}_{t+T_n})}{b_{-}({\mathbf{Y}}_{t+T_{n}})}\right|^{2p}\mathbb{E}_{}\left[\left|\sum_{s=1}^{\tau_{n+1}}\rho^{|t-s|}{\mathrm{osc}}\{S(\cdot,\cdot, {\mathbf{Y}}_{s+T_n})\}\right|^{\bar p}\middle|\mathcal{F}_{n,t-1}^{N}\right]^{p/\bar p}\right]{\;}.\end{gathered}$$ [^1]: This work is partially supported by the French National Research Agency, under the programs ANR-07-ROBO-0002 and ANR-08-BLAN-0218. [^2]: LTCI, CNRS and TELECOM ParisTech, 46 rue Barrault 75634 Paris Cedex 13, France

--- abstract: 'In this paper we propose a scheme for partially teleporting entangled atomic states. Our scheme can be implemented using only four two-level atoms interacting either resonantly or off-resonantly with a single cavity-QED. The estimative of losses occurring during this partial teleportation process is accomplished through the phenomenological operator approach technique.' author: - 'W. B. Cardoso' - 'N. G. de Almeida' title: Partial teleportation of entangled atomic states --- Introduction ============ Quantum teleportation, first suggested by Bennett *et al.* [Bennett93]{}, is one of the cornerstones of quantum information and computation [@Brassard98; @Gottesman99; @Nielsen00]. The crucial ingredient characterizing this phenomenon is the transfer of information between noninteracting systems at the expense of a quantum channel. This issue has received great attention since its pioneer proposal, mainly after its experimental realizations from 1997 onwards [Bouwmeester97,Boschi98,experiments,experiments2,experiments3]{}. In the meantime, various proposals have been suggested for implementing teleportation, for instance, teleportation of trapped wave fields inside high-Q microwave cavities [Davidovich94,Cirac94,Almeida00,Pires04,Yang06,Cardoso08]{}, teleportation of running wave fields [@BraunsteinPRL98; @Villas-BoasPRA99; @LeePRA00], teleportation of trapped field states inside a single bimodal cavity [Iara07]{}, nonprobabilistic teleportation of a field state via cavity QED [@Garreau07], and teleportation of the angular spectrum of a single-photon field [@Matos07], among others. Since the pioneering work by Bennet *et al.* [@Bennett93], several schemes for teleportation differing from this original protocol (OP) have appeared in the literature. For example, in Ref. [@LeePRA00] the authors show how to partially teleport an entanglement of zero and one photon state in the running wave domain. By partial teleportation (PT) it is mean that teleportation will occur by changing one of the partners of the entangled state to be teleported. PT can be detailed, step by step, in the following sequence: i) particles $A$ and $B$ are previously prepared in the state that we want to teleport; ii) an entangled state of particles $C$ and $% D$ is created; iii) a joint measurement on particles $B$ and $C$ is accomplished such that particles $A$ and $D$ assumes the previous entangled state from particles $A$ and $B$. Note that the term PT is also used in literature to deal with teleportation of an unknown state with the generation of its clone [@FilipPRA04; @ZhaoPRL05]. Other interesting protocol is the so called entanglement swapping [ZukowskiPRL93,BosePRA98,LuPLA00]{}. In a standard entanglement swapping, there are usually three spatially separate users: Alice, Bob, and Charlie. Alice and Bob share pairs of entangled particles with Charlie. Initially, the particles with Alice and Bob are not entangled. Then, Charlie makes a Bell-state measurement on his two particles, leading to the entanglement of the two particles with Alice and Bob. In this paper we present a scheme for partial teleportation in the cavity QED domain using entanglement swapping in only one particle, as is done in Ref. [@LeePRA00]. Our scheme uses four two-level atoms interacting either on or off resonantly with a single mode of a high-Q cavity, Ramsey zones, and selective atomic state detectors. To estimate losses occurring during the partial teleportation process, we used the phenomenological operator approach technique (POA) [@poa]. Ideal Teleportation Process =========================== We assume atom $1$ previously entangled with atom $2$ in the following state, which is the state we want to teleport $$\left\vert \phi \right\rangle _{12}=C_{0}\left\vert g\right\rangle _{1}\left\vert e\right\rangle _{2}+C_{1}\left\vert e\right\rangle _{1}\left\vert g\right\rangle _{2}\text{,} \label{in}$$where $C_{0}$ and $C_{1}$ are unknown coefficients obeying $\left\vert C_{0}\right\vert ^{2}+\left\vert C_{1}\right\vert ^{2}=1$, and $\left\vert g\right\rangle $ ($\left\vert e\right\rangle $) is the atomic ground (excited) state. The state (\[in\]) can be prepared, for instance, by the method presented in Ref. [@zhengPRL00], where two two-level atoms interact simultaneously with a single mode of a cavity-field. The Hamiltonian describing the atom-field interaction, in the interaction picture, is $$H_{I}=\hbar \lambda \left( a^{\dagger }\sigma ^{-}+a\sigma ^{+}\right), \label{hon}$$ when the atom-field interaction is resonant, and $$H_{I}=\frac{\hbar\lambda^{2}}{\delta}a^{\dagger}a \sigma^{+}\sigma^{-}, \label{hid}$$ when the atom-field interaction is off-resonant. This condition is valid provided that $\overline{n}\lambda ^{2}\ll \delta ^{2}$ $+\gamma ^{2}$, where $\overline{n}$ is the mean photon number and $\gamma $ is the damping rate for the cavity-field. Here $a^{\dagger }$ and $a$ are the creation and annihilation operators for the cavity field mode, $\sigma ^{+}$ and $\sigma ^{-}$ are the raising and lowering operators for the atom, $\lambda $ is the atom-field coupling constant, and $\delta =\omega -\omega _{0}$ is the detuning between the cavity field frequency $\omega $ and the atomic frequency $\omega _{0}$. To compose the nonlocal channel, a third atom (initially prepared in the excited state $|e\rangle _{3}$) interacts resonantly with the cavity field mode $A$ (in vacuum state $|0\rangle _{A}$), according to Eq. (\[hon\]) (see Fig. 1a). Adjusting the atom-field interaction time to $t=\pi /4\lambda $, the nonlocal channel will be given by$$\left\vert \psi \right\rangle =\frac{1}{\sqrt{2}}\left( \left\vert e\right\rangle _{3}\left\vert 0\right\rangle _{A}-i\left\vert g\right\rangle _{3}\left\vert 1\right\rangle _{A}\right) .$$At this point, Alice has the atom $1$ and the cavity, while atoms $2$ and $3$ are sent to Bob. The state of the whole system composed by the two-level atoms and the cavity mode field is $$\begin{aligned} \left\vert \psi \right\rangle _{total} &=&\frac{1}{2}\left[ \left\vert \Psi ^{+}\right\rangle _{1A}\left( C_{0}\left\vert g\right\rangle _{3}\left\vert e\right\rangle _{2}+C_{1}\left\vert e\right\rangle _{3}\left\vert g\right\rangle _{2}\right) \right. \notag \\ &+&\left. \left\vert \Psi ^{-}\right\rangle _{1A}\left( C_{0}\left\vert g\right\rangle _{3}\left\vert e\right\rangle _{2}-C_{1}\left\vert e\right\rangle _{3}\left\vert g\right\rangle _{2}\right) \right. \notag \\ &+&\left. \left\vert \Phi ^{+}\right\rangle _{1A}\left( C_{0}\left\vert e\right\rangle _{3}\left\vert e\right\rangle _{2}-C_{1}\left\vert g\right\rangle _{3}\left\vert g\right\rangle _{2}\right) \right. \notag \\ &+&\left. \left\vert \Phi ^{-}\right\rangle _{1A}\left( C_{0}\left\vert e\right\rangle _{3}\left\vert e\right\rangle _{2}+C_{1}\left\vert g\right\rangle _{3}\left\vert g\right\rangle _{2}\right) \right] , \label{3}\end{aligned}$$where, for convenience, we have defined the Bell states $\left\vert \Psi ^{\pm }\right\rangle _{1A}$ and $\left\vert \Phi ^{\pm }\right\rangle _{1A}$ as $$\left\vert \Psi ^{\pm }\right\rangle _{1A}=\frac{1}{\sqrt{2}}\left( -i\left\vert g\right\rangle _{1}\left\vert 1\right\rangle _{A}\pm \left\vert e\right\rangle _{1}\left\vert 0\right\rangle _{A}\right) \text{,}$$$$\left\vert \Phi ^{\pm }\right\rangle _{1A}=\frac{1}{\sqrt{2}}\left( \left\vert g\right\rangle _{1}\left\vert 0\right\rangle _{A}\pm i\left\vert e\right\rangle _{1}\left\vert 1\right\rangle _{A}\right) \text{.}$$ As in the OP, the teleportation is completed after Alice measuring on particle $1$ and cavity $A$ and sending her result to Bob, whom will know which unitary operation to accomplish on its particles in order to recover the entangled state that Alice wanted to teleport. Note, however, that different from the OP, when comparing the teleported state resulting from Eq. (\[3\]) with the state to be teleported, Eq. (\[in\]), we see that partner $1$ was replaced by particle $3$. This explains the partial teleportation term used. The experimental setup is shown, step by step, in Fig. \[scheme\]. Next, we show how Alice must proceed to perform the Bell state measurements. {width="4.5cm"} {width="6cm"} {width="5cm"} Bell State Measurements ----------------------- First, atom $1$ crosses a Ramsey zone $R$, adjusted to produce the following evolutions $$\left\vert e\right\rangle \rightarrow \frac{1}{\sqrt{2}}\left( \left\vert g\right\rangle +\left\vert e\right\rangle \right) \label{ev1}$$and $$\left\vert g\right\rangle \rightarrow \frac{1}{\sqrt{2}}\left( \left\vert g\right\rangle -\left\vert e\right\rangle \right) . \label{ev2}$$Then, atom $1$ crosses the cavity interacting off-resonantly with mode $A$, with the interaction time adjusted to $\chi t=\pi $ (with $\chi =\lambda ^{2}/\delta $), resulting in the evolutions $\left\vert g\right\rangle _{1}\left\vert 0\right\rangle _{A}\rightarrow \left\vert g\right\rangle _{1}\left\vert 0\right\rangle _{A},$ $\left\vert g\right\rangle _{1}\left\vert 1\right\rangle _{A}\rightarrow \left\vert g\right\rangle _{1}\left\vert 1\right\rangle _{A},$ $\left\vert e\right\rangle _{1}\left\vert 0\right\rangle _{A}\rightarrow \left\vert e\right\rangle _{1}\left\vert 0\right\rangle _{A},$ and $\left\vert e\right\rangle _{1}\left\vert 1\right\rangle _{A}\rightarrow -\left\vert e\right\rangle _{1}\left\vert 1\right\rangle _{A}$. Next, atom $1$ crosses another Ramsey zone $R^{\prime }$ adjusted like the Ramsey zone $R$ (see Eqs. (\[ev1\]-\[ev2\])). As a consequence, the states of the Bell basis evolve as $$\left\vert \Psi ^{\pm }\right\rangle _{1A}\rightarrow \frac{1}{\sqrt{2}}% \left\vert g\right\rangle _{1}\left( -i\left\vert 1\right\rangle _{A}\pm \left\vert 0\right\rangle _{A}\right) ,$$$$\left\vert \Phi ^{\pm }\right\rangle _{1A}\rightarrow \frac{1}{\sqrt{2}}% \left\vert e\right\rangle _{1}\left( \left\vert 0\right\rangle _{A}\pm i\left\vert 1\right\rangle _{A}\right) .$$Fig. 1b shows the passage of the atoms 1 and 2 in the schematic setup. By selective atomic state detection on atom $1$ it is possible to know if the joint state is $\left\vert \Psi \right\rangle _{1A}$ or $\left\vert \Phi \right\rangle _{1A}.$ Next, we have to discern the phases $(\pm )$ of the Bell states. With the Ramsey zone $R$ turned off, another two-level atom (atom $4$) in the ground state $\left\vert g\right\rangle _{4}$ is sent through the cavity to interact resonantly with mode $A$ (see Fig. 1c) as indicated by Eq. (\[hon\]), with the interaction time $t=\pi /2\lambda $, thus resulting in the following evolutions: $|g\rangle _{4}|0\rangle _{A}\rightarrow |g\rangle _{4}|0\rangle _{A}$ and $|g\rangle _{4}|1\rangle _{A}\rightarrow -i|e\rangle _{4}|0\rangle _{A}$. Next, the atom $4$ crosses the Ramsey zone $R^{\prime }$ (according to Eqs. (\[ev1\]-\[ev2\])) such that the Bell-states are written as $$\left\vert \Psi ^{\pm }\right\rangle _{1A}\left\vert g\right\rangle _{4}\rightarrow \left\{ \begin{array}{c} \left\vert g\right\rangle _{1}\left\vert e\right\rangle _{4}\left\vert 0\right\rangle _{A}\text{ \ \ if (}+\text{)} \\ \left\vert g\right\rangle _{1}\left\vert g\right\rangle _{4}\left\vert 0\right\rangle _{A}\text{ if (}-\text{)}% \end{array}% \right. ,$$$$\left\vert \Phi ^{\pm }\right\rangle _{1A}\left\vert g\right\rangle _{4}\rightarrow \left\{ \begin{array}{c} \left\vert e\right\rangle _{1}\left\vert g\right\rangle _{4}\left\vert 0\right\rangle _{A}\text{ \ \ if (}+\text{)} \\ \left\vert e\right\rangle _{1}\left\vert e\right\rangle _{4}\left\vert 0\right\rangle _{A}\text{ \ \ if (}-\text{)}% \end{array}% \right. \text{.}$$Thus, by measuring atom $4$ we will be able to distinguish between the phase $\left( -\right) $ or $\left( +\right) $. The perfect discrimination between the four states composing the Bell base can be accomplished by Alice through the detection of the atoms $1$ and $4$, separately. After that, Alice sends a sign to Bob, whom accomplishes an appropriate rotation in the states of the atoms $2$ and $3$ to complete the partial teleportation with $100\%$ of fidelity and success probability, in the ideal case. The unitary operations required by Bob are summarized in Table $1$. -------------------------------------------------------------------------------------------------------------------------- BSM $|\psi \rangle _{32}$ Unitary operation ------------------------------------------- ----------------------------------------- ------------------------------------ $\left\vert \Psi ^{+}\right\rangle _{1A}$ $C_{0}|g\rangle _{3}|e\rangle $\mathbb{I}_{3}\otimes \mathbb{I}% _{2}+C_{1}|e\rangle _{3}|g\rangle _{2}$ _{2}$ $\left\vert \Psi ^{-}\right\rangle _{1A}$ $C_{0}|g\rangle _{3}|e\rangle $\sigma _{3z}\otimes \mathbb{I}% _{2}-C_{1}|e\rangle _{3}|g\rangle _{2}$ _{2} $ $\left\vert \Phi ^{+}\right\rangle _{1A}$ $C_{0}|e\rangle _{3}|e\rangle $\sigma _{3y}\otimes \mathbb{I}% _{2}-C_{1}|g\rangle _{3}|g\rangle _{2}$ _{2} $ $\left\vert \Phi ^{-}\right\rangle _{1A}$ $C_{0}|e\rangle _{3}|e\rangle $\sigma _{3x}\otimes \mathbb{I}% _{2}+C_{1}|g\rangle _{3}|g\rangle _{2}$ _{2} $ -------------------------------------------------------------------------------------------------------------------------- : Results of the teleportation scheme. BSM denotes the resulting measurement on atom $1$ and field mode $A$. Unitary operation denotes the required operation by Bob after Alice communicating her results. The $% \protect\sigma _{3j}$ is the Pauli operator $\protect\sigma _{j}$ acting on atom $3$. Decay of the free atomic excited state ====================================== Phenomenological operator approach (POA) ---------------------------------------- Here we observe that the coupling of the atomic states to a surrounding environment $\mathcal{E}$ can be described by the relations [@poa] $$|g\rangle |\mathcal{E}\rangle \overset{U_{t}}{\longrightarrow }|g\rangle \hat{\mathcal{T}}_{0}|\mathcal{E}\rangle , \label{1d}$$$$|e\rangle |\mathcal{E}\rangle \overset{U_{t}}{\longrightarrow }|e\rangle \hat{\mathcal{T}}_{e}^{\dagger }|\mathcal{E}\rangle +|g\rangle \hat{\mathcal{% T}}_{g}^{\dagger }|\mathcal{E}\rangle \text{,} \label{1e}$$where $|\mathcal{E}\rangle $ denotes the initial state of the environment, the operators $\hat{\mathcal{T}}$, acting on this state, account for the atom-environment coupling, and $U_{t}$ denotes an unitary operation mixing the atom to its environment. We will assume the environment $|\mathcal{E}% \rangle $ in the vacuum state, which is an excellent approximation for high-Q cavities in the microwave domain [@Brune96]. Accordingly, we assume that $\hat{\mathcal{T}}_{0}^{\dagger }=\mathbf{1}$, $\hat{\mathcal{T}}% _{e}^{\dagger }=\mathbf{f}(t)=e^{-\kappa t}\mathbf{1}$, $\hat{\mathcal{T}}% _{g}^{\dagger }=\sum_{j}\mathbf{g}_{j}(t)\hat{b}_{j}^{\dagger }$, with $% \sum_{j}|\mathbf{g}_{j}(t)|^{2}=1-e^{-2\kappa t}$, $\kappa $ denoting the spontaneous decay rate of the atomic excited state, $\mathbf{1}$ is the identity operator, $b_{j}^{\dagger }$ is the creation operator, having a corresponding annihilation operator $b_{j}$, of the $j$th oscillator mode of the environment, and $t$ is the time elapsing after the atom suffering a given excitation. With these assumptions, it is straightforward to verify that the superposition $\left( \left\vert g\right\rangle +\left\vert e\right\rangle \right) /\sqrt{2}$ leads to the reduced density operator $$\begin{aligned} \rho &=&\frac{1}{2}\left\{ \exp (-2\kappa t)|e\rangle \langle e|+\left[ 2-\exp (-2\kappa t)\right] |g\rangle \langle g|\right. \notag \\ &+&\left. \exp (-\kappa t)\left( |e\rangle \langle g|+|g\rangle \langle e|\right) \right\} \text{.} \label{1c}\end{aligned}$$Note that the evolution (\[1d\]) and (\[1e\]) are consistent with the well-known result that an unstable atomic state decays exponentially. In this case, the phenomenological-operator evolution leads to the same atomic density operator as the one we obtain using an *ab-initio* master equation approach. Moreover, due to recent advances in high-Q cavities [harochenature]{}, we will neglect the damping time of the mode $A$. Decay of the teleported state ----------------------------- To estimate the losses, we assume the whole state starting to decay after the preparation of the quantum channel. In the first step, the phenomenological operators used to introduce damping effects, Eqs. (\[1d\]) and (\[1e\]), are applied to the whole system until the time $t$. Then, for each excitement suffered by the atoms during the teleportation process, a new phenomenological operator is included, which modifies the decay probability of the atomic states, and as a consequence, the fidelity of the whole teleportation process. Summarizing the applications of the phenomenological operators from the beginning, *i.e.,* since the preparation of the quantum channel until the end of our teleportation protocol, which occurs at the instant that the fourth atom is detected, we have to apply them soon after *i)* the atom $1$ crossing the first Ramsey zone ($t_{1}$); *ii)* the atom $1$ crossing the second Ramsey zone ($t_{2}$); *iii)* the atom $4$ interacting resonantly with the mode field cavity ($t_{3}$); *iv)* the atom $4$ crossing the Ramsey zone ($t_{4}$). After the inclusion of the decay via POA, the state of the whole system by the time the teleportation is concluded becomes a mixture, being represented by a reduced operator density when the reservoir is traced out. In our estimative, we take the case of the teleported state in Bob hands when Alice measures the Bell state $\left\vert \Psi ^{+}\right\rangle _{1A}$ ($\left\vert g\right\rangle _{1}\left\vert e\right\rangle _{4}$). The corresponding fidelity is shown in Fig. \[fid\]. Note that at the time the teleportation is completed ( $t_{4}$) the fidelity rounds $1$, indicating that we can safely neglect losses occurring during the teleportation process. In fact, taking $t_{1}=2\mu s$, as reported in [@RaimondRMP01], we will have $t_{2}\simeq 5\times 10^{-4}s+2\mu s+t_{1}$, which is the necessary time for the atom $1$ to interact dispersively with the cavity field and to cross the Ramsey zone, $t_{3}=10^{-4}s+t_{2}$, which is the necessary time for the atom-field resonant interaction, and $t_{4}=2\mu s+t_{3}\simeq 6,06\times 10^{-4}s$, which is much shorter than the atomic decay $\kappa ^{-1}\cong 10^{-2}s$, being the fidelity at this time $0.99$ as can be seen from Fig. \[fid\]. However, as the time goes on, the decay becomes faster and the fidelity is reduced to $2/3$ at the instant $% t_{f}=5,78\times 10^{-3}s$. Therefore, the effective time during which the teleported state is at our disposal for further operations is $t_{f}-t_{4}=$ $5,17\times 10^{-3}s$.  . Fluctuation effect in the interaction time ========================================== In this section we show that the fluctuation effect in the interaction time due to the uncertainty in the atomic speed is not relevant as compared to the decay effects of the teleported state presented in Section above. First, we consider the impossibility of sharply fixing the atom-field interaction time. The method adopted here is the same presented in Ref. [MessinaEPJD02]{}. We introduce the probability density $f_{j}(t_{j};% \widetilde{t_{j}})$, where $t$ is the true duration of the interaction between the atom $j$ ($j=1,2$) and the cavity. We assume that $f_{j}(t_{j};% \widetilde{t_{j}})$ is a Gaussian distribution centered around $\widetilde{% t_{j}}$, e.g.,$$f_{j}(t_{j};\widetilde{t_{j}})=\frac{1}{\Delta _{j}\sqrt{2\pi }}\exp \left( -% \frac{\left( t_{j}-\widetilde{t_{j}}\right) ^{2}}{2\Delta _{j}^{2}}\right) ,$$where $\Delta _{j}=x\widetilde{t_{j}}$ and $x$ is a parameter related to the uncertainty in the atomic velocity (around $0.5\%$ according to recent experiments [@RauschenbeutelPRL99]), and therefore in the requested interaction times $\widetilde{t_{j}}$. Thus, the density operator of the system including the fluctuation effect is written as $$\rho =\int \int \int_{-\infty }^{\infty }\left[ \prod% \limits_{j=1}^{N}f_{j}(t_{j};\widetilde{t_{j}})dt_{j}\right] \left\vert \psi \right\rangle _{total}\left\langle \psi \right\vert .$$Here, for simplicity, we consider $N=1,2,3$ to describe the three interactions between atoms $1$ and $4$ and the cavity. Following the steps in Ref. [@MessinaEPJD02], we obtain the fidelity given by $$\begin{aligned} F&=&N^{2}\left[ 1/2\,{{C_{0}}}^{4} \left( {{\rm e}^{3/2\,{x}^{2}{\pi }^{2}}}+{{\rm e} ^{1/2\,{x}^{2}{\pi }^{2}}}+2\,{{\rm e}^{{x}^{2}{\pi }^{2}}} \right) { {\rm e}^{-3/2\,{x}^{2}{\pi }^{2}}}+ \left( 1-{{C_{0}}}^{2} \right) \left( 2-2\,{{C_{0}}}^{2} \right) \right. \nonumber \\ &&- \left. 2\,{{C_{0}}}^{2} \left( -{ {\rm e}^{1/2\,{x}^{2}{\pi }^{2}}}-1+{{C_{0}}}^{2}{{\rm e}^{1/2\,{x}^{ 2}{\pi }^{2}}}+{{C_{0}}}^{2} \right) {{\rm e}^{-3/4\,{x}^{2}{\pi }^{2 }}} \right],\end{aligned}$$ with $$\begin{aligned} N&=&\left( \sqrt { 2\,{{C_{0}}}^{2}{{\rm e}^{-1/2\,{x}^{2}{\pi }^{2}}}+3- 2\,{{C_{0}}}^{2}-{{\rm e}^{-1/2\,{x}^{2}{\pi }^{2}}} } \right) ^{-1}.\end{aligned}$$ The plot of the fidelity is shown in Fig. \[fvel\]. Note that the fidelity does not suffer a significant modification when considering up to $3\%$ of uncertainty in the interaction time. Comments and Conclusion ======================= Since the teleportation protocol by Bennett *et al*. [Bennett93]{}, several other proposals have appeared, modifying slightly or substantially the original protocol. In this paper we have explored a kind of teleportation named by *partial teleportation* [@LeePRA00]. In our scheme, Alice has an atomic state to be teleported, given by an entanglement of particles $1$ and $2$. Besides, Alice shares with Bob a nonlocal channel composed by the joint state of a particle $A$ (represented by a single mode of a high Q cavity) and a particle $3$ (represented by an atomic state). After Alice performing a Bell measurement on the states of particles $A$ and $1$, and informing Bob her result, the following interesting result emerges, after the usual rotation by Bob: particle $3$ takes exactly the role of particle $1$ in the entanglement addressed to Alice, but in Bob location. As the entanglement between the particles $1$ and $2$ is broken and a new entanglement between the particles $3$ and $2$ is created in a different place, this characterizes a partial teleportation. Note that different from Ref. [@LuPLA00], in our scheme the teleportation occurs in only one particle of the entangled pair. To estimate losses occurring during and after the teleportation process, we have used the phenomenological operator approach (POA), as introduced in Ref. [poa]{}. The fluctuation effect in the atom-field interaction time due to the uncertainty in atomic velocities was also considered, showing a small variation in the fidelity. Taking experimental parameters from recent experiments in QED-cavity [@harochenature], we believe that this scheme can experimentally be accomplished using nowadays technology. Acknowledgments {#acknowledgments .unnumbered} --------------- We thanks CNPq and CAPES, Brazilian Agency, and PROPE/UCG - Universidade Católica de Goiás, for supporting this work. [99]{} C. H. Bennett, G. Brassard, C. Crepeau, R. Jozsa, A. Peres, W. K. Wootters, Phys. Rev. Lett. 70 (1993) 1895. G. Brassard, S. L. Braunstein, R. Cleve, Physica D 120 (1998) 43. D. Gottesman, I. L. Chuang, Nature 402 (1999) 390; Knill E, R. Laamme, G. J. Milburn, Nature 409 (2001) 46. M. A. Nielsen, I. L. Chuang, *Quantum Computation and Quantum Information* (Cambridge University Press, Cambridge, England, 2000). D. Bouwmeester *et al.*, Nature 390 (1997) 575. D. Boschi *et al*., Phys. Rev. Lett. 80 (1998) 1121. A. Furusawa *et al*., Science 282 (1998) 706. E. Lombardi *et al.*, Phys. Rev. Lett. 88 (2002) 070402. H. de Riedmatten *et al.*, Phys. Rev. Lett. 92 (2004) 047904. L. Davidovich *et al.*, Phys. Rev. A 50 (1994) R895. J. I. Cirac, A. S. Parkins, Phys. Rev. A 50 (1994) R4441. N.G. de Almeida, R. Napolitano, M.H.Y. Moussa, Phys. Rev A. 62 (2000) 010101(R). G. Pires *et al.*, Phys. Rev. A 70 (2004) 025803. Z. B. Yang, J. Phys. B: At. Mol. Opt. Phys. 39 (2006) 603; W. B. Cardoso *et al.*, J. Phys. B: At. Mol. Opt. Phys. 40 (2007) 1089; W. B. Cardoso, Int. J. Theor. Phys. 47 (2008) 977. S. L. Braunstein, H. J. Kimble, Phys. Rev. Lett. 80 (1998) 869. C. J. Villas-Bøas, N. G. de Almeida, M. H. Y. Moussa, Phys. Rev. A 60 (1999) 2759. H. W. Lee, J. Kim, Phys. Rev. A 63 (2000) 012305. G. Pires *et al.*, Phys. Rev. A 71 (2005) 060301; Iara P. de Queirós *et. al.*, Phys. Rev. A 76 (2007) 034101. C. R. Carvalho *et al.*, J. Phys. B: At. Mol. Opt. Phys. 40 (2007) 1271. S. P. Walborn *et al.*, Phys. Rev. A 76 (2007) 033801. R. Filip, Phys. Rev. A 69 (2004) 052301. Z. Zhao *et al.*, Phys. Rev. Lett. 95 (2005) 030502. M. Żukowski *et al.*, Phys. Rev. Lett. 71 (1993) 4287. S. Bose, V. Vedral, P.L. Knight, Phys. Rev. A 57 (1998) 822; Phys. Rev. A 60 (1999) 194. H. Lu, G. C. Guo, Phys. Lett. A 276 (2000) 209; Z. Y. Xue *et al.*, Opt. Commun. 258 (2006) 315. N. G. de Almeida, R. Napolitano, M. H. Y. Moussa, Phys. Rev. A 62 (2000) 033815. S. B. Zheng and G. C. Guo, Phys. Rev. Lett. 85 (2000) 2392. M. Brune *et al.*, Phys. Rev. Lett. 77 (1996) 4887. S. Gleyzes *et. al*, Nature 446 (2007) 297. J. M. Raimond, M. Brune, S. Haroche, Rev. Mod. Phys. 73 (2001) 565. A. Messina, Eur. Phys. J. D 18 (2002) 379. A. Rauschenbeutel et al., Phys. Rev. Lett. 83 (1999) 5166.

--- abstract: 'Random Forests are one of the most popular classifiers in machine learning. The larger they are, the more precise is the outcome of their predictions. However, this comes at a cost: their running time for classification grows linearly with the number of trees, i.e. the size of the forest. In this paper, we propose a method to aggregate large Random Forests into a single, semantically equivalent decision diagram. Our experiments on various popular datasets show speed-ups of several orders of magnitude, while, at the same time, also significantly reducing the size of the required data structure.' author: - Frederik Gossen - Bernhard Steffen bibliography: - 'literature.bib' title: | Large Random Forests:\ Optimisation for Rapid Evaluation --- Introduction ============ Random Forests are one of the most widely known classifiers in machine learning [@Ho1995; @Breiman2001]. The method is easy to understand, implement, and at the same time achieves impressive classification accuracies in many applications. Compared to other methods such as neural networks, Random Forests are fast to train and they are clearly more suitable for smaller datasets. In contrast to a single decision tree, Random Forests, a collection of many trees, do not overfit as easily on a dataset and their variance decreases with their size. On the other hand, their running time for classification linearly grows with this size, which is critical as forests may well consist of thousands of trees – a problem especially for applications with a high throughput [@facebook]. In this paper, we present an optimisation method that is based on radical aggregation: Forests are transformed into a single decision diagram in a semantics-preserving fashion, which, in particular, also preserves the learner’s variance and accuracy. Being a post-process, also the ease of Random Forest training is maintained. The great advantage of these decision diagrams is their absence of redundancy: during classification every predicate is considered at most once. Our experiments with popular data sets from the UCI Machine Learning Repository [@uci-repo] showed performance gains of several order of magnitude (cf. Fig. \[fig:running-time\] and Table \[tab:running-time\]). A potential problem is only an explosion in size which can, in principle, be exponential for decision diagrams. However, this problem did not arise in our experiments. On the contrary, we even observed drastic size reductions (cf. Fig. \[fig:size\] and Table \[tab:size\]). Key to our approach are Algebraic Decision Diagrams (ADDs) [@Bahar1993]. Their algebraic structure supports aggregation, abstraction, and reduction operations that we utilise to optimise both, the running time needed for classification and the size of the final data structure. When combined with a reduction that exploits the unfeasibility of paths in decision diagrams, this leads to impressive performance gains in our experiments: - Using basic algebraic operations, such as concatenation and addition, allows to aggregate a Random Forest into a single ADD. - Abstracting results to the essence, in this case the outcome of a majority vote, leads to quite dramatic reductions, both of classification times, as well as of size requirements. - Subsequent elimination of unfeasible paths (path with contradicting predicates) finally achieves our goal: high performance gains (cf. Fig. \[fig:running-time\] and Table \[tab:running-time\]) and, additionally, significant reductions of space requirements (cf. Fig. \[fig:size\] and Table \[tab:size\]). Please note that these results are achieved for a widely used classifier and for utterly standard datasets. This indicates the generality of our approach to aggregate Random Forests with thousands of trees into a compact decision diagram for rapid classification – faster by multiple orders of magnitude. **Related work:** Runtime performance of Random Forests has been addressed, e.g., via optimising code generation with moderate success [@treelite; @weka; @facebook; @Browne2019], and, with a greater performance impact, via model simplification which, however, changed the semantics [@Kargupta2004]. Others applied semantic aggregation [@Mulvaney2003; @Kargupta2004; @Giabbanelli2015; @Bonfietti2015] to Random Forests, however, without explicitly addressing the runtime performance, while the authors in [@Peterson2009; @Painsky2019] were focusing solely on the memory footprint, all with moderate success. The only paper on Random Forests we know of that uses decision diagrams similar our ADDs is [@Nakahara2017]. However, they use these diagrams only to compact the individual tree and not to aggregate an entire forest. In fact, the reported speedup by a factor of up to $61$ seems more to rely on technical and even hardware details than on the use of decision diagrams. In contrast, our approach focuses on the decision diagram-based holistic aggregation of entire Random Forests, which, due to its globality, has a much greater impact. In fact, we obtain speed-ups already at the hardware-independent level that are orders of magnitude higher than in [@Nakahara2017]. After a short introduction to Random Forests in Section \[sec:random-forest\], we present our approach to their aggregation in Section \[sec:label-vector\], which is subsequently refined in two steps: by compositional and non-compositional abstraction in Section \[sec:abstraction\], and by the elimination of redundant predicates from the decision diagrams in Section \[sec:unsatisfiable-path-elimination\]. The impact on the classification time and size of the new decision diagrams is evaluated in Section \[sec:evaluation\]. The paper closes with conclusions and direction to future work in Section \[sec:conclusion\]. Random Forests {#sec:random-forest} ============== Random Forests is one of the most widely known classifiers in machine learning. The algorithm is relatively simple and yields good results for many real-world applications. Its decision model generalises a training dataset that holds examples of input data labelled with the desired output, also called *class*. As its name suggests, a Random Forest consists of some number of decision trees. Each of these trees is itself a classifier that was learned from a random sample of the training dataset. Consequently, all trees are different in structure, they represent different decision functions, and can yield different decisions for the very same input data. To apply a Random Forest to previously unseen input data, every decision tree is evaluated separately: Tracing the trees from their root down to one of the leaves yields one decision per tree, i.e. the predicted class. The overall decision of the Random Forest is then derived as the most frequently chosen class, an aggregation commonly referred to as *majority vote*. Key advantage of this approach is the, compared to single decision trees, reduced variance. A detailed introduction to Random Forests, decision trees, and their learning procedures can be found in [@Ho1995; @Breiman2001; @Quinlan1986]. In this paper, we use Weka [@weka] as our reference implementation of Random Forests. However, our approach does not depend on implementation details and can be easily adapted to other implementations. {width=".9\textwidth"} Figure \[fig:random-forest\] shows a small Random Forests that was learned from the popular Iris dataset [@Fisher1936]. The dataset lists dimensions of Iris flowers’ sepals and petals for three different species. Using this forest to decide the species on the basis of given measurements requires to first evaluate the three trees individually and to subsequently determine the majority vote. This effort clearly grows linearly with the size of the forest. In the following we use this example to illustrate our approach of forest aggregation and its great effect on the required evaluation effort. Rapid Evaluation of Random Forests {#sec:label-vector} ================================== To evaluate a Random Forest, we are forced to evaluate every single tree separately. While our illustrative example consists of only three trees there is essentially no limit to the number of trees in a forest. In fact, increasing its size can only improve the classifier and will, in contrast to other classifiers, not lead to overfitting [@Breiman2001]. However, with a growing number of trees comes increasing computational cost for its construction and, more importantly, for its classification process. In this paper we focus on the costs associated with the model’s classification process (the classification time), while accepting additional effort for their construction. This focus reflects the fact that constructed decision structures, once deployed, are often meant to be used by millions of users in parallel. Key idea behind our approach is to partially evaluate the Random Forests at construction time which, we will see, has an enormous impact on the classification performance and the corresponding space requirements. This is, not the least, due to the fact that the individual trees of a Random Forest typically share some similarities. E.g., in our accompanying Iris flower example (cf. Fig. \[fig:random-forest\]) the predicate $petalwidth~<~1.65$ is used in all three trees. This can easily lead to cases where the same predicate is evaluated many times in the classification process. The partial evaluation proposed in this paper transforms Random Forests into decision structures where such redundancies are totally eliminated. An adequate data structure to achieve this goal for binary decisions are Binary Decision Diagrams [@Bryant1986; @Akers1978; @Lee1959] (BDDs): For a given predicate ordering, they constitute a normal form where each predicate is evaluated at most once, and only if required to determine the final outcome. Algebraic Decision Diagrams (ADDs) [@Bahar1993] generalise BDDs to capture functions of the type $\mathbb{B}^\mathcal{P} \rightarrow \mathcal{C}^n$ which are exactly what we need to specify the semantics of Random Forests for a classification domain $\mathcal{C}$. Moreover, in analogy to BDDs, which inherit the algebraic structure of their co-domain $\mathbb{B}$, ADDs also inherit the algebraic structure of their co-domains if available. We exploit this property during the partial evaluation of Random Forest by considering two algebraic co-domains, the class word co-domain $\mathcal{C}*$ (cf. Sec. \[subsec:label-vector-monoid\]) and the class vector co-domain (cf. Sec. \[subsec:abstraction-label-multiset\]). The aggregation to achieve the corresponding optimised decision structures is then a straightforward consequence of the used ADD technology. Algebraic Structure for Random Forest Results {#subsec:label-vector-monoid} --------------------------------------------- Let us put aside the evaluation procedure for now and focus only on its outcome, i.e. the final result. The unprocessed result of a Random Forest evaluation is an ordered sequence with one decision per tree. This sequence preserves all information and remains independent of any particular aggregation method, e.g. *majority vote*. To formalise this, let $\mathcal{C}$ be the set of classes, e.g. the three Iris flower species. An individual tree’s decision is one class $c \in \mathcal{C}$, a Random Forest’s decision is a word over these classes $\mathbf{c} \in \mathcal{C}^*$. Note that we can describe the results of any Random Forests this way, no matter its size. In particular, we can represent the decision made by the empty Random Forest with the empty word $\epsilon$ and the results of a single decision tree with a word of length one. Moreover, this representation naturally allows for composition: we can simply concatenate the results of two distinct Random Forests, maintaining a one-to-one association between the word’s symbols, i.e. the classes, and the corresponding tree in the forest. With that in mind, we can define the algebraic structure *class words* as a string monoid to represent the results of any Random Forest: $$\begin{aligned} W &:= (\mathcal{C}^*, \circ, \epsilon) \textrm{.}\end{aligned}$$ The classes $\mathcal{C}$ form its alphabet, concatenation $\circ$ is its associative join operation, and the empty word $\epsilon$ serves as a neutral element of the monoid. Semantics-preserving Transformation {#subsec:label-vector-transform} ----------------------------------- With *class words*, we have discussed a means to represent only the outcome of a Random Forest classification. To transform the entire Random Forest to an Algebraic Decision Diagram (ADD), however, we need to transform not only the results but also the decision structure. To this aim, we must guarantee the unique properties of decision diagrams: - they enforce an order of predicates along all paths, and - they are directed acyclic graphs that share common substructures where possible. With the compositionality of the algebraic structure $W$ and the corresponding ADDs $\mathcal{D}_W$, we can transform any Random Forest incrementally. Starting with the empty Random Forest, we consider one tree after the other, aggregating a growing sequence of decision trees until the entire forest is entailed in the new decision diagram. We will first find a semantically equivalent decision diagram for the empty Random Forest, the neutral element of this aggregation procedure. Subsequently, we will describe a semantics-preserving transformation for single decision trees and a join operation to incorporate these decision diagrams into the overall aggregation. No matter the input it was given, the empty Random Forest with $0$ trees can only result in one outcome: the empty word $\epsilon$. Hence, it resembles the constant function, also denoted $\epsilon$ for brevity, that is semantically equivalent to the constant decision diagram with $\epsilon$ as its only terminal node. This diagram forms the neutral element of our aggregation procedure. To transform a single decision tree, we can build upon the well-known ADD construction operation $ite$. For a predicate $p$ and two decision diagrams $f$ and $g$, $ite(p, f, g)$ constructs the diagram that evaluates to $f$ if $p$ holds and to $g$ otherwise. We derive the decision diagram recursively along the tree structure, effectively delegating the entire model transformation to well-known and efficient algorithms in a service-oriented fashion. The algorithm implementing $ite$ ensures a strict predicate order and automatically shares substructures where possible. In fact, the resulting decision diagram is a canonical representation of the function for a given predicate order. Formally, this defines a function $d_W: \mathcal{T} \rightarrow \mathcal{D}_W$ mapping decision trees to decision diagrams over *class words*: $$\begin{aligned} d_W(t) &:= \begin{cases} t_{val} & \textrm{if } t \textrm{ is leaf,} \\ ite(t_{pred}, d_W(t_{then}), d_W(t_{else})) & \textrm{otherwise.} \end{cases}\end{aligned}$$ Having transformed every decision tree individually leaves us with the task to compose the resulting sequence of decision diagrams. This is where the above-mentioned characteristic properties of ADDs come into play: they inherit the algebraic structure of their co-domains, and they even come with efficient algorithms for computing the required operations. In this case, this concerns only the concatenation $\circ$ of words over the carrier set $\mathcal{C}^*$. To ease readability, we denote also this terminal-wise concatenation of decision diagrams with the same symbol $\circ$. The desired decision diagram, aggregating an entire sequence of decision trees $t_0, t_1, \dots, t_n$, can now simply be defined as follows: $$\begin{aligned} d_W(t_0, t_1, \dots, t_{n - 1}) &:= d_W(t_0) \circ d_W(t_1) \circ \dots \circ d_W(t_{n - 1}) \textrm{.}\end{aligned}$$ {width=".9\textwidth"} Figure \[fig:label-vector-dd\] shows the aggregation of our exemplary Random Forest (cf. Fig. \[fig:random-forest\]). Already, for this extremely small example, the average running time for classification is reduced. Its true impact, however, becomes apparent only with increasing forest size (cf. Sec. \[sec:evaluation\]). Partial Evaluation through Abstraction {#sec:abstraction} ====================================== *Class words* faithfully represent the information about the decisions of each individual tree in the forest. This is far more than necessary for determining the final decision of the forest, the majority vote. In this section we exploit this leeway as follows: - We first abstract from the information which tree was responsible for which decision by moving from the domain of class words to the domain $\mathbb{N}^{|\mathcal{C}|}$ of *class vectors* which simply record the frequency with which each class has been proposed. $(\mathbb{N}^{|\mathcal{C}|,+)}$, where addition $+$ is defined component-wise, is again a monoid. The structure can again be lifted to the ADD-level in order to guarantee the compositional aggregation of forest as illustrated in the previous section for the class word monoid. - Subsequently, we abstract the aggregated result to a decision diagram which only reflects the majority vote. This abstraction is not compositional and can therefore only be applied at the very end of the aggregation process as an additional optimisation. The following subsections are devoted to these two steps, respectively. From Class Words to Class Frequencies {#subsec:abstraction-label-multiset} ------------------------------------- As mentioned above, the precise knowledge about the individual decisions of the trees in a Random Forest is unnecessary for determining the final decision of a Random Forest. The knowledge about the frequency with which each class has been proposed suffices. This information can elegantly be represented as *class vectors*, where each component represents one class and its value the frequency with with the class was chosen. Formally, the domain of *class vectors* forms a monoid $$\begin{aligned} V &:= (\mathbb{N}^{|\mathcal{C}|}, +, \mathbf{0}) \end{aligned}$$ where addition $+$ is defined component-wise and $\mathbf{0}$ is the neutral element. Based on this structure we can replay the development of the previous section by replacing $W$ by $V$, $\circ$ by $+$, and $\epsilon$ by $\mathbf{0}$. Indexing the class vectors directly with the class labels $c \in \mathcal{C}$ rather than integers this reads as follows: The required representation of the empty decision is again provided by the neutral element, here the $\mathbf{0}$ vector, and a single class $c \in \mathcal{C}$ can be represented naturally by a vector $\mathbf{i}(c)$ that is $0$ everywhere except for position $c$, where it is $1$. Also the construction of the ADDs $\mathcal{D}_V$ with class vectors as their terminal values can simply be transcribed, as can the new transformation function $d_V: \mathcal{T} \rightarrow \mathcal{D}_V$, which differs only in its mapping from the tree’s leaves to the new carrier set: $$\begin{aligned} d_V(t) &:= \begin{cases} \mathbf{i}(t_{val}) & \textrm{if } t \textrm{ is leaf,} \\ ite(t_{pred}, d_V(t_{then}), d_V(t_{else})) & \textrm{otherwise.} \end{cases}\end{aligned}$$ Having adopted the underlying algebraic structure, all operations are seamlessly applicable to the corresponding decision diagrams as well. In this case, vector summation $+$ is lifted to the new decision diagrams over vectors and we can again easily aggregate the Random Forest incrementally: $$\begin{aligned} d_V(t_0, t_1, \dots, t_{n - 1}) &:= \sum_{i = 1}^{n - 1} d_V(t_i) \textrm{.}\end{aligned}$$ The new transformation abstracts from the order of class labels but maintains all the information required to construct and aggregate decision diagrams incrementally. Abstracting from the order of class labels has two important advantages: 1. **Memory** is saved as many leaf nodes that differed only in the order of class labels are now unified. In fact, this effect can ripple up the entire decision diagram, i.e. the structure can partially collapse. Moreover, the vector representation itself also becomes more compact. 2. **Classification time** of the decision diagram is reduced as a result of the partial collapse of the structure: Where a predicate was previously needed to differentiate between two class words that differed only in the order of their class labels, this evaluation step becomes redundant. Moreover, the final aggregation step reduces to finding the maximal component of a single class vector. ![Class vector abstraction of aggregated Random Forest.[]{data-label="fig:label-multiset-dd"}](./figures/inf-rand-forest-example/labelMultisetDD.pdf){width="\textwidth"} Figure \[fig:label-multiset-dd\] shows the result of the class frequency abstraction for our running example. Majority Vote at Compile Time {#subsec:abstraction-most-frequent-label} ----------------------------- As mentioned before, just maintaining the information about the result of the majority votes is not compositional. In fact, knowing the result of the majority votes for two Random Forest gives no clue about the majority vote of the combined forest. Thus the most frequent class abstraction can only be applied at the very end, after the entire aggregation has been computed compositionally. In fact, the class frequency abstraction provides the most concise compositional abstraction. Any further reduction directly leads to potential compositionality violations, or as we say, the class frequency abstraction is *fully abstract* for this scenario. Thus taking the formerly defined model transformation $d_V$ to iteratively aggregate the trees of a Random Forest is provably the best one can do. [^1] The result is a decision diagram $d_V(\mathbf{t}) \in \mathcal{D}_V$ with class vectors in its terminal nodes. Only the subsequent monadic transformation $mv : \mathcal{D}_V \rightarrow \mathcal{D}_C$ remains to be defined. With ADDs we can simply define this on the carrier set, i.e. on the class vectors. For any class vector $\mathbf{v} \in \mathbb{N}^{|\mathcal{C}|}$ the majority vote is defined as $$\begin{aligned} mv(\mathbf{v}) &:= \argmax_{c \in \mathcal{C}} \mathbf{v}_{c} \textrm{.}\end{aligned}$$ Just like binary operations of the algebraic structure $V$ are lifted to Algebraic Decision Diagrams, so can monadic operations [@Bahar1993]. Note that $mv$ does not project into the same carrier set but rather from one algebraic structure $V$ into another $C$. However, these transformations can be applied to the corresponding decision diagrams in the very same way. We can therefore define the final transformations as $$\begin{aligned} d_C(\mathbf{t}) &:= mv(d_V(\mathbf{t})) \textrm{.}\end{aligned}$$ Post-processing vector decision diagrams in this way has again quite some effect: Both memory and classification time are reduced for the same reasons as for the class frequency abstraction. However, the coarser abstraction leads to stronger reductions. Moreover, the aggregation step for determining the final decision of the Random Forest, the majority vote, is no longer necessary. ![Most frequent label abstraction of aggregated Random Forest (majority vote).[]{data-label="fig:label-dd"}](./figures/inf-rand-forest-example/labelDDDerivedFromLabelMultisetDD.pdf){width="\textwidth"} Fig. \[fig:label-dd\] shows the result of the most frequent class abstraction for our running example. Our evaluation in Sec. \[sec:evaluation\] shows that all the abstractions proposed so far are insufficient to guarantee true scalability, and this despite the fact that they are optimal. This is due to the fact that they all are symbolic and do not take the semantics of predicates into account. The following section sketches how this shortcoming can be overcome and finally provides us with the scalability results we desired. Please note, however, that the preceding abstractions are necessary to enable the potential of unfeasible path elimination (cf. Sec. \[fig:running-time\]). Unsatisfiable Path Elimination {#sec:unsatisfiable-path-elimination} ============================== ![Most frequent label abstraction of aggregated Random Forest (majority vote) without semantically redundant nodes.[]{data-label="fig:label-dd-2"}](./figures/inf-rand-forest-example/labelDDDerivedFromLabelMultisetDDWithMergedUnsatPaths.pdf){width="\textwidth"} When aggregating the trees of a Random Forest they all use varying sets of predicates. In contrast to simple Boolean variables, predicates are not independent on one another, i.e. evaluation of one predicate may yield some degree of knowledge about other predicates. E.g., the predicate $petallength~<~2.45$ induces knowledge about other predicates that reason about $petallength$: When the petal length is smaller than $2.45$ it cannot possibly be greater or equal to $2.7$ at the same time. This is not taken care of by the symbolic treatment of predicates we followed until now. Unsatisfiable path elimination, as illustrated by the difference between Figure \[fig:label-dd\] and Figure \[fig:label-dd-2\] for our running example, leverages the potential of a semantic treatment of predicates with significant effect: - The **size** of decision diagrams is drastically reduced (cf. the cases marked with ’\*’ in Fig. \[fig:size\]), and even - the **classification times** further improve because semantically redundant decisions are eliminated (cf. the cases marked with ’\*’ in Fig. \[fig:running-time\]). Please note that unfeasible path elimination - depends on previous powerful abstraction: The trees in the original Random Forest have no unfeasible paths by construction. They are introduced in the course of our *symbolic* aggregation, which is insensitive to semantic properties. - is compositional and can therefore be applied during the stepwise transformation and before the final most frequent label abstraction and at the very end. This avoids that intermediate decision diagrams grow too large which would inhibit the scalability. As can be seen in Figure \[fig:size\], without this effect, our approach would hardly scale to forests beyond the size of 100 trees. - does not support normal forms. Thus our approach may yield different decision diagrams depending on the order of tree aggregation. It is guaranteed, however, that the resulting decision diagrams are minimal. Unsatisfiable path elimination is a hard problem in general.[^2] Our corresponding implementation uses SMT-solving to eliminate all unsatisfiable paths. An in-depth discussion of unsatisfiable path elimination is a topic in its own and beyond the scope of this paper. Evaluation {#sec:evaluation} ========== ![Average running time for classification over all examples in the Iris dataset [@iris-dataset].[]{data-label="fig:running-time"}](./figures/inf-rand-forest-size-rt/avgrt.pdf){width=".8\textwidth"} **Dataset** **Random Forest** **Final DD** --------------- ------------------- ----------------- Balance Scale 80,277.03 8.16 (-99.99%) Breast Cancer 130,361.20 17.73 (-99.99%) Lenses 43,883.79 3.67 (-99.99%) Iris 44,043.89 7.01 (-99.98%) Tic-Tac-Toe 107,300.69 14.18 (-99.99%) Vote 69,216.62 8.30 (-99.99%) : Running time improvements for classification with Random Forests of size 10.000 for other datasets [@uci-repo].[]{data-label="tab:running-time"} Our three tree accompanying example is useful to explain the concepts but inadequate to illustrate the impact of our radical aggregation technology. This section therefore provides a careful quantitative analysis on the basis of a number of popular data sets that illustrates the performance differences between the semantically equivalent representations of the original Random Forest. The diagrams in this section show the results concerning quantitative extensions of the Iris flower set which, in the small, also served for our running example. The tables summarise the results for other popular data sets to indicate the generality of our approach. All the reported classification time and size results were determined as the average over the entire corresponding data sets. For the Iris flower example these are 150 records, a number also explaining the quite smooth result graphs. Our implementation relies on the standard Random Forest implementation in Weka [@weka] and on the ADD implementation of the ADD-Lib [@add-lib; @lde; @lde-stress]. Please note that the considered data sets have been developed with evaluations of this kind in mind, by independent parties, and that we are not using any additional data for our transformation. Thus our analysis can be considered unbiased. Optimising the classification time is the primary goal of our approach. As wall clock time measurements are very sensitive to implementation details and machine profiles, we decided for the, in our eyes more objective measure of step count for our performance analysis. As steps we consider here the steps through the corresponding data structures, and in cases where the most frequent class must be computed at runtime, we account one additional step per read. For both, the original Random Forest and the word-based decision diagram these are $n$ additional steps and the class vector variant needs $|\mathcal{C}|$ additional steps. Figure \[fig:running-time\] shows the average evaluation times of the decision models for Random Forests of up to 10,000 trees. The evaluation time of the original Random Forest grows linearly as expected: every new tree contributes approximately the same running time. Due to the large number of trees relative to their individual sizes our measurements appear as an almost straight line. Already, the word-based diagrams (cf. Class word DD in Fig. \[fig:running-time\]) reduce the classification time significantly in comparison to the original Random Forest. This is due to the suppression of redundant predicate evaluations. In fact, the overall classification time is dominated by the linearly growing time to compute the most frequent class in each terminal word. The reduction to just $|\mathcal{C}|$ terminal nodes of the class vector-based variants has two effects: - A partial collapse of the decision diagram: it is no longer essential which tree proposes which class, unifying all cases where the various classes are equally often proposed. - Reduction to a constant overhead for the final aggregation step, in this case $|\mathcal{C}|$. The evaluation time reductions are again quite significant, only the space requirement got, like for the word-based variant, out of hand very soon (cf. Fig. \[fig:size\]), explaining the cut-off in Fig. \[fig:running-time\]. Whereas the previous two model structures can directly be computed compositionally, the most frequent label abstraction, i.e. the evaluation of the *majority vote* at compile time, can only be applied at the very end. Thus its construction has the same limitation as the class vector variant, and its impact on the size of the corresponding decision model is moderate (cf. Fig. \[fig:size\]). Its impact on the evaluation time is, however, quite substantial (cf. Fig. \[fig:running-time\]): Many of the internal decision nodes have become redundant by just focusing on the results of the majority vote. ![Sizes of the Random Forest and its semantically equivalent decision diagrams.[]{data-label="fig:size"}](./figures/inf-rand-forest-size-rt/size.pdf){width=".8\textwidth"} **Dataset** **Random Forest** **Final DD** --------------- ------------------- ----------------- Balance Scale 2,158,330 144 (-99.99%) Breast Cancer 5,494,682 3,760 (-99.93%) Lenses 136,986 11 (-99.99%) Iris 135,952 1,267 (-99.07%) Tic-Tac-Toe 5,670,532 1,529 (-99.97%) Vote 988,358 1,148 (-99.88%) : Decision diagram sizes for Random Forests of size 10.000 for other datasets [@uci-repo].[]{data-label="tab:size"} Breathing semantics into the decision diagrams by unsatisfiable path elimination overcomes the scalability problems that are due to the enormous space requirements. In fact, it avoids the exponential blow-up in size in all three variants, with DD\* even becoming significantly smaller than the original Random Forest (cf. Fig. \[fig:size\]). Moreover, also the classification times are drastically reduced in all three cases (cf. Fig. \[tab:running-time\]). In fact, the classification times eventually stabilise for DD\*, illustrating the key feature of Random Forests, the reduction of the learner’s variance. [^3] As sketched in Tables \[tab:running-time\] and \[tab:size\] these observations carry over to other popular data sets in the UCI Machine Learning Repository [@uci-repo]. Conclusion {#sec:conclusion} ========== In this paper, we have presented an approach to aggregate large Random Forests into a single and compact decision diagram using the machinery of Algebraic Decision Diagrams. This radical transformation allows for rapid classification and decision-making while, at the same time, preserving the original semantics. We have refined our method in multiple steps: by compositional abstraction, pre-evaluation at compile time, and the elimination of redundant predicates. As a result, we could achieve running time reductions by factors of thousands on multiple popular datasets, proving the method’s relevance for real world problems. At the same time, and beyond our original expectations, also the size requirements could be reduced by orders of magnitude. The results reported in this paper concern a widely used classifier and standard datasets and were achieved with clean aggregation, abstraction, and reduction criteria: - ADD-based aggregation is canonical as soon as an order of predicates has been fixed. Thus the freedom of choice here reduces to the choice of an adequate variable ordering, a task heuristically taken care of by the corresponding frameworks [@Somenzi2001]. - *Class frequency* abstraction is the coarsest, compositional abstraction that still allows one to faithfully represent the classification function of the original Random Forest. - Unfeasible path elimination does not support normal forms, but the results are minimal, meaning that the resulting structures cannot be reduced further without changing the semantics of the classification function. In essence, the variability here is a consequence of the freedom of choice where to *root* unfeasible paths. It can be seen as a generalisation of the classical problem of minimising Boolean functions with *don’t cares*. - *Most frequent class* abstraction reduces the final compositionally reduced decision diagrams to the smallest diagram that still represents the original classification function. Thus our approach is optimal relative to two well-known conceptual hurdles, the choice of variable ordering for decision diagrams and the treatment of *don’t cares*. In particular, our approach does not exploit any peculiarities of certain classifiers or data sets. Of corse, its impact may still strongly depend on the structure of the concrete considered scenario. We are therefore currently investigating how easily these results can be adopted to other data sets and classifiers. While we expect the transfer of these ideas to be relatively straight forward for some discrete classifiers, e.g. Decision Jungles [@Shotton2013], this will be more difficult for others. For more complex classifiers such as neural networks, compromises may be unavoidable. In these cases, rapid classification may come at the cost of semantic approximation. With this approach being so successful for the very specific domain of Random Forests, we are also interested in its generalisation ability. Rather than the optimisation of what can be considered a domain-specific program, it will be interesting to see its potential in the context of general-purpose programming languages. First results have shown that we can indeed use very similar techniques to optimise more general programs [@addCompiler]. [^1]: A corresponding proof via contraposition is quite straightforward but beyond the scope of this paper. [^2]: For the cases considered here it is polynomial, but there are of course theories for which it becomes exponentially hard or even undecidable. [^3]: Remember, being just a different representation of the original Random Forest, DD\* has the same variance.

--- abstract: | A system of renewal equations on a graph provides a framework to describe the exploitation of a biological resource. In this context, we formulate an optimal control problem, prove the existence of an optimal control and ensure that the target cost function is polynomial in the control. In specific situations, further information about the form of this dependence is obtained. As a consequence, in some cases the optimal control is proved to be necessarily bang–bang, in other cases the computations necessary to find the optimal control are significantly reduced. **Keywords:** Management of Biological Resources; Optimal Control of Conservation Laws; Renewal Equations. **2010 MSC:** 35L50, 92D25 author: - 'Rinaldo M. Colombo$^1$' - Mauro Garavello$^2$ title: Polynomial Profits in Renewable Resources Management --- Introduction ============ A biological resource is grown to provide an economical profit. Up to a fixed age $\bar a$, this population consists of *juveniles* whose density $J (t,a)$ at time $t$ and age $a$ satisfies the usual renewal equation [@PerthameBook Chapter 3] $$\partial_t J + \partial_a\left(g_J (t,a) \, J\right) = d_J (t,a) \, J \qquad \qquad a \in [0, \bar a]\,,$$ $g_J$ and $d_J$ being, respectively, the usual growth and mortality functions, see also [@ColomboGaravello2014; @ColomboGaravello2015; @GaravelloHYP2014]. For further structured population models, we refer for instance to [@MR1699033; @MR2264557; @MR1624188; @MR2285538; @MR772205]. At age $\bar a$, each individual of the $J$ population is selected and directed either to the market to be sold or to provide new juveniles through reproduction. Correspondingly, we are thus lead to consider the $S$ and the $R$ populations whose evolution is described by the renewal equations $$\begin{array}{r@{\;}c@{\;}l} \partial_t S + \partial_a\left(g_S (t,a) \, S\right) & = & d_S (t,a) \, S \\ \partial_t R + \partial_a\left(g_R (t,a) \, R\right) & = & d_R (t,a) \, R \end{array} \qquad \qquad a \geq \bar a \,,$$ with obvious meaning for the functions $g_S, g_R, d_S, d_R$. Here, the selection procedure is described by a parameter $\eta$, varying in $[0,1]$, which quantifies the percentage of the $J$ population directed to the market, so that $$\begin{aligned} g_S (t, \bar a) \, S (t, \bar a) & = & \eta \, g_J (t, \bar a) \, J (t,\bar a) \\ g_R (t, \bar a) \, R (t, \bar a) & = & (1-\eta) \, g_J (t, \bar a) \, J (t,\bar a).\end{aligned}$$ The overall dynamics is completed by the description of reproduction, which we obtain here through the usual age dependent fertility function $w = w (a)$ using the following nonlocal boundary condition $$g_J (t, 0)\, J (t, 0) = \int_{\bar a}^{+\infty} w (\alpha) \, R (t,\alpha) {\mathinner{\mathrm{d}{\alpha}}} \,.$$ In this connection, we recall the related results [@Ackleh2009; @Ackleh2012; @MR2251787] in structured populations that take into consideration a juvenile–adult dynamics. \[fig:stru2\] Once the biological evolution is defined, we introduce the income and cost functionals as follows. The income is related to the withdrawal of portions of the $S$ population at given stages of its development. More precisely, we assume there are fixed ages $\bar a_1, \ldots, \bar a_N$, with $\bar a < \bar a_1 < \bar a_2 < \cdots < \bar a_N$, where the fractions ${\vartheta}_1, \ldots, {\vartheta}_N$ of the $S$ population are kept, while the portions $(1-{\vartheta}_1), \ldots, (1-{\vartheta}_N)$ are sold. A very natural choice is to set $\bar {\vartheta}_N\equiv 0$, meaning that nothing is left unsold after age $\bar a_N$. The dynamics of the whole system has then to be completed introducing the selection $$S (t, \bar a_i+) = {\vartheta}_i \, S (t, \bar a_i-)$$ that takes place at the age $\bar a_i$, for $i=1, \ldots, N$. Summarizing, the dynamics of the structured $(J,S,R)$ population is thus described by the following nonlocal system of balance laws, see also Figure \[fig:stru2\]: $$\label{eq:7} \left\{ \begin{array}{l@{\qquad}r@{\,}c@{\,}l} \partial_t J + \partial_a\left(g_J (t,a) \, J\right) = d_J (t,a) \, J & (t,a) & \in & {{\mathbb{R}}}^+ \times [0, \bar a] \\ \partial_t S + \partial_a\left(g_S (t,a) \, S\right) = d_S (t,a) \, S & (t,a) & \in & {{\mathbb{R}}}^+ \times \left( \left[\bar a, +\infty\right[ \setminus \{\bar a_1, \ldots, \bar a_N\} \right) \\ \partial_t R + \partial_a\left(g_R (t,a) \, R\right) = d_R (t,a) \, R & (t,a) & \in & {{\mathbb{R}}}^+ \times \left[\bar a, +\infty\right[ \\[6pt] g_S (t, \bar a) \, S (t, \bar a) = \eta \, g_J (t, \bar a) \, J (t,\bar a) & t & \in & {{\mathbb{R}}}^+ \\ g_R (t, \bar a) \, R (t, \bar a) = (1-\eta) \, g_J (t, \bar a) \, J (t,\bar a) & t & \in & {{\mathbb{R}}}^+ \\[6pt] g_J (t, 0)\, J (t, 0) = \int_{\bar a}^{+\infty} w (\alpha) \, R (t,\alpha) {\mathinner{\mathrm{d}{\alpha}}} & t & \in & {{\mathbb{R}}}^+ \\[6pt] S (t, \bar a_i+) = {\vartheta}_i \, S (t, \bar a_i-) & t & \in & {{\mathbb{R}}}^+ \,, \quad i=1, \ldots, N \\[6pt] J (0, a) = J_o (a) & a & \in & [0,\bar a] \\ S (0, a) = S_o (a) & a & \in & \left[\bar a, +\infty\right[ \\ R (0, a) = R_o (a) & a & \in & \left[\bar a, +\infty\right[ \,, \end{array} \right.$$ where we inserted the initial data $(J_o, S_o, R_o)$. Our key result is the proof that for all $t$ and all $a$, the quantities $J (t,a)$, $S (t,a)$ and $R (t,a)$ are *polynomial* in the values attained by the control parameters $\eta$ and ${\vartheta}$. We now pass to the introduction of the expressions of cost and income. To this aim, we first fix a time horizon $T$, with $T>0$. Then, a reasonable expression for the income is $$\label{eq:IncomeJSR} \mathcal{I} (\eta,{\vartheta}; T) = \int_0^{\bar a} P \left(a, J (T, a) \right) {\mathinner{\mathrm{d}{a}}} + \sum_{i=1}^N \int_0^T P_i \left(t, \left(1-{\vartheta}_i (t)\right) \, S (t, \bar a_i-) \right) {\mathinner{\mathrm{d}{t}}}\,.$$ The latter term above is the sum of the incomes due to the selling of the $S$ individuals at the ages $\bar a_1, \ldots, \bar a_N$. Typically, each value function $s \to P_i (t,s)$ can be chosen linear in its second argument, but the present framework applies also to the more general polynomial case. The former term in the right hand side of , namely $\int_0^{\bar a} P \left(a, J (T, a) \right) {\mathinner{\mathrm{d}{a}}}$, accounts for the total amount of the $J$ population at time $T$ and it can also be seen as the capital consisting of the biological resource at time $T$. Neglecting this term obviously leads to optimal strategies that leave no juveniles at the final time $T$. The value function $j \to P (t,j)$ is also assumed to be polynomial, see Section \[subs:SS\]. To model the various costs, we use a general integral functional of the form $$\label{eq:CostJSR} \begin{array}{rcl} \mathcal{C} (\eta,{\vartheta};T) & = & \displaystyle \int_0^T \int_0^{\bar a} C_J \left(t, a, J (t,a) \right) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} + \int_0^T \int_{\bar a}^{+\infty} C_S \left(t, a, S (t,a) \right) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} \\[12pt] & &\displaystyle + \int_0^T \int_{\bar a}^{+\infty} C_R \left(t, a, R (t,a) \right) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} \,. \end{array}$$ The cost functions $w \to C_u(t,a,w)$, for $u \in \left\{J,S,R\right\}$, are assumed to be polynomial in $w$, for all $a$ and $t$. In the simplest case of *linear* cost and income, and  reduce to $$\begin{aligned} \label{eq:2_1} \mathcal{I} (\eta,{\vartheta}; T) & = & \displaystyle \int_0^{\bar a} p (a) \, J (T, a) {\mathinner{\mathrm{d}{a}}} + \sum_{i=1}^N \int_0^T p_i (t) \, \left(1-{\vartheta}_i (t)\right) \, S (t, \bar a_i-) {\mathinner{\mathrm{d}{t}}}\,. \\ \nonumber \mathcal{C} (\eta,{\vartheta};T) & = & \displaystyle \int_0^T \int_0^{\bar a} c_J (t, a) \, J (t,a) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} + \int_0^T \int_{\bar a}^{+\infty} c_S (t, a) \, S (t,a) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} \\ \label{eq:2_2} & & \displaystyle + \int_0^T \int_{\bar a}^{+\infty} c_R (t, a) \, R (t,a) {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} \,.\end{aligned}$$ Here, $p (a)$ is the unit value of juveniles of age $a$, while $p_i(t)$ is the price at time $t$ per each individual of the population $S$ sold at maturity $\bar a_i$. Similarly, the quantity $c_u (t, a)$, for $u \in \{ J, S, R\}$, is the unit cost related to the keeping of individuals of the population $u$, of age $a$, at time $t$. Below, we provide the essential tools to establish effective numerical procedures able to actually compute the profit $$\label{eq:5} \mathcal{P} (\eta, {\vartheta}; T) = \mathcal{I} (\eta, {\vartheta}; T) - \mathcal{C} (\eta, {\vartheta}; T) \,.$$ as a function of the (open loop) control parameters $\eta$ and ${\vartheta}$. In particular, this also allows to find choices of the time dependent control parameters $\eta$ and ${\vartheta}$ that allow to maximize $\mathcal{P}$. Moreover, the procedures presented below provide an alternative to the use of *bang-bang* controls. For a comparison between the two techniques we refer to Section \[subs:SS\]. The next section presents the main results of this note, while specific examples are deferred to paragraphs \[subs:Gen\], \[subs:Periodic\] and \[subs:SS\]. All analytic proofs are in Section \[sec:TD\]. Main Results {#sec:Main} ============ Throughout we denote ${{\mathbb{R}}}^+ = \left[0, +\infty\right[$, while $\chi_{\strut A}$ is the usual characteristic function of the set $A$, so that $\chi_{\strut A} (x) =1$ if and only if $x \in A$, whereas $\chi_{\strut A}$ vanishes outside $A$. The positive integers $\kappa, m$ and $N$ are fixed throughout, as also the positive strictly increasing real numbers $\bar a$, $\bar a_1, \ldots, \bar a_N$. It is also of use to introduce the real intervals $I_J = [0, \bar a]$, $I_S = I_R = \left[\bar a, +\infty\right[$, and $I_T = [0,T]$. Below, for a real valued function $u$ defined on an interval $I$, we call ${\mathop\mathrm{TV}}(u)$ its total variation, while ${\mathbf{BV}}(I; {{\mathbb{R}}})$ is the set of real valued functions with finite total variation, namely: $$\begin{array}{c} \displaystyle {\mathop\mathrm{TV}}(u) = \sup \left\{ \sum_{i=1}^N {{\left|u (t_i) - u (t_{i-1})\right|}} \colon N \in {{\mathbb{N}}},\; t_1, \ldots, t_N \in I \mbox{ and } t_{i-1}< t_i \mbox{ for all } i \right\} \\ \displaystyle \!\!\! {\mathbf{BV}}(I;{{\mathbb{R}}}) = \left\{ u \colon I \to {{\mathbb{R}}}\colon {\mathop\mathrm{TV}}(u) < +\infty \right\} \mbox{ and } {\mathbf{BV}}(I;{{\mathbb{R}}}^+) = \left\{ u \colon I \to {{\mathbb{R}}}^+ \colon {\mathop\mathrm{TV}}(u) < +\infty \right\} . \end{array}$$ We posit the following assumptions: (A) : For $u=J, S, R$, the growth rate $g_u$ and mortality rate $d_u$ satisfy $$ \begin{array}{lcl} g_u \in ({\mathbf{C^{1}}} \cap {\mathbf{L^\infty}}) (I_T \times I_u; \left[\check g_u, +\infty\right[) & \quad \mbox{ and } \quad & \displaystyle \sup_{t \in {{\mathbb{R}}}^+} {\mathop\mathrm{TV}}\left(g_u (t,\cdot)\right) < +\infty \,, \vspace{.2cm} \\ d_u \in ({\mathbf{C^{1}}} \cap {\mathbf{L^\infty}}) (I_T \times I_u; {{\mathbb{R}}}) & \quad \mbox{ and } \quad & \displaystyle \sup_{t \in {{\mathbb{R}}}^+} {\mathop\mathrm{TV}}\left(d_u (t,\cdot)\right) < +\infty \,, \end{array}$$ for a suitable $\check g_u > 0$, while the fertility function $w$ satisfies $w \in {\mathbf{C_c^{1}}} (\left[\bar a, +\infty\right[; {{\mathbb{R}}}^+)$. (ID) : $J_o \in {\mathbf{BV}}(I_J; {{\mathbb{R}}}^+)$, $S_o \in ({\mathbf{L^1}} \cap {\mathbf{BV}}) (I_S; {{\mathbb{R}}}^+)$ and $R_o \in ({\mathbf{L^1}} \cap {\mathbf{BV}}) (I_R; {{\mathbb{R}}}^+)$. (P) : $P \in {\mathbf{L^{\infty}_{loc}}} ([0,\bar a] \times {{\mathbb{R}}}^+; {{\mathbb{R}}})$ and $P_i \in {\mathbf{L^{\infty}_{loc}}} (I_T\times {{\mathbb{R}}}^+; {{\mathbb{R}}})$ for $i = 1, \ldots, N$. Moreover, the map $j \to P (a,j)$, respectively $s \to P_i (t,s)$ for $i=1, \ldots, N$, is a polynomial of degree at most $\kappa$ in $j$ for all $a \in [0, \bar a]$, respectively in $s$ for $t \in I_T$. (C) : $C_u \in {\mathbf{L^{1}_{loc}}} (I_T \times I_u \times {{\mathbb{R}}}; {{\mathbb{R}}})$ and the map $v \to C_u (t,a,v)$ is a polynomial of degree at most $\kappa$ in $v$, for $u = J, S, R$. Above, the restriction to ${{\mathbb{R}}}^+$ of the initial data is not necessary from the analytic point of view, but it is justified by the biological meaning of the variables. Clearly, the extension to the case of polynomials with different degrees is essentially a mere problem of notation. Recall, as in [@ColomboGaravello2015; @GaravelloHYP2014], the strictly increasing sequence of *generation times* $T_\ell$ recursively defined for $\ell \in {{\mathbb{N}}}$, by $$\label{eq:Tk} T_0 = 0 \quad \mbox{ and } \quad \mathcal{A}_J (T_\ell; T_{\ell-1},0) = \bar a \quad \mbox{ or, equivalently, } \quad \mathcal{T}_J (\bar a; T_{\ell-1},0) = T_\ell \,,$$ the characteristic functions $\mathcal{A}_J$ and $\mathcal{T}_J$ being defined in  for $u=J$. If $g_J$ satisfies **(A)**, then the sequence $T_\ell$ is well defined and $T_\ell \to +\infty$ as $\ell \to +\infty$. The interval $[T_{\ell-1}, T_\ell]$ is the time period when the juveniles of the $\ell$-th generation are born. The following results apply to the case of a constant $\eta$ and a constant ${\vartheta}$, when system  fits into [@ColomboGaravello2014 Theorem 2.4] and turns out to be well posed in ${\mathbf{L^1}}$. Let **(A)** hold. For every $\eta \in [0,1]$, ${\vartheta}\in [0,1]^{N}$ and every initial data $(J_o , S_o, R_o)$ as in **(ID)**, system (\[eq:7\]) admits a unique solution $\left(J, S, R\right)$ such that $$\begin{array}{l@{\qquad}l@{\quad}l} J(t,a) \ge 0, & t \in I_T, & a \in I_J, \\ S(t,a) \ge 0, & t \in I_T, & a \in I_S, \\ R(t,a) \ge 0, & t \in I_T, & a \in I_R, \end{array}$$ and the stability estimates in [@ColomboGaravello2014 Theorem 2.4 and Theorem 2.5] hold. In order to exhibit the existence and to actually find a value of $\eta$ and ${\vartheta}$ that maximizes $\mathcal{P}$ as defined in , we first investigate the regularity of $\mathcal{I}$ and $\mathcal{C}$, defined in  and , as functions of the control parameters $\eta$ and ${\vartheta}$. Let **(A)** hold. Let $C_J, C_S, C_R$ satisfy **(C)** and the functions $P$ and $P_i$ satisfy **(P)**. For every $T>0$, every $\eta \in [0,1]$, every ${\vartheta}\in [0,1]^N$ and every initial data $(J_o , S_o, R_o)$ as in **(ID)**, 1. the maps $\eta \to J(T, \cdot)$, $\eta \to S(T, \cdot)$, $\eta \to R(T, \cdot)$, and $\eta \to \mathcal{I} (\eta,{\vartheta};T)$ are all polynomials in $\eta$; 2. the maps ${\vartheta}\to J(T,\cdot)$, ${\vartheta}\to S(T,\cdot)$, ${\vartheta}\to R(T,\cdot)$ are affine in each component ${\vartheta}_i$ of ${\vartheta}$, separately, while the map ${\vartheta}\to \mathcal{I} (\eta,{\vartheta};T)$ is polynomial in each component ${\vartheta}_i$ of ${\vartheta}$. Hence, all the maps $(\eta,{\vartheta}) \to \mathcal{C} (\eta,{\vartheta};T)$, $(\eta,{\vartheta}) \to \mathcal{I} (\eta,{\vartheta};T)$, and $(\eta,{\vartheta}) \to \mathcal{P} (\eta,{\vartheta};T)$ are continuously differentiable in both $\eta$ and ${\vartheta}$. When the control parameters are time dependent, the well posedness of  follows from [@ColomboGaravello2015 Theorem 2.1], which we recall here for completeness. \[thm:time\] Pose conditions **(A)**, **(ID)**. For any $\eta \in {\mathbf{BV}}(I_T;[0,1])$ and ${\vartheta}\in {\mathbf{BV}}(I_T; [0,1]^N)$, system  admits a unique solution. Moreover, $$\begin{array}{l@{\qquad}l@{\quad}l} J(t,a) \ge 0, & t \in I_T, & a \in I_J, \\ S(t,a) \ge 0, & t \in I_T, & a \in I_S, \\ R(t,a) \ge 0, & t \in I_T, & a \in I_R, \end{array}$$ and there exists a function $\mathcal{K} \in {\mathbf{C^{0}}} (I_T; {{\mathbb{R}}}^+)$, with $\mathcal{K} (0) = 0$, dependent only on $g_J$, $g_S$, $g_R$, $d_J$, $d_S$, $d_R$ and $w$ such that for all initial data $(J_o',S_o',R_o')$ and $(J_o'', S_o'', R_o'')$ and for all controls $\eta'$, $\eta''$, ${\vartheta}'$ and ${\vartheta}''$, the corresponding solutions $(J',S',R')$ and $(J'', S'', R'')$ to  satisfy, for every $t \in I_T$, the following stability estimate: $$\begin{aligned} & & {{\left\|J' (t)-J'' (t)\right\|}}_{{\mathbf{L^1}} (I_J; {{\mathbb{R}}})} + {{\left\|S' (t) - S'' (t)\right\|}}_{{\mathbf{L^1}} (I_S; {{\mathbb{R}}})} + {{\left\|R' (t) - R'' (t)\right\|}}_{{\mathbf{L^1}} (I_R; {{\mathbb{R}}})} \\ & \leq & \mathcal{K} (t) \left( {{\left\|J'_o - J''_o\right\|}}_{{\mathbf{L^1}} (I_J; {{\mathbb{R}}})} + {{\left\|S'_o - S''_o\right\|}}_{{\mathbf{L^1}} (I_S; {{\mathbb{R}}})} + {{\left\|R'_o - R''_o\right\|}}_{{\mathbf{L^1}} (I_R; {{\mathbb{R}}})} \right) \\ & & + t \, \mathcal{K} (t) \left( {{\left\|J'_o - J''_o\right\|}}_{{\mathbf{L^\infty}} (I_J; {{\mathbb{R}}})} + {{\left\|S'_o - S''_o\right\|}}_{{\mathbf{L^\infty}} (I_S; {{\mathbb{R}}})} + {{\left\|R'_o - R''_o\right\|}}_{{\mathbf{L^\infty}} (I_R; {{\mathbb{R}}})} \right) \\ & & + \mathcal{K} (t) \left( {{\left\|\eta' - \eta''\right\|}}_{{\mathbf{L^\infty}} ([0,t]; {{\mathbb{R}}})} + {{\left\|{\vartheta}' - {\vartheta}''\right\|}}_{{\mathbf{L^\infty}} ([0,t]; {{\mathbb{R}}}^N)} \right)\,. \end{aligned}$$ Recall the following definition, which allows us to describe the form of the cost, income, and profit as functions of the controls. A map $f \colon {{\mathbb{R}}}^n \to {{\mathbb{R}}}$ is *multiaffine* if $f (\eta)$ is affine as a function of each $\eta_l$, for $l=1, \ldots, n$ , (keeping all other $\eta_k$ fixed). The elementary property below of multiaffine functions plays a key role in selecting those situations where a bang–bang control may yield the optimal profit. Its proof is deferred to Section \[sec:TD\]. \[lem:base\] Let $n \in {{\mathbb{N}}}$ and $f \colon {{\mathbb{R}}}^n \to {{\mathbb{R}}}$ be multiaffine and not constant. Then, $f$ admits neither points of strict local minimum, nor points of strict local maximum. Hence, $\max_{[0,1]^n} f$ is attained on a vertex of $[0,1]^n$. The two theorems below constitute the main results of the present work. \[thm:main\_eta\] Pose conditions **(A)**, **(ID)**. Introduce times $\tau_0, \tau_1, \ldots, \tau_m$ such that $$\label{eq:1eta} \tau_0 = 0 \,,\qquad \tau_{k-1} < \tau_k \mbox{ for }k=1, \ldots, m \,,\qquad T_\ell \not\in\left]\tau_{k-1}, \tau_k \right[ \mbox{ for } \begin{array}{r@{\;}c@{\;}l} k & = & 1, \ldots, m \,, \\ \ell & \in & {{\mathbb{N}}}\end{array}$$ and control parameters $\eta_k \in [0,1]$ for $k = 1, \ldots, m$. Let $(J,S,R)$ be the solution to  corresponding to the control $$\label{eq:etai} \eta (t) = \sum_{k=1}^m \eta_k \, \chi_{\strut[\tau_{k-1}, \tau_k[} (t)\,.$$ Then, for all $t$ and $a$, the quantities $J (t,a)$, $R (t,a)$ and $S (t,a)$ are multiaffine in $(\eta_1, \ldots, \eta_m)$. Remark that the latter condition $T_\ell \not\in \left]\tau_{k-1}, \tau_k \right[$ in  can always be met, through a suitable splitting of the intervals $[\tau_{k-1}, \tau_k]$. \[thm:main\_theta\] Pose conditions **(A)**, **(ID)**. Introduce times $\tau_0, \tau_1, \ldots, \tau_m$ such that $$ \tau_0 = 0 \,,\qquad \tau_{k-1} < \tau_k \quad \mbox{ for } \quad k=1, \ldots, m$$ and control parameters ${\vartheta}_i^k \in [0,1]$ for $k = 1, \ldots, m$ and $i = 1,\ldots, N-1$. Let $(J,S,R)$ be the solution to  corresponding to the controls $$\label{eq:thetai} {\vartheta}_i (t) = \sum_{k=1}^m {\vartheta}_i^k \, \chi_{\strut[\tau_{k-1}, \tau_k[} (t) \quad \mbox{ for } i = 1, \ldots, N-1 \,.$$ Then, for all $i = 1, \ldots, N-1$, if $a \in \,]\bar a_i, \bar a_{i+1}[\,$, the quantity $S (t,a)$ is multiaffine in the variables $({\vartheta}_1^1, \ldots, {\vartheta}_1^m, \ldots, {\vartheta}_i^1, \ldots, {\vartheta}_i^m)$. \[cor:1\] Pose conditions **(A)**, with $g_J$ constant in time, **(ID)**, **(P)** and **(C)**. Choose controls $\eta$ as in – and ${\vartheta}$ as in . Then, the net profit $\mathcal{P}$ defined in  is polynomial in $\eta$ and ${\vartheta}$ of degree at most $\kappa$ in each of the (scalar) variables $\eta_1, \ldots, \eta_m, {\vartheta}_1^k, \ldots, {\vartheta}_{N-1}^k$ separately. Moreover, globally, it is a polynomial of degree at most $\kappa \, m$ in $\eta_1, \ldots, \eta_m$ and of degree at most $\kappa \, m \, (N-1)$ in ${\vartheta}_1^k, \ldots, {\vartheta}_{N-1}^k$. Thanks to the form of the costs and of the gains ensured by **(P)** and **(C)**, the proof is an immediate consequence of Theorem \[thm:main\_eta\] and Theorem \[thm:main\_theta\]. A direct consequence of Corollary \[cor:1\] in the case – of linear gains and costs, thanks to Lemma \[lem:base\], is that optimal controls ${\vartheta}(t) = ({\vartheta}_1, \ldots, {\vartheta}_{N-1}) (t)$, among those of the form (\[eq:thetai\]), can be found restricting the search to only bang–bang controls, i.e., to those assuming only the values $0$ and $1$. Nevertheless, in [@ColomboGaravello2015 Theorem 1.8], it is proved that bang-bang controls well approximate the optimal ones, found in the class of ${\mathbf{BV}}\left(I_T; [0,1]\right)$ for $\eta$ and of ${\mathbf{BV}}\left(I_T; [0,1]^N\right)$ for ${\vartheta}$, provided the cost and income are linear, i.e. in the form (\[eq:2\_1\])-(\[eq:2\_2\]). Examples ======== The examples in paragraphs \[subs:Gen\] and \[subs:Periodic\] rely on several numerical integrations of . They were accomplished using the explicit formula . To compute the gains and the costs –, we used the standard trapezoidal rule. For simplicity, we assume throughout that at age $\bar a_N$ all the population $S (t, \bar a_N)$ is sold; this corresponds to the case ${\vartheta}_N \equiv 0$. A Generational Control {#subs:Gen} ---------------------- We particularize Theorem \[thm:time\] to the case of $\eta$ as in – with $\tau_\ell = T_\ell$, so that $\eta$ is constant on each generation. On the other hand, we keep ${\vartheta}$ constant. \[cor:2Gen\] Pose conditions **(A)**, **(ID)**, **(P)** and **(C)**. Choose linear gains and costs as in –. Let $T_0, T_1, \ldots, T_n$ be as in (\[eq:Tk\]). Set $$\label{eq:3} \eta (t) = \displaystyle \sum_{\ell=1}^n \eta_\ell \, \chi_{\strut[T_{\ell-1}, T_\ell[} (t)$$ and let ${\vartheta}$ be constant. Then, the net profit $\mathcal{P}$ defined in  is multiaffine in $(\eta_1, \ldots, \eta_n)$. Therefore, the optimal profit can be obtained through a bang–bang control. In the present case (\[eq:3\]) there are $2^n$ distinct bang–bang controls: Corollary \[cor:2Gen\] ensures that one of them yields the maximum profit. At the same time, the profit ${\cal P}$ is a multiaffine function in $\eta$, so that it contains at most $2^n$ terms. Therefore, the integration of $2^n$ suitable instances of (\[eq:7\]) permits to obtain all the coefficients in the expression of ${\cal P}$ as a function of $\eta$ and, hence, to compute ${\cal P}$ for *all* (i.e., not necessarily bang–bang) possible controls (\[eq:3\]). Consider the situation $n = 2$ corresponding to the time interval $[0,\, T_2]$, we have $$\eta (t) = \eta_1 \, \chi_{\strut[0, \,T_1]} (t) + \eta_2 \, \chi_{\strut[T_1, \, T_2]} (t)$$ and Corollary \[cor:2Gen\] ensures that the profit $\mathcal{P}$ defined at –– is actually a multiaffine function of $\eta \equiv (\eta_1,\, \eta_2)$, so that $$\begin{aligned} \mathcal{P} (\eta_1, \eta_2) & = & \mathcal{P} (0,0) + \left(\mathcal{P} (1,0) - \mathcal{P} (0,0)\right) \, \eta_1 + \left(\mathcal{P} (0,1) - \mathcal{P} (0,0)\right) \, \eta_2 \\ & & + \left(\mathcal{P} (1,1) - \mathcal{P} (1,0) - \mathcal{P} (0,1) + \mathcal{P} (0,0)\right) \, \eta_1 \, \eta_2 \,.\end{aligned}$$ In other words, thanks to the qualitative information provided by Corollary \[cor:2Gen\], computing $\mathcal{P}$ only $4$ times allows to obtain the expression of $\mathcal{P} (\eta_1, \eta_2)$ valid for all $(\eta_1,\eta_2) \in [0,\,1]^2$. As an example, we consider the setting –– defined by the choices: $$\begin{array}{c} \begin{array}{r@{\;}c@{\;}l@{\qquad}r@{\;}c@{\;}l@{\qquad}r@{\;}c@{\;}l@{\qquad}r@{\;}c@{\;}l} g_J (t,a) & = & 1.00 & d_J (t,a) & = & 1.50 & c_J (t,a) & = & 0.25 & J_o (a) & = & 1.00 \\ g_S (t,a) & = & 1.00 & d_S (t,a) & = & 0.50 & c_S (t,a) & = & 0.00 & S_o (a) & = & 0.00 \\ g_r (t,a) & = & 1.00 & d_R (t,a) & = & 0.75 & c_R (t,a) & = & 0.00 & R_o (a) & = & 0.00 \end{array} \\ \begin{array}{r@{\;}c@{\;}l@{\qquad}r@{\;}c@{\;}l @{\qquad}r@{\;}c@{\;}l@{\qquad}r@{\;}c@{\;}l @{\qquad}r@{\;}c@{\;}l} \bar a & = & 1.00 & \bar a_1 & = & 1.50 & N & = & 1 \\ p (a) & = & 0.00 & p_1 (t) & = & 8.00 & w (a) & = & 120.00 \, \chi_{\strut[1.00,\,4.00]} (a). \end{array} \end{array}$$ Using the expression  of the exact solution to  we obtain (up to the second decimal digit) $$P (0,0) = -19.97\,,\qquad P (1,0) = 3.13 \,,\qquad P (0,1) = 8.22 \,,\qquad P (1,1) = 3.13 \,,$$ so that $$\label{eq:4} \mathcal{P} (\eta_1,\eta_2) = -19.97 + 23.10\, \eta_1 +28.18 \, \eta_2 -28.18 \, \eta_1 \, \eta_2 \,.$$ {width="\linewidth"} {width="0.5\linewidth"}\ {width="0.5\linewidth"} {width="0.5\linewidth"} \[fig:2Gen\] Coherently with the results above, the maximum of $\mathcal{P}$ is attained at the bang–bang control $(\eta_1, \eta_2) = (0,1)$, see Figure \[fig:2Gen\]. This strategy amounts to first keep all juveniles for reproduction and then sell them all. A Periodic Control {#subs:Periodic} ------------------ We now consider the case of a growth function $g_J$ independent of time. Then, with reference to , we have $T_\ell = \ell \, T_1$ for all $\ell \in {{\mathbb{N}}}$. It is then natural to consider a piecewise constant control which is periodic and with period $T_1$: $$\label{eq:8} \begin{array}{@{}r@{\;}c@{\;}l@{\quad}l@{}} \eta (t) & = & \eta (\tau) & \mbox{whenever } (t - \tau) / T_1 \in {{\mathbb{N}}}\\ \eta (t) & = & \displaystyle \sum_{h=1}^m \eta_h \, \chi_{\strut[\tau_{h-1}, \tau_h[} (t) & \mbox{if } 0 \leq \tau_{h-1} < \tau_h \leq T_1 \mbox{ for } h = 1, \ldots, m \mbox{ and } t \in [0, T_1] \,. \end{array}$$ \[cor:Periodic\] Pose conditions **(A)**, **(ID)**, **(P)** and **(C)**. Assume that the growth function $g_J$ is independent of time. Choose $\eta$ as in  with $T = T_n$ for a given $n \in {{\mathbb{N}}}\setminus\{0\}$ and let ${\vartheta}$ be constant. Then, the net profit $\mathcal{P}$ defined in –– is a polynomial of degree at most $n$ in $(\eta_1, \eta_2, \ldots, \eta_m)$. The proof is a direct consequence of Theorem \[thm:main\_eta\] and is hence omitted. Observe that a polynomial of degree $n$ in $m$ variables contains at most $\nu = \left( \begin{array}{@{\,}c@{\,}} n+m\\n \end{array} \right)$ terms. Therefore, Corollary \[cor:Periodic\] reduces the problem of maximizing  along the solutions to  to: 1. the computation of $\nu$ solutions to , 2. the solution to a linear system of $\nu$ equations in $\nu$ variables, 3. the maximization of a polynomial. Consider the following example. In the case $m = 2$ in , choosing the time interval $[0, T_2]$, i.e., $n = 2$, we set $$\label{eq:poly_control} \eta (t) = \eta_1 \, \chi_{\strut[0.0, 0.5]} (t) + \eta_2 \, \chi_{\strut[0.5, 1.0]} (t) + \eta_1 \, \chi_{\strut[1.0, 1.5]} (t) + \eta_2 \, \chi_{\strut[1.5, 2.0]} (t) \,,$$ corresponding to $\tau_0=0.0$, $\tau_1 = 0.5$ and $\tau_2 = 1.0$. Corollary \[cor:Periodic\] ensures that $\mathcal{P}$ is a polynomial of degree at most $2$ separately in $\eta_1$ and $\eta_2$, so that $$\label{eq:poly_form} \mathcal{P} (\eta_1, \eta_2) = c_0 + c_1 \, \eta_1 + c_2 \, \eta_2 + c_3 \, \eta_1 \, \eta_2 + c_4 \, \eta_1^2 + c_5 \, \eta_2^2$$ and $\nu = 6$ numerical integrations of  with the consequent evaluation of  allow to obtain the coefficients $c_0, \ldots, c_5$ and, hence, the full knowledge of the profit as a function of the control parameters. We consider now the setting –– defined by the choices: $$\begin{array}{@{}r@{\;}c@{\;}l@{\quad} r@{\;}c@{\;}l@{\quad}r@{\;}c@{\;}l@{\quad} r@{\;}c@{\;}l@{\quad}r@{\;}c@{\;}l@{}} \bar a & = & 1.00 & d_J (t,a) & = & 0.50 & c_J (t,a) & = & 0.25 & p (a) & = & 1.00 & J_o (a) & = & 1.00 \\ N & = & 1 & d_S (t,a) & = & 1.00 & c_S (t,a) & = & 0.25 & p_1 (t) & = & 8.20 & S_o (a) & = & 0.00 \\ \bar a_1 & = & 1.50 & d_R (t,a) & = & 1.50 & c_R (t,a) & = & 0.25 & w (a) & = & 10.00 \, \chi_{\strut[1.00,\,4.00]} (a) & R_o (a) & = & 0.00. \end{array}$$ Using the expression of the exact solution to  we obtain (up to the second decimal digit) $$\label{eq:poly_coeff} c_0 = 3.65 \,,\ c_1 = 0.46 \,,\ c_2 = -0.88 \,,\ c_3 = 1.11 \,,\ c_4 = - 1.06 \,,\ c_5 = 0.46 \,.$$The resulting profit is plotted in Figure \[fig:periodic\] as a function of $(\eta_1, \eta_2)$. Remark that the resulting optimal control is *not* bang–bang. At the times $t = 0.50, \, 1.00,\, 1.50$ the sharp changes in the graphs of $J, \, S$ and $R$ are due to the sharp changes in the control, as prescribed in (\[eq:poly\_control\]). {width="\linewidth"} {width="0.5\linewidth"}\ {width="0.5\linewidth"} {width="0.5\linewidth"} \[fig:periodic\] A Stabilizing Strategy {#subs:SS} ---------------------- As a further example, we consider the case of a nonlinear profit. A justification for this choice can be the necessity to stabilize the juvenile population to reduce the running costs caused by the $J$ population. Therefore, we consider system , with an income function of the type  and a nonlinear cost for the $J$ population given by $$\label{eq:quadCost} \mathcal{C} (\eta, {\vartheta}; T) = -\int_0^T \int_0^{\bar a} \left(J (t,a) - \bar J\right)^2 {\mathinner{\mathrm{d}{a}}} {\mathinner{\mathrm{d}{t}}} \,.$$ Here, the fixed parameter $\bar J$ can be seen as the juvenile density that, say, minimizes the running costs. We are thus lead to the maximization of the profit , with linear income  and cost . Let $T_\ell$ be as in  and consider a generational control $\eta$ as in , and piecewise constant controls ${\vartheta}_i$ ($i \in \left\{1, \ldots, N-1\right\}$) as $$\label{eq:thetai_bis} {\vartheta}_i (t) = \sum_{\ell=1}^n {\vartheta}_i^\ell \, \chi_{\strut[T_{\ell-1}, T_\ell[} (t )\,,$$ where ${\vartheta}_i^\ell \in [0, 1]$ for every $i \in \left\{1, \ldots, N-1\right\}$ and $\ell \in \left\{1, \ldots, n\right\}$. Then, by the analysis in Section \[sec:Main\], we can assert that the profit  is a second order polynomial in $\eta_1, \ldots, \eta_n$ whose first and zeroth order terms are multiaffine in ${\vartheta}_1^\ell, \ldots, {\vartheta}_{N-1}^\ell$: $$\label{eq:profit-stabilizing-strategy} \begin{array}{rcl} \mathcal{P} (\eta, {\vartheta}) & = & \displaystyle \sum_{\lambda \in \{0,1\}^n} \sum_{\boldsymbol\ell \in \{1, \ldots, n\}^{N-1}} \sum_{\beta \in \{0,1\}^{N-1}} c_{\lambda,\boldsymbol\ell, \beta} \; \eta^\lambda \; ({\vartheta}_1^{\boldsymbol\ell_1})^{\beta_1} \cdots ({\vartheta}_{N-1}^{\boldsymbol\ell_{N-1}})^{\beta_{N-1}} \vspace{.2cm}\\ & & \displaystyle + \sum_{\lambda \in \{0,1,2\}^n \colon \max \lambda = 2} c'_\lambda \; \eta^\lambda \end{array}$$ which is a polynomial defined by $$\label{eq:1} \nu = n^{N-1} \, 2^{n+N-1} + 3^n - 2^n$$ real coefficients. {width="9cm"} \[fig:compa\] Thus, solving $\nu$ times the renewal equations , computing the corresponding $\nu$ profits (\[eq:profit-stabilizing-strategy\]), solving a $\nu \times \nu$ linear system to get the $\nu$ coefficients allows to obtain an expression for $\mathcal{P}$ valid for *all* possible control parameters $\eta \in [0,1]^n$, ${\vartheta}\in [0,1]^{n (N-1)}$. As a comparison, we remark that the total number of bang–bang controls in the present case is $\nu_{bb} = 2^{n\,N}$ and there is no guarantee that the optimal control is of bang–bang type. For a comparison between $\nu$ and $\nu_{bb}$, refer to Figure \[fig:compa\]. Technical Details {#sec:TD} ================= As in [@ColomboGaravello2014; @ColomboGaravello2015; @PerthameBook], we recall that the initial – boundary value problem for the renewal equation $$\label{eq:4old} \left\{ \begin{array}{l} \partial_t u + \partial_a \left(g_u (t,a) \, u\right) = d_u (t,a) \,u \\ u (0, a) = u_o (a) \\ g_u (t,a_u) \, u (t, a_u+) = b (t) \end{array} \right. \qquad \begin{array}{r@{\;}c@{\;}l} t & \geq & 0 \\ a & \geq & a_u \end{array}$$ admits a unique solution that can be explicitly computed integrating along characteristics as $$\label{eq:12} \!\!\! u (t,a) = \left\{ \begin{array}{lr@{\,}c@{\,}l@{}} u_o \left(\mathcal{A}_u (0;t,a)\right) \; \psi_u (0,t,a) & a & \geq & \mathcal{A}_u (t;0,a_u) \\ \frac{b\left(\mathcal{T}_u (a_u;t,a)\right)}{g_u\left(\mathcal{T}_u (a_u;t,a),a_u\right)} \; \psi_u \! \left(\mathcal{T}_u (a_u;t,a),t,a\right) & a & < & \mathcal{A}_u (t;0,a_u) \,, \end{array} \right.$$ where the maps $t \to \mathcal{A}_u (t, t_o, a_o)$ and $a \to \mathcal{T}_u (a;t_o,a_o)$, with $t \in {{\mathbb{R}}}^+$ and $a,a_o \in I_u$, are defined as $$\label{eq:AT} \begin{array}{ll} t \to \mathcal{A}_u (t;t_o,a_o) & \mbox{ is the solution to} \quad \left\{ \begin{array}{l} \dot a = g_u (t,a) \\ a (t_o) = a_o \end{array} \right. \quad \mbox{ and} \\ a \to \mathcal{T}_u (a;t_o, a_o) & \mbox{ is its inverse, i.e., }\quad \mathcal{A}_u\left(\mathcal{T}_u (a;t_o,a_o);t_o,a_o\right) = a \quad \mbox{ for all } a \in I_u\,, \end{array}$$ while the map $\psi_u$ is given by $$\label{eq:psi} \psi_u (t_1,t_2,a) = \exp \int_{t_1}^{t_2} \left( d_u\left(s,\mathcal{A}_u (s;t_2,a)\right) - \partial_a g_u \left(s,\mathcal{A}_u (s;t_2,a)\right) \right) {\mathinner{\mathrm{d}{s}}}.$$ Clearly, the knowledge of the maps $\mathcal{A}_u$, $\mathcal{T}_u$ and $\psi_u$ does not require knowledge of the solution to  but relies only on the solution to the ordinary differential equation . [Lemma \[lem:base\]]{} The proof is by induction on $n$. If $n=1$, then $f (\eta) = a + b\, \eta$ and the proof follows by basic calculus. Let now $n>1$. Assume that $\bar \eta \in {{\mathbb{R}}}^{n+1}$ is a point of strict local maximum or minimum for the multiaffine function $f \colon {{\mathbb{R}}}^{n+1} \to {{\mathbb{R}}}$. Then, one can write $$f (\eta_1, \ldots, \eta_{n+1}) = a (\eta_1, \ldots, \eta_n) + b(\eta_1, \ldots, \eta_n) \, (\eta_{n+1}-\bar\eta_{n+1})$$ for suitable multiaffine functions $a,b \colon {{\mathbb{R}}}^n \to {{\mathbb{R}}}$. Since $a (\eta_1, \ldots, \eta_n) = f (\eta_1, \ldots, \eta_n, \bar\eta_{n+1})$ has a point of strict local maximum or minimum at $(\bar \eta_1, \ldots, \bar \eta_n)$, by the inductive assumption the map $a$ is constant. Since, the map $\eta_{n+1} \to b(\bar \eta_1, \ldots, \bar \eta_n) \, (\eta_{n+1}-\bar\eta_{n+1})$ may not attain a strict local maximum or minimum at $\eta_{n+1} = \bar\eta_{n+1}$, the proof is completed. [Theorem \[thm:main\_eta\] and Proof of Theorem \[thm:main\_theta\]]{} Fix an arbitrary time $\tau \geq 0$. Lengthy but elementary computations based on Figure \[fig:proof\] show that the $J$ component of the solution to  admits the following representation, for $t \in [\tau, \tau + \bar a]$ and where we used –– for $u=J,S,R$: $$\label{eq:J} \!\!\!\!\! J (t,a) = \left\{ \begin{array}{@{}l@{\qquad}r@{\;}c@{\;}l@{}} J \left(\tau, \mathcal{A}_J (\tau;t,a) \right) \; \psi_J (\tau, t, a) & a & \in & [\mathcal{A}_J (t;\tau,0), \bar a] \\ \\ \frac{1}{g_J\left(\mathcal{T}_J (0;t,a)\right)} \int_{\bar a}^{\mathcal{A}_R (\mathcal{T}_J (0;t,a); \tau, a))} w (\alpha) \\ \qquad \times R \left( \tau, \mathcal{A}_R(\tau,\mathcal{T}_J (0;t,a); \tau, \alpha) \right) \; \psi_R (\tau,\mathcal{T}_J (0;t,a); \tau, \alpha) {\mathinner{\mathrm{d}{\alpha}}} \\ \quad + \frac{1}{g_J\left(\mathcal{T}_J (0;t,a)\right)} \int_{\mathcal{A}_R (\mathcal{T}_J (0;t,a); \tau, a))}^{+\infty} w (\alpha) & a & \in & [0, \mathcal{A}_J (t;\tau,0)] \\ \qquad \times \left( 1 - \eta \left( \mathcal{T}_R (\bar a; \mathcal{T}_J (0;t,a),\alpha) \right) \right) \, \frac{g_J \left( \mathcal{T}_R (\bar a; \mathcal{T}_J (0;t,a),\alpha), \bar a \right)} {g_R \left( \mathcal{T}_R (\bar a; \mathcal{T}_J (0;t,a),\alpha), \bar a \right)} \\ \qquad \times J \left( \tau, \mathcal{A}_J \left( \tau; \mathcal{T}_R (\bar a; \mathcal{T}_J (0;t,a),\alpha), \bar a \right) \right) \\ \qquad \times \psi_J\left( \tau, \mathcal{T}_R (\bar a;\mathcal{T}_J (0;t,a),\alpha), \bar a \right) \\ \qquad \times \psi_R\left( \mathcal{T}_R (\bar a; \mathcal{T}_J (0;t,a), \alpha), t, \alpha \right) {\mathinner{\mathrm{d}{\alpha}}}. \end{array} \right. \!\!\!\!\!$$ The $R$ population is given by $$\label{eq:R} R (t,a) = \left\{ \begin{array}{l@{\qquad}r@{\;}c@{\;}l} R \left(\tau, \mathcal{A}_R(\tau,t,a)\right) \; \psi_R (\tau,t,a) & a & \geq & \mathcal{A}_R (t;\tau, \bar a) \\ \\ \left(1-\eta \left(\mathcal{T}_R (\bar a; t,a) \right) \right) \, \frac{g_J \left(\mathcal{T}_R (\bar a; t,a), \bar a\right)} {g_R \left(\mathcal{T}_R (\bar a; t,a), \bar a\right)} \\ \quad \times J \left(\tau, \mathcal{A}_J\left(\tau; \mathcal{T}_R (\bar a; t,a), \bar a\right) \right) & a & \in & [\bar a, \mathcal{A}_R (t; \tau, \bar a)] \\ \quad \times \psi_J\left(\tau, \mathcal{T}_R (\bar a;t,a), \bar a\right) \; \psi_R\left(\mathcal{T}_R (\bar a; t, a), t, a\right) \end{array} \right.$$ and, finally, the $S$ population for $a \in [\bar a, \bar a_1]$ is $$\label{eq:S} S (t,a) = \left\{ \begin{array}{l@{\qquad}r@{\;}c@{\;}l} S \left(\tau, \mathcal{A}_S(\tau,t,a)\right) \; \psi_S (\tau,t,a) & a & \geq & \mathcal{A}_S (t;\tau, \bar a) \\ \\ \eta \left(\mathcal{T}_S (\bar a; t,a) ; \right) \, \frac{g_J \left(\mathcal{T}_S (\bar a; t,a), \bar a\right)} {g_S \left(\mathcal{T}_S (\bar a; t,a), \bar a\right)} \\ \quad \times J \left(\tau, \mathcal{A}_J\left(\tau; \mathcal{T}_S (\bar a; t,a), \bar a\right) \right) & a & \in & [\bar a, \mathcal{A}_S (t; \tau, \bar a)] \\ \quad \times \psi_J\left(\tau, \mathcal{T}_S (\bar a;t,a), \bar a\right) \; \psi_S\left(\mathcal{T}_S (\bar a; t, a), t, a\right). \end{array} \right.$$ The expression of $S$ for $a \geq \bar a_1$ directly follows. Note that the right hand side in the explicit expression above depends only on the values attained by $(J,S,R)$ at $t = \tau$. \[fig:proof\] Fix now an index $k$. Clearly, $J (t,a)$, $S (t,a)$ and $R (t,a)$ are all independent of $\eta_k$ for $t \in [0, \tau_{k-1}]$. Consider the time interval $[\tau_{k-1}, \mathcal{T}_J (\bar a; \tau_{k-1}, 0)]$. By , see also Figure \[fig:proof\], it is clear that $J (t,a)$ is independent of $\eta_k$ for $$(t, a) \in \left\{ (\tau,\alpha) \colon \tau \in [\tau_{k-1}, \mathcal{T}_J (\bar a; \tau_{k-1}, 0)] \mbox{ and } \alpha \geq \mathcal{A}_J (\tau; \tau_{k-1}, 0) \right\} \,.$$ Clearly, $S (t,a)$, respectively $R (t,a)$, is independent of $\eta_k$ whenever $a \geq \mathcal{A}_S (t; \tau_{k-1}, \bar a)$, respectively $a \geq \mathcal{A}_R (t; \tau_{k-1}, \bar a)$. On the strip $\left\{(t,a) \colon t \in [\mathcal{T}_S (a; \tau_{k-1}, \bar a), \mathcal{T}_S (t; \tau_k, \bar a)] \mbox{ and } a \geq \bar a \right\}$, the quantity $S (t,a)$ is linear in $\eta_k$ by . Similarly, on $\left\{(t,a) \colon t \in [\mathcal{T}_R (a; \tau_{k-1}, \bar a), \mathcal{T}_R (t; \tau_k, \bar a)] \mbox{ and } a \geq \bar a \right\}$, by  $R (t,a)$ is linear in $(1-\eta_k)$. Again by  and , $S (t,a)$, respectively $R (t,a)$, is independent of $\eta_k$ for $t \in [\mathcal{T}_S (a; \tau_k, \bar a), \mathcal{T}_S (a; \mathcal{T}_J (\bar a; \tau_{k-1}, 0))]$ and $a \geq \bar a$, respectively $t \in [\mathcal{T}_R (a; \tau_k, \bar a),$ $ \mathcal{T}_R (a; \mathcal{T}_J (\bar a; \tau_{k-1}, 0))]$ and $a \geq \bar a$. Finally, the above considerations and  ensure that $J (t,a)$ is affine in $\eta_k$ for $t \in [\mathcal{T}_J (a; \tau_{k-1}, 0), \mathcal{T}_J (a; \tau_k, 0)]$ and $a \in [0, \bar a]$. The proof is thus completed for $t \in [\tau_{k-1}, \mathcal{T}_j (\bar a, \tau_{k-1}, 0)]$. On the basis of ––, a straightforward iterative procedure allows to complete the proof related to the dependence of $(J,S,T) (t,a)$ on $\eta_k$. The proof concerning the dependence of $S (t,a)$ on ${\vartheta}_i^k$ directly follows from . [Corollary \[cor:2Gen\]]{} Apply Corollary \[cor:1\] with $T = T_\ell$, use the assumption  and Lemma \[lem:base\] to complete the proof. **Acknowledgment:** This work was partially supported by the 2015–INDAM–GNAMPA project *Balance Laws in the Modeling of Physical, Biological and Industrial Processes*. [10]{} A. S. Ackleh and K. Deng. A nonautonomous juvenile-adult model: well-posedness and long-time behavior via a comparison principle. , 69(6):1644–1661, 2009. A. S. Ackleh, K. Deng, and X. Yang. Sensitivity analysis for a structured juvenile–adult model. , 64(3):190–200, 2012. . Calsina and J. Salda[ñ]{}a. Global dynamics and optimal life history of a structured population model. , 59(5):1667–1685, 1999. . Calsina and J. Salda[ñ]{}a. Basic theory for a class of models of hierarchically structured population dynamics with distributed states in the recruitment. , 16(10):1695–1722, 2006. R. M. Colombo and M. Garavello. Stability and optimization in structured population models on graphs. , 12(2):311–335, 2015. R. M. Colombo and M. Garavello. Control of biological resources on graphs. , 2016. J. M. Cushing. A juvenile-adult model with periodic vital rates. , 53(4):520–539, 2006. O. Diekmann, M. Gyllenberg, J. A. J. Metz, and H. R. Thieme. On the formulation and analysis of general deterministic structured population models. [I]{}. [L]{}inear theory. , 36(4):349–388, 1998. J. Z. Farkas and T. Hagen. Stability and regularity results for a size-structured population model. , 328(1):119–136, 2007. J. H. Gallier. . Morgan Kaufmann series in computer graphics and geometric modeling. Morgan Kaufmann Publishers, free web version edition, 2015. M. Garavello. Optimal control in renewable resources modeling. , 47(1):347–357, 2016. B. Perthame. . Frontiers in Mathematics. Birkhauser Verlag, Basel, 2007. G. F. Webb. , volume 89 of [*Monographs and Textbooks in Pure and Applied Mathematics*]{}. Marcel Dekker, Inc., New York, 1985.

--- abstract: 'We study the non-Gaussian features in single-field slow-roll inflationary scenario where inflation is preceded by a radiation era. In such a scenario both bispectrum and trispectrum non-Gaussianities are enhanced. Interestingly, the trispectrum in this scenario does not depend up on the slow-roll parameters and thus $\tau_{NL}$ is larger than $f_{NL}$ which can be a signature of such a pre-inflationary radiation era.' address: 'Tata Institute of Fundamental Research, Mumbai 400005, India' author: - Suratna Das title: 'Effects of Radiation on Primordial Non-Gaussianity' --- Introduction ============ Study of non-Gaussian features in primordial perturbations generated during inflation has become a subject of great importance as the precise determination of these primordial non-Gaussianities can quantify the dynamics of the early universe [@matarrese]. In generic single-field slow-roll inflationary scenario the preferred initial vacuum chosen for the inflaton perturbations is the Bunch-Davies vacuum. It is shown in [@prl1], that if inflation is preceded by a radiation era then the inflaton fluctuations will have an initial thermal distribution where the initial vacuum will depart from the standard Bunch-Davies one. The presence of pre-inflationary radiation era enhances the power spectrum of scalar modes by an extra temperature depended factor $\coth(k/2T)$. The enhanced power spectrum is in accordance with the observations if the comoving temperature $T$ of the primordial perturbations is less than $10^{-3}$ Mpc$^{-1}$ [@prl1]. In this talk we will show that presence of pre-inflationary radiation era not only enhances the power spectrum but also generates large bispectrum and trispectrum and these non-Gaussianities will carry signatures of such pre-inflationary radiation era [@suratna] which will be discussed in detail. Bispectrum and Trispectrum in single field slow-roll inflation ============================================================== The derivations shown in this talk are done in [*spatially flat gauge*]{}. This gauge is preferred in the derivations as in this gauge the comoving curvature perturbation ${\mathcal R}(t,{\mathbf x})$ is proportional to the inflaton fluctuations $\delta\phi(t,{\mathbf x})$ as $$\begin{aligned} {\mathcal R}(t,{\mathbf x})=\frac{H}{\dot{\phi}}\delta\phi(t,{\mathbf x}),\end{aligned}$$ where $H$ is the Hubble parameter and the overdot represents derivative w.r.t. cosmic time $t$. Thus, in this gauge, the comoving curvature power spectrum, i.e. the two-point correlation function of ${\mathcal R}(t,{\mathbf k})$ in Fourier space, is directly related to inflaton’s power spectrum as $$\begin{aligned} {\cal P}_{\cal R}(k)=\frac{k^3}{2\pi^2}\langle{\mathcal R}(k){\mathcal R}(k)\rangle\longleftrightarrow\left(\frac{H}{\dot{\phi}}\right)^2\langle\delta\phi(k)\delta\phi(k)\rangle.\end{aligned}$$ The comoving curvature power spectrum is measured through observations of the $TT$ anisotropy spectrum of CMBR which is nearly scale-invariant. Bispectrum, the non-vanishing three-point correlation function of primordial fluctuations, is the lowest order departure from Gaussianity of those primordial perturbations. The non-Gaussianity arising from bispectrum is quantified by a non-linear parameter, $f_{NL}$, which is constrained by several experiments as : (i) the WMAP 5yr data yields $-151 < f_{NL}^{eq}<253$ ($95 \%$ CL) [@Komatsu], (ii) the PLANCK mission is sensitive to probe bispectrum upto $f_{NL} \sim 5$ [@spergel] and (iii) in future experiments if the primordial non-Gaussianities imprinted in 21 cm background is measured then $f_{NL}< 0.1$ can be probed [@cooray; @cooray-prd]. In a free theory, as the primordial perturbations are Gaussian in nature, the three-point correlation function vanishes yielding no non-Gaussianity. It is shown in [@raghu] that self-interactions of inflaton field of the kind $V(\phi)=\lambda\phi^3$ generates non-vanishing bispectrum proportional to $\lambda/H$ but the non-Gaussianity is too small $(\sim\mathcal{O}(10^{-7}))$ to be probed by any existing or future experiments. On the other hand, in a generic single-field slow-roll inflationary model non-linearities in the evolution of primordial perturbations ${\mathcal R}(t,{\mathbf k})$ can also generate primordial non-Gaussianities in CMBR. In the non-linear limit one can write $$\begin{aligned} {\mathcal R}_{NL}(t,{\mathbf x})=\frac{H}{\dot{\phi}}\delta\phi_L(t,{\mathbf x})+\frac12\frac{\partial}{\partial\phi}\left(\frac{H}{\dot{\phi}}\right)\delta\phi_L^2(t,{\mathbf x})+{\mathcal O}(\delta\phi_L^3), \label{non-linear}\end{aligned}$$ which yields a non-vanishing three-point correlation function of ${\mathcal R}_{NL}$ in terms of four-point correlation function of inflaton perturbations $\delta\phi_L$ as $$\begin{aligned} \langle{\mathcal R}_{NL}{\mathcal R}_{NL}{\mathcal R}_{NL}\rangle\simeq\left(\frac{H}{\dot{\phi}}\right)^2\frac12\frac{\partial}{\partial\phi}\left(\frac{H}{\dot{\phi}}\right)\langle\delta\phi_L\delta\phi_L\delta\phi_L^2\rangle, \label{three-pt}\end{aligned}$$ even when the initial perturbations $\delta\phi_L$ are Gaussian in nature. The four-point correlation function on the R.H.S. can be written in terms of product of two two-point correlation functions and defining $f_{NL}$ as $$\begin{aligned} \left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3)\right\rangle=(2\pi)^{-\frac32}\delta^3({\mathbf k_1}+{\mathbf k_2}+{\mathbf k_3})\frac65 f_{NL}\left(\frac{P_{\cal R}(k_1)}{k_1^3}\frac{P_{\cal R}(k_2)}{k_2^3}+2\,\,{\rm perms.}\right),\end{aligned}$$ one can show that the non-linear parameter is of the order of slow-roll parameters $f_{NL}=\frac56(\delta-\epsilon)$ [@wands], which is also too small $\left(\mathcal{O}(10^{-2})\right)$ to be detected by any present or forthcoming experiments. The delta-function in the above equation ensures that the three momenta form a triangle and $f_{NL}$ is determined in several such triangle configurations, some of them which we will consider in this talk are : (i) [ *Squeezed configuration*]{} $(|{\mathbf k}_1|\approx|{\mathbf k}_2|\approx k\gg|{\mathbf k}_3|)$, (ii) [*Equilateral configuration*]{} $(|{\mathbf k}_1|=|{\mathbf k}_2|=|{\mathbf k}_3|=k)$ and (iii) [*Folded configuration*]{} $(|{\mathbf k}_1|=|{\mathbf k}_3|=\frac12|{\mathbf k}_2|=k)$. The connected part of four-point correlation function of primordial fluctuations is called the trispectrum $$\begin{aligned} \left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3){\mathcal R}({\mathbf k}_4)\right\rangle_c&\equiv&\left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3){\mathcal R}({\mathbf k}_4)\right\rangle-\left(\left\langle{\mathcal R}_{L}({\mathbf k}_1){\mathcal R}_{L}({\mathbf k}_2)\right\rangle\left\langle{\mathcal R}_{L}({\mathbf k}_3){\mathcal R}_{L}({\mathbf k}_4)\right\rangle\right.\nonumber\\ &&+2\,\,{\rm perm}\left.\right).\end{aligned}$$ The non-linear parameter $\tau_{NL}$ quantifies the non-Gaussianity arising from trispectrum and it is constrained by observations as (i) WMAP constraints trispectrum as $\left|\tau_{NL}\right|<10^8$ [@WMAP-tri], (ii) PLANCK is expected to reach the sensitivity upto $\left|\tau_{NL}\right|\sim 560$ [@PLANCK-tri] and (iii) future 21cm experiments can probe trispectrum up to the level $\tau_{NL} \sim 10$ [@cooray-prd]. A free-scalar theory yields vanishing trispectrum like vanishing bispectrum. But non-linear evolution of $\mathcal{R}$ as given in Eq. (\[non-linear\]) yields a trispectrum where $\tau_{NL}=\left(\frac65 f_{NL}\right)^2$ [@wands], which being proportional to the square of slow-roll parameters is too small to be detected by any present or future experiments. It is to be noted that the generic single-field slow-roll inflation predicts a trispectrum which smaller than the bispectrum by orders of magnitude. Inflation with prior radiation era and enhanced non-Gaussianity =============================================================== If inflation is preceded by a radiation era then the inflaton field will have an initial thermal distribution where the thermal vacuum $|\Omega\rangle$ will have finite occupation as $N_k|\Omega\rangle=n_k|\Omega\rangle$ (the number operator $N_k\equiv a^\dagger_{\mathbf k}a_{\mathbf k}$). Also there will be a probability of the system to be in an energy state $\varepsilon_k\equiv n_k k$ as $$\begin{aligned} p(\varepsilon_k)\equiv\frac{e^{-\beta n_k k}}{\sum_{n_k}e^{-\beta n_k k}}=\frac{e^{-\beta n_k k}}{z}.\end{aligned}$$ Due to this probability distribution the correlation functions have to be thermal averaged. Taking into account the initial thermal distributions of the primordial fluctuations and the probability distribution due to pre-inflationary radiation era the thermal averaged inflaton’s power spectrum will have an enhancement factor $1+2f_B(k)$ where $f_B(k)$ is the distribution function of primordial perturbations. For inflaton (scalar) perturbations the distribution function will be Bose-Einstein distribution function $\left(f_B(k)\equiv\frac{1}{e^{\beta k}-1}\right)$ and thus the enhancement factor will be $1+2f_B(k)=\coth(\beta k/2)$ where $\beta\equiv\frac1T$ [@prl1; @suratna]. This enhanced power spectrum is in accordance with the observations when $T<10^{-3}$ Mpc$^{-1}$ [@prl1]. As the two-point correlation function is thermal averaged due to the effects of pre-inflationary radiation era, the other higher-point correlation functions have also to be thermal averaged in a similar way. For three-point correlation function this will be $$\begin{aligned} \langle{\mathcal R}_{NL}{\mathcal R}_{NL}{\mathcal R}_{NL}\rangle_{\beta}\simeq\left(\frac{H}{\dot{\phi}}\right)^2\frac12\frac{\partial}{\partial\phi}\left(\frac{H}{\dot{\phi}}\right)\langle\delta\phi\delta\phi\delta\phi^2\rangle_{\beta},\end{aligned}$$ but now the probability of occupancy of a state with four energies $\epsilon_r$ will be $$\begin{aligned} p(k_1,k_2, k_3,k_4)\equiv\frac{\prod_re^{-\beta n_{k_r}k_r}}{\prod_r\sum_{n_k}e^{-\beta n_{k_r}k_r}}.\end{aligned}$$ Due to thermal averaging $f_{NL}$ is enhanced. We compute the non-Gaussianity in this scenario arising from bispectrum in different triangle configurations : (i) [*Squeezed configuration*]{} : in this configuration the enhanced non-Gaussianity is $f_{NL}^{\rm th}=f_{NL}\times2\left(1+3.72\coth\left(\frac{\beta k}{2}\right)\right)$ where $f_{NL}$ is enhanced by a factor of 64.82, (ii) [*Equilateral configuration*]{} : in this configuration the enhanced non-Gaussianity is $f_{NL}^{\rm th}=f_{NL}\times\left(3+\frac{5}{4\sinh^2\left(\frac{\beta k}{2}\right)}\right)$ where $f_{NL}$ is enhanced by a factor of 90.85, (iii) [*Folded configuration*]{} : in this configuration the enhanced non-Gaussianity is $f_{NL}^{\rm th}=f_{NL}\times\left(3+\frac{1}{\sinh^2\left(\frac{\beta k}{2}\right)}\right)$ where $f_{NL}$ is enhanced by a factor of 73.28. It can be seen that $f_{NL}$, which is enhanced due to effects of pre-inflationary radiation era, is within the sensitivity of future 21-cm experiments where primordial non-Gaussianities can be detected [@cooray; @cooray-prd]. It is also to be noted that the maximum non-Gaussianity can arise in the Equilateral configuration when inflation is preceded by a radiation era. In a later work [@parker] similar analysis is done when perturbations are already present in the initial vacuum. It is found in [@parker] too that initial presence of quanta in the vacuum can significantly enhance non-Gaussianity arising from bispectrum which is in agreement with the results presented here and in [@suratna]. It is very interesting to note at this point that due to thermal averaging the four-point correlation function is not equal to the product of two two-point correlation functions which can yield a non-vanishing connected part as $$\begin{aligned} \left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3){\mathcal R}({\mathbf k}_4)\right\rangle_c&\neq&\left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3){\mathcal R}({\mathbf k}_4)\right\rangle_\beta\nonumber \\ &-&\left(\left\langle{\mathcal R}_{L}({\mathbf k}_1){\mathcal R}_{L}({\mathbf k}_2)\right\rangle_\beta\left\langle{\mathcal R}_{L}({\mathbf k}_3){\mathcal R}_{L}({\mathbf k}_4)\right\rangle_\beta+2{\rm perm}\right).\end{aligned}$$ Thus defining the trispectrum in such a situation as $$\begin{aligned} \left\langle{\mathcal R}({\mathbf k}_1){\mathcal R}({\mathbf k}_2){\mathcal R}({\mathbf k}_3){\mathcal R}({\mathbf k}_4)\right\rangle_c=\tau_{NL}\left[\frac{P_{\mathcal R}(k_1)}{k_1^3}\frac{P_{\mathcal R}(k_2)}{k_2^3}\delta^3({\mathbf k_1}+{\mathbf k_3})\delta^3({\mathbf k_2}+{\mathbf k_4})+2\,\,{\rm perm.}\right],\end{aligned}$$ we see that, as the linear perturbations can generate non-vanishing connected part due to thermal averaging, the non-linear parameter $|\tau_{NL}|$ will not depend up on slow-roll parameters and can be as large as 42.58 [@suratna] which is within the detection range of future 21-cm background anisotropy experiments [@cooray-prd]. Hence, we see that the presence of pre-inflationary radiation era yields larger trispectrum non-Gaussianity than bispectrum. Conclusion ========== The talk was focused on non-Gaussian features in a single-field slow-roll inflationary model where inflation is preceded by a radiation era. In a generic single-field slow-roll model of super-cool inflation non-linear evolution of primordial fluctuations generate bispectrum non-Gaussianity which is proportional to the slow-roll parameters [@wands] and thus too small to be detected by any present or future experiments. Non-linear evolution of primordial fluctuations also generates trispectrum non-Gaussianity where $\tau_{NL}$ is proportional to the square of slow-roll parameters [@wands]. Thus, this generic inflationary scenario predicts trispectrum non-Gaussianity which is much smaller than the bispectrum non-Gaussianity. We showed that if such a generic inflationary scenario is preceded by a radiation era it can yield large bispectrum and trispectrum non-Gaussianities [@suratna] which are within the range of detection of future 21-cm background anisotropy experiments [@cooray; @cooray-prd]. Due to presence of pre-inflationary radiation era the initial vacuum will contain thermal fluctuations and also the energy states will have a probability distribution. Accordingly, the thermal averaged three-point correlation function generates large non-Gaussianity where the enhancement of $f_{NL}$ over the generic scenario is largest in the equilateral configuration. An interesting situation arises in the case of trispectrum as the thermal averaged four-point correlation function is not equal to the product of two thermal-averaged two-point correlation functions. Thus the linear primordial perturbations can generate a non-vanishing connected part of four-point correlation function due to thermal averaging and $\tau_{NL}$ in such a case will not depend up on the slow-roll parameters. We compute that in such a scenario $|\tau_{NL}|$ can be as large as 43 [@suratna]. Thus a significant signature of such pre-inflationary radiation era is that it yields larger trispectrum than bispectrum. This signature can distinguish between an inflationary scenario preceded by a radiation era and the generic scenario of single-field slow-roll super-cool inflation. References {#references .unnumbered} ========== [99]{} Bartolo N, Matarrese S and Riotto A 2005 [*Nucl. Phys. Proc. Suppl. *]{} [**148**]{} 56 Bhattacharya K, Mohanty S and Rangarajan R 2006 [*Phys. Rev. Lett. *]{} [**96**]{} 121302 \[arXiv:hep-ph/0508070\] Das S and Mohanty S 2009 [*Phys. Rev. *]{} D [**80**]{} 123537 \[arXiv:0908.2305 \[astro-ph.CO\]\] Komatsu E [*et al.*]{} \[WMAP Collaboration\] 2009 [*Astrophys. J. Suppl. * ]{}[**180**]{} 330 \[arXiv:0803.0547 \[astro-ph\]\] Komatsu E and Spergel D N 2001 [*Phys. Rev. * ]{}D [**63**]{} 063002 \[arXiv:astro-ph/0005036\] Cooray A 2006 [*Phys. Rev. Lett. *]{} [**97**]{} 261301 \[arXiv:astro-ph/0610257\] Cooray A, Li C and Melchiorri A 2008 [*Phys. Rev. * ]{}D [**77**]{} 103506 \[arXiv:0801.3463 \[astro-ph\]\] Falk T, Rangarajan R and Srednicki M 1993 [*Astrophys. J. *]{} [**403**]{} L1 \[arXiv:astro-ph/9208001\] Byrnes C T, Sasaki M and Wands D 2006 [*Phys. Rev. *]{} D [**74**]{} 123519 \[arXiv:astro-ph/0611075\] Boubekeur L and Lyth D H 2006 [*Phys. Rev. *]{} D [**73**]{} 021301 \[arXiv:astro-ph/0504046\] Kogo N and Komatsu E 2006 [*Phys. Rev. *]{} D [**73**]{} 083007 \[arXiv:astro-ph/0602099\] Agullo I and Parker L 2011 [*Phys. Rev. *]{} D [**83**]{} 063526 \[arXiv:1010.5766 \[astro-ph.CO\]\]

--- abstract: 'I present a density-matrix renormalization-group (DMRG) method for calculating dynamical properties and excited states in low-dimensional lattice quantum many-body systems. The method is based on an exact variational principle for dynamical correlation functions and the excited states contributing to them. This dynamical DMRG is an alternate formulation of the correction vector DMRG but is both simpler and more accurate. The finite-size scaling of spectral functions is discussed and a method for analyzing the scaling of dense spectra is described. The key idea of the method is a size-dependent broadening of the spectrum. The dynamical DMRG and the finite-size scaling analysis are demonstrated on the optical conductivity of the one-dimensional Peierls-Hubbard model. Comparisons with analytical results show that the spectral functions of infinite systems can be reproduced almost exactly with these techniques. The optical conductivity of the Mott-Peierls insulator is investigated and it is shown that its spectrum is qualitatively different from the simple spectra observed in Peierls (band) insulators and one-dimensional Mott-Hubbard insulators.' author: - Eric Jeckelmann title: 'Dynamical density-matrix renormalization-group method' --- Introduction \[sec:intro\] ========================== The density-matrix renormalization group (DMRG) [@steve; @dmrgbook] is a very successful numerical methods for calculating static properties of ground states and low-lying eigenstates in quantum many-body systems. For low-dimensional strongly correlated systems DMRG is as accurate as exact diagonalization techniques but can be used to study much larger systems than with those techniques (currently, up to $\sim 10^3$ sites). Using a finite-size scaling analysis it is thus possible to determine the static properties of a system in the thermodynamic limit with great accuracy. The calculation of dynamical properties and higher energy excitations with DMRG has proved to be more difficult. Several approaches have been proposed but calculations have been carried out successfully for few problems only. The simplest of these methods is the Lanczos DMRG. [@karen; @till] In practice, this method gives accurate results for the first few moments of a dynamical spectrum. Therefore, it works well for simple discrete spectra made of a few peaks but it usually fails for more complicated spectra. An alternate method for calculating dynamical properties is the correction vector DMRG. [@pati; @till] Contrary to the Lanczos DMRG, this method can describe complex or dense spectra accurately. Nevertheless, there have been few applications [@till2; @brune] of correction vector DMRG until now because this method is more difficult and requires significantly more computer resources than the usual DMRG method for calculating static properties at low energy. In this paper I describe a simple and efficient method, called the dynamical DMRG (DDMRG), for calculating dynamical properties and excited states with DMRG. This approach is based on a variational principle for dynamical correlation functions and the related excited states. The variational principle is essentially an elegant formulation of the correction vector technique. Because of the variational formulation, however, the DDMRG method is easier to use and significantly more accurate than the correction vector DMRG method. While the spectrum of a finite system is necessarily discrete, continuous excitation bands are often found in the thermodynamic limit. It is possible to broaden finite-size spectra to simulate the continuum of an infinite-system spectrum. Usually, the broadening is arbitrarily large and no systematic quantitative analysis of finite-size effects is performed for the spectrum. Here I show that dynamical properties of infinite systems can be obtained reliably using an appropriate finite-size scaling analysis. The key to the analysis is the use of a broadening which scales systematically with the system size. The DDMRG method and the finite-size scaling technique for dynamical spectrum have already been successfully used to investigate the optical properties of one-dimensional Mott insulators. [@eric; @fabian] Here I apply these techniques to the calculation of the optical conductivity in the one-dimensional Peierls-Hubbard model of conjugated polymers. [@dionys] Much effort has been devoted to understanding the optical properties of these materials. In particular, the optical conductivity of the Peierls-Hubbard model has been studied extensively and analytical results have been obtained for various special cases, such as the Mott-Hubbard insulator and the Peierls (band) insulator limits. Leaving out these special limits the Peierls-Hubbard model describes a Mott-Peierls insulator and its optical properties are still poorly understood. In this paper I show that DDMRG can accurately reproduce the known analytical results for the optical spectrum in the thermodynamic limit. Then I investigate the optical conductivity of a Mott-Peierls insulator using DDMRG and show that it displays specific features. The paper is organized as follows: The variational principle for dynamical correlation functions and related excited states is presented in the next section. I describe the dynamical DMRG method in Sec. \[sec:ddmrg\]. The finite-size scaling analysis is presented in Sec. \[sec:scaling\]. I report and discuss the results for the optical conductivity of the Peierls-Hubbard model in Sec. \[sec:model\]. Finally, I conclude in the last section. Variational principle \[sec:principle\] ======================================= The dynamic response of a quantum system to a time-dependent perturbation is often given by dynamical correlation functions such as $$G_{A}(\omega + i \eta) = - \frac{1}{\pi} \langle \psi_0| A^{\dag} \frac{1}{E_0+\omega + i \eta - H} A |\psi_0\rangle \; , \label{dynamCF}$$ where $H$ is the time-independent Hamiltonian of the system, $E_0$ and $|\psi_0 \rangle$ are its ground-state energy and wavefunction, $A$ is the quantum operator corresponding to the physical quantity which is analyzed, and $A^{\dag}$ is the Hermitian conjugate of $A$. A small real number $\eta > 0$ is used in the calculation to shift the poles of the correlation function into the complex plane. (I set $\hbar=1$ in sections \[sec:principle\] to \[sec:scaling\].) In general, we are interested in calculating the imaginary part of the correlation function $$\begin{aligned} I_{A}(\omega + i\eta) &=& \text{Im} \ G_{A}(\omega + i \eta) \\ & = & \frac{1}{\pi} \langle \psi_0 | A^{\dag} \frac{\eta}{(E_0+\omega -H)^2 + \eta^2} A |\psi_0 \rangle \; \nonumber \end{aligned}$$ in the limit $\eta \rightarrow 0$ $$I_{A}(\omega) = \lim_{\eta \rightarrow 0} I_{A}(\omega + i \eta) \; \geq 0 \;.$$ It should be noted that the spectrum $I_{A}(\omega + i \eta)$ for any finite $\eta > 0$ can be calculated from the spectrum $I_{A}(\omega)$ by convolution with a Lorentzian distribution $$I_{A}(\omega + i \eta) = C_{\eta}[I_{A}(\omega)] \; > 0 \; ,$$ where I use the notation $C_{\eta}[f(\omega)]$ to represent the convolution of a spectral function $f(\omega)$ with a Lorentzian distribution of width $\eta$ $$C_{\eta}[f(\omega)] = \int_{-\infty}^{+\infty} d\omega' f(\omega') \frac{1}{\pi}\frac{\eta}{(\omega-\omega')^2+\eta^2} \; .$$ The moments of the spectrum $I_{A}(\omega)$ fulfill sum rules such as $$\begin{aligned} \int_{-\infty}^{+\infty} d\omega I_{A}(\omega) & = & \langle \psi_0 | A^{\dag} A | \psi_0 \rangle \; , \nonumber \\ % \int_{-\infty}^{+\infty} d\omega I_{A}(\omega) \omega & = & \langle \psi_0 | A^{\dag} [H,A] | \psi_0 \rangle \; , \label{sumrules} \\ % \int_{-\infty}^{+\infty} d\omega I_{A}(\omega) \omega^2 & = & \langle \psi_0 | [A^{\dag},H] [H,A] | \psi_0 \rangle \; , \nonumber\end{aligned}$$ where $[A,B] = AB-BA$. A dynamical correlation function (\[dynamCF\]) can be calculated using the correction vector method. The correction vector associated with $G_A(\omega + i \eta)$ is defined by $$|\psi_A(\omega + i \eta) \rangle = \frac{1}{E_0+\omega + i \eta - H} | A \rangle \; ,$$ where $| A \rangle = A | \psi _0 \rangle$. If the correction vector is known, the dynamical correlation function can be calculated directly $$G_A(\omega + i \eta) = -\frac{1}{\pi} \langle A|\psi_A(\omega + i \eta) \rangle \; . \label{dynamCF2}$$ To calculate a correction vector one first solves an inhomogeneous linear equation $$\left [ (E_0+\omega-H)^2+\eta^2 \right ] | \psi \rangle = -\eta | A \rangle \; , \label{CVequation1}$$ which always has a unique solution $| \psi \rangle = | Y_A(\omega + i \eta) \rangle$ for $\eta \neq 0$. The correction vector is then given by $$|\psi_A(\omega + i \eta) \rangle = | X_A(\omega + i \eta) \rangle + i | Y_A(\omega + i \eta) \rangle$$ with $$| X_A(\omega + i \eta) \rangle = \frac{H-E_0-\omega}{\eta} | Y_A(\omega + i \eta) \rangle \; . \label{CVequation2}$$ One should note that the states $| X_A(\omega + i \eta) \rangle$ and $| Y_A(\omega + i \eta) \rangle$ are complex if the state $|A\rangle$ is not real, but they always determine the real part and imaginary part of the dynamical correlation function $G_A(\omega + i \eta)$, respectively, $$\begin{aligned} \text{Re} \ G_A(\omega + i \eta) & = & -\frac{1}{\pi} \langle A|X_A(\omega + i \eta) \rangle \; , \label{ReDynamCF} \\ \text{Im} \ G_A(\omega + i \eta) & = & -\frac{1}{\pi} \langle A|Y_A(\omega + i \eta) \rangle \; . \label{ImDynamCF}\end{aligned}$$ The derivatives of the real and imaginary parts can also be calculated from these states $$\begin{aligned} \frac{d}{d\omega} \text{Re} G_A(\omega + i \eta) & = & \frac{1}{\pi} [ \langle X_A(\omega + i\eta)|X_A(\omega + i\eta) \rangle \nonumber \\ & &-\langle Y_A(\omega + i\eta)|Y_A(\omega + i\eta) \rangle] \label{derivatives}\\ \frac{d}{d\omega} \text{Im} G_A(\omega + i \eta) & = & \frac{2}{\pi} \langle X_A(\omega + i\eta)|Y_A(\omega + i\eta) \rangle \; . \nonumber\end{aligned}$$ A well-established approach for solving an inhomogeneous linear equation (\[CVequation1\]) is to formulate it as a minimization problem. One considers the functional $$\begin{aligned} W_{A,\eta}(\omega, \psi) & = & \langle \psi | (E_0+\omega-H)^2+\eta^2 | \psi \rangle \nonumber \\ & & + \eta \langle A | \psi \rangle + \eta \langle \psi | A \rangle \; . \label{functional}\end{aligned}$$ For any $\eta \neq 0$ and a fixed frequency $\omega$ this functional has a well-defined and non-degenerate minimum for the quantum state which is solution of Eq. (\[CVequation1\]) $$| \psi_{\text{min}} \rangle = | Y_A(\omega + i \eta) \rangle \; .$$ It is easy to show that the value of the minimum is related to the imaginary part of the dynamical correlation function $$W_{A,\eta}(\omega, \psi_{\text{min}}) = -\pi\eta I_A(\omega + i \eta).$$ Therefore, the calculation of spectral functions can be formulated as a minimization problem. To determine $I_A(\omega + i \eta)$ at any frequency $\omega$ and for any $\eta > 0$, one minimizes the corresponding functional $W_{A,\eta}(\omega, \psi)$. Once this minimization has been carried out, the real part of the correlation function $G_A(\omega + i \eta)$ can be calculated using Eqs. (\[CVequation2\]) and (\[ReDynamCF\]) if necessary. This is the variational principle for dynamical correlation functions. It is clear that if we can calculate $| Y_A(\omega + i \eta) \rangle$ exactly, this variational formulation is completely equivalent to the correction vector method. However, if we can only calculate an approximate solution with an error of the order $\epsilon \ll 1$, $|\psi\rangle = | Y_A(\omega + i \eta) \rangle + \epsilon |\phi\rangle$ with $\langle \phi|\phi \rangle=1$, the variational formulation is more accurate. In the correction vector method the error in the spectrum $I_A(\omega + i \eta)$ calculated with Eq. (\[ImDynamCF\]) is also of the order of $\epsilon$. In the variational approach it is easy to show that the error in the value of the minimum $W_{A,\eta}(\omega, \psi_{\text{min}})$, and thus in $I_A(\omega + i \eta)$, is of the order of $\epsilon^2$. With both methods the error in the real part of $G_A(\omega + i \eta)$ is of the order of $\epsilon$. One can write the function $I_A(\omega)$ in the spectral form (or Lehmann representation) $$I_A(\omega) = \sum_{n} |\langle \psi_n|A|\psi_0\rangle|^2 \delta(\omega + E_0-E_n) \; ,$$ where $|\psi_0 \rangle$ is the ground state, $|\psi_n\rangle, n > 1, $ denotes the other eigenstates of $H$, and $E_0, E_n$ are their respective eigenenergies. Obviously, only the eigenstates with a finite matrix element $\langle \psi_n|A|\psi_0\rangle \neq 0$ contribute to the spectrum and here we are only interested in those excited states. In the correction vector method the excitation energies $E_n-E_0$ and the spectral weights $|\langle \psi_n|A|\psi_0\rangle|^2$ can be obtained from the poles of $G_A(\omega + i \eta)$. The corresponding wavefunctions $|\psi_n\rangle$ can be calculated by taking the $\eta \rightarrow 0$ limit of the correction vectors $$|\psi_n\rangle \propto \lim_{\eta \rightarrow 0} | Y_A(E_n-E_0 + i \eta) \rangle \;.$$ The excited states contributing to $G_A(\omega + i \eta)$ correspond to the local maxima of the spectrum $I_A(\omega + i \eta)$ for small enough $\eta > 0$. Therefore, they can also be obtained by minimization of the functional $W_{A,\eta}(\omega,\psi)$ with respect to both $\omega$ and $\psi$. The local minima of $W_{A,\eta}(\omega,\psi)$ are given by the conditions $$\begin{aligned} \omega_{\text{min}} + E_0 & = & \frac{\langle \psi_{\text{min}} | H | \psi_{\text{min}} \rangle} {\langle \psi_{\text{min}}| \psi_{\text{min}} \rangle} \; , \nonumber \\ % |\psi_{\text{min}}\rangle &=& | Y_A(\omega_{\text{min}} + i \eta) \rangle \; .\end{aligned}$$ In the limit $\eta \rightarrow 0$, $\omega_{\text{min}}+E_0$ tends to the energy $E_n$ of an eigenstate with finite spectral weight, $|\psi_{\text{min}}\rangle$ is equal to the corresponding eigenstate $|\psi_n\rangle$ up to a normalization constant, and $-W_{A,\eta}(\omega_{\text{min}},\psi_{\text{min}})$ tends to the spectral weight $|\langle \psi_n|A|\psi_0\rangle|^2$. This is the variational principle for excited states contributing to a dynamical correlation function $G_A(\omega +i\eta)$. Again this variational formulation is completely equivalent to the correction vector method if $|Y_A(\omega +i\eta)\rangle$ can be calculated exactly. In an approximate calculation, however, errors in the eigenenergies and spectral weights are of the order of $\epsilon$ with the correction vector method, while they are of the order of $\epsilon^2$ with the variational formulation. Dynamical DMRG \[sec:ddmrg\] ============================ DMRG is a numerical method for calculating the properties of lattice quantum many-body systems. It is described in detail in several publications (for instance, see Refs.  and ). DMRG can be considered as a variational approach. The system energy $$E(\psi)= \frac{\langle \psi |H|\psi \rangle}{\langle \psi |\psi \rangle} \label{energyfunc}$$ is minimized in a variational subspace (the DMRG basis) of the system Hilbert space to find the ground-state wavefunction $|\psi_0\rangle$ and energy $E_0 = E(\psi_0)$. If the ground-state wavefunction is calculated with an error of the order of $\epsilon \ll 1$ (i.e., $|\psi\rangle = |\psi_0\rangle + \epsilon |\phi\rangle$ with $\langle \phi | \phi \rangle = 1$), the energy obtained is an upper bound to the exact result and the error in the energy is of the order of $\epsilon^2$ (as in all variational approaches). In principle, the DMRG energy error is proportional to the weight of the density-matrix eigenstates discarded in the renormalization procedure. This discarded weight can be reduced by increasing the number $m$ of density-matrix eigenstates kept in the calculation, which corresponds to an increase of the variational subspace dimension. Therefore, the energy error systematically decreases with increasing $m$ in a DMRG calculation. The DMRG procedure used to minimize the energy functional (\[energyfunc\]) can also be used to minimize the functional $W_{A,\eta}(\omega,\psi)$ and thus to calculate the dynamical correlation function $G_A(\omega+i\eta)$. I call this approach the dynamical DMRG method. The minimization of the functional $W_{A,\eta}(\omega,\psi)$ is easily integrated into the usual DMRG algorithm. At every step of a DMRG sweep through the system lattice, a superblock representing the system is built and the following calculations are performed in the the superblock subspace: 1. The energy functional $E(\psi)$ is minimized using a standard iterative algorithm for the eigenvalue problem. This yields the ground state vector $|\psi_0\rangle$ and its energy $E_0$ in the superblock subspace. 2. The state $|A \rangle$ is calculated. 3. The functional $W_{A,\eta}(\omega,\psi)$ is minimized using an iterative minimization algorithm. This gives the first part of the correction vector $|Y_{A}(\omega+ i\eta)\rangle$ and the imaginary part $I_A(\omega + i \eta)$ of the dynamical correlation function. 4. The second part $|X_{A}(\omega+ i\eta)\rangle$ of the correction vector is calculated using Eq. (\[CVequation2\]). 5. The real part and the derivatives of the dynamical correlation function can be calculated from Eqs. (\[ReDynamCF\]) and (\[derivatives\]), respectively. 6. The four states $|\psi_0\rangle$, $|A \rangle$, $|Y_{A}(\omega+ i\eta)\rangle$, and $|X_{A}(\omega+ i\eta)\rangle$ are included as target in the density-matrix renormalization to build a new superblock at the next step. The robust finite-system DMRG algorithm must be used to perform several sweeps through a lattice of fixed size. Sweeps are repeated until the procedure has converged to the minimum of both functionals $E(\psi)$ and $W_{A,\eta}(\omega,\psi)$. To obtain the dynamical correlation function $G_{A}(\omega+ i\eta)$ over a range of frequencies, one has to repeat this calculation for several frequencies $\omega$. If the DDMRG calculations are performed independently, the computational effort is roughly proportional to the number of frequencies. It is also possible to carry out a DDMRG calculation for several frequencies simultaneously, including several states $|X_{A}(\omega+ i\eta)\rangle$ and $|Y_{A}(\omega+ i\eta)\rangle$ with different frequencies $\omega$ as target. The optimal number of different frequencies to be included in a single calculation depends strongly on the problem studied and the computer used. As calculations for different frequencies are essentially independent, it would be easy and very efficient to perform these calculations on a parallel computer. If one performs a DDMRG calculation for two close frequencies $\omega_1$ and $\omega_2$ simultaneously, it is possible to calculate the dynamical correlation function for additional frequencies $\omega$ between $\omega_1$ and $\omega_2$ without including the corresponding states $|X_{A}(\omega+ i\eta)\rangle$ and $|Y_{A}(\omega+ i\eta)\rangle$ as target in the density-matrix renormalization. This approach can significantly reduce the computer time necessary to determine the spectrum over a frequency range but the results obtained for $\omega \neq \omega_1, \omega_2$ are less accurate and not always reliable, as the DMRG basis is optimized for the frequencies $\omega_1$ and $\omega_2$ only. (A similar technique is the calculation of spectra with the Lanczos algorithm in the DMRG basis optimized for a pair of correction vectors, see Ref. .) Alternatively, between the frequencies for which $G_A(\omega + i\eta)$ is determined directly with DDMRG, we can calculate the dynamical correlation function by interpolation using the DDMRG data for the function and its derivative. If a complete spectrum $I_A(\omega + i\eta)$ has been obtained, it is possible to calculate the moments of the spectral distribution \[the left-hand-side of Eq. (\[sumrules\])\]. The first few moments \[the right-hand-side of Eq. (\[sumrules\])\] can be calculated accurately using the Lanczos DMRG method. [@karen; @till] This provides an independent check of DDMRG results. \[Note that only the first sum rule (\[sumrules\]) is satisfied exactly for $\eta > 0$.\] To calculate individual excited states contributing to the spectrum in a given frequency range $(\omega_1,\omega_2$), one includes a minimization of $W_{A,\eta}(\omega,\psi)$ with respect to $\omega$ ($\omega_1 < \omega < \omega_2$) in the third step of the DDMRG algorithm described above. In this case, $|Y_{A}(\omega_{\text{min}} + i\eta)\rangle$ and $|X_{A}(\omega_{\text{min}}+ i\eta)\rangle$ are included as target in the sixth step. The parameter $\eta$ must be much smaller than the distance $E_{n+1} - E_{n}$ between two successive eigenstates contributing to the dynamical correlation function or must decrease during the calculation until the desired accuracy is obtained. To make the procedure robust it is necessary to simultaneously target a second correction vector $|\psi_{A}(\omega+ i\eta)\rangle$ with a fixed frequency and a parameter $\eta$ of the order of the frequency range. Typically, I use $\omega=(\omega_1+\omega_2)/2$ and $\eta=(\omega_2-\omega_1)/4$. Because of the variational principle one naively expects that the DDMRG results for $I_A(\omega +i\eta)$ must converge monotonically from below to the exact result as the number $m$ of density-matrix eigenstates is increased. In practice, the convergence is not always regular because of two approximations made to calculate the functional $W_{A,\eta}(\omega,\psi)$ in a DMRG basis. First, the ground-state energy $E_0$ and the state $|A\rangle$ used in the definition (\[functional\]) of $W_{A,\eta}(\omega,\psi)$ are not known exactly but calculated with DMRG. If the number $m$ of density matrix eigenstates is increased, $E_0$ and $|A\rangle$ are modified (they become progressively more accurate) and the functional $W_{A,\eta}(\omega,\psi)$ is changed, which can affect its minimum arbitrarily. We also note that errors of the order of $\epsilon$ in $E_0$ or $|A\rangle$ result in errors of the same order in $I_A(\omega +i\eta)$. Therefore, to observe a regular convergence with increasing $m$ and to obtain accurate results for $I_A(\omega +i\eta)$, it is necessary in the first place to determine the ground state and the state $|A\rangle$ with great precision (and thus to include the state $|A\rangle$ as target). To calculate the functional $W_{A,\eta}(\omega,\psi)$ in the third step of the DDMRG algorithm, one needs an effective representation of the operator $(H-E_0-\omega)^2$ in the superblock subspace $$[(H-E_0-\omega)^2]_{\text{eff}} = O^{\dag} (H-E_0-\omega)^2 O \; ,$$ where the operator $O$ represents the projection onto the superblock subspace. For a typical many-body Hamiltonian $H$ such a calculation is excessively complicated and computationally intensive. Therefore, I calculate an effective representation of $H$ only, $H_{\text{eff}} = O^{\dag} H O$, and assume that $$[(H-E_0-\omega)^2]_{\text{eff}} \approx (H_{\text{eff}}-E_0-\omega)^2 \label{substitution}$$ to calculate $W_{A,\eta}(\omega,\psi)$ in the superblock subspace. This second approximation can cause a violation of the variational bound $W_{A,\eta}(\omega,\psi) \geq -\pi \eta I_A(\omega +i\eta)$. Fortunately, the substitution (\[substitution\]) has no significant effect on the minimum of $W_{A,\eta}(\omega,\psi)$ if the state $(H-E_0-\omega) |Y_{A}(\omega + i\eta)\rangle \propto |X_{A}(\omega + i\eta)\rangle$ is accurately represented in the DMRG basis \[i.e., if $O |X_{A}(\omega + i\eta)\rangle \approx |X_{A}(\omega + i\eta)\rangle$ for all superblock subspaces\]. Therefore, to use the substitution (\[substitution\]) without loss of accuracy it is necessary and sufficient to include the state $|X_{A}(\omega + i\eta)\rangle$ as a target in a DDMRG calculation, even if the real part of the dynamical correlation function is not calculated. In practice, for sufficiently large $m$ I have found that the absolute values of errors in a spectrum $I_A(\omega +i \eta)$ decrease systematically with increasing $m$. Therefore, it is possible to estimate the accuracy of a DDMRG calculation from the results obtained for different values of $m$ as one can do for static properties calculated with DMRG. Moreover, DDMRG results for $I_A(\omega +i \eta)$ tend to be smaller than the exact result for almost all frequencies although they can exceed it occasionally. Obviously, DDMRG is very similar to the correction vector DMRG. [@till; @pati] The same DMRG basis is built because the same target states are used in both methods. As numerical errors are usually dominated by the DMRG basis truncation, the correction vector parts $|X_{A}(\omega + i\eta)\rangle$ and $|Y_{A}(\omega + i\eta)\rangle$ are calculated with the same precision $\epsilon$ in both methods for a given number $m$ of density-matrix eigenstates kept per block. Nevertheless, the variational formulation has two significant advantages. First, the errors in the spectrum $I_A(\omega+i\eta)$, the excitation energies $E_n-E_0$, and the spectral weights are of the order of $\epsilon^2$ instead of $\epsilon$ in the correction vector method, as explained in Sec. \[sec:principle\]. If one uses the Lanczos algorithm instead of Eq. (\[dynamCF2\]) in the correction vector DMRG, errors become even larger. Therefore, DDMRG results are more accurate than those obtained with the correction vector DMRG for a given number $m$ of density-matrix eigenstates. Second, a DDMRG calculation is essentially an application of the standard DMRG algorithm to the minimization of a different functional. In particular, the numerical accuracy and computational effort are controlled by the sole parameter $m$ in an optimized DDMRG calculation as in a ground-state DMRG calculation. The correction vector DMRG [@till; @pati] and Lanczos DMRG [@karen; @till] are significantly more complicated than the standard DMRG method. In particular, the numerical accuracy and computational effort depend significantly and sometimes unpredictably on the specific states (Lanczos vectors and correction vectors) included as target in the density-matrix renormalization. Therefore, it is easier to implement and use DDMRG than the correction vector DMRG or the Lanczos DMRG. Finite-size scaling \[sec:scaling\] =================================== DDMRG allows us to calculate spectral functions of a correlated electron (or spin) system on a finite lattice with a broadening given by the parameter $\eta > 0$. They have the generic form $$I_{N, \eta}(\omega) = \frac{1}{\pi} \sum_{n} A_n(N) \frac{\eta}{(\omega_n(N)-\omega)^2 + \eta^2} \; ,$$ where $\omega_n(N)$ denotes the excitation energy and $A_n(N) > 0$ the spectral weight of the system eigenstates, and $N$ is the number of lattice sites. Such spectra are discrete for $\eta \rightarrow 0$ because there is only a finite number of eigenstates in a finite system. In the thermodynamic limit a spectral function $$I(\omega) = \lim_{\eta \rightarrow 0} \lim_{N \rightarrow \infty} I_{N, \eta}(\omega) \label{inflim}$$ can contain discrete and continuous parts. (It should be noted that the order of limits in the above formula is important.) To determine the properties of a dynamical spectrum $I(\omega)$ in the thermodynamic limit one has to analyze the scaling of the corresponding spectra $I_{N, \eta}(\omega)$ as a function of system size. Here I present a finite-size scaling technique for spectral functions calculated with a numerical method such as DDMRG. Computing both limits in Eq. (\[inflim\]) from numerical results for $I_{N, \eta}(\omega)$ requires a lot of accurate data for different values of $\eta$ and $N$ and can be the source of large extrapolation errors. A much better approach is to use a broadening $\eta(N) >0$ which decreases with increasing $N$ and vanishes in the thermodynamic limit. The dynamical spectrum is then given by $$\begin{aligned} I(\omega) & = & \lim_{N \rightarrow \infty} I_{N, \eta(N)}(\omega) \label{scalinglimit} \\ & = & \lim_{N \rightarrow \infty} \frac{1}{\pi} \sum_{n} A_n(N) \frac{\eta(N)}{(\omega_n(N)-\omega)^2 + \eta^2(N)} \; . \nonumber \end{aligned}$$ From the existence of both limits in Eq. (\[inflim\]) it can be demonstrated that there exists a minimal broadening $\eta_0(N) \geq 0$, which decreases as a function of $N$ and converges to zero for $N \rightarrow \infty$, such that the above equation is valid for all functions $\eta(N)$ with $\eta(N) > \eta_0(N)$ and $\lim_{N \rightarrow \infty} \eta(N) = 0$. The function $\eta_0(N)$ depends on the frequency $\omega$ considered. For a finite lattice with $N$ sites, let $M_{\omega, \epsilon}(N)$ be the number of excited states contributing to the spectral function in a small interval of width $\epsilon$ around the frequency $\omega$ (i.e., $|\omega_n(N)-\omega| < \epsilon/2$). If $M_{\omega, \epsilon}(N)$ remains finite for any $\epsilon > 0$ as $N \rightarrow \infty$, the spectral function $I(\omega)$ is discrete at $\omega$ and $\eta_0(N)=0$. Equivalently, one can take the $\eta \rightarrow 0$ limit first in Eq. (\[inflim\]). If $M_{\omega, \epsilon}(N)$ diverges for all $\epsilon > 0$ as $N \rightarrow \infty$, the spectrum is dense at $\omega$ and a minimal broadening $\eta_0(N) > 0$ is required for Eq. (\[scalinglimit\]) to be valid. For instance, $\eta_0(N)$ must be larger than the distance $\delta\omega = \omega_{n+1}(N)-\omega_n(N)$ between two consecutive excited states in the spectrum. Note that while a continuous spectrum is obviously dense, a dense spectrum can be continuous or discrete. For instance, an infinite number of excited states with $A_n(N) > 0$ can converge to the same energy as $N \rightarrow \infty$. This seems to happen for the optical conductivity associated with an exciton in a open chain. [@fabian] The function $\eta_0(N)$ depends naturally on the specific problem studied \[i.e., the scaling of the energies $\omega_n(N)$ and spectral weights $A_n(N)$\]. For the optical conductivity of one-dimensional correlated electron systems such as the Peierls-Hubbard model, I have found numerically that a sufficient condition for all frequencies $\omega$ in a dense part of the optical spectrum is $$\eta \geq \frac{c}{N} \; , \label{etacondition}$$ where the constant $c$ is comparable to the width of the dynamical spectrum $I(\omega)$, which is finite in such lattice models. Usually, one wants to keep the broadening $\eta$ as small as possible because it reduces the resolution of the spectrum. Therefore, I use $$\eta(N) = \frac{c}{N} \label{etascaling}$$ in Eq. (\[scalinglimit\]) to analyze the finite-size scaling of spectral functions calculated with DDMRG and to extrapolate the finite-size results to the thermodynamic limit. The condition (\[etacondition\]) has a very simple physical interpretation. The spectral function $I_{N, \eta}(\omega)$ represents the dynamical response of the system over a time period $\sim 1/\eta$ after one has started to apply an external force. Typically, in a lattice model the spectral width is proportional to the velocity of the excitations involved in the system response. Thus the condition (\[etacondition\]) means that excitations are too slow to travel the full length $\sim N$ of the system in the time interval $\sim 1/\eta$ and do not “sense” that the system is finite. An additional benefit of a broadening satisfying the condition (\[etacondition\]) is that the finite-system spectrum $I_{N,\eta}(\omega)$ becomes indistinguishable from the infinite-system spectrum with the same broadening $\eta$ for relatively small $N$, $$I_{N,\eta}(\omega) \approx C_{\eta}[I(\omega)]. \label{comparison}$$ Therefore, if one knows an exact or conjectured spectral function $I(\omega)$ for an infinite system, its convolution with a Lorentzian of width $\eta$ can be compared directly with the numerical results for the finite-system spectrum $I_{N,\eta}(\omega)$. This approach has been applied successfully to the optical conductivity of one-dimensional Mott insulators in Refs.  and , and additional examples are presented in the next section. In practice, the extrapolation scheme (\[scalinglimit\]) works well at fixed frequency for the continuous parts and the non-dense discrete parts of a spectrum $I(\omega)$ only. To detect singularities in $I(\omega)$ and determine their properties, it is generally easier to analyze the scaling of maxima in $I_{N,\eta}(\omega)$ or in its derivative as $N \rightarrow \infty$ and $\eta \rightarrow 0$. To perform this scaling analysis one can use a size-dependent broadening $\eta(N)$ such that Eq. (\[scalinglimit\]) is valid and $I_{N,\eta}(\omega)$ is a good approximation of $C_{\eta}[I(\omega)]$ around the maximum. Then the scaling of a maximum in $I_{N,\eta}(\omega)$ for $\eta(N) \rightarrow 0$ gives the scaling of the corresponding maximum in $C_{\eta}[I(\omega)]$ for $\eta \rightarrow 0$. Here I discuss some examples of this technique which are useful for the analysis of the Peierls-Hubbard model optical conductivity presented in the next section. \[One should also note that to detect the presence of a gap between two bands in an infinite-system spectrum $I(\omega)$, it is often faster and more reliable to investigate the scaling of minima in $I_{N,\eta}(\omega)$ for $\eta(N) \rightarrow 0$ than to perform extrapolations at fixed frequencies.\] First, we consider an infinite-system spectrum with a peak in a continuous band, $$I(\omega) = I_0 \delta(\omega - \omega_0) + I_{\text{cont}}(\omega)$$ for $|\omega - \omega_0| < \Lambda$, where $I_0 > 0$ and $I_{\text{cont}}(\omega)$ is a continuous function. It is easy to show that for $\eta \ll \Lambda$ the maximum of $C_{\eta}[I(\omega)]$ diverges as $I_0/(\pi \eta)$ and that the position of the maximum converges to $\omega_0$ for $\eta \rightarrow 0$. The maximum of the corresponding finite-system spectra $I_{N,\eta}(\omega)$ has the same scaling properties for $\eta(N) \rightarrow 0$. Therefore, it is possible to detect such a $\delta$-peak and determine its weight $I_0$ in an infinite-system spectrum, even if $I_0$ is only a small fraction of the total spectral weight of $I_{\text{cont}}$. Second, we consider an infinite-system spectrum with a power-law divergence at the band edge $$I(\omega) = I_0 \theta(\omega - \omega_0) |\omega - \omega_0|^{-\alpha}$$ for $|\omega - \omega_0| < \Lambda$, where $\theta(x) = 0$ for $x < 0$ and $\theta(x) =1$ for $x > 0$, $I_0 > 0$, and $0 < \alpha < 1$. One can show that for $\eta \ll \Lambda$ the maximum of $C_{\eta}[I(\omega)]$ diverges as $\eta^{-\alpha}$ and that the position of the maximum converges to $\omega_0$ from above for $\eta \rightarrow 0$. Again, the maximum of the corresponding finite-system spectra $I_{N,\eta}(\omega)$ has the same scaling properties for $\eta(N) \rightarrow 0$. Thus it is possible to detect such a singularity and determine the exponent $\alpha$ from the finite-system numerical data. Third, we consider a continuous infinite-system spectrum with a singularity in its derivative $$I(\omega) = I_0 \theta(\omega - \omega_0) |\omega - \omega_0|^{\alpha}$$ for $|\omega - \omega_0| < \Lambda$, where the function $\theta(x)$ and the constants $I_0$ and $\alpha$ are as in the previous example. Here the derivative of $C_{\eta}[I(\omega)]$ has a maximum which diverges as $\eta^{\alpha-1}$ for $\eta \ll \Lambda$ while its position converges to $\omega_0$ from above as $\eta \rightarrow 0$. The maximum in the derivative of $I_{N,\eta}(\omega)$ has the same scaling properties for $\eta(N) \rightarrow 0$. Therefore, it is possible to determine the exponent $\alpha$ from the finite-system numerical data in this case too. Finally, we consider a special function representing a continuous spectrum above a gap and a truncated divergence close to the band edge $$I(\omega) = I_0 \theta(\omega -\omega_0) \frac{2\sqrt{\omega_0 |\omega -\omega_0|}} {\gamma \omega_0 +|\omega -\omega_0|} \label{typicalOC}$$ for $|\omega - \omega_0| < \Lambda$, where $\gamma$ is a constant such that $0 \leq \gamma<\Lambda/\omega_0$. The function $\theta(x)$ and the other constants are as in the previous examples. For $\gamma >0$ this spectrum vanishes as a square-root at the band edge $\omega_0$, goes trough a maximum $I_0/\sqrt{\gamma}$ at $\omega = (1+\gamma) \omega_0$ then decreases monotonically. For $\gamma \ll 1$ the maximum is very sharp and close to the band edge, and $I(\omega)$ appears to diverge as $1/\sqrt{|\omega -\omega_0|}$ at higher energy $ \omega > (1+\gamma) \omega_0$. The continuous vanishing of $I(\omega)$ at the band edge is apparent only in a small frequency range $\omega_0 < \omega < (1+\gamma) \omega_0$. As $\gamma \rightarrow 0$ this maximum becomes a square-root singularity at $\omega_0$. A qualitatively similar behavior is often found in the optical conductivity of one-dimensional insulators (see the discussion in the next section). Obviously, the maximum of $C_{\eta}[I(\omega)]$ tends to $I_0/\sqrt{\gamma}$ and its position converges to $\omega = (1+\gamma) \omega_0$ for $\eta \rightarrow 0$. For $\gamma < 1$, however, the convergence of the maximum becomes apparent only for $\eta \ll \gamma \omega_0$. For larger $\eta$ the maximum appears to diverge as $1/\sqrt{\eta}$. Similarly, the maximum of the derivative diverges as $1/\sqrt{\eta}$ for $\eta \rightarrow 0$ as discussed the previous example. In the present case, however, this scaling is not observed as soon as $\eta \ll \Lambda$ but only if $\eta \ll \gamma \omega_0$. The finite-system spectrum $I_{N,\eta}(\omega)$ and its derivative have the same scaling properties for $\eta(N) \rightarrow 0$. Therefore, with the scaling analysis of finite-system spectra it is possible to distinguish an infinite-system spectrum with a truncated divergence above the band edge from a spectrum with a real divergence at the band edge, provided that one can do calculations with a resolution $\eta(N) \ll \gamma \omega_0$. In summary, the dynamical spectrum of an infinite system can be determined accurately and efficiently from numerical data for finite-system spectra using a size-dependent broadening $\eta(N)$. The broadening $\eta(N)$ must be larger than a minimal broadening $\eta_0(N)$, which depends on the system investigated and can vary with the frequency. Often this broadening conceals the finite-size effects and one can directly compare finite-system spectra to analytical results for infinite systems using a convolution with a Lorentzian distribution, see Eq. (\[comparison\]). If this comparison is not possible or not sufficient, specific points of the spectrum can be extrapolated to the thermodynamic limit using Eq. (\[scalinglimit\]). Finally, the scaling of maxima in finite-system spectra or their derivatives \[for $\eta(N) \rightarrow 0$\] allows us to find and analyze singularities in the infinite-system spectrum. For one-dimensional correlated electron systems a sufficient condition for the minimal broadening is given by Eq. (\[etacondition\]) and one can use a size-dependent broadening (\[etascaling\]). Optical conductivity of the Peierls-Hubbard model \[sec:model\] =============================================================== In this section I apply the DDMRG method and the finite-size scaling analysis to the optical conductivity of the one-dimensional Peierls-Hubbard model. [@dionys] This model is defined by the Hamiltonian $$H = T + U \sum_{l=1}^N \left(n_{l,\uparrow}-\frac{1}{2}\right) \left(n_{l,\downarrow}-\frac{1}{2}\right) \label{hamiltonian}$$ with $$T = - \sum_{l;\sigma} \left(t-(-1)^l \frac{\Delta}{2} \right ) \left(c_{l,\sigma}^+c_{l+1,\sigma} + c_{l+1,\sigma}^+c_{l,\sigma}\right) \; . \label{kinetic}$$ It describes electrons with spin $\sigma=\uparrow,\downarrow$ which can hop between neighboring sites in a lattice with an even number $N$ of sites. In Eq. (\[kinetic\]) the index $l$ runs from 1 to $N-1$ for an open chain and from 1 to $N$ if periodic boundary conditions are used. Here $c^+_{l,\sigma}$, $c_{l,\sigma}$ are creation and annihilation operators for electrons with spin $\sigma$ at site $l$ and $n_{l,\sigma}= c^+_{l,\sigma}c_{l,\sigma}$ are the corresponding density operators. The hopping integral $t > 0$ gives rise to a single-electron band of width $4t$. The dimerization parameter $0 \leq |\Delta| \leq 2t$ determines the strength of the periodic lattice potential generated by the Peierls instability. (For a finite open chain I only use $\Delta \geq 0$ to avoid spurious excitations at the chain ends.) The Coulomb repulsion is mimicked by a local Hubbard interaction $U\geq 0$. The number of electrons equals the number of lattice sites. The ground state, single-particle charge gap and spin gap of this system can be calculated with great accuracy on lattices with up to $N \sim 10^3$ sites using DMRG. The single-particle charge gap is given by $$E_c(N) = E_0(N+1) + E_0(N-1) - 2 E_0(N) \; , \label{chargegap}$$ where $E_0(M)$ denotes the ground-state energy for $M$ electrons in a $N$-site system. For an even number $N$ of sites and electrons the Peierls-Hubbard model ground state is a singlet [@lieb] and the spin gap is given by $$E_s(N) = E_0(S_z=\hbar) - E_0(S_z=0) \; ,$$ where $S_z$ is the the $z$-component of the total spin and $E_0(S_z)$ is the ground-state energy for a fixed value of $S_z$. Spectroscopy with electromagnetic radiation is a common experimental probe of solid-state materials. The linear optical absorption is proportional to the real part $\sigma_1(\omega)$ of the optical conductivity. For $\omega > 0$, $\sigma_1(\omega)$ is related to the imaginary part of the current-current correlation function $G_{J} (\hbar \omega+ i \eta)$ by $$\begin{aligned} \sigma_1(\omega) & = & \frac{\pi}{N a \omega} \lim_{\eta \rightarrow 0} \text{Im} \ G_{J} (\hbar \omega+ i \eta) \label{sigma1} \\ & = & \frac{\pi}{Na \omega} \sum_n |\langle \psi_0 |J| \psi_n\rangle|^2 \ \delta(\hbar\omega+E_0- E_n) \; . \nonumber\end{aligned}$$ Here $|\psi_0\rangle$ is the ground state of the Hamiltonian $H$, $|\psi_n\rangle (n > 1)$ are the other eigenstates of $H$, and $E_0$, $E_n$ are their respective eigenenergies. In this model the current operator is $$J=\frac{iae}{\hbar} \sum_{l;\sigma} \left(t-(-1)^l \frac{\Delta}{2} \right ) \left( c_{l,\sigma}^+c_{l+1,\sigma} - c_{l+1,\sigma}^+c_{l,\sigma} \right) \; , %\label{currentop}$$ where $a$ is the lattice constant, $-e$ is the charge of an electron, and the index $l$ takes the same values depending on the boundary conditions as in Eq. (\[kinetic\]). Note that this is the natural definition of the current operator for both types of boundary conditions. The Parzen filter used for open boundary conditions in other works [@till2; @brune] is not necessary and thus not used in this work. In an open chain the optical absorption is also related to the dynamical polarizability $\alpha(\omega)$, which is given by the imaginary part of the dipole-dipole correlation function $G_{D} (\hbar \omega+ i \eta)$, $$\begin{aligned} \alpha(\omega) & = & \frac{\pi}{N a} \lim_{\eta \rightarrow 0} \text{Im} \ G_{D} (\hbar \omega+ i \eta) \label{alpha} \\ & = & \frac{\pi}{Na} \sum_n |\langle \psi_0 |D| \psi_n\rangle|^2 \ \delta(\hbar\omega+E_0- E_n) \; . \nonumber \end{aligned}$$ For a constant lattice spacing $a$ the dipole operator is $$D=-ea \sum_{l=1}^N l \left(n_{l} - 1 \right ) \; %\label{dipoleop}$$ with $n_{l} = n_{l, \uparrow} + n_{l, \downarrow}$. Using the relation $J = -\frac{i}{\hbar} [D,H]$ one easily proves that $$\sigma_1(\omega) = \omega \alpha(\omega) \label{sigmaalpha}$$ and $$\sigma_1(\omega) = \frac{\pi}{N a \hbar} \lim_{\eta \rightarrow 0} \text{Im} \{ (\hbar \omega+ i\eta) \ G_{D} (\hbar \omega+ i \eta)\} \; . \label{sigma1bis}$$ The optical conductivity can be calculated with the DDMRG method described in this paper. For an open chain Eqs. (\[sigma1\]),(\[alpha\]), and (\[sigma1bis\]) provide us with three different approaches. First, one can calculate the imaginary part of the current-current correlation function with DDMRG and use Eq. (\[sigma1\]) to obtain the convolution of the reduced optical conductivity $$\begin{aligned} C_{\eta}[\omega \sigma_1(\omega)] & = & \frac{\pi}{N a} \text{Im} \ G_{J} (\hbar \omega+ i \eta) \label{Crsigma} \\ & = & \frac{1}{Na} \sum_n \frac{\eta |\langle \psi_0 |J| \psi_n\rangle|^2} {(\hbar\omega+E_0- E_n)^2+\eta^2} \; . \nonumber \end{aligned}$$ This is also the only approach possible with periodic boundary conditions. Second, one can calculate the imaginary part of the dipole-dipole correlation function with DDMRG and use Eq. (\[alpha\]) to obtain the convolution of the dynamical polarizability $$\begin{aligned} C_{\eta}[\alpha(\omega)] & = & \frac{\pi}{N a} \text{Im} \ G_{D} (\hbar \omega+ i \eta) \label{Calpha} \\ & = & \frac{1}{Na} \sum_n \frac{\eta |\langle \psi_0 |D| \psi_n\rangle|^2} {(\hbar\omega+E_0- E_n)^2+\eta^2} \; . \nonumber \end{aligned}$$ The optical conductivity is then given by the relation (\[sigmaalpha\]). Finally, one can calculate the complete dipole-dipole correlation function with DDMRG and use Eq. (\[sigma1bis\]) to obtain the convolution of $\sigma_1(\omega)$ directly $$\begin{aligned} C_{\eta}[\sigma_1(\omega)] & = & \frac{\pi}{N a \hbar} \text{Im} \{ (\hbar \omega+ i\eta) \ G_{D} (\hbar \omega+ i \eta)\} \label{Csigma1} \\ & = & \frac{1}{Na\hbar} \sum_n \frac{\eta |\langle \psi_0 |D| \psi_n\rangle|^2 (E_n -E_0)} {(\hbar\omega+E_0- E_n)^2+\eta^2} \; . \nonumber \end{aligned}$$ $C_{\eta}[\sigma_1(\omega)]$ can also be formulated in terms of the current matrix elements $|\langle \psi_0 |J| \psi_n\rangle|^2$ \[see Eq. (4) of Ref. \]. Although the real part of a dynamical correlation function is used in Eq. (\[Csigma1\]) to calculate the optical conductivity, its relative contribution to $C_{\eta}[\sigma_1(\omega)]$ is of the order $(\eta/t)^2$. Therefore, the numerical precision is not significantly reduced by the lower accuracy of DDMRG for the real part of dynamical correlation functions. Clearly, all three approaches give the same spectrum $\sigma_1(\omega)$ for $\eta \rightarrow 0$. In DDMRG calculations with $\eta > 0$, however, they are not equivalent. First, I have found that it is easier to calculate the dipole-dipole correlation function $G_{D} (\hbar \omega+ i \eta)$ than the current-current correlation function $G_{J} (\hbar \omega+ i \eta)$, except for very strong coupling $U \gg t$. Second, the finite-size scaling \[using a size-dependent broadening (\[etascaling\])\] is different for the three optical spectra $C_{\eta}[\omega \sigma_1(\omega)]/\omega$, $\omega C_{\eta}[\alpha(\omega)]$, and $C_{\eta}[\sigma_1(\omega)]$, especially for very small and very large frequencies $\omega$. Usually, $C_{\eta}[\sigma_1(\omega)]$ is the best approximation to $\sigma_1(\omega)$ but at low energy ($\hbar \omega < t$) it can be more convenient to use $\omega C_{\eta}[\alpha(\omega)]$ while at high energy ($\hbar \omega \gg t$) I prefer $C_{\eta}[\omega \sigma_1(\omega)]$. To calculate the optical spectrum of the Peierls-Hubbard model I have used the third approach, Eq. (\[Csigma1\]), in most cases. Thus $C_{\eta}[\sigma_1(\omega)]$ is shown in the figures of this paper unless I state explicitly otherwise. Only optical spectra calculated with open boundary conditions are presented. As the size-dependent broadening (\[etascaling\]) conceals most of the finite-size effects, spectra calculated with periodic boundary conditions would be almost identical. In all figures showing optical spectra I set $a=e=\hbar=t=1$. Thus $\sigma_1(\omega)$ is shown in units of $e^2a/\hbar$, $\omega \sigma_1(\omega)$ in units of $e^2a t/\hbar^2$, and the frequency $\omega$ in units of $\hbar/t$. The sum rules (\[sumrules\]) take a simple form for the optical conductivity in the Peierls-Hubbard model with open boundary conditions \[OCsums\] $$\begin{aligned} \frac{\hbar}{\pi} \int_0^\infty d\omega \omega \sigma_1(\omega)&=& \frac{1}{Na} \ \langle \psi_0|J^2|\psi_0\rangle , \label{sumJ} \\ % \frac{\hbar}{\pi} \int_0^\infty d\omega \sigma_1(\omega) \; \; &=& \frac{-a e^2}{2N\hbar} \ \langle \psi_0|T|\psi_0\rangle , \label{sumT}\\ % \frac{\hbar}{\pi} \int_0^\infty d\omega \frac{\sigma_1(\omega)}{\omega} \; \, &=& \frac{1}{Na} \, \langle \psi_0| D^2 |\psi_0\rangle\ . \label{sumD}\end{aligned}$$ To prove the second sum rule (\[sumT\]) one uses the relation $[D,J] = -i a^2 e^2 T/\hbar $. With periodic boundary conditions, only the first two sum rules remain valid. In the second sum rule, however, one must take into account the coherent part of the conductivity at $\omega =0$ and the proof is more complicated than for an open chain. The right-hand-side of Eq. (\[OCsums\]) can be calculated accurately with the ground-state or Lanczos DMRG method. Note that for $\eta > 0$ an optical spectrum calculated from Eq. (\[Crsigma\]), (\[Csigma1\]), or (\[Calpha\]) exactly fulfills the sum rule (\[sumJ\]), (\[sumT\]), or (\[sumD\]), respectively. For the DDMRG results presented in this paper, the sum rules are fulfilled within a few percents. The optical gap $E_{\text{opt}}(N)$ is defined as the excitation energy ($E_n - E_0$) of the lowest eigenstate contributing to the optical conductivity (i.e., $\langle \psi_n|J| \psi_0\rangle \neq 0$) in a $N$-site system. $E_{\text{opt}}(N)$ can be calculated with the DDMRG method for individual excited states described in Sec. \[sec:ddmrg\]. As the Peierls-Hubbard Hamiltonian (\[hamiltonian\]) has a particle-hole symmetry the optical gap can also be determined using the symmetrized DMRG method. [@ramasesha] As expected, both approaches give the same results for $E_{\text{opt}}(N)$ within numerical errors. In the thermodynamic limit ($N \rightarrow \infty$) I have found that the optical gap $E_{\text{opt}}$ is equal to the single-particle charge gap $E_c$ (\[chargegap\]) for all $U \geq 0$ and $2t > \Delta \geq 0$. \[In the dimer limit ($\Delta=2t$) the Hamiltonian (\[hamiltonian\]) describes independent dimers: the optical weight is concentrated in a single peak corresponding to Frenkel excitons localized on a dimer and the other (delocalized) excitations above and below this peak carry not optical weight, thus $E_c < E_{\text{opt}}$.\] In a finite system or with additional electronic interactions [@fabian] the single-particle charge gap $E_c$ can be different from the optical gap $E_{\text{opt}}$. All DMRG methods have a truncation error which is reduced by increasing the number $m$ of retained density matrix eigenstates (for more details, see Refs.  and ). Varying $m$ allows one to compute physical quantities for different truncation errors and thus to obtain error estimates on these quantities. For some quantities, especially eigenenergies, it is possible to extrapolate the results to the limit of vanishing truncation error and thus to achieve a greater accuracy. I have systematically used these procedures to estimate the precision of my numerical calculations and adjusted the maximal number $m$ of density matrix eigenstates to reach a desired accuracy. This is especially important for DDMRG calculations as truncation errors in dynamical spectra can greatly vary as a function of the frequency $\omega$ for fixed $m$. In this work the largest number of density matrix eigenstates used is $m=600$. For all numerical results presented here DMRG truncation errors are negligible. In the following three subsections I demonstrate the finite-size scaling technique and the accuracy of DDMRG on three special limits of the Peierls-Hubbard model. Then the optical conductivity of a Mott-Peierls insulator is presented and discussed in the last subsection. Peierls insulator \[sec:peierls\] --------------------------------- For $U=0$ the Hamiltonian (\[hamiltonian\]) describes a system of independent electrons, which can be solved exactly for any value of $\Delta$, boundary conditions, or lattice size. This provides us with a perfect test case for the DDMRG method. I have checked that DDMRG can reproduce the optical spectrum of this system on lattices with several hundred sites, for any frequency $\omega$, and with relative errors as small as $10^{-4}$ using only a few hundred density-matrix eigenstates. This demonstrates that one can obtain almost exact results for the optical conductivity of finite one-dimensional systems such as the Peierls-Hubbard model using DDMRG. In the thermodynamic limit the Hamiltonian (\[hamiltonian\]) describes a Peierls (band) insulating phase for $\Delta \neq 0$ and $U=0$. The optical gap $E_{\text{opt}}$, the charge gap $E_c$, and the spin gap $E_s$ equal $2|\Delta|$. Optical excitations are made of one hole in the valence band and one electron in the conduction band. The optical conductivity is given by $$\sigma_1(\omega) = \frac{a e^2 (2\Delta)^2(4t)^2} {2 \hbar(\hbar \omega)^2 \sqrt{[(\hbar \omega)^2-(2\Delta)^2] [(4t)^2-(\hbar \omega)^2]}} \label{peierlsOC}$$ for $2|\Delta| < \hbar \omega <4t$ and is zero elsewhere. [@florian] This optical spectrum contains a single band of width $4t-2|\Delta|$ with square-root divergences at both band edges. These divergences are a typical feature of a one-dimensional band insulator. The convolution of Eq. (\[peierlsOC\]) with a Lorentzian distribution of width $\eta/t=0.05$ is shown in Fig. \[fig1\] for $\Delta=0.6t$. Both divergences are replaced by maxima at $\hbar \omega \approx 2 \Delta = 1.2t$ and $\hbar \omega \approx 4t$. In Fig. \[fig1\] I also show the optical conductivity calculated with DDMRG on a 128-site lattice with the same broadening. We see that the finite-system optical spectrum is indistinguishable from the infinite-system spectrum. The broadening $\eta/t=0.05$ satisfies the condition (\[etacondition\]) and thus conceals the finite-size effects as discussed in Sec \[sec:scaling\]. In this case a broadening $\eta(N)/t=6.4/N$ is enough because the spectrum band width is smaller than $4t$. ![ \[fig1\] Optical conductivity of a Peierls insulator with $\Delta=0.6t$ and a broadening $\eta/t=0.05$. Both the DDMRG result for a 128-site chain and the exact result (\[peierlsOC\]) in the thermodynamic limit are shown. ](fig1.eps){width="8cm"} With this size-dependent broadening $\eta(N)$ one can use Eq. (\[scalinglimit\]) to extrapolate the finite-size DDMRG results to the thermodynamic limit. For instance, for $\hbar\omega=2.6t$ I have obtained $\sigma_1(\omega)= 0.245$ (in units of $ae^2/\hbar$) using data for systems with up to $N=256$ sites \[i.e., with a broadening down to $\eta(N)/t = 0.025$\], in excellent agreement with the exact result 0.243. If we did not know the exact result (\[peierlsOC\]), we could nevertheless determine the existence of square-root divergences at both band edges using a scaling analysis of the maxima in the DDMRG spectra. For instance, the height of the low-energy maximum (close to $\hbar \omega = 1.2t$) diverges as $1/\sqrt{\eta}$ for $\eta(N) \rightarrow 0$ \[see Fig. \[fig2\](a)\]. Moreover, the position of the maximum tends from above to the optical gap $E_{\text{opt}} = 2 \Delta = 1.2 t$ for $N \rightarrow \infty$ \[see Fig. \[fig2\](b)\]. As explained in Sec. \[sec:scaling\] these scaling properties correspond to a square-root divergence at the band edge. Figure \[fig2\](b) also shows the finite-size optical gaps $E_{\text{opt}}(N)$ calculated with the DDMRG method for individual excited states. They tend to the exact result $E_{\text{opt}} = 1.2 t$ for $N \rightarrow \infty$ as expected. Mott-Hubbard insulator \[sec:hubbard\] -------------------------------------- ![ \[fig2\] (a) Maximum $\sigma_{\text{max}}$ of the optical spectrum $\sigma_1(\omega)$ as a function of the broadening $\eta(N)$. (b) Position of the spectrum maximum $E_{\text{max}} = \hbar \omega_{\text{max}}$ (square) and optical gap $E_{\text{opt}}$ (circle) as a function of system size $N$. In both figures filled symbols correspond to the Mott-Hubbard insulator \[$U=4t, \eta(N)N=12.8t$\] and open symbols to the Peierls insulator \[$\Delta=0.6t, \eta(N)N= 6.4t$\]. ](fig2a.eps "fig:"){width="7cm"} ![ \[fig2\] (a) Maximum $\sigma_{\text{max}}$ of the optical spectrum $\sigma_1(\omega)$ as a function of the broadening $\eta(N)$. (b) Position of the spectrum maximum $E_{\text{max}} = \hbar \omega_{\text{max}}$ (square) and optical gap $E_{\text{opt}}$ (circle) as a function of system size $N$. In both figures filled symbols correspond to the Mott-Hubbard insulator \[$U=4t, \eta(N)N=12.8t$\] and open symbols to the Peierls insulator \[$\Delta=0.6t, \eta(N)N= 6.4t$\]. ](fig2b.eps "fig:"){width="7cm"} For $\Delta = 0$ the Peierls-Hubbard model (\[hamiltonian\]) becomes the one-dimensional Hubbard model at half filling. For $U > 0$ this model describes a Mott-Hubbard insulator with gapless spin excitations. [@lieb2] The optical conductivity of this system has recently been determined using DDMRG and analytical methods. [@eric] Here I only summarize the most important results and give more information about the finite-size scaling analysis carried out in this previous work. In the half-filled Hubbard model an optical excitation is made of a pair of spinless bosonic excitations carrying opposite charges in the lower (holon) and upper (doublon or anti-holon) Hubbard bands, respectively. As in a Peierls insulator, the optical spectrum consists of a single band but its width is larger, about $8t$. A second distinctive feature of this spectrum is a square-root vanishing $\sigma_1(\omega) \sim \sqrt{\hbar \omega - E_{\text{opt}}}$ at the band threshold $E_{\text{opt}}$. There is also a tiny peak in the middle of the band, at least for $U \geq 4t$. In Ref.  it is shown that for weak coupling ($U \leq 3t$) and in the strong-coupling limit ($U/t \rightarrow \infty$) the finite-system optical spectra calculated with DDMRG agree perfectly with the analytical results obtained in the thermodynamic limit using a field-theoretical approach and a strong-coupling analysis, respectively. For instance, Fig. \[fig3\] shows the low-energy parts of DDMRG spectra calculated for three different lattice sizes at $U=3t$ and the corresponding field-theoretical spectrum for an infinite system. A size-dependent broadening $\eta(N)/t=12.8/N$ is used in this case. One clearly sees the convergence of the finite-size spectra toward the field-theoretical result as $\eta(N)$ decreases. To make a quantitative comparison one can calculate the convolution of the field-theoretical spectrum with a Lorentzian of width $\eta$ satisfying the condition (\[etacondition\]) as discussed in Sec. \[sec:scaling\]. One finds then that finite-size effects are completely concealed by the broadening even for relatively small system sizes. For instance, it is shown in Fig. 3 of Ref.  that the low-energy optical spectrum calculated on a 128-site lattice for $U=3t$ is indistinguishable from the field-theoretical spectrum with the same broadening $\eta/t=0.1$. In the strong-coupling limit DDMRG and analytical results agree even better and finite-size effects are no longer visible for systems as small as $N=32$. ![ \[fig3\] Optical conductivity of the Hubbard model with $U=3t$ for several values of $\eta$. Results for $\eta(N) > 0$ have been calculated with DDMRG on $N$-site lattices with $\eta(N) N =12.8t$. For $\eta=0$ the field-theoretical result for an infinite system is shown. ](fig3.eps){width="8cm"} For other coupling strengths ($4 \leq U/t < \infty$) it is necessary to analyze the scaling of the finite-system DDMRG spectra to determine the optical conductivity of the Hubbard model in the thermodynamic limit. Using numerical results for lattices with up to $N=256$ sites \[i.e., a resolution $\eta(N)/t$ down to 0.05\], I have found that for all $U/t$ the optical conductivity at the lower band edge has the qualitative behavior described by Eq. (\[typicalOC\]): $\sigma_1(\omega)$ vanishes as $\sqrt{\hbar \omega - E_{\text{opt}}}$ at the band threshold and there is a maximum in $\sigma_1(\omega)$ at a frequency $\omega = (1 + \gamma) E_{\text{opt}}/\hbar$, where $\gamma$ is a small number. Field theory predicts the same behavior with $\gamma \approx 0.24$ in the weak-coupling limit while the strong-coupling analysis gives a maximum at $\hbar \omega = U = E_{\text{opt}} + 4t$, and thus $\gamma$ vanishes as $t/U$ for $U \gg t$. The distance $\gamma E_{\text{opt}}$ between the spectrum threshold and the maximum increases with $U/t$. For $U \geq 4t$ this distance is large enough to determine the finite-size scaling of the lower band edge using systems with up to $N=256$ sites. As an example, Fig. \[fig2\](a) shows the low-energy maximum in the optical spectrum $\sigma_1(\omega)$ calculated with DDMRG for $U=4t$ as a function of $\eta(N)$. The contrast between the Mott-Hubbard insulator and the Peierls insulator is striking and the maximum in the Mott-Hubbard insulator optical spectrum clearly tends to a constant for $\eta(N) \rightarrow 0$. \[For $U=4t$ the optical gap of the Hubbard model is comparable to that of the Peierls insulator with $\Delta=0.6t$, so that a direct comparison of both systems is relevant.\] In Fig. \[fig2\](b) one sees that the finite-size optical gaps calculated with DDMRG converge to the exact result [@lieb2] $E_{\text{opt}} = 1.287t$ in the thermodynamic limit, but the maximum tends to a higher energy $\hbar \omega \approx 1.7t$. Therefore, one can conclude that there is no divergence at the optical conductivity threshold $\hbar \omega = E_{\text{opt}}$. Moreover, it is possible to confirm that $\sigma_1(\omega)$ vanishes as $\sqrt{\hbar \omega - E_{\text{opt}}}$ at the lower band edge using either a similar scaling analysis for the derivative of DDMRG spectra or a direct comparison with the convolution of functions such as Eq. (\[typicalOC\]) or the field-theoretical optical spectrum. [@eric] For very weak coupling one would need to calculate $\sigma_1(\omega)$ for very large system sizes in order to perform the same scaling analysis. Because $E_{\text{opt}}$ vanishes exponentially with $U/t$ and the scaling analysis must be performed in the asymptotic regime $\eta(N) < \gamma E_{\text{opt}}$, the required system sizes $N$ increases exponentially as $t/E_{\text{opt}}$ for $U/t \rightarrow 0$. Fortunately, it is not necessary to carry out this analysis for the Hubbard model because the optical conductivity of the weak-coupling field theory is already in excellent agreement with the optical conductivity of the lattice model for $U = 3t$. Strong-coupling limit \[sec:strong\] ------------------------------------ In this section I discuss the special case of a Mott-Peierls insulator ($\Delta \neq 0, U > 0$) in the strong-coupling limit $U/t \rightarrow \infty$, for which the shape of the optical spectrum is known analytically. [@florian2] In this limit there is exactly one electron on each site in the ground state of the Peierls-Hubbard Hamiltonian (\[hamiltonian\]). An optical excitation moves one electron from a site to another and thus creates a double occupation (doublon) and an empty site (holon). Therefore, the optical gap is of the order of $U$. These elementary charge excitations are spinless bosons as in the Hubbard model. The properties of the spin degrees of freedom are determined by an effective Heisenberg model with alternating exchange couplings $J_1 \sim (t+\Delta/2)^2/U$ and $J_2 \sim (t-\Delta/2)^2/U$. The spin gap $E_s$ vanishes in the limit $U \rightarrow \infty$. However, as there is a gap in the spin excitation spectrum for any finite $U/t$ (see also next section), the structure of the spin ground state in the limit $U/t \rightarrow \infty$ is actually similar to that of a gapped state. For instance, the antiferromagnetic spin-spin correlations decreases exponentially with increasing distance. Thus, this strong-coupling limit of the Peierls-Hubbard model is different from the two limiting cases discussed previously and from the general case presented in the next section. ![ \[fig4\] Reduced optical conductivity $C_{\eta}[\omega \sigma_1(\omega)]$ as a function of $\hbar \Omega = \hbar \omega - U$ in the strong-coupling limit $U \gg t$ for $\Delta=0.6t$. The solid line is the DDMRG result, Eq. (\[Crsigma\]), for a 128-site lattice with a broadening $\eta/t=0.1$. The dot-dashed line is a Lorentzian distribution of width $\eta/t=0.1$ centered at $\hbar \omega=U$. The two dashed lines represent the analytical result (\[strongOC1\]) for the continuum of an infinite system ($\eta=0$). Note the logarithmic scale of the ordinate axis. ](fig4.eps){width="8cm"} In the thermodynamic limit the optical conductivity can be calculated analytically using some reasonable assumptions. [@florian2] If $0 < |\Delta| < 2t$, the spectrum consists of two bands for $2|\Delta| \leq |\hbar\omega -U| \leq 2t$ $$\sigma_1(\omega) = \frac{g_0 e^2 a}{8\hbar} \frac{\sqrt{[(\hbar \Omega)^2-(2\Delta)^2] [(4t)^2 -(\hbar \Omega)^2]}}{\hbar \omega |\hbar\Omega|} \; , \label{strongOC1}$$ where $\hbar \Omega = \hbar \omega - U$, and a $\delta$-peak at $\hbar \omega = U$ $$\sigma_1(\omega) = \frac{\pi g_{\pi} e^2 a t^2}{\hbar U} \delta(\hbar \omega-U) \label{strongOC2}$$ in the middle of the gap $4 |\Delta|$ separating the bands. For $\Delta \rightarrow 0$ one recovers the optical spectrum of the Hubbard model in the strong coupling limit, which consists of a single band and a $\delta$-peak in the middle of this band. [@eric; @florian2] The prefactors $g_0$ and $g_{\pi}$ are spin form factors given by ground-state spin correlation functions. They are functions of the effective exchange coupling ratio $J_2/J_1$ and thus of $\delta = |\Delta/2t|$. Assuming a dimer spin ground state (i.e., $J_1 > 0$ and $J_2=0$) one obtains $g_0 = 9/4$ and $$g_{\pi} = \frac{1+32 \delta +62 \delta^2 + 32 \delta^3 + \delta^4} {4 (1 + \delta)^2} \; .$$ This result becomes exact in the dimer limit $|\Delta|=2t$, where $g_{\pi}=8$. For $\Delta\rightarrow 0$ the dimer spin ground state does not give the correct form factors because it is known exactly that $g_0 + g_{\pi} = 4 \ln(2)$ and it was found numerically that $g_{\pi}/g_0 \approx 10^{-2}$ (see Ref. ). Figure \[fig4\] shows the reduced optical conductivity $C_{\eta}[\omega \sigma_1(\omega)]$ calculated using DDMRG on a 128-site lattice for $\Delta=0.6t$ and $\eta/t=0.1$. A logarithmic scale is used to make visible the weak bands on both sides of the strong central peak. In Fig. \[fig4\] one can recognize the spectral shape predicted by the strong-coupling analysis. To make a quantitative comparison, however, it is first necessary to determine $g_0$ and $g_{\pi}$ using the finite-size scaling analysis of Sec. \[sec:scaling\]. Here I use a size-dependent broadening $\eta(N)/t = 12.8/N$ as for the Hubbard model because the spectral width is also of the order of $8t$. For $\hbar \omega = U-2t$ DDMRG results for $C_{\eta}[\omega \sigma_1(\omega)]$ tend to 0.78 (in units of $e^2at/\hbar^2$) for $N \rightarrow \infty$. Comparison with Eq. (\[strongOC1\]) then yields $g_0\approx 2.2$. In Fig. \[fig4\] I also show the two bands (\[strongOC1\]) with this value of $g_0$ (without broadening). The agreement with the finite-system DDMRG spectrum is excellent. The small deviations visible close to the band edges are due to the different broadening used for the numerical result ($\eta/t=0.1$) and for the analytical result ($\eta=0$). They vanish if the same broadening is used in both calculations. Once more this confirms that a broadening satisfying Eq. (\[etacondition\]) hides most finite-size effects in this model as already shown by other examples in Sec. \[sec:peierls\] of this paper and in Refs.  and . In the DDMRG spectra $C_{\eta}[\omega \sigma_1(\omega)]$ the height of the central peak diverges as $2.42t/\eta$ (in units of $e^2at/\hbar^2$) for increasing $\eta(N)$ but its position does not change. This confirms that it corresponds to a $\delta$-peak at $\hbar \omega =U$ and gives an estimate $g_{\pi} \approx 2.42$. This $\delta$-peak broadened with a Lorentzian distribution of width $\eta/t=0.1$ is also shown in Fig. \[fig4\]. One sees that the agreement with the DDMRG result is perfect. A similar finite-size scaling was performed to determine the form factor $g_{\pi}$ in the Hubbard model. [@eric] One notes the surprisingly good agreement between the form factors determined numerically with DDMRG ($g_0 \approx 2.2$ and $g_{\pi} \approx 2.42$) and those obtained using the approximation of a dimer spin ground state ($g_0 = 2.25$ and $g_{\pi} \approx 2.52$). For the value $\Delta = 0.6t$ used in this example, the ratio $J_2/J_1 \approx 0.29$ of the effective exchange coupling is already quite small and thus the dimer spin ground state is probably a very good approximation of the actual spin ground state. Mott-Peierls insulator \[sec:mottpeierls\] ------------------------------------------ The optical conductivity of the Peierls-Hubbard model is not known for general interaction parameters $U > 0 $ and $2t > |\Delta| > 0$. In this regime the system is in a Mott-Peierls insulating phase [@dionys]: both a periodic lattice potential (i.e., the alternating hopping terms) and electronic correlations contribute to the formation of a charge gap $E_c > 0$ and there is a finite spin gap $E_s > 0$. Numerical investigations of the charge and spin gaps and of static correlation functions reveal no phase transition at finite $U$ and intermediate $\Delta$ (see also Ref.  and references therein). Thus the entire parameter space ($0 < U/t < \infty$, $0 < |\Delta| < 2t$) belongs to a single Mott-Peierls insulating phase. Figure \[fig5\] shows charge and spin gaps as a function of $U$ and $\Delta$. These gaps have been calculated on lattices with up to $N=400$ sites using DMRG and extrapolated to the thermodynamic limit. The charge gap of the Mott-Peierls insulator is always larger than the gap of the Mott-Hubbard and Peierls insulators in the $\Delta=0$ and $U=0$ limits, respectively. The spin gap is always smaller than the charge gap in the Mott-Peierls phase. ![ \[fig5\] Charge (diamond) and spin (circle) gaps extrapolated to the thermodynamic limit. (a) As a function of $U$ for $\Delta=0.6t$. (b) As a function of $\Delta$ for $U=4t$. ](fig5a.eps "fig:"){width="7cm"} ![ \[fig5\] Charge (diamond) and spin (circle) gaps extrapolated to the thermodynamic limit. (a) As a function of $U$ for $\Delta=0.6t$. (b) As a function of $\Delta$ for $U=4t$. ](fig5b.eps "fig:"){width="7cm"} In the thermodynamic limit the optical gap $E_{\text{opt}}$ is equal to the charge gap. The nature of the optical excitations in the Mott-Peierls insulator is not well understood. Despite the obvious difference between charge and spin excitation energies, $E_s < E_c$, it is not even known if there is a spin-charge separation for single-particle excitations. Optical excitations could consist of a pair of fermionic quasi-particles with opposite spins $\pm \sigma$ and opposite charges $\pm e$ as in a Peierls insulator (Sec. \[sec:peierls\]). They could as well be made of two spinless bosonic excitations carrying opposite charges $\pm e$ as in the Mott-Hubbard insulator (Sec. \[sec:hubbard\]) and in the strong-coupling limit (Sec. \[sec:strong\]). The investigation of spin and charge gaps and static correlation functions clearly shows that the three special cases described in the previous sections are singular limits of the Peierls-Hubbard model. Unsurprisingly, I have found that the optical conductivity in the Mott-Peierls phase is unlike the simple spectrum observed in these limits. \[All optical spectra presented in this section have been calculated using DDMRG and the finite-size scaling analysis has always been performed using a size-dependent broadening $\eta(N)/t=12.8/N$.\] For large but finite $U$ the optical spectrum consists of three bands: a narrow band with a strong singularity around $\hbar \omega = U$ and one weak band on each side of this central peak. The singularity seems to be made of two very close power-law divergences which merge to form the single isolated $\delta$-peak (\[strongOC2\]) in the $U/t \rightarrow \infty$ limit. The optical spectrum starts as $\sqrt{\hbar\omega-E_{\text{opt}}}$ at the lower band edge $E_{\text{opt}}$ for all $|\Delta| < 2t$. Figure \[fig6\] shows the reduced optical conductivity $C_{\eta}[\omega \sigma_1(\omega)]$ for $U=40t$ and $\Delta=0.6t$. The spectrum looks very similar to the spectrum for $U/t \rightarrow \infty$, which is again shown in this figure. A finite-size-scaling analysis shows however that the strong central peak is not a $\delta$-function for $U=40t$ but a narrow band with at least one singularity diverging as $\eta^{-0.8}$. The spectra in Fig. \[fig6\] are made of three bands: the gaps between the bands appear as local minima on both sides of the central peak because of the relatively large broadening used ($\eta/t=0.1$). The finite-size scaling analysis confirms the existence of these gaps. For decreasing parameters $U$ or $|\Delta|$ first the lower gap, then the upper gap close. Therefore, the number of bands in the optical spectrum of the Mott-Peierls insulator is not constant but depend on the interaction parameters $U$ and $\Delta$. ![ \[fig6\] Reduced optical conductivity $C_{\eta}[\omega \sigma_1(\omega)]$ calculated with DDMRG \[see Eq. (\[Crsigma\])\] on a 128-site lattice ($\eta/t=0.1$) in the strong-coupling regime for $\Delta=0.6t$ as a function of $\hbar \Omega= \hbar \omega-U$. Note the logarithmic scale of the ordinate axis. ](fig6.eps){width="8cm"} The evolution of the optical conductivity as a function of $U$ is very interesting. For decreasing $U/t$ one observes that the central peak breaks into two peaks appearing as local maxima in the broadened spectrum of finite-size systems. The first peak (at the lowest energy) takes over most of the optical weight of the central peak. Its weight decreases progressively with decreasing $U/t$ but remains strong even for small $U$. In Fig. \[fig7\], it is clearly visible (at $\hbar \omega > 4t$) even for $U=2t$ (with $\Delta=0.6t$). This peak corresponds to a power-law divergence within a band with an exponent that tends to $-1/2$ for $U \rightarrow 0$. The peak position moves to lower energy as $U$ decreases and reaches $\hbar \omega = 4t$ for $U=0$. Therefore, the central peak observed at strong coupling $U \gg t$ corresponds to the upper square-root divergence in the Peierls insulator spectrum (\[peierlsOC\]). (In Fig. \[fig7\] this divergence is barely visible as a local maximum at $\hbar \omega \approx 4t$ because of the relatively large broadening $\eta/t =0.2$ used.) The second peak has very little optical weight (it is not visible in Fig. \[fig7\]) and I have not been able to determine its structure. ![ \[fig7\] Optical conductivity $\sigma_1(\omega)$ calculated with DDMRG on a 64-site lattice ($\eta/t=0.2$) for $\Delta=0.6t$ and several values of $U$. ](fig7.eps){width="8cm"} For $U \gg t$ the optical weight is distributed symmetrically around the central peak, as seen in Fig. \[fig6\]. As $U$ decreases, there is a progressive transfer of optical weight from high frequency (above the central peak) to low frequency (below the central peak), see Fig. \[fig7\]. The high-frequency spectrum becomes very weak for small $U/t$ but it completely disappears only at $U=0$. It is difficult to determine the spectral width for general parameters because the optical conductivity is very weak and vanishes smoothly at high frequency. I estimate that the width of the spectrum lies between $4t$ and $8t$ for $U > 0$. The smallest width is reached for large $\Delta$ and small $U$, the largest for small $\Delta$ and large $U$. The low-frequency spectrum becomes stronger as $U$ diminishes. The local maximum below the central peak (seen in Fig. \[fig6\]) progressively rises, moves closer to the lower band edge, and transforms into a strong narrow peak, visible in the spectra shown in Fig. \[fig7\] (at $\hbar \omega < 4t$). For small enough $U/t$ this low-energy peak contains more optical weight than the central peak. For $U \rightarrow 0$ the low energy peak becomes the square-root divergence of the Peierls insulator spectrum (\[peierlsOC\]) at the band threshold $E_{\text{opt}}$. For $U>0$, however, my results suggest that the optical spectrum always vanishes smoothly at the optical gap. I think that the low-energy optical spectrum of the Peierls-Hubbard model at weak coupling has a qualitative behavior similar to that of the Hubbard model: as $\hbar \omega - E_{\text{opt}}$ decreases, $\sigma_1(\omega)$ first appears to diverge as $(\hbar \omega - E_{\text{opt}})^{-1/2}$, then goes through a maximum just above the optical gap $E_{\text{opt}}$, and vanishes smoothly for $\hbar \omega \rightarrow E_{\text{opt}}$. For large enough $U$ it is possible to carry out a finite-size scaling analysis similar to the one performs for the Hubbard model (see Sec. \[sec:hubbard\]). Thus it is possible to check that the low-energy spectrum maximum is finite and lies at a higher energy than the optical gap, and to show explicitly that $\sigma_1(\omega) \sim \sqrt{\hbar \omega - E_{\text{opt}}}$ for $\hbar \omega - E_{\text{opt}} \rightarrow 0^+$. ![ \[fig8\] Optical conductivity $\omega C_{\eta}[\alpha(\omega)]$ calculated with DDMRG \[see Eq. (\[Calpha\])\] on a 128-site lattice ($\eta/t=0.1$) in the small gap regime: Mott-Hubbard insulator with $U=3t$ (dashed), Peierls insulator with $\Delta=0.3t$ (dot-dashed), and Peierls-Hubbard insulator with $U=2.3t$ and $\Delta=0.15t$ (solid). The optical gaps are $E_{\text{opt}}= 0.631t$, $0.6t$, and $0.704t$, respectively. ](fig8.eps){width="8cm"} For smaller $U$ it becomes increasingly difficult to distinguish a smooth spectrum with a truncated divergence from a true divergence. For instance, Fig. \[fig8\] shows the low-frequency optical conductivity $\omega C_{\eta}[\sigma_1(\omega)/\omega]$ for $U=2.3t$ and $\Delta=0.15t$ with a broadening $\eta/t=0.1$ ($N =128$ sites). For comparison, I also show the spectra in two limits discussed previously, the Mott-Hubbard insulator \[$\sigma_1(\omega) \sim \sqrt{\hbar \omega - E_{\text{opt}}}\ $\] and the Peierls insulator \[$\sigma_1(\omega) \sim 1/ \sqrt{\hbar \omega - E_{\text{opt}}} \ $\], with similar optical gaps ($E_{\text{opt}}/t=0.6-0.7$) and the same broadening $\eta$. Clearly, the Mott-Peierls insulator spectrum looks like an intermediate case between the spectra observed in both limiting cases. The position of the maximum in the Mott-Peierls insulator spectrum tends to $0.78t$ for $\eta(N) \rightarrow 0$ while the charge gap (and optical gap) equals $0.704t$ in the thermodynamic limit. Certainly, there is no divergence in the low-energy optical spectrum. However, the maximum seems to diverge as $1/\sqrt{\eta}$ even for the smallest broadening I have used ($\eta/t = 0.05$). Thus this spectrum seems to be qualitatively similar to a function such as Eq. (\[typicalOC\]), but the maximum is so close to the optical gap that broadenings $\eta$ significantly smaller than $0.05t$ (i.e., system sizes much larger than $N =256$) would be necessary to reach the asymptotic regime as discussed in Sec. \[sec:scaling\]. For the same reason it is not possible to determine how the spectrum vanishes for $\hbar\omega - E_{\text{opt}} \rightarrow 0^+$ in such a case In the Hubbard model it is possible to confirm the absence of a singularity and the square-root vanishing at the band threshold even if the optical gap is as small as $E_{\text{opt}}=0.4t$, because we know the optical spectrum of an infinite system for $E_{\text{opt}} \rightarrow 0$ from field theory. [@eric] The field theory approach does not only apply to the Hubbard model, but more generally, gives the low-energy optical spectrum of one-dimensional Mott insulators with small Mott gaps. [@controzzi; @fabian] The different spectral functions depend only on an interaction parameter $\beta^2 \leq 1$. In addition, the optical gap $E_{\text{opt}}>0$ and a normalization constant set the frequency scale and the conductivity scale. For $1/2 < \beta^2 \leq 1$ these optical spectra described truncated square-root divergence with a square-root vanishing at the band threshold as in Eq. (\[typicalOC\]). For $\beta^2=1$ (Hubbard model) the spectrum has the shape shown in Fig. \[fig3\] with a maximum at $1.24 E_{\text{opt}}$. As $\beta^2$ decreases the peak becomes sharper and the maximum moves closer to the band edge. For $\beta^2=1/2$ the optical spectrum is similar to that of a Peierls band insulator with a square-root divergence at the band threshold. ![ \[fig9\] Optical conductivity $\sigma_1(\omega)$ of a Mott-Peierls insulator calculated with DDMRG on a 128-site lattice ($\eta/t=0.1$) for $U=2.3t$ and $\Delta=0.15t$ (dashed) and the fitted field-theoretical spectrum for Mott insulators with $\beta^2=0.58$ and the same broadening $\eta/t$ (solid). ](fig9.eps){width="8cm"} Therefore, this field theory [@fabian; @controzzi] can describe the optical spectrum in both limiting cases (Mott-Hubbard and Peierls insulators) of the Peierls-Hubbard model in the small gap regime, and the field-theoretical spectrum evolves continuously from one limit to the other with $\beta^2$ going from 1 to 1/2. Using $1 \geq \beta^2 \geq 1/2$, the optical gap $E_{\text{opt}}$, and the normalization constant as fit parameters, I have compared the low-energy optical conductivity calculated with DDMRG for small gaps ($E_{\text{opt}} < 0.71t$) to field-theoretical spectra with similar broadening as explained in Sec. \[sec:scaling\]. For instance, I show in Fig. \[fig9\] the DDMRG spectrum for the lattice model (\[hamiltonian\]) with $U=2.3t$ and $\Delta=0.15t$ and the fitted field-theoretical spectrum with $\beta^2=0.58$. Both spectra agrees up to $\hbar \omega = 1.2t \approx 1.7 E_{\text{opt}}$. Generally, I have found that the optical spectrum of a Mott-Peierls insulator can be fitted by a field-theoretical spectrum with $\beta^2 > 1/2$ over a finite frequency range, from $\omega =0$ to a frequency $\omega$ which lies between the position of the low-energy maximum and $2E_{\text{opt}}/\hbar$. (Naturally, for $U\rightarrow 0$ the best fit is always obtained with $\beta^2=1/2$.) Therefore, I think that for any $U>0$ (and $0 \leq |\Delta| < 2t$) the optical spectrum vanishes as $\sqrt{\hbar \omega - E_{\text{opt}}}$ for $\hbar \omega - E_{\text{opt}} \rightarrow 0^+$. Note that I do not assume that the field-theoretical calculations in Refs.  and  are also valid for the Peierls-Hubbard model with general interaction parameters. Actually, there are visible discrepancies starting at rather low energy between field theory and DDMRG results for the lattice model, as shown in Fig. \[fig9\]. The agreement between field theory and DDMRG results in the region of the band threshold simply means that the optical spectra in Mott-Peierls insulators and in one-dimensional Mott insulators have similar shapes just above the optical gap. Finally, it is interesting to examine the evolution of the optical spectrum from weak to strong bond alternation for fixed $U$. It has been shown in Ref.  (see also Sec. \[sec:hubbard\]) that for $\Delta=0$ (Hubbard model) the spectrum consists of a single band with a maximum close to the lower band edge and a tiny peak in the center (at least for $U \geq 4t$). If $|\Delta|$ increases one observes in Fig. \[fig10\](a) that the maximum moves closer to the optical gap and corresponds to a sharper peak. The optical spectrum still starts as $\sqrt{\hbar \omega - E_{\text{opt}}}$ at the band threshold as discussed above. The central peak, which is too weak to be seen in Fig. \[fig10\](a) for $\Delta=0$, becomes rapidly stronger as $|\Delta|$ increases and is clearly visible for $\Delta=0.4t$. As discussed previously this peak becomes a $\delta$-function in the strong-coupling limit $U\gg t$ and corresponds to the upper square-root divergence of the Peierls insulator if $U$ vanishes. For moderate $|\Delta|$ the ratio between the hopping integrals $r(\Delta) = (t-|\Delta|/2)/(t+|\Delta|/2)$ is not too small and the optical weight is mostly concentrated below the central peak. If this ratio becomes small, however, the central peak becomes the spectrum dominant feature, see Fig. \[fig10\](b). The proportion of the optical weight which is in the central peak increases as $1-r^2(\Delta)$ for $r(\Delta) \rightarrow 0$. A finite-size scaling analysis confirms however that this peak is not a $\delta$-function but is still a power-law divergence within an excitation band. Only in the dimer limit $|\Delta|=2t$ \[$r(\Delta)=0$\] the optical spectrum is made of a single $\delta$-peak, which corresponds to the excitation of Frenkel excitons localized on a dimer. ![ \[fig10\] Optical conductivity $\sigma_1(\omega)$ calculated with DDMRG on a 64-site lattice ($\eta/t=0.2$) for $U=4t$ and several values of $\Delta$. (a) For $r(\Delta) > 0.5$. (b) For $r(\Delta) < 0.5$. ](fig10a.eps "fig:"){width="7cm"} ![ \[fig10\] Optical conductivity $\sigma_1(\omega)$ calculated with DDMRG on a 64-site lattice ($\eta/t=0.2$) for $U=4t$ and several values of $\Delta$. (a) For $r(\Delta) > 0.5$. (b) For $r(\Delta) < 0.5$. ](fig10b.eps "fig:"){width="7cm"} In summary, the optical spectrum of a Mott-Peierls insulator consists of one or more bands with a total spectral width ranging from $4t$ to $8t$. The distinctive features of the spectrum are a square-root vanishing of $\sigma_1(\omega)$ at the lower band edge and a peak due to a power-law singularity around the middle of the spectrum. For strong couplings \[$U \gg t$ and $r(|\Delta|) \ll 1$\] most of the optical weight is in the central peak, while for weak couplings \[$U \ll 4t$ and $r(|\Delta|) > 1/2$\] it is mostly concentrated in a narrow peak just above the optical gap. In the limit of a vanishing gap ($U \rightarrow 0$ and $\Delta \rightarrow 0)$ this narrow peak becomes a Drude peak at $\omega=0$. For intermediate couplings most of the optical weight is distributed over a broad frequency range between the optical gap and the central peak and $\sigma_1(\omega)$ goes through a maximum in this range. The central peak always appears at an energy larger than the bare band width $4t$. For parameters which are realistic for conjugated polymers,[@dionys] most of the optical weight lies below this peak. Therefore, I think that it is not possible to observe such a structure in the optical spectrum of conjugated polymers, because it occurs at a too high energy ($>$ 10 eV) and its intensity is too weak. The optical spectrum of Mott-Peierls insulators is unlike that of Mott-Hubbard and Peierls insulators. However, most of its main features are found in the strong-coupling limit investigated in Ref.  and discussed in Sec. \[sec:strong\]. This suggests that the optical excitations of a Mott-Peierls insulator could be made of a pair a spinless bosonic excitations with opposite charges as in the strong-coupling limit (and in a Mott-Hubbard insulator). Nevertheless, understanding the nature of the system elementary excitations requires the study of additional dynamical properties, such as the one-particle Green’s functions. The DDMRG method will enable us to carry out this further investigation. Conclusion \[sec:conclusion\] ============================= In this paper I have presented a dynamical DMRG method which allows one to calculate the optical conductivity of one-dimensional correlated electron systems on large lattices with great accuracy. The DDMRG approach to the calculation of dynamical properties is essentially an application of the standard DMRG algorithm for ground-state calculations. Therefore, both methods have the same advantages but also the same limitations. In particular, DDMRG will directly benefit from recent and future improvements of DMRG such as the use of additional symmetries. With DDMRG it is possible to calculate the dynamical response of correlated systems with hundreds of sites and particles. Relative errors of the order of $10^{-4}$ can be achieved for the optical spectrum of finite systems with $N \sim 10^2$ sites. Using a finite-size scaling analysis based on a size-dependent broadening of the discrete finite-system spectra, one can then calculate a dynamical spectrum in the thermodynamic limit with a resolution of the order of 1% of the spectral width and even investigate singularities in a continuum. The DDMRG approach can be used for various dynamical quantities, such as dynamical spin-spin correlation functions or single-particle Green’s functions. It can also be applied to other lattice quantum many-body models, in higher dimension, including spin or boson degrees of freedom, or long-range interactions. The correction vector DMRG method has been used to calculate non-linear dynamic response functions, such as third-order dynamical polarizabilities. [@pati] Similarly, the variational principle presented in Sec. \[sec:principle\] can be generalized to dynamical correlation functions describing these non-linear responses. Thus I believe that it is possible to develop an efficient DDMRG method for calculating these quantities. A limitation of the DDMRG approach is the restriction to zero temperature. It would be desirable to extend the variational principle and the DDMRG approach to finite-temperature dynamical properties. The computational resources used by the DDMRG method are relatively modest. For instance, all calculations presented in this paper were carried out on workstations with a single 500 MHz Alpha processor and required less than 1 GByte of memory. It would be easy and very efficient to run DDMRG on a parallel computer as calculations for different frequencies are almost independent. This would permit one to investigate much larger or more complicated systems than in this work. In summary, the DDMRG method and the finite-size scaling technique for dynamical spectra appear extremely accurate and versatile. They provide a powerful approach for investigating the dynamical properties in low-dimensional lattice quantum many-body systems. Finally, it should be kept in mind that the variational principle for dynamical correlation functions and their related excited states is completely independent from the DMRG method. Therefore, it is possible to combine this principle with other variational methods to calculate dynamical properties. For instance, one could build a trial wavefunction $|\psi(\{\lambda_i\})\rangle = R(\{\lambda_i\}) |\psi_0\rangle$, where $R(\{\lambda_i\})$ is an operator depending on a few parameters $\lambda_i$, such that the calculation of $W(\{\lambda_i\})= W_{A,\eta}(\omega,\psi(\{\lambda_i\}))$ reduces to the evaluation of ground-state correlation functions. Then the minimization of $W(\{\lambda_i\})$ with respect to the variational parameters $\{\lambda_i\}$ would give a lower bound and an approximate value of the dynamical spectrum $I_{A}(\omega+ i \eta)$. I am grateful to F. Essler and F. Gebhard for many stimulating conversations. I also acknowledge useful discussions with T.D. Kühner and S.R. White about the correction vector DMRG method. [99]{} S.R. White, **69**, 2863 (1992);  **48**, 10345 (1993). *Density-Matrix Renormalization*, edited by I. Peschel, X. Wang, M. Kaulke, and K. Hallberg (Springer, Berlin, 1999). K.A. Hallberg, **52**, R9827 (1995). T.D. Kühner and S.R. White, **60**, 335 (1999). S.K. Pati, S. Ramasesha, Z. Shuai, and J.L. Brédas, **59** 14827 (1999). T.D. Kühner, S.R. White, and H. Monien, **61**, 12474 (2000). P. Brune, G.I. Japaridze, and A.P. Kampf, cond-mat/0106007 (unpublished). E. Jeckelmann, F. Gebhard, and F.H.L. Essler, **85**, 3910 (2000). F.H.L. Essler, F. Gebhard, and E. Jeckelmann, **64**, 125119 (2001). D. Baeriswyl, D.K. Campbell, and S. Mazumdar, in *Conjugated Conducting Polymers*, edited by H. Kiess, (Springer, Berlin, 1992). E.H. Lieb, **62**, 1201 (1989). S. Ramasesha, S.K. Pati, H.R. Krishnamurthy, Z. Shuai, and J.L. Brédas,  **54**, 7598 (1996). F. Gebhard, K. Bott, M. Scheidler, P. Thomas, and S.W. Koch, Philos. Mag. B **75**, 1 (1997). E.H. Lieb and F.Y. Wu, **20**, 1445 (1968). F. Gebhard, K. Bott, M. Scheidler, P. Thomas, and S.W. Koch, Philos. Mag. B **75**, 47 (1997). D. Controzzi, F.H.L. Essler, and A.M. Tsvelik, **86**, 680 (2001).

--- abstract: 'We study survival properties of inhomogeneous Galton-Watson processes. We determine the so-called branching number (which is the reciprocal of the critical value for percolation) for these random trees (conditioned on being infinite), which turns out to be an a.s. constant. We also shed some light on the way the survival probability varies between the generations. When we perform independent percolation on the family tree of an inhomogeneous Galton-Watson process, the result is essentially a family of inhomogeneous Galton-Watson processes, parametrized by the retention probability $p$. We provide growth rates, uniformly in $p$, of the percolation clusters, and also show uniform convergence of the survival probability from the $n$-th level along subsequences. These results also establish, as a corollary, the supercritical continuity of the percolation function. Some of our results are generalisations of results by Lyons (1992).' author: - 'Erik Broman[^1]' - 'Ronald Meester[^2]' date: 'June 5, 2008' title: 'Survival of inhomogeneous Galton-Watson processes' --- [**AMS subject classification:**]{} 60K37, 60J80, 60K35. Introduction and main results ============================= We start by defining the main object of study in this paper, namely inhomogeneous Galton-Watson processes. Start with a root $o$ and let $L_1$ be the distribution of the (random) number of offspring of the root. Proceed by letting each child (if any) of the root have an i.i.d. number of offspring with distribution $L_2,$ and also let these offsprings be independent of the number of children of the root. Given a sequence $\{L_n\}_{n=1}^{\infty},$ we let $L_n$ be the offspring distribution of every individual of generation $n-1.$ Sometimes we will treat $L_n$ as a random variable rather than as a distribution, this is standard abuse of notation. The root is considered to be generation 0. Observe that if the distributions $\{L_n\}_{n=1}^{\infty}$ all are the same we get a regular Galton-Watson process. Observe also that if ${\mathbb P}(L_n=l_n)=1$ for every $n$ and some sequence of numbers $\{l_n\}_{n=1}^{\infty},$ then we a.s. get a (deterministic) spherically symmetric tree, that is, a rooted tree in which any two vertices in the same generation have the same degree. We denote the random family tree of such an inhomogeneous Galton-Watson process by $T$. We will let $\overline{T}$ be a tree with distribution equal to $T$ conditioned on survival, and we will also let $I \subset \overline{T}$ be the tree that consists of those vertices $x \in \overline{T}$ (and the edges between them) that have infinitely many descendents in $\overline{T}$. We will denote by $T_n$, $\overline{T}_n$ and $I_n$ the number of points in the $n$th generation of $T$, $\overline{T}$ and $I$ respectively. It is well known (see e.g. [@Ly2]) and not hard to see that $I$ is itself the family tree of an inhomogeneous Galton-Watson process; we will use this fact later on. For inhomogeneous Galton-Watson processes we define the survival probability $\theta_n$ from the $n$th generation, that is, $$\theta_n:=\lim_{m \rightarrow \infty}{\mathbb P}(T_m>0 \Big{|} T_n=1).$$ For an infinite tree, a [*cutset*]{} $\pi$ is defined to be a finite set of edges such that every infinite path starting at the origin must contain at least one edge of the cutset. We denote by $\Pi$ the set of all such cutsets. Any infinite tree $S$ has a so-called [*branching number*]{} which is defined as follows. \[def1\] The [*branching number*]{} of an infinite tree $S$ with root $o$ is denoted by ${\rm br} S$ and defined by $${\rm br} S:=\sup\left\{\lambda ; \inf_{\pi \in \Pi} \sum_{e \in \pi} \lambda^{-|e|}>0 \right\}.$$ The branching number is a very important property for trees (see [@R]). For instance it is known that (see [@Lyons90]) the critical density $p_c(S)$ for independent percolation (we are assuming that the reader is familiar with the concept of percolation, otherwise please see [@G]) on $S$ is the reciprocal of the branching number, that is, $$p_c(S)= 1/{\rm br}S.$$ Closely related to the branching number is the [*lower growth number*]{} $\underline{\rm gr} S$ which is defined by $$\underline{\rm gr} S:=\liminf_{n \rightarrow \infty} S_n^{1/n},$$ where $S_n$ denotes the number of vertices in the $n$th generation of $S$. It is not hard to see that we always have ${\rm br} S \leq \underline{\rm gr} S,$ while equality is not always true. It is however well-known that if $S$ is spherically symmetric, then ${\rm br} S = \underline{\rm gr} S$. We start with the following simple survival criterion. This result is essentially contained in Proposition 4.15 of [@Ly2], but we do give a different proof based on even earlier work in [@Agresti]. The reason is that some of the elements in the proof will be used again later in this paper. \[lemma4\] For any inhomogeneous Galton-Watson process with offspring distributions $\{L_n\}_{n=1}^{\infty}$ we have that $$\liminf_{n \rightarrow \infty}({\mathbb E}[T_n])^{1/n}<1 \Rightarrow \lim_{n \rightarrow \infty}{\mathbb P}(T_n>0)=0.$$ Furthermore, if $$\label{eqn11} \sup_n {\mathbb E}[L^2_n]=C_1< \infty$$ and $$\label{eqn12} \inf_n {\mathbb E}[L_n]=C_2>0,$$ then we also have that $$\liminf_{n \rightarrow \infty}({\mathbb E}[T_n])^{1/n}>1 \Rightarrow \lim_{n \rightarrow \infty}{\mathbb P}(T_n>0)>0.$$ Next we have a result concerning the branching number of $\overline{T}$. A priori this is a random variable, but it turns out that ${\rm br} \overline{T}$ is an almost sure constant (under mild conditions). \[thm4\] Consider an inhomogeneous Galton-Watson process with offspring distributions $\{L_n\}_{n=1}^{\infty}$ satisfying (\[eqn11\]) and (\[eqn12\]). Assume also that $$\label{eqn13} \liminf_{n \rightarrow \infty}{\mathbb E}[T_{n}]^{1/n}>1.$$ Then we have that ${\rm br} \overline{T} = \liminf_{n \rightarrow \infty} {\mathbb E}[T_n]^{1/n}$, $[\overline{T}]$-a.s. We make some remarks about this result. 1. In [@Ly2], it is proved that a.s., $${\rm br}\overline{T}=\liminf_{n \rightarrow \infty} {\mathbb E}[T_n]^{1/n}$$ under the assumption that $\sup_{n}||L_n||_{\infty}<\infty$. It is claimed in [@Ly2] that this assumption cannot be weakened much; our results show that if one adds the very natural condition (\[eqn13\]), then in fact one can significantly weaken the assumptions. 2. Naively one might believe that this result would follow from easy arguments. For instance one might try the following approach: Define a new inhomogeneous Galton-Watson tree $T'$ by performing percolation on $T$ with probability for being open equal to $p$. Depending on whether $p$ was smaller or greater than $1/(\liminf E[T_n]^{1/n})$, we get from Proposition \[lemma4\] that $T'$ dies out a.s./survives with positive probability (respectively), concluding the argument. However one then misses the point that the fact that $T'$ survives with positive probability if $p>1/(\liminf E[T_n]^{1/n})$ does not lead to the conclusion that ${\rm brT} \geq \liminf E[T_n]^{1/n}$. Indeed, it is imaginable that with positive probability ${\rm brT}=\liminf E[T_n]^{1/n}-\delta$ for some positive $\delta$ and with positive probability ${\rm brT}=\liminf E[T_n]^{1/n}.$ If this were true, $T$’ would still survive with positive probability for the indicated $p$. The following result about the behaviour of $\theta_n$ will be needed in the proof of Theorem \[thm4\] but is also quite interesting in its own right. It is not to be expected that $\theta_n$ is in general bounded away from 0 since one can always insert any finite number of generations of degree one in the tree. However, it is the case that there is a subsequence along which $\theta_n$ is bounded away from 0. \[lemma5\] Consider an inhomogeneous Galton-Watson process with offspring distributions $\{L_n\}_{n=1}^{\infty}$ satisfying (\[eqn11\]), (\[eqn12\]) and (\[eqn13\]). Then there exists a sequence $\{n_k\}_{k=1}^{\infty}$ of increasing integers and a constant $C>0$ such that for all $k \geq 1$, $$\theta_{n_k} \geq C.$$ Next, we study bond percolation on $I$. Note that $p_c(I)=p_c(\overline{T})$, since pruning a tree does not change its critical probability. We already noted that $I$ itself is the family tree of an inhomogeneous Galton-Watson process, and when we perform independent bond percolation on $I$, the resulting component of the origin, to be denoted by $I^p$, also constitutes a family tree of an inhomogeneous Galton-Watson process. Therefore, general results about inhomogeneous Galton-Watson processes automatically apply to percolation on $I$. However, being equipped with a parameter $p$ now, we will derive survival estimates [*uniformly*]{} in $p$. We remark that a special case of inhomogeneous Galton-Watson processes results from starting with a deterministic spherically symmetric tree and performing percolation on that tree. One more piece of notation: the number of vertices in $I^p$ at distance $n$ from the root is denoted by $I_n^p$. Also, in this paper, we use various coupling constructions. To facilitate this, all processes, for all values of $p$, are jointly constructed in the obvious way. Consequently, as in the previous example, we will express the $p$-dependence in the events rather than in the measure. In light of Theorem \[thm4\], one might expect that for any $\epsilon>0,$ $$\lim_{n \rightarrow \infty}{\mathbb P}(0<I_n^p<((1-\epsilon){\rm br} \overline{I^p} )^n)=0.$$ In fact, we have the next, much stronger statement. \[propunifconv\] Consider an inhomogeneous Galton-Watson process satisfying (\[eqn11\]), (\[eqn12\]) and (\[eqn13\]), with family tree $T$, and let $\epsilon>0.$ For $p_c(\overline{T}) <p_1 \leq 1$ it is the case that $$\lim_{n \rightarrow \infty}{\mathbb P}(0<I_n^p<((1-\epsilon){\rm br}\overline{I^{p_1}})^n)=0,$$ uniformly in $p \in [p_1,1].$ Note that the pointwise (in $p$) convergence in Theorem \[propunifconv\] is almost a triviality. The whole point of the theorem is proving the uniform convergence. Theorem \[propunifconv\] combined with Proposition \[lemma5\] will in turn lead us to our next result. Here we define $$\theta(p):={\mathbb P}(|I^p|=\infty).$$ \[thm2\] Consider an inhomogeneous Galton-Watson process satisfying (\[eqn11\]), (\[eqn12\]) and (\[eqn13\]), and let $p_1>p_c(\overline{T}).$ Then there exists a sequence of increasing integers $\{n_k\}_{k=1}^{\infty}$ such that $$\theta(p)=\lim_{k \to \infty} \P(I_{n_k}^p >0),$$ uniformly on $[p_1,1].$ This result also leads to continuity of the percolation function above $p_c$ for random trees. \[corr1\] Consider an inhomogeneous Galton-Watson process satisfying (\[eqn11\]), (\[eqn12\]) and (\[eqn13\]). Then the percolation function $\theta(p)$ is continuous above $p_c(\overline{T}).$ In particular, on any spherically symmetric tree $S$ with uniformly bounded degrees, the percolation function is continuous above $p_c(S)$. In fact, one can also use Theorem \[thm4\] to construct a more or less classical proof of this result. As an interesting side remark, we mention that the route via Proposition \[thm2\] also has a counterpart on $\mathbb{Z}^d$, and gives a new proof for the continuity of the percolation function in that context. This proof does in fact give a rate of convergence for the natural approximations of the percolation function; we discuss these continuity matters in the last section. In contrast to our last corollary, we have the following example of a tree for which the percolation function is not continuous above $p_c$. To construct such a tree, we use a result in [@Ly2], a special case of which says that there is percolation with positive probability on a spherically symmetric tree $S$ with parameter $p$, if and only if $$\sum_{n=1}^{\infty} \frac{p^{-n}}{S_n} < \infty.$$ To construct an example, we first take a spherically symmetric tree $S$ which is such that $S_n$ is of the order $2^n n^2$. It follows from the above that $p_c(S)=1/2$ and that $\theta_{S}(1/2)>0$. Next, we take a regular tree $S'$ with common degree 4. It is well-known that $p_c(S')=1/3$. We then construct a tree $S''$ by joining the roots of $S$ and $S'$ by a single edge. It is easy to see that $p_c(S'')=1/3$ and that $\theta_{S''}$ is discontinuous at $1/2$. Theorem \[thm4\] along with Propositions \[lemma4\] and \[lemma5\] will be proved in the next section. All the other results are proved in Section \[sec3\]. The issues about continuity of the percolation function are discussed in Section \[sec4\]. Proof of Proposition \[lemma4\], Theorem \[thm4\] and Proposition \[lemma5\] {#secbr} ============================================================================ We start by defining a useful probability generating function by $$h(n,s):=\sum_{j=0}^{\infty}{\mathbb P}(L_n=j)s^j, \ \forall n \geq 1.$$ It is known (see [@Agresti]) that if $h''(n,1)<\infty,$ for every $n,$ then for all $n \geq 1$ we have $$\label{eqn10} \left[{\mathbb E}[T_n]^{-1}+\sum_{j=1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}]^{-1}\right]^{-1} \leq {\mathbb P}(T_n>0).$$ Of course we have $$h'(n,1)=\sum_{j=0}^{\infty}j{\mathbb P}(L_n=j)={\mathbb E}[L_n],$$ and $$h''(n,1)=\sum_{j=0}^{\infty}j(j-1){\mathbb P}(L_n=j)={\mathbb E}[L^2_n]-{\mathbb E}[L_n].$$ We can now proceed with the proof of Proposition \[lemma4\]. [**Proof of Proposition \[lemma4\].**]{} The proof of the first statement is easy. Assume that $$\liminf_{n \rightarrow \infty}({\mathbb E}[T_n])^{1/n}=a<1,$$ then we get that for any $\epsilon>0$ such that $a(1+\epsilon)<1,$ there exists a sequence $\{n_k\}_{k=1}^{\infty}$ such that $${\mathbb P}(T_{n_k}>0) \leq {\mathbb E}[T_{n_k}] \leq (a(1+\epsilon))^{n_k},$$ so that $$\lim_{n \rightarrow \infty}{\mathbb P}(T_n>0)=0.$$ For the second statement we start by observing that condition (\[eqn11\]) gives us that $h''(n,1)={\mathbb E}[L^2_n]-{\mathbb E}[L_n]<\infty$ for every $n.$ Of course this does not require the full statement of equation (\[eqn11\]) which will be needed later. In turn, this gives us that inequality (\[eqn10\]) is valid for every $n,$ and therefore we need to show that $$\begin{aligned} \label{eqn15} \lefteqn{\limsup_{n \rightarrow \infty} \left[ {\mathbb E}[T_n]^{-1}+\sum_{j=1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}]^{-1}\right]^{-1}} \\ & & =\limsup_{n \rightarrow \infty} \left[{\mathbb E}[T_n]^{-1}+\sum_{j=1}^{n} \frac{{\mathbb E}[L^2_j]-{\mathbb E}[L_j]}{{\mathbb E}[L_j]} {\mathbb E}[T_{j}]^{-1}\right]^{-1} >0. \nonumber\end{aligned}$$ To this end, we observe that by equations (\[eqn11\]) and (\[eqn12\]) $$\sup_{j}\frac{{\mathbb E}[L^2_j]-{\mathbb E}[L_j]}{{\mathbb E}[L_j]} \leq \frac{C_1}{C_2}=C<\infty.$$ Since $\liminf_{n \rightarrow \infty}({\mathbb E}[T_n])^{1/n}>1$ there exists a constant $b>1$ and an $N$ such that for all $n \geq N$, $${\mathbb E}[T_n]>b^n.$$ Therefore, for some constant $D<\infty,$ $$\begin{aligned} \lefteqn{ {\mathbb E}[T_n]^{-1}+\sum_{j=1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}]^{-1}} \\ & & \leq {\mathbb E}[T_n]^{-1}+C\sum_{j=1}^{n} {\mathbb E}[T_{j}]^{-1} %\leq D+ C\sum_{j=N}^{n} {\mathbb E}[T_{j}]^{-1} \\ %& & \leq D+C\sum_{j=N}^{n} b^{-j} \leq D+C\sum_{j=N}^{\infty} b^{-j}< \infty.\end{aligned}$$ Since the right hand side of the above inequality is independent of $n,$ inequality (\[eqn15\]) is valid and that concludes the proof. We continue by proving Proposition \[lemma5\]. [**Proof of Proposition \[lemma5\].**]{} Let $\{X_i\}_{i\geq 1}$ be i.i.d. with distribution according to $T_n$ conditioned on the event that $T_{\ell}=1.$ Observe that for $n \geq \ell$ $$T_n=\sum_{k=1}^{T_{\ell}}X_k,$$ so that (using Wald’s lemma) $$\label{eqn16} {\mathbb E}[X_1]={\mathbb E}[T_n \Big{|} T_{\ell}=1]=\frac{{\mathbb E}[T_n]}{{\mathbb E}[T_{\ell}]}.$$ Observe that by condition (\[eqn11\]) we can use inequality (\[eqn10\]) to conclude that for $n \geq \ell,$ $$\label{eqn19} \left[{\mathbb E}[T_n\Big{|} T_{\ell}=1]^{-1} +\sum_{j=\ell+1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}\Big{|} T_{\ell}=1]^{-1}\right]^{-1} \leq {\mathbb P}(T_n>0\Big{|} T_{\ell}=1).$$ We will show that there exists a sequence $\{n_k\}_{k=1}^{\infty}$ of increasing integers and a constant $C<\infty$ such that for all $k \geq 1$ and for all $n \geq n_k$, $$\label{eqn13a} {\mathbb E}[T_n\Big{|} T_{n_k}=1]^{-1} +\sum_{j=n_k+1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}\Big{|} T_{n_k}=1]^{-1}\leq C.$$ This will give us that for all $k \geq 1$ we have $$\lim_{n \rightarrow \infty} {\mathbb P}(T_n>0\Big{|} T_{n_k}=1) \geq \frac{1}{C},$$ proving the lemma. To that end, observe that as in the proof of Lemma \[lemma4\] there exists a constant $C_3$ such that for $ n \geq \ell$, $$\begin{aligned} \label{eqn18} \lefteqn{{\mathbb E}[T_n\Big{|} T_{\ell}=1]^{-1} +\sum_{j=\ell+1}^{n}\frac{h''(j,1)}{h'(j,1)} {\mathbb E}[T_{j}\Big{|} T_{\ell}=1]^{-1}} \\ & & \leq {\mathbb E}[T_n\Big{|} T_{\ell}=1]^{-1} \nonumber +C_3\sum_{j=\ell+1}^{n} {\mathbb E}[T_{j}\Big{|} T_{\ell}=1]^{-1} \\ & & \leq (C_3+1)\sum_{j=\ell+1}^{n} {\mathbb E}[T_{j}\Big{|} T_{\ell}=1]^{-1} \nonumber =(C_3+1){\mathbb E}[T_{\ell}]\sum_{j=\ell+1}^{n} \frac{1}{{\mathbb E}[T_{j}]},\end{aligned}$$ where we use equation (\[eqn16\]) in the last equality. Therefore, showing that there exists a sequence $\{n_k\}_{k=1}^{\infty}$ of increasing integers and a constant $C<\infty$ such that for all $k$ we have $${\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]} \leq C$$ will give us equation (\[eqn13a\]). We divide the proof into three cases. First however, define $$m:=\liminf_{n \rightarrow \infty}{\mathbb E}[T_{n}]^{1/n}>1.$$ In the first case, we have that ${\mathbb E}[T_{n}]^{1/n}<m$ for infinitely many $n.$ We can then conclude that there exists $n_1$, defined to be the largest integer such that ${\mathbb E}[T_{n_1}]^{1/n_1}=\min_{n \geq 1} {\mathbb E}[T_{n}]^{1/n}.$ Having defined $n_k,$ we can then define $n_{k+1}$ to be the largest integer greater than $n_k$ such that ${\mathbb E}[T_{n_{k+1}}]^{1/n_{k+1}}=\min_{n >n_k} {\mathbb E}[T_{n}]^{1/n}.$ Let $\epsilon_k$ be defined through ${\mathbb E}[T_{n_k}]^{1/n_k}=m(1-\epsilon_k).$ Observe that by definition of $n_k,$ ${\mathbb E}[T_{n}]^{1/n}\geq m (1-\epsilon_k)$ for every $n \geq n_k$ and also that $\epsilon_k >0$ for every $k,$ and finally that $\epsilon_k \rightarrow 0,$ as $ k \rightarrow \infty.$ Therefore, $$\begin{aligned} \lefteqn{{\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]} \leq {\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty}\frac{1}{(m(1-\epsilon_k) )^{j}}} \\ & & = (m(1-\epsilon_k) )^{n_k}\sum_{j=1}^{\infty}\frac{1}{(m(1-\epsilon_k) )^{n_k+j}} =\sum_{j=1}^{\infty}\frac{1}{(m(1-\epsilon_k) )^{j}}.\end{aligned}$$ There exists a $K$ such that $m(1-\epsilon_k)>1$ for every $k \geq K.$ For $k \geq K,$ the right hand side of the above equation is then bounded by some constant $D_k< \infty.$ Furthermore, we can take $D_k \geq D_{k+1}$ and conclude that for all $k \geq K$, $${\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]}\leq D_K < \infty.$$ For the second and third case, we have that ${\mathbb E}[T_{n}]^{1/n}<m$ for only finitely many $n.$ We can therefore find $N$ large enough so that ${\mathbb E}[T_{n}]^{1/n}\geq m$ for every $n \geq N.$ We have that for every $n,$ ${\mathbb E}[T_{n}]^{1/n}=m(1+a(n)),$ where the sequence of numbers $\{a(n)\}_{n=1}^{\infty}$ is such that $a(n) \geq 0$ for every $n \geq N.$ The second case is if $\liminf_{n \rightarrow \infty}(1+a(n))^n=C_4$ for some constant $C_4< \infty.$ Then there exists a sequence of strictly increasing integers $\{n_k\}_{k=1}^{\infty}$ such that $(1+a(n_k))^{n_k} \leq 2C_4$ for every $k \geq 1.$ By also requiring that $n_1 \geq N,$ we get that $$\begin{aligned} \lefteqn{{\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]} \leq {\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{m^{j}}} \\ & & =m^{n_k}(1+a(n_k))^{n_k}\sum_{j=1}^{\infty} \frac{1}{m^{n_k+j}} \leq 2C_4\sum_{j=1}^{\infty} \frac{1}{m^{j}} < \infty.\end{aligned}$$ The third case is if $\lim_{n \rightarrow \infty}(1+a(n))^n=\infty.$ We can then find a sequence $\{n_k\}_{k=1}^{\infty}$ (much as in the first case) such that for every $k \geq 1,$ $(1+a(n))^n \geq (1+a(n_k))^{n_k}$ for every $n \geq n_k.$ By again requiring that $n_1 \geq N,$ we get that $$\begin{aligned} \lefteqn{{\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]} \leq {\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty}\frac{1}{(m(1+a(n_k)))^{j}}} \\ & & =\frac{(m(1+a(n_k)))^{n_k}} {(m(1+a(n_k)))^{n_k}}\sum_{j=1}^{\infty}\frac{1}{(m (1+a(n_k)))^{j}} \leq \sum_{j=1}^{\infty}\frac{1}{m^{j}} < \infty.\end{aligned}$$ We can therefore conclude that there exists a constant $C=C(\{L_n\}_{n=1}^{\infty})<\infty$ and a sequence of strictly increasing integers $\{n_k\}_{k=1}^{\infty}$ such that for all $k \geq 1$, $${\mathbb E}[T_{n_k}]\sum_{j=n_k+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]}\leq C.$$ This concludes the proof. We are now ready to prove Theorem \[thm4\]. [**Proof of Theorem \[thm4\].**]{} Using that $p_c(I)^{-1}=p_c(\overline{T})^{-1}={\rm br \overline{T}}$, we need to show that $$p_c(I)^{-1}= \liminf_{n \rightarrow \infty} ({\mathbb E}[T_n])^{1/n}.$$ We will do this by first proving that $p_c(I)^{-1}= \liminf_{n \rightarrow \infty} (\theta_{n}{\mathbb E}[T_n])^{1/n}$ and then proving that $$\ \liminf_{n \rightarrow \infty} (\theta_{n}{\mathbb E}[T_n])^{1/n} =\liminf_{n \rightarrow \infty} {\mathbb E}[T_n]^{1/n}.$$ Consider the offspring distribution $L'_1$ of the root of $I.$ Let $T^i$ be the tree consisting of child number $i\in\{1,\ldots,L_1\}$ of the root of $T$ and all the descendents of this child. Define also $N_{1,\infty}=|\{T^i: |T^i|=\infty, i=1,\ldots,L_1\}|.$ It is not hard to see that for $k \geq 1,$ $${\mathbb P}(L'_1=k)={\mathbb P}(N_{1,\infty}=k \Big{|} N_{1,\infty} \geq 1) =\frac{{\mathbb P}(N_{1,\infty}=k)}{\theta}.$$ Furthermore, letting $Y_i$ be i.i.d. ${\rm Bin}(1,\theta_1)$ random variables and using Wald’s lemma we get that $${\mathbb E}[L'_1]=\frac{1}{\theta}{\mathbb E}[N_{1,\infty}] =\frac{1}{\theta}{\mathbb E} \left[\sum_{i=1}^{L_1} Y_i\right] =\frac{1}{\theta}{\mathbb E} \left[Y_1\right]{\mathbb E} \left[L_1\right] =\frac{\theta_1}{\theta}{\mathbb E}[L_1].$$ Furthermore, this argument holds for any generation $n$ and therefore we have for all $ n \geq 1$, $$\label{generations} {\mathbb E}[L'_n]=\frac{\theta_n}{\theta_{n-1}}{\mathbb E}[L_n].$$ Now, perform independent percolation on $I$ with parameter $p,$ thus creating a random graph that we denote by ${\cal I}^p.$ Recall that $I^p$ is the component of the root of this graph. Obviously, $I^p$ is the family tree of an inhomogeneous Galton-Watson process with some offspring distributions $\{L''_n\}_{n=1}^{\infty}.$ Furthermore, trivially $${\mathbb E}[L''_n]=p{\mathbb E}[L'_n]=p\frac{\theta_n}{\theta_{n-1}}{\mathbb E}[L_n] \ \ \forall n\geq 1.$$ Recall that $I_n^p$ is the number of vertices in $I^p$ at distance $n$ from the root and recall that we defined $I_n$ similarly. We have, using a standard result from the theory of branching processes and (\[generations\]), that $${\mathbb E}[I_n^p]=p^n{\mathbb E}[I_n] =p^n \prod_{i=1}^{n}{\mathbb E}[L'_i] =p^n \prod_{i=1}^{n}\frac{\theta_i}{\theta_{i-1}}{\mathbb E}[L_i] =p^n \frac{\theta_{n}}{\theta}{\mathbb E}[T_n].$$ Therefore, $$\label{eqn14} \liminf_{n \rightarrow \infty}({\mathbb E}[I_n^p])^{1/n} =p\liminf_{n \rightarrow \infty}(\frac{\theta_{n}}{\theta}{\mathbb E}[T_n])^{1/n} =p\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}.$$ We would like to use Proposition \[lemma4\] and Proposition \[lemma5\] on $I^p.$ However, before we can do that we need to show that the offspring distributions $\{L''_n\}_{n=1}^{\infty}$ satisfies conditions (\[eqn11\]) and (\[eqn12\]). When we use Proposition \[lemma5\] we will assume that condition (\[eqn13\]) is satisfied; the details will become clear. For some vertex $x$ in generation $n-1,$ let $T^i_x$ be the tree consisting of child number $i\in\{1,\ldots,L_n\}$ of $x$ and all the descendents of this child. Define $N_{n,\infty}=|\{T^i_x: |T^i_x|=\infty, i=1,\ldots,L_n\}|,$ and observe that the distribution of this random variable is trivially independent of the specific choice of $x$ in generation $n-1.$ Let $Y_i^n$ be i.i.d. ${\rm Bin}(1,\theta_n)$ and observe that $$\begin{aligned} {\mathbb E}[(L''_n)^2] & \leq & {\mathbb E}[(L'_n)^2]=\sum_{j=1}^{\infty}j^2{\mathbb P}(L'_n=j) \\ & = & \sum_{j=1}^{\infty}j^2{\mathbb P}(N_{n,\infty}=j \Big{|}N_{n,\infty} \geq 1 ) \\ & = & \frac{{\mathbb E}[N_{n,\infty}^2]}{\theta_{n-1}} =\frac{{\mathbb E}\left[{\mathbb E}\left[\left(\sum_{i=1}^{L_n}Y_i^n\right)^2 \Big{|} L_n\right]\right]}{\theta_{n-1}} \\ & \leq & \frac{{\mathbb E}\left[{\mathbb E}\left[L_n \sum_{i=1}^{L_n}(Y_i^n)^2 \Big{|} L_n\right]\right]}{\theta_{n-1}} \\ & = & \frac{{\mathbb E}\left[L_n \sum_{i=1}^{L_n}{\mathbb E}\left[Y_i^n \Big{|} L_n\right]\right]}{\theta_{n-1}} =\frac{{\mathbb E}\left[L_n^2 \theta_n \right]}{\theta_{n-1}}.\end{aligned}$$ In the second inequality we use that for any real numbers $a_1, \ldots, a_n$ we have that $(a_1+\cdots + a_n)^2 \leq n(a_1^2+\cdots+a_n^2).$ Obviously we must also have that $$\theta_{n-1} \geq {{\mathbb P}(L_n>0)\theta_{n}},$$ and we can use Cauchy-Schwarz to see that $${\mathbb E}[L_n]^2= {\mathbb E}[L_n I_{\{L_n>0\}}]^2\leq {\mathbb P}(L_n>0) {\mathbb E}[L_n^2].$$ Therefore, $$\frac{{\mathbb E}\left[L_n^2 \theta_n \right]}{\theta_{n-1}} \leq \frac{{\mathbb E}\left[L_n^2 \right]}{{\mathbb P}(L_n>0)} \leq {\mathbb E}\left[L_n^2 \right] \frac{{\mathbb E}\left[L_n^2 \right]}{{\mathbb E}\left[L_n \right]^2} \leq \frac{C_1^2}{C_2^2}<\infty.$$ Furthermore $$\inf_{n}{\mathbb E}[L''_n]=p\inf_{n}{\mathbb E}[L'_n] \geq p,$$ since ${\mathbb E}[L'_n] \geq 1$ for every $n.$ We can now proceed to use Proposition \[lemma4\] with equation (\[eqn14\]) to see that $I^p$ survives with positive probability if $p>(\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n})^{-1}$ while it dies out a.s. if $p<(\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n})^{-1}.$ This is not quite enough for our purposes: it could be the case that with positive probability, $I$ is such that $I^p$ a.s. dies out. Since we want to make a statememt about almost all trees $I$, we argue that in fact, if $p>1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}$, then ${\cal I}^p$ contains an infinite component with probability 1 as our next argument shows. Assume therefore that $\liminf_{n \rightarrow \infty}({\mathbb E}[I^p_n])^{1/n}= p\liminf_{n \rightarrow \infty}(\theta_n{\mathbb E}[T_n])^{1/n}>1.$ This is condition (\[eqn13\]) for $I^p.$ Construct the tree $I^p$ by letting $I^p_1$ have distribution equal to $L_1''.$ Proceed by letting $I^p_2$ be the sum $\sum_{i=1}^{I^p_1}L''_{2,i},$ where $\{L''_{2,i}\}_{i=1}^{\infty}$ are i.i.d. with distribution equal to $L_2''$ and let them also be independent of everything else. Continuing in this fashion, we have two possibilities. First we may find that $I_n^p>0$ for every $n.$ Second, we might instead find that for some $n$, we have $I_n^p=0.$ If this is the case, there exists some integer $n_{k_1} >n$ in the subsequence dictated by Proposition \[lemma5\]. However, since $I$ is infinite we must have that ${\cal I}^p$ contains a subtree (possibly consisting of only one vertex ) with the root being some vertex at level $n_{k_1}.$ Construct this subtree in the same way as we constructed $I^p$ above. This subtree has some probability to survive which is by Proposition \[lemma5\] uniformly bounded away from 0. It is also easy to see that the event of survival of this subtree is conditionally independent of the part of ${\cal I}^p$ examined so far (up to generation $n$). If again we find that this subtree is finite, we continue in the same way. Since all the subtrees that we pick have uniformly positive probability to survive by Proposition \[lemma5\] and the survival of them are conditionally independent we see that ${\cal I}^p$ must contain an infinite component with probability 1. We therefore conclude that $${\mathbb P}({\cal I}^p \textrm{ has an infinite component}) =\left\{ \begin{array}{cc} 1, & p>1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}, \\ 0, & p < 1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}. \end{array} \right.$$ This is the same as saying that for almost every $I,$ we will after performing percolation with parameter $p$ on $I$ a.s. get an infinite component if $p>1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}$ while we will a.s. not get an infinite component if $p<1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}.$ It follows that for almost every $I$ the probability that the component of the root is infinite is positive if $p>1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}$ while it is 0 if $p<1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}.$ This gives us that $p_c(I)=1/\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}$ from which it follows that ${\rm br}I=\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}$ (recall that $p_c(I)=1/{\rm br}I$). We now proceed with the final step in proving that $$\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n}=\liminf_{n \rightarrow \infty}{\mathbb E}[T_n]^{1/n}.$$ Obviously, $\theta_{n}{\mathbb E}[T_n] \leq {\mathbb E}[T_n]$ for every $n,$ so we only need to show that $$\liminf_{n \rightarrow \infty}(\theta_{n}{\mathbb E}[T_n])^{1/n} \geq \liminf_{n \rightarrow \infty}{\mathbb E}[T_n]^{1/n}.$$ As before, let $m=\liminf_{n \rightarrow \infty}{\mathbb E}[T_n]^{1/n}>1$ and choose $\epsilon>0,$ so that $m (1-\epsilon)>1.$ Furthermore, we can choose an $N$ such that ${\mathbb E}[T_n]^{1/n} \geq m (1-\epsilon)$ for every $n \geq N.$ Using inequalities (\[eqn19\]) and (\[eqn18\]) we get that for some constant $C$ and $m \geq n,$ $${\mathbb P}(T_m>0 \Big{|} T_n=1) \geq \left[C{\mathbb E}[T_{n}]\sum_{j=n+1}^{m} \frac{1}{{\mathbb E}[T_{j}]}\right]^{-1} \geq \left[C{\mathbb E}[T_{n}]\sum_{j=n+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]}\right]^{-1}.$$ Therefore, for $n \geq N,$ $$\begin{aligned} \lefteqn{\theta_n {\mathbb E}[T_{n}]= \lim_{m \rightarrow \infty}{\mathbb P}(T_m>0 \Big{|} T_n=1){\mathbb E}[T_{n}]} \\ & & \geq \left[C\sum_{j=n+1}^{\infty} \frac{1}{{\mathbb E}[T_{j}]}\right]^{-1} \geq \left[C\sum_{j=n+1}^{\infty} \frac{1}{(m(1-\epsilon))^{j}}\right]^{-1} \\ & & =\left[\frac{C}{(m(1-\epsilon))^n}\sum_{j=1}^{\infty} \frac{1}{(m(1-\epsilon))^{j}}\right]^{-1} =(m(1-\epsilon))^n C',\end{aligned}$$ where $C'>0.$ Therefore, for all $n \geq N$ we have $$(\theta_n {\mathbb E}[T_{n}])^{1/n} \geq m(1-\epsilon) C'^{1/n},$$ so that $$\liminf_{n \rightarrow \infty}(\theta_n {\mathbb E}[T_{n}])^{1/n} \geq m(1-\epsilon).$$ Since $\epsilon>0$ can be choosen arbitrarily small, we are done. [**Remark**]{} In fact, the proof of Theorem \[thm4\] shows that if the family tree $T$ of a Galton-Watson process satisfies (\[eqn11\]), (\[eqn12\]) and (\[eqn13\]), and $p>p_c(\overline{T})$ , then so does the family tree associated with the $I^p$ process. Proof of Theorem \[propunifconv\] and Proposition \[thm2\] {#sec3} ========================================================== Before we can prove Theorem \[propunifconv\], we need the following domination lemmas. The first one appears (without proof) in [@BHS]. The proof we give is due to Olle Häggström (unpublished). \[lem:conditioned\_binomial\] For $k \geq 1$, $p \in (0,1)$ and $0 \leq m \leq k$, write $\rho_{k,p,m}$ for the distribution of a Binomial$(k,p)$ random variable conditioned on taking value at least $m$. For $p_1 \leq p_2$, we have $$\rho_{k,p_1,m} \, \preceq \, \rho_{k,p_2,m} \, ,$$ where $\preceq$ denotes stochastic domination. [**Proof.**]{} For $i=1,2$, let $Y_i$ be a Bin$(k, p_i)$ random variable, and let $X_i$ be a random variable with distribution $\rho_{k, p_i, m}$. Since $x/(1-x) < y/(1-y)$ for $0 < x < y <1$, it is enough to show that for any $n\in\{m+1, \ldots, k\}$ we have $$\frac{{\mathbb P}(X_1 \geq n)}{{\mathbb P}(X_1 < n)} \, \leq \, \frac{{\mathbb P}(X_2 \geq n)}{{\mathbb P}(X_2 < n)},$$ which is the same as showing that $$\label{eq:need_to_show} \frac{{\mathbb P}(X_2 \geq n)}{{\mathbb P}(X_1 \geq n)} \cdot \frac{{\mathbb P}(X_1 < n)}{{\mathbb P}(X_2 < n)} \, \geq \, 1 \, .$$ Writing $Z_1$ and $Z_2$ for the probabilities that $Y_1 \geq m$ and $Y_2 \geq m$, respectively, the left-hand side of (\[eq:need\_to\_show\]) becomes $$\label{eq:first_rewrite} \frac{\frac{1}{Z_2} \sum_{j=n}^k {k \choose j} p_2^j(1-p_2)^{k-j}}{\frac{1}{Z_1} \sum_{j=n}^k {k \choose j} p_1^j(1-p_1)^{k-j}} \cdot \frac{\frac{1}{Z_1} \sum_{j=m}^{n-1} {k \choose j} p_1^j(1-p_1)^{k-j}}{\frac{1}{Z_2} \sum_{j=m}^{n-1} {k \choose j} p_2^j(1-p_2)^{k-j}} \, .$$ Cancelling the $Z_i$’s and introducing the notation $\phi_i=\frac{p_i}{1-p_i}$ for $i=1,2$, the expression in (\[eq:first\_rewrite\]) may further be rewritten as $$\begin{aligned} \nonumber \lefteqn{ \mbox{ } \hspace{-20mm} \frac{p_2^n(1-p_2)^{k-n} \sum_{j=n}^k {k \choose j} \phi_2^{j-n}}{p_1^n(1-p_1)^{k-n} \sum_{j=n}^k {k \choose j} \phi_1^{j-n}} \cdot \frac{p_1^n(1-p_1)^{k-n} \sum_{j=m}^{n-1} {k \choose j} \phi_1^{j-n}}{p_2^n(1-p_2)^{k-n} \sum_{j=m}^{n-1} {k \choose j} \phi_2^{j-n}} = } \\ & = & \frac{ \sum_{j=n}^k {k \choose j} \phi_2^{j-n}}{ \sum_{j=n}^k {k \choose j} \phi_1^{j-n}} \cdot \frac{ \sum_{j=m}^{n-1} {k \choose j} \phi_1^{j-n}}{ \sum_{j=m}^{n-1} {k \choose j} \phi_2^{j-n}} \, . \label{eq:second_rewrite}\end{aligned}$$ Now note that $\phi_1 \leq \phi_2$, so that $$\sum_{j=n}^k {k \choose j} \phi_2^{j-n} \, \geq \, \sum_{j=n}^k {k \choose j} \phi_1^{j-n}$$ and $$\sum_{j=m}^{n-1} {k \choose j} \phi_1^{j-n} \, \geq \, \sum_{j=m}^{n-1} {k \choose j} \phi_2^{j-n} \, .$$ Hence, the expression in (\[eq:second\_rewrite\]) is greater than or equal to $1$, so (\[eq:need\_to\_show\]) is verified and the lemma is established. We proceed with the following lemma. We will in fact only use it in the case $m=1$, but we nevertheless provide a proof of the general statement. \[more\_bin\_cond2\] In the notation of Lemma \[lem:conditioned\_binomial\], it is the case that for any $1\leq k \leq l$ and $0 \leq m \leq k$ $$\rho_{k,p,m} \preceq \rho_{l,p,m},$$ for all $0 < p < 1.$ [**Proof.**]{} It is obvious that we only need to prove the lemma in the case $l=k+1.$ Therefore, let $Y_1,\ldots, Y_{k+1}$ and $X_1,\ldots, X_k$ be i.i.d. Bernoulli random variables with expectation $p$ and let $Y=\sum_{i=1}^{k+1} Y_i$ and $X=\sum_{j=1}^k X_j$. We need to show that $\P(X \geq n|X \geq m) \leq \P(Y \geq n|Y \geq m)$, for all $n=m,m+1,\ldots, k$. To this end we write $$\begin{aligned} \P(Y \geq n|Y\geq m) &=& \P(Y\geq n|Y\geq m, Y_{k+1}=0) \P(Y_{k+1}=0|Y\geq m)\\ & & + \, \P(Y \geq n |Y\geq m, Y_{k+1}=1) \P(Y_{k+1}=1|Y\geq m)\\ &=& \P(X \geq n | X\geq m)\P(Y_{k+1}=0|Y\geq m) \\ & & + \, \P(X \geq n-1| X \geq m-1 )\P(Y_{k+1}=1|Y\geq m).\end{aligned}$$ Therefore, we need to show that for $n >m$, $$\P(X \geq n-1| X \geq m-1 ) \geq \P(X \geq n| X \geq m),$$ or equivalently, $$\P(X \geq n| X \geq n-1 ) \leq \P(X \geq m| X \geq m-1 ).$$ It is easy to see that it suffices to prove this for $m=n-1,$ or to simplify notation, to show that $$\P(X \geq n+1| X \geq n ) \leq \P(X \geq n| X \geq n-1 ).$$ Since $$\P(X \geq n+1| X \geq n )=1-\P(X =n| X \geq n ),$$ we need to show that $$\P(X =n-1| X \geq n-1 ) \leq \P(X =n| X \geq n ).$$ Writing $p_n:=\P(X=n)$ we rewrite this as $$\frac{p_n+\cdots+p_k}{p_{n-1}+\cdots+p_k} \leq \frac{p_n}{p_{n-1}},$$ or equivalently that $$\label{eqnnew2} p_{n-1}(p_{n+1}+\cdots+p_k) \leq p_n (p_{n}+ \cdots+ p_k).$$ It suffices to show that $p_{n-1}p_{n+j} \leq p_n p_{n+j-1}$, for $1 \leq j \leq k-n$. This however is easily checked by a straightforward calculation. We are now ready to prove Theorem \[propunifconv\]. [**Proof of Theorem \[propunifconv\].**]{} For the purpose of this proof, we introduce a new stochastic process $\tilde{I}_n^p$, indexed by $n=1,2,\ldots$ as follows. $\tilde{I}_1^p$ is distributed as the number of points in $I_1^p$. If this number of points is 0 however, we resample according to the same distribution, and repeat this until the total number of offspring is at least 1. If we do [*not*]{} resample at this first generation, we define $R_0:=1$; if we do resample, we set $R_0=0.$ In an inductive fashion, having defined $\tilde{I}_n^p$, we consider all points in $\tilde{I}_n^p$ and give each of them a random number of offspring distributed as $L_{n+1}''$, independently of each other. However, if the total number of offspring is 0, we resample [*all*]{} offsprings using the same distributions, until the total number of offspring is at least 1. If we do not have to resample, we define $R_n:=1$; if we do resample, we set $R_n=0$. Of course, the distribution of the number of points in $\tilde{I}_n^p$ given $\tilde{I}_{n-1}^p=k$ for some $k \geq 1$ is the same as the distribution of the number of points in $I_n^p$ given $I_{n-1}^p=k$ conditioned on being at least one. We can now write, for any $M,$ $$\begin{aligned} \label{belangrijk} \lefteqn{\P(0 < I_n^p <M) = \P(\prod_{i=0}^{n-1}R_i =1, 0 < \tilde{I}_n^p < M)} \nonumber\\ & & \leq \P(0 < \tilde{I}_n^p < M ) = \P(\tilde{I}_n^p < M).\end{aligned}$$ Now let $p_c(\overline{T}) < p < q$. We claim that $$\tilde{I}_n^p \preceq \tilde{I}_n^q.$$ To see this, we note that the offspring distributions of $I^p$ can be realised by first drawing from the appropriate $L_n'$, and then keep all points in the offspring with probability $p$, independently of each other. Now the combination of Lemma \[lem:conditioned\_binomial\] and Lemma \[more\_bin\_cond2\] implies that for $k \leq \ell$ and $p \leq q$ we have $$\label{ttrr} \rho_{k, p, 1} \preceq \rho_{\ell, q, 1}.$$ Clearly, we can couple $\tilde{I}_1^p$ and $\tilde{I}_1^q$ so that $\tilde{I}_1^p \leq \tilde{I}_1^q$, since we can use the same offspring $L_1'$ for them to get $I_1$ and then the domination follows from Lemma \[lem:conditioned\_binomial\]. Let $\{L'_{2,i}\}_{i=1}^{\tilde{I}_1^q}$ be i.i.d. with distribution equal to $L'_2$ and independent of everything else. We can now get $\tilde{I}_2^p$ by letting it be a ${\rm Bin}(\sum_{i=1}^{\tilde{I}_1^p}L'_{2,i},p)$ conditioned on being at least one. Similarly we get $\tilde{I}_2^q$ by letting it be a ${\rm Bin}(\sum_{i=1}^{\tilde{I}_1^q}L'_{2,i},q)$ conditioned on being at least one. The fact that we can couple $\tilde{I}_2^q$ and $\tilde{I}_2^p$ so that $\tilde{I}_2^q \leq \tilde{I}_2^p$ now follows from (\[ttrr\]). Repeating this procedure at every level gives that $$\label{eqn20} \P(\tilde{I}_n^p<M) \leq \P(\tilde{I}_n^{p_1}<M),$$ for all $p > p_1$, and this is where the uniformity in $p$ comes from. Of course letting $M$ above depend on $n$ does not change the validity of the argument. According to (\[belangrijk\]) and (\[eqn20\]) it therefore suffices to show that $$\P(\tilde{I}_n^{p_1} < ((1-\epsilon){\rm br} \overline{I^{p_1}})^n) \to 0,$$ as $ n \to \infty.$ For this, we use Theorem \[thm4\] and Proposition \[lemma5\]. Consider the subsequence $\{n_k\}$ and the constant $C>0$ dictated by applying Proposition \[lemma5\] to $I^{p_1}$. This is allowed according to the remark following the proof of Theorem \[thm4\]. Since each element in the $n_1$th generation of the $I^{p_1}$ process has a probability at least $C$ to survive, there is at least probability $C>0$ that no resampling is ever going to be necessary in the $\tilde{I}^{p_1}$ process after time $n_1$. There are now two possibilities. Either, at some point resampling is needed, or no resampling is ever needed after time $n_1$. In the latter case, we have that $\tilde{I}_n^{p_1}$ is at least as large as the number of points in a surviving copy of an $I^{p_1}$ tree with only one vertex at generation $n_1.$ It follows from Theorem \[thm4\] that this surviving tree has branching number ${\rm br} \overline{I^{p_1}}.$ Using that the lower growth number is at least as large as the branching number we are done in this case. On the other hand, if resampling is needed, then we take the first element in the subsequence $\{n_k\}$ after the first resampling, and repeat the reasoning from there. It follows that a.s., $\liminf_{n \rightarrow \infty} (\tilde{I}_n^{p_1})^{1/n} \geq {\rm br} \overline{I^{p_1}}$, and the proof is complete. We can now prove Proposition \[thm2\] [**Proof of Proposition \[thm2\]**]{}. We write $$\theta(p)=\P(I_n^p>0)-\P(I_n^p>0, |I^p| < \infty),$$ recall that $I^p$ denotes the component of the root. We will prove that along a subsequence, the last term tends to zero uniformly in $p \in [p_1, 1]$, where $p_1 > p_c(\overline{T})$, from which the result follows. Since the $p$-dependence is important now, we write $\theta_n(p)$ for $\theta_n$ in the context of the Galton-Watson process associated with $I^p$. For any $M >0$ we write, for $p_1 \leq p \leq 1$, $$\begin{aligned} \P(I_n^p>0, |I^p| < \infty) & \leq & \P(0 < I_n^p < M) + \P(I_n^p \geq M, |I^p| < \infty) \\ & \leq & \P(0 < I_n^p < M) + (1-\theta_n(p))^M \\ & \leq & \P(0 < I_n^p < M) + (1-\theta_n(p_1))^M.\end{aligned}$$ Let $\epsilon >0$ be arbitrary. We want to apply Proposition \[lemma5\] to $I^p.$ According to the remark after the proof of Theorem \[thm4\], all the assumptions of Proposition \[lemma5\] holds for $I^p$ since $p>p_c(\overline{T}).$ Now let $C$ be the constant in Proposition \[lemma5\] when we apply it to $I^{p_1}$. We choose $M$ so large that $(1-C)^M < \epsilon/2$. Next choose $n$ in the appropriate subsequence of Proposition \[lemma5\] and at the same time so large that the first term at the right hand side is at most $\epsilon/2$; this is possible according to Theorem \[propunifconv\] above. The right hand side is then bounded above by $\epsilon$, uniformly in $p \in [p_1,1]$. In summary, for any $\epsilon >0$ we can find $K$ such that $$\theta(p) \geq \P(I_{n_k}^p>0 )-\epsilon$$ for every $p \in [p_1,1]$ and every $n_k$ in the subsequence dictated by Proposition \[lemma5\] with $k \geq K.$ We see that for all $k \geq K$ and for all $p \in [p_1,1]$, $$|\theta(p)-\P(I_{n_k}^p>0)| \leq \epsilon,$$ which concludes the argument. Continuity of the percolation function {#sec4} ====================================== The supercritical continuity of $\theta(p)$ (Corollary \[corr1\]) follows immediately from Proposition \[thm2\]. We point out however that it is possible to obtain the same result by combining Theorem \[thm4\] with a modified version of the classical argument found in [@BK]. We provide a sketch. [**Sketch of proof of Corollary \[corr1\] from Theorem \[thm4\].**]{} We start by drawing an $I$ from the correct distribution. Associate to every edge $e$ in $I$ an independent $U([0,1])$ random variable, denoted by $U_e.$ For $p_c<q<p,$ create ${\cal I}^q$ and ${\cal I}^p$ by keeping every vertex of $I$ and those edges $e \in I$ such that $U_e \leq q,p$ respectively. Consider any infinite subtree $J$ in ${\cal I}^p.$ Theorem \[thm4\] gives us that $p_c(J)=1/{\rm br}J=1/p\liminf_{n \to \infty} {\mathbb E}[I_n]^{1/n}=p_c(I)/p$ a.s. Therefore, performing further percolation on $J$ with density $q/p>p_c(I)/p$ will result in a new graph containing an infinite subgraph a.s. Of course, the distribution of this new graph must be the same as $J \cap {\cal I}^q.$ Furthermore this holds in particular if $J=I^p$ showing that if $|I^p|=\infty$ then there exists a.s. an infinite subtree of $I^p \cap {\cal I}^q.$ It is now possible to proceed as in [@BK]. The non-classical way to conclude continuity of the percolation function has an interesting analogy on $\mathbb{Z}^d$. Define $B_n:= [-n,n]^d$ and write $\partial B_n$ for the (inner) boundary of $B_n.$ Letting $\{0 \leftrightarrow \partial B_n\}$ denote the event that the origin is connected to $\partial B_n$ by a path of open edges, define $$\varphi_n(p):={\mathbb P}_p(0 \leftrightarrow \partial B_n).$$ Clearly, $$\label{convergence} \theta(p)=\lim_{n \to \infty} \varphi_n(p),$$ for all $0 \leq p \leq 1$. The inequality of the following equation (valid for every $n \geq 1$) is a part of Theorem 8.18 of [@G]: $$\label{eqn22} \varphi_n(p)-\theta(p)=\P_p(0 \leftrightarrow B_n, |C| <\infty) \leq A(p,d)n^de^{-n \sigma(p)},$$ where we can take $$\label{eqn23} A(p,d)=\frac{d^2}{p^2(1-p)^{d-2}}.$$ Furthermore, according to Theorem 8.21 of [@G] we can take $\sigma(p)$ to be uniformly bounded away from $0$ on any closed sub-interval of $(p_c,1).$ We point out the following corollary and sketch the proof. The percolation function $\theta(p)$ on ${\mathbb Z}^d,$ $d \geq 2$ is continuous for $p>p_c.$ [**Sketch of proof.**]{} Choose $p_c<p_1<p_2<1.$ Combining equations (\[eqn22\]), (\[eqn23\]) and Theorem 8.21 of [@G] explained directly above, it is straightforward to prove that there exists constants $C=C(p_1,p_2)<\infty$ and $\delta=\delta(p_1,p_2)>0$ such that for any $p \in [p_1,p_2]$ and any $n \geq 1,$ $$\varphi_n(p)-\theta(p) \leq Ce^{-n \delta}.$$ Since trivially $$\theta(p) \leq \varphi_n(p),$$ it follows that $\varphi_n(p) \to \theta(p)$ uniformly on any closed subinterval of $(p_c,1),$ from which the statement follows. [99]{} Agresti, A. [*On the extinction times of varying and random environment branching processe*]{}, J. Appl. Prob. [**12**]{}, 39-46 (1975). Van den Berg, J. and Keane, M., [*On the continuity of the percolation probability function*]{}, Conference in modern analysis and probability, R. Beats et al (ed), 61 - 65 AMS, Providence, RI (1982). Broman E.I., Häggström O. and Steif J. E., [*Refinements of Stochastic Domination*]{}, Probab. Theory and Rel. Fields [**136**]{} No. 4, 587-603 (2006). Grimmett, G., [*Percolation*]{}, Second edition, Springer-Verlag, Berlin (1999). Häggström, O. and Peres, Y., [*Monotonicity of uniqueness for percolation on Cayley graphs: all infinite clusters are born simultaneously*]{}, Probab. Theory and Rel. Fields [**113**]{}, 273 - 285 (1999). Lyons, R., [*Random walks and percolation on trees*]{}, Ann. Probab. [**18**]{} no. 3, 931–958 (1990). Lyons, R., [*Random walks, capacity and percolation on trees*]{}, Ann. Prob. [**20**]{} no. 4, 2043-2088 (1992). Lyons, R., [*Probability on trees and networks*]{}, In progress, URL: http://mypage.iu.edu/ rdlyons/prbtree/prbtree.html. Schonmann, R., [*Stability of infinite clusters in supercritical percolation*]{}, Probab. Theory and Rel. Fields [**113**]{}, 287 - 300 (1999). [^1]: Chalmers University of Technology, [broman@math.chalmers.se]{} [^2]: VU University Amsterdam, Dept. of mathematics, De Boelelaan 1081a, 1081 HV Amsterdam, The Netherlands, [rmeester@few.vu.nl]{}

--- abstract: 'We derive the electrical and thermal conductivities of the $d$-density-wave state in an external magnetic field $B$ in the low-temperature regime and in the presence of impurities. We show that in the zero-temperature limit, $T \to 0$, the Wiedemann-Franz (WF) law remains intact irrespectively of the value of the applied field and chemical potential $\mu$ as soon as there is scattering from impurities. For finite $T \lesssim |\mu|$ the WF law violation is possible and it is enhanced by the external field. Depending on the relative values of $B$ and $T$ the electrical conduction may dominate the heat conduction and vice versa. For $\mu \gg T$ the WF is restored even in the presence of the external field.' author: - 'S.G. Sharapov$^{1}$' - 'V.P. Gusynin$^{1,2}$' - 'H. Beck$^1$' title: 'Transport properties in the $d$-density-wave state in an external magnetic field: The Wiedemann-Franz law' --- Introduction ============ A recent experiment of Hill [*et al.*]{} [@Hill:2001:Nature] that measured electrical and thermal conductivities of the optimally doped copper oxide superconductor Pr$_{2-x}$Ce$_x$CuO$_4$ (PCCO) in its normal state found striking deviations from the Wiedemann-Franz (WF) law. The electron-doped PCCO compound with $T_c \simeq 20 \mbox{K}$ near the optimal $x =0.15$ doping is the analog of the hole doped material La$_{2-x}$Sr$_x$CuO$_{4-y}$ (LSCO), which is also optimally doped at approximately $x=0.15$ with a maximum $T_{c} \simeq 40 \mbox{K}$. Lower $T_c$ in PCCO makes it particularly convenient from a technical point of view since an external magnetic field is necessary to destroy superconductivity to study its low temperature transport properties in the normal state. While for LSCO the upper critical field is $\simeq 50 \mbox{T}$, in the optimally doped PCCO the superconductivity is already destroyed at $8 \mbox{T}$ [@Hill:2001:Nature]. The choice of overdoped compounds that can be driven into the normal state by a relatively small magnetic field $\sim 13 \mbox{T}$ is wider and the hole overdoped system Tl$_2$Ba$_2$CuO$_{6+\delta}$ (Tl-2201) with $T_c = 15 \mbox{K}$ was studied very recently by Proust [*et al.*]{} [@Proust:2002:PRL]. They verified that in the overdoped Tl-2201 WF law holds perfectly. The WF law is one of the basic properties of a Fermi liquid, reflecting the fact that the ability of a quasiparticle to transport heat is the same as its ability to transport charge, provided it cannot lose energy through collisions. The WF law states that the ratio of the heat conductivity $\kappa$ to the electrical conductivity $\sigma$ of a metal is a universal constant: $$L_0 \equiv \frac{\kappa}{\sigma T} = \frac{\pi^2}{3} \left( \frac{k_B}{e}\right)^2,$$ where $k_B$ is the Boltzmann constant, $e$ is electron charge, and $L_0 = 2.45 \times 10^{-8} \mbox{W} \Omega \mbox{K}^{-2}$ is Sommerfeld’s value for the Lorenz ratio $L \equiv \kappa/(\sigma T)$. To be more precise, one should also specify the temperature range where the WF law holds. Strictly speaking this law is proven only in the limit $T \to 0$ and for a small concentration of impurities [@Langer:1962:PR]. In a less strict sense it is often implied that the WF law is valid for $k_B T \ll \mu$, where $\mu$ is chemical potential. In this case one can provide some qualitative arguments [@Taylor.book] that if the scattering from impurities does not strongly depend on the energy the WF will still be valid. Moreover, for $k_B T \ll \mu$ and elastic electron scattering these arguments can also be extended to the case of a strong ($\hbar \omega_L \gg k_B T$, $\hbar \omega_L \gg \Gamma$) quantizing magnetic field [@Smrcka:1977:JPC], where $\hbar \omega_L$ is the distance between Landau levels and $\Gamma$ is the impurity scattering rate. Thus, if these conditions are fulfilled, any violation of the WF law would indicate that there is a breakdown of the picture based on the Fermi liquid theory. One of the possible theoretical interpretations of the WF law breakdown is that the quasiparticle fractionalizes into separate spin and charge. This separation can be investigated using various models and approaches, so here we mention only the most recent studies done in the context of the WF law violation in cuprates. Specifically, the WF law violation has been studied for a large-$N$ limit of the $t-J$ model on the square lattice [@Houghton:2002:PRB]. Another examination of the WF law was done by Yang and Nayak (YN) [@Yang:2002:PRB] and also by Kim and Carbotte (KC) [@Kim:2002:PRB] within the phenomenological $d$-density-wave (DDW) picture. The DDW scenario proposed in Ref. [@Chakravarty:2001:PRB] is based on the assumption that the pseudogap phenomenon [@Timusk:1999:RPP] in high-$T_c$ cuprates is the result of the development of another order parameter called the DDW order [@Nayak:2000:PRB] that has $d$-wave symmetry: $$\label{DDW.def} \langle c_{s}^{\dagger} (t,\mathbf{k} + \mathbf{Q}) c_{s^{\prime}}(t,\mathbf{k}) \rangle = i \Phi_{\mathbf{Q}} f(\mathbf{k}) \delta_{s s^{\prime}}$$ where $s, s^{\prime} = \pm$ is the spin index, $f(\mathbf{k}) = \cos k_x a - \cos k_y a $, $\mathbf{Q} = (\pi/a, \pi/ a)$, and $a$ is the lattice constant. In comparison to the $d$-wave superconducting order the DDW order parameter appears to be rather exotic because it breaks the time reversal, translation by one lattice spacing and rotation by $\pi/2$ symmetries, but respects any combination of two of these. The underlying reason for this breaking is that there are countercirculating currents in the DDW ground state. The schematic phase diagram for the DDW pseudogap scenario is shown in Fig. \[fig:1\]. For a hole doping larger than the critical value $x_{c} \approx 0.19$ [@Tallon:2001:PC], the DDW order is presumed to disappear, whereas for the underdoped system DDW order coexists with $d$-wave superconductivity ($d$SC). One of the unusual features of the DDW state is that for a half-filled system the chemical potential of the nodal quasiparticles participating in the electrical and thermal transport can be small or even zero, i.e., $|\mu| < k_B T$, which violates the usual conditions of the WF law validity. Indeed, exactly in the limit $\mu =0$ the WF law is strongly violated in the extremely clean limit [@Yang:2002:PRB]. There is no WF law violation in the $T =0$ limit for finite $\mu$ and/or $\Gamma$ [@Yang:2002:PRB; @Kim:2002:PRB]. For finite temperatures the WF violation depends on the impurity scattering [@Kim:2002:PRB]: in the Born limit (for a constant impurity scattering rate) there is no change in the WF law, but in the unitary limit for the frequency dependent scattering rate the WF law is violated, but only for $|\mu|$ smaller that the DDW gap. When $\mu$ is increased sufficiently, the Lorenz number becomes approximately equal to its conventional value and its temperature dependence is small. While in general the validity of the DDW pseudogap scenario is still questionable, it is important to scrutinize all its theoretical consequences. One of the opportunities is to study possible WF law violations using the DDW model, so that these results can be compared with the experimental results of Refs. [@Hill:2001:Nature; @Proust:2002:PRL]. Although this investigation to a large extent had already been done by YN [@Yang:2002:PRB] and KC [@Kim:2002:PRB], both these papers do not consider [*the presence of the external magnetic field*]{} which is an essential ingredient of the experiments [@Hill:2001:Nature; @Proust:2002:PRL]. Indeed, in these experiments the DDW state appears in the magnetic field in the underdoped regime at low $T$ when the superconductivity is destroyed. In this paper we study the WF law for the DDW model [*including*]{} the external magnetic field for a constant impurity scattering rate, paying special attention to the regime $|\mu| \lesssim k_B T$ where the violation of the WF law is expected. Although there are some formal similarities between the DDW and $d$-wave superconducting states, the present problem is much simpler because there are no vortices and no superflows surrounding them in the DDW state. The $U(1)$ symmetry in the DDW state also remains unbroken, so that the electromagnetic field enters into the effective low-energy theory in the same way as in QED$_{2+1}$. Thus we avoid the famous problem of [*Dirac Landau-level mixing*]{} (see Refs. [@Vafek:2001:PRB; @Mel'nikov:2002:book]). We begin by presenting in Sec. \[sec:model\] the model DDW Hamiltonian and rewriting it using two-component spinors. Since there is a discrepancy between the expressions for the electrical current used by YN [@Yang:2002:PRB] and KC [@Kim:2002:PRB], we also derive the electrical current operator and show that in the nodal approximation it reduces to the current of Ref.  [@Yang:2002:PRB]. Then we obtain the low-energy effective action (Lagrangian) for the nodal quasiparticles in the presence of an external electromagnetic field, consider its discrete symmetries, and rewrite the Lagrangian in the Dirac form. In Sec. \[sec:Green.magnetic\] we consider the Green’s function of nodal quasiparticles in an applied field (the calculational details are given in Appendix \[sec:A\]). In Appendix \[sec:B\], using this Green’s function, we derive a generalized polarization bubble that can be applied to the calculation of electrical and thermal conductivities. In Secs. \[sec:electrical\] and \[sec:thermal\] we obtain general expressions for electrical and thermal conductivities in the external field. These general expressions are further analyzed in detail by analytical and numerical methods. In Sec. \[sec:WF\] we investigate the implications for the WF law that follow from the results presented in the previous sections. Conclusions are discussed in Sec. \[sec:conclusions\], where we give a concise summary of the obtained results and compare them with experiment. Some useful integrals are given in Appendix \[sec:C\]. -density wave Hamiltonian and its effective low energy form {#sec:model} =========================================================== We start with the mean-field Hamiltonian for the DDW state [@Nayak:2000:PRB; @Yang:2002:PRB; @Nersesyan:1989:JLTP] $$\label{Hamiltonian.DDW} H^{\mathrm{DDW}} = \int_{\mathrm{BZ}} \frac{d^2 k}{(2 \pi)^2} [(\varepsilon(\mathbf{k}) - \mu) c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k}) + i D(\mathbf{k}) c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k} + \mathbf{Q}) ],$$ where the single particle energy is $\varepsilon({\mathbf{k}}) = - 2 t (\cos k_x a + \cos k_y a) - 4 t^{\prime} \cos k_x a \cos k_y a$ with $t$, $t^{\prime}$ being the hopping parameters, $\mu$ is the chemical potential, $D(\mathbf{k}) = \frac{D_0}{2}(\cos k_x a - \cos k_y a )$ is the $d$-density-wave gap and $\mathbf{Q} = (\pi/a, \pi/ a)$ is the wave vector at which the density-wave ordering takes place. The integral is over the full Brillouin zone. Throughout the paper $\hbar = k_{B} = c= 1$ units are chosen, unless stated explicitly otherwise. Due to the presence of the factor $i$ in the second term of the Hamiltonian (\[Hamiltonian.DDW\]), it breaks time-reversal symmetry, where the time-reversal operation is defined as [@Enz.book] $$\label{time.reversal.Enz} c_{s}(\mathbf{k}) \to \mathcal{T} c_{s}(\mathbf{k}) \mathcal{T}^{-1} = i s c_{-s}(-\mathbf{k}), \qquad c_{s}^{\dagger}(\mathbf{k}) \to \mathcal{T} c_{s}^{\dagger}(\mathbf{k}) \mathcal{T}^{-1} = - i s c_{-s}^{\dagger}(-\mathbf{k})$$ and the time-reversal operator $\mathcal{T}$ is unitary and antilinear. Nevertheless the Hamiltonian (\[Hamiltonian.DDW\]) preserves a combined symmetry of the time reversal and a translation by one lattice spacing $\tilde{\mathcal{T}} = \mathcal{T} T_{\mathbf{a}}$: $$\label{time-translation} c_s(\mathbf{k}) \to \tilde{\mathcal{T}} c_s(\mathbf{k}) \tilde{\mathcal{T}}^{-1} = i e^{-i \mathbf{k} \mathbf{a}} c_{-s}(-\mathbf{k}).$$ Since the order parameter also breaks translational symmetry by one lattice spacing, it is convenient to halve the Brillouin zone as shown in Fig. \[fig:2\] and form a two-component electron operator $$\label{Nambu.variables} \chi_s(t, \mathbf{k}) = \left( \begin{array}{c} c_{s}(t, \mathbf{k}) \\ c_{s}(t, \mathbf{k} + \mathbf{Q}) \end{array} \right), \qquad \chi_s^{\dagger}(t, \mathbf{k}) = \left( \begin{array}{cc} c_{s}^{\dagger}(t, \mathbf{k}) \quad c_{s}^{\dagger}(t, \mathbf{k + Q}) \end{array} \right).$$ Then the mean-field Hamiltonian (\[Hamiltonian.DDW\]) in terms of $\chi$ becomes $$\label{Hamiltonian.DDW.spinor} H^{\mathrm{DDW}} = \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \chi_{s}^{\dagger}(t, \mathbf{k}) \left[ \frac{1}{2}[\varepsilon(\mathbf{k}) + \varepsilon(\mathbf{k} + \mathbf{Q})] - \mu + \frac{1}{2}[\varepsilon(\mathbf{k}) - \varepsilon(\mathbf{k} + \mathbf{Q})] \sigma_3 - D(\mathbf{k}) \sigma_2 \right] \chi_{s}(t, \mathbf{k}),$$ where $\sigma_i$ are Pauli matrices and the integral is over the reduced Brillouin zone (RBZ). The symmetry operation (\[time-translation\]), for example, is $$\label{time-reversal.spinor} \chi_s(t, \mathbf{k}) \to \tilde{\mathcal{T}} \chi_s(t, \mathbf{k}) \tilde{\mathcal{T}}^{-1} = - e^{-i \mathbf{k} \mathbf{a}} \sigma_2 \chi_{-s}(-t, -\mathbf{k} - \mathbf{Q}).$$ In what follows we will consider the more simple case $t^{\prime} =0$, so that after employing the nesting property $\varepsilon(\mathbf{k} + \mathbf{Q}) = - \varepsilon(\mathbf{k})$ the Hamiltonian (\[Hamiltonian.DDW.spinor\]) takes a simple form $$\label{Hamiltonian.DDW.matrix} H^{\mathrm{DDW}} = \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \chi_{s}^{\dagger}(t, \mathbf{k}) \left[ H_0(\mathbf{k}) - \mu \right] \chi_{s}(t, \mathbf{k}),$$ where $$\label{H} H_0(\mathbf{k}) =\varepsilon(\mathbf{k}) \sigma_3 - D(\mathbf{k}) \sigma_2.$$ Electrical and energy current operators --------------------------------------- As we already mentioned the expressions for the electrical current used in Refs. [@Yang:2002:PRB] and [@Kim:2002:PRB] are different. Although quantitatively this difference for high-temperature superconductors is expected to be rather small \[see after Eq. (\[Green.nodal\])\], the extra term in the current present in Ref. [@Yang:2002:PRB] is of qualitative importance, so we need to discuss its origin. To find out electrical current operator $\mathbf{j}$ for Hamiltonian (\[Hamiltonian.DDW\]) we apply the method similar to that of Refs. [@Durst:2000:PRB] and [@Kim:2002:PRB], where it was used to determine the heat current in the $d$-wave superconducting and DDW states, respectively. The idea of the derivation is to make the expression for the current operator compatible with the charge conservation law, $$\label{conserve} \frac{\partial \rho(t, \mathbf{r})}{\partial t} + \nabla \cdot \mathbf{j}(\mathbf{t, \mathbf{r}}) =0,$$ and with the equations of motion for field operators, $$\label{motion} \begin{split} i \frac{\partial c_{s}^{\dagger}(\mathbf{k})}{\partial t} = [ c_{s}^{\dagger}(\mathbf{k}), H] = -c_{s}^{\dagger}(\mathbf{k}) [\varepsilon(\mathbf{k}) - \mu] - c_{s}^{\dagger}(\mathbf{k} - \mathbf{Q}) i D(\mathbf{k} - \mathbf{Q}). %\\ % i \frac{\partial c_{s}(\mathbf{k} + \mathbf{q})}{\partial t} = % [ c_{s}(\mathbf{k} + \mathbf{q}), H] = % c_{s}(\mathbf{k} + \mathbf{q}) (\varepsilon(\mathbf{k} + \mathbf{q}) - \mu) %+ c_{s}(\mathbf{k} + \mathbf{q}+ \mathbf{Q}) i D(\mathbf{k} + \mathbf{q}). \end{split}$$ Writing Eq. (\[conserve\]) in the momentum space, one obtains $$\label{conserve.momentum} \frac{\partial \rho(t,\mathbf{\mathbf{q}})}{\partial t} + i \mathbf{q} \cdot \mathbf{j}(t, \mathbf{q}) =0,$$ where the Fourier transform of the charge density is $$\rho(t, \mathbf{q}) = e \int_{\mathrm{BZ}} \frac{d^2 k}{(2 \pi)^2} c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k} + \mathbf{q}).$$ For its derivative we obtain $$\begin{split} \frac{\partial \rho(t, \mathbf{q})}{\partial t} = -i e \int_{\mathrm{BZ}} \frac{d^2 k}{(2 \pi)^2} \left[ c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k} + \mathbf{q}) \mathbf{q} \frac{\partial \varepsilon(\mathbf{k})}{\partial \mathbf{k}} + c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k} + \mathbf{q}+ \mathbf{Q}) i \mathbf{q} \frac{\partial D(\mathbf{k})}{\partial \mathbf{k}} \right], \qquad \mathbf{q} \to 0, \end{split}$$ so that the electrical current is $$\label{electric.current.DDW} \begin{split} \mathbf{j} (t, \mathbf{0}) & = e \int_{\mathrm{BZ}} \frac{d^2 k}{(2 \pi)^2} \left[ c_{s}^{\dagger}(\mathbf{k}) c_{s}(\mathbf{k}) \frac{\partial \varepsilon(\mathbf{k})}{\partial \mathbf{k}} + c_{s}^{\dagger}(\mathbf{k} ) c_{s}(\mathbf{k} + \mathbf{Q}) i \frac{\partial D(\mathbf{k})}{\partial \mathbf{k}} \right] \\ & = e \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \chi_s^{\dagger}(\mathbf{k}) \mathbf{V}(\mathbf{k}) \chi_s(\mathbf{k}), \end{split}$$ where we introduced the generalized velocity operator that depends of the Fermi, $\mathbf{v}_F$, and gap, $\mathbf{v}_D$, velocities $$\label{velocity.general} \mathbf{V}(\mathbf{k}) = \mathbf{v}_F(\mathbf{k}) \sigma_3 + \mathbf{v}_D(\mathbf{k}) \sigma_2, \qquad \mathbf{v}_F(\mathbf{k}) \equiv \frac{\partial \epsilon(\mathbf{k})}{\partial \mathbf{k}}, \quad \mathbf{v}_D(\mathbf{k}) \equiv -\frac{\partial D(\mathbf{k})}{\partial \mathbf{k}}.$$ In the last equality in Eq. (\[electric.current.DDW\]) we used the two-component form and relied on the nesting property $\varepsilon(\mathbf{k} + \mathbf{Q}) = - \varepsilon(\mathbf{k})$ and $D(\mathbf{k} + \mathbf{Q}) = - D(\mathbf{k})$. This form of the electrical current operator reduces to the current used by YN [@Yang:2002:PRB] after the nodal approximation is made. The last term of Eq. (\[electric.current.DDW\]), $\sim \mathbf{v}_D$, is due to the fact that the DDW gap in the mean-field Hamiltonian (\[Hamiltonian.DDW\]) is $\mathbf{k}$-dependent. The expression for the electrical current used by KC contains only one term $\sim \mathbf{v}_F$. In this case, calculating electrical conductivity in the bare bubble approximation is not consistent with charge-current conservation, since such a bare vertex does not satisfy the Ward identity when quasiparticle self-energy has a nontrivial momentum dependence. Instead one should use a vertex that complies with the Ward identity. The derivation of the energy current operator from the Hamiltonian (\[Hamiltonian.DDW\]) is quite similar to the derivation of the electrical current operator (\[electric.current.DDW\]), so here we only outline the main steps. This current is calculated from the corresponding continuity equation for the energy density and the equations of motion (\[motion\]) (we set there $\mu =0$ because the energy density is considered) that give $$\label{heat.current} \mathbf{j}^E (t, \mathbf{0}) = \int_{\mathrm{BZ}} \frac{d^2 k}{(2 \pi)^2} \left[\varepsilon(\mathbf{k}) \mathbf{v}_F(\mathbf{k}) + D(\mathbf{k}) \mathbf{v}_D(\mathbf{k}) \right] c_s^{\dagger}(\mathbf{k}) c_s(\mathbf{k}),$$ or in the spinor form $$\mathbf{j}^E (t, \mathbf{0}) = \frac{1}{2} \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \chi^{\dagger}(t, \mathbf{k}) \{H_0(\mathbf{k}), \mathbf{V}(\mathbf{k}) \} \chi(t, \mathbf{k}).$$ Then using again the equations of motion we finally get $$\label{heat.current2.B=0} \mathbf{j}^E (\Omega, \mathbf{0}) = \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \int d \omega \, \left(\omega + \frac{\Omega}{2} \right) \chi^{\dagger}(\omega, \mathbf{k}) \mathbf{V}(\mathbf{k}) \chi(\omega+ \Omega, \mathbf{k}).$$ One can easily see that Eq. (\[heat.current2.B=0\]) agrees with the corresponding energy current obtained in Ref. [@Kim:2002:PRB]. Effective low-energy nodal action --------------------------------- The action $S$ corresponding to the Hamiltonian (\[Hamiltonian.DDW.matrix\]) is $$\begin{split} S & = - \int \limits_{0}^{\beta} d \tau \left[ \int d^2 r \chi_{s}^{\dagger} (\tau,\mathbf{r}) \partial_{\tau} \chi_{s} (\tau,\mathbf{r}) + H^{\mathrm{DDW}}(\tau) \right] \\ & = - \int \limits_{0}^{\beta} d \tau \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \chi_{s}^{\dagger} (\tau,\mathbf{k}) \left[ \partial_{\tau} - \mu + \varepsilon(\mathbf{k}) \sigma_3 - D(\mathbf{k}) \sigma_2 \right] \chi_{s} (\tau,\mathbf{k}), \end{split}$$ where $\tau$ is the imaginary time. Hence the Green’s function of the DDW state reads $$\label{Green.common} G(i\omega, \mathbf{k}) = \frac{(i\omega + \mu)\hat{I} + \varepsilon(\mathbf{k}) \sigma_3 - D(\mathbf{k}) \sigma_2} {(i \omega + \mu)^2 - E^{2}(\mathbf{k})}, \qquad E(\mathbf{k}) = \sqrt{\varepsilon^2(\mathbf{k}) + D^2(\mathbf{k})},$$ where $i \omega = i (2n+1) \pi T $ is fermionic (odd) Matsubara frequency. Linearizing the spectrum about the four nodes $\mathbf{N} = (\pm \pi/2a, \pm \pi/2a )$ at half-filling ($\mu =0$), one obtains the Green’s function (GF) for a given node (we choose the local nodal coordinate systems as shown in Fig. \[fig:2\]) $$\label{Green.nodal} G_{\mathrm{node}}(i\omega, \mathbf{k}) = \frac{(i\omega + \mu)\hat{I} + v_F k_x \sigma_3 + v_D k_y \sigma_2} {(i \omega + \mu)^2 - E^2(\mathbf{k})}, \qquad E(\mathbf{k}) = \sqrt{v_F^2 k_x^2 + v_D^2 k_y^2},$$ where the Fermi velocity for half-filling $v_F = |\partial \varepsilon(\mathbf{k})/\partial \mathbf{k} |_{\mathbf{k} = \mathbf{N}} | = 2 \sqrt{2} t a$ and the DDW gap velocity $v_{D} = |\partial D(\mathbf{k})/\partial \mathbf{k} |_{\mathbf{k} = \mathbf{N}} | = \frac{1}{\sqrt{2}} D_0 a$. It is important to stress that this way of linearization is different from the nodal approximation used to describe $d$SC [@Durst:2000:PRB] when the expansion is done around the nodal (“Dirac”) points defined as an intersection of the Fermi surface $\varepsilon (\mathbf{k}) - \mu =0$ and the nodal lines $D(\mathbf{k}) = 0$. For the DDW case one always expands around half-filling $\varepsilon (\mathbf{k}) =0$ and $D(\mathbf{k}) = 0$, so that for nonzero $\mu$ the nodes transform into small pockets (see Fig. \[fig:2\]) which are cross sections of the “Dirac cone”. Thus only at exactly half-filling is there a correspondence between DDW and $d$SC cases. There are the following estimates [@Morr:2002:PRL] of the model parameters, $t = 300$ meV and $D_0 = 50$meV, which give the ratio $v_{F}/v_{D} = 24$. This shows that the second term of the electrical current (\[electric.current.DDW\]), which is proportional to $v_{D}$, is indeed small. The value of the chemical potential corresponding to a hole doping of 10% is [@Morr:2002:PRL] $\mu = 0.91 t$. The linearized Dirac action that corresponds to the Green’s function (\[Green.nodal\]) can be written as $$\label{S.nodal} \begin{split} S_{\mathrm{node}} = - \int \limits_{0}^{\beta} d \tau \int d^2 r \chi_{s}^{\dagger}(\tau, \mathbf{r}) [\hat{I} (\partial_{\tau} - \mu) - v_F \sigma_3 i \partial_x - v_D \sigma_2 i \partial_y ] \chi_{s}(\tau, \mathbf{r}) . \end{split}$$ Dealing with Eqs. (\[Green.nodal\]) and (\[S.nodal\]), one should not forget that all physical quantities involve the summation over the 4 nodes present in the original Green’s function (\[Green.common\]). In what follows we will use the nodal action (\[S.nodal\]) instead of the original DDW Hamiltonian (\[Hamiltonian.DDW\]). This approximation is well justified for $T \ll D_0$. As we will see below, this approximation and the fact that the electromagnetic field enters the theory through the minimal coupling prescription will allow us to apply the results of QED$_{2+1}$. Discrete symmetries of the nodal Lagrangian ------------------------------------------- Since the action (\[S.nodal\]) does not directly correspond to the common QED$_{2+1}$ form used in the literature [@Dittrich.book] (see also Ref. [@Jackiw:1981:PRD]), we present below the transformations for the discrete symmetries. Let us consider only one node and suppress the spin index $s$, so that in real time $t = - i \tau$ the action $S$ reads $$S_{\mathrm{node}} = i \int d t \int d^2 r \mathcal{L},$$ with $$\label{Dirac.Lagrangian.real} \mathcal{L} = \chi^{\dagger}(t, \mathbf{r}) \left[i \hat{I} (\partial_{t} - i e A_0(t, \mathbf{r})) + \mu + i v_F \sigma_3 \left(\partial_x - i \frac{e}{c}A_1(t, \mathbf{r})\right) + i v_D \sigma_2 \left( \partial_y - i\frac{e}{c}A_2(t, \mathbf{r})\right) \right] \chi (t, \mathbf{r}),$$ where $A_{\nu}(t, \mathbf{r})$, $\nu =0,1,2$ is the electromagnetic field vector potential. This is a [*highly nontrivial fact*]{} in that the vector potential can be inserted in Eq. (\[Dirac.Lagrangian.real\]) using the [*minimal coupling prescription*]{} of QED. Only due to this fact can one apply the results of QED$_{2+1}$ for the description of the DDW state. Moreover, this is only correct for the DDW state [@Yang:2002:PRB] but not for the $d$SC state, where the last term of the Lagrangian (\[Dirac.Lagrangian.real\]) does not couple with the electromagnetic field. It has to be admitted that the way of inserting the electromagnetic field is not just related to the nodal approximation and can be traced back to the expression (\[electric.current.DDW\]) for the electrical current that can only be obtained when the vector potential enters both $\varepsilon(\mathbf{k})$ and $D(\mathbf{k})$ terms in Eq. (\[Hamiltonian.DDW\]). Finally we note that the symmetries of the Lagrangian have to be analyzed in real time where one can distinguish time and spatial coordinates. ### Parity In $2+1$ dimensions, parity corresponds to inverting one axis (since the inversion of both axes would be rotation by $\pi$): $P \mathbf{r} = \mathbf{r}^{\prime}$, where $\mathbf{r} = (x,y)$ and $\mathbf{r}^{\prime} = (x, -y)$. The corresponding operation on the two-component spinor and on the gauge field is $$\begin{split} & \chi(\tau, \mathbf{r}) \to \mathcal{P} \chi(\tau, \mathbf{r}) \mathcal{P}^{-1} = \sigma_3 \chi(\tau, \mathrm{r}^{\prime}), \qquad \chi^{\dagger}(t, \mathbf{r}) \to \mathcal{P} \chi^{\dagger}(t, \mathbf{r}) \mathcal{P}^{-1} = \chi^{\dagger}(t, \mathrm{r}^{\prime}) \sigma_3, \\ & A_0(t, \mathbf{r}) \to \mathcal{P} A_{0}(t, \mathbf{r}) \mathcal{P}^{-1} = A_{0}(t, \mathbf{r}^{\prime}), \quad A_1(t, \mathbf{r}) \to \mathcal{P} A_{1}(t, \mathbf{r}) \mathcal{P}^{-1} = A_{1}(t, \mathbf{r}^{\prime}), \quad A_2(t, \mathbf{r}) \to \mathcal{P} A_{2}(t, \mathbf{r}) \mathcal{P}^{-1} = -A_{2}(t, \mathbf{r}^{\prime}). \end{split}$$ One can verify that the Lagrangian (\[Dirac.Lagrangian.real\]) is invariant under this parity transformation while possible mass terms with $\sigma_1$ and $\sigma_2$ break parity: $$\begin{split} & \chi^{\dagger}(t, \mathbf{r}) \sigma_1 \chi(t, \mathbf{r}) \to \mathcal{P}\chi^{\dagger}(t, \mathbf{r}) \sigma_1 \chi(t, \mathbf{r}) \mathcal{P}^{-1} = - \chi^{\dagger}(t, \mathbf{r}^{\prime}) \sigma_1 \chi(t, \mathbf{r}^{\prime}); \\ & \chi^{\dagger}(t, \mathbf{r}) \sigma_2 \chi(t, \mathbf{r}) \to \mathcal{P}\chi^{\dagger}(t, \mathbf{r}) \sigma_2 \chi(t, \mathbf{r}) \mathcal{P}^{-1} = - \chi^{\dagger}(t, \mathbf{r}^{\prime}) \sigma_2 \chi(t, \mathbf{r}^{\prime}). \end{split}$$ ### Charge conjugation Charge conjugation, $$\begin{split} & \chi (t, \mathbf{r}) \to \mathcal{C} \chi(t, \mathbf{r}) \mathcal{C}^{-1} = \sigma_3 [\chi^{\dagger}(t,\mathbf{r})]^T, \qquad \chi^{\dagger} (t, \mathbf{r}) \to \mathcal{C} \chi^{\dagger} (t, \mathbf{r}) \mathcal{C}^{-1} = [\chi(t,\mathbf{r})]^T \sigma_3, \\ & A_\nu(t, \mathbf{r}) \to \mathcal{C} A_{\nu}(t, \mathbf{r}) \mathcal{C}^{-1} = -A_{\nu}(t, \mathbf{r}), \qquad \nu =0,1,2 \end{split}$$ leaves the equations of motion invariant. It is easy to check that the Lagrangian (\[Dirac.Lagrangian.real\]) is invariant under this transformation only at half-filling ($\mu = 0$) while away from half-filling the term with the chemical potential breaks $C$ $$\mu \chi^{\dagger}(t, \mathbf{r}) \hat{I}\chi(t, \mathbf{r}) \to \mu \mathcal{C}\chi^{\dagger}(t, \mathbf{r}) \hat{I}\chi(t, \mathbf{r}) \mathcal{C}^{-1} = - \mu \chi^{\dagger}(t, \mathbf{r}) \hat{I} \chi(t, \mathbf{r}).$$ One can also check that the mass term with $\sigma_2$ also breaks $C$, whereas $\sigma_1$ mass term does not break $C$. ### Time-reversal symmetry The operation of time reversal for the electromagnetic field is defined as $$A_0(t, \mathbf{r}) \to \mathcal{T} A_{0}(t, \mathbf{r}) \mathcal{T}^{-1} = A_{0}(-t, \mathbf{r}), \qquad A_{1,2}(t, \mathbf{r}) \to \mathcal{T} A_{1,2}(t, \mathbf{r}) \mathcal{T}^{-1} = -A_{1,2}(-t, \mathbf{r}).$$ We determine a matrix $T$ responsible for the time-reversal transformation of spinors $$\label{time.reversal.Dirac} \chi (t, \mathbf{r}) \to \mathcal{T} \chi(t, \mathbf{r}) \mathcal{T}^{-1} = T \chi(-t,\mathbf{r}), \qquad \chi^{\dagger} (t, \mathbf{r}) \to \mathcal{T} \chi^{\dagger} (t, \mathbf{r}) \mathcal{T}^{-1} = \chi^{\dagger}(-t,\mathbf{r}) T$$ demanding that the current components transform as follows $$\chi^{\dagger}(t, \mathbf{r}) \alpha_{\nu} \chi(t, \mathbf{r}) \to \chi^{T \dagger}(-t, \mathbf{r}) \alpha_{\nu}^{\ast} \chi^{T}(-t, \mathbf{r}) = \chi^{\dagger}(-t, \mathbf{r}) T \alpha_{\nu}^{\ast} T \chi(-t, \mathbf{r}) = \chi^{\dagger}(-t, \mathbf{r}) \tilde{\alpha}_{\nu} \chi(-t, \mathbf{r}),$$ where $$\label{alpha.nu} \alpha_{\nu} = (\hat{I}, \sigma_3, \sigma_2), \qquad \tilde{\alpha}_{\nu} = (\alpha_0, - \alpha_1, -\alpha_2).$$ The time-reversal operation (\[time.reversal.Dirac\]) can be written down using $T = \sigma_2$ and one can see that the Lagrangian $\mathcal{L}$ is invariant with respect to this symmetry $$\mathcal{L}[\chi^{\dagger}(t, \mathbf{r}), \chi(t, \mathbf{r}), A_{\nu}(t, \mathbf{r})] \to \mathcal{L}[\chi^{T \dagger}(-t, \mathbf{r}), \chi^{T}(-t, \mathbf{r}), A_{\nu}^{T}(-t, \mathbf{r})] = \mathcal{L}[\chi^{\dagger}(t, \mathbf{r}), \chi(t, \mathbf{r}), A_{\nu}(t, \mathbf{r})],$$ while the $\sigma_1$ and $\sigma_2$ mass terms would break it $$\chi^{\dagger}(t, \mathbf{r}) \sigma_{1,2} \chi(t, \mathbf{r}) \to \mathcal{T}\chi^{\dagger}(t, \mathbf{r}) \sigma_{1,2} \chi(t, \mathbf{r}) \mathcal{T}^{-1} = - \chi^{\dagger}(-t, \mathbf{r}) \sigma_{1,2} \chi(-t, \mathbf{r}).$$ One can check that if we use notations corresponding to those of Ref. [@Jackiw:1981:PRD], $P$ and $T$ operations introduced here agree with the transformations given in Ref. [@Jackiw:1981:PRD]. The time-reversal symmetry operation (\[time.reversal.Dirac\]) for the continuum nodal Lagrangian should be identified with the combined symmetry operation (\[time-reversal.spinor\]) that leaves the Hamiltonian (\[Hamiltonian.DDW.spinor\]) invariant. We note that, in principle, the time-reversal transformation should also flip the spin $s \to - s$, but since all considered expressions assume summation over two spin components, we do not include this spin flip in the transformation of Dirac spinors. Parity and time-reversal symmetries can still be broken by the mass (gap) term $\sim \sigma_1$. In particular, this mass term may be generated in an external magnetic field due to the so called [*magnetic catalysis*]{} phenomenon [@Gusynin:1995:PRD]. Final form of the nodal Lagrangian ---------------------------------- We choose the vector potential for the external magnetic field $\mathbf{B}$ in the symmetric Poincaré gauge $\mathbf{A} = (-\frac{B}{2}x_2 , \frac{B}{2} x_1 )$, so that the field $\mathbf{B}$ is perpendicular to CuO$_2$ planes. To consider also a possibility of magnetic catalysis [@Gusynin:1995:PRD] we add to the Lagrangian (\[Dirac.Lagrangian.real\]) the interaction term originating from particle-hole attraction $$\label{L.int} \mathcal{L}_{\mathrm{int}} = \frac{g}{2} (\chi^{\dagger}(x) \sigma_1 \chi(x))^2, \qquad x = (t, \mathbf{r}).$$ The simplest way to treat the interaction (\[L.int\]) is to introduce Hubbard-Stratonovich field $\varphi(x) = - g \chi^{\dagger}(x) \sigma_1 \chi(x)$, so that the Lagrangian (\[Dirac.Lagrangian.real\]) becomes $$\label{Lagrangian.alpha} \mathcal{L} = \chi^{\dagger}(x) [i \alpha_{\nu} D_{\nu} -\sigma_1 \varphi(x) ]\chi (x)- \frac{\varphi^2(x)}{2g},$$ where the covariant derivatives are $$\label{long.derivative} D_{\nu} = \begin{cases} \hbar \partial_t - i e A_{0}(x), & \nu =0,\\ v_{F} \left(\hbar \partial_x - i \frac{e}{c} A_{1}(x) \right), & \nu =1, \\ v_{\Delta} \left( \hbar \partial_y - i \frac{e}{c} A_{2}(x)\right), & \nu =2, \end{cases}$$ and $\alpha$ matrices were defined in Eq. (\[alpha.nu\]). Finally, to simplify further calculations and to make a direct link with QED$_{2+1}$ it is convenient to introduce Dirac conjugated spinor $\bar{\chi}(x) = \chi^{\dagger}(x) \sigma_1$ and rewrite the Lagrangian (\[Lagrangian.alpha\]) as $$\label{Dirac.conjugated.Lagrangian} \mathcal{L} = \bar{\chi}(x) [i \gamma^{\nu} D_{\nu} - \varphi(x) ]\chi (x)- \frac{\varphi^2(x)}{2g},$$ where the $\gamma$ matrices are $$\gamma^{\nu} = (\sigma_1, - i \sigma_2, i \sigma_3), \qquad \{\gamma^\mu, \gamma^{\nu}\} = 2 \hat{I} g^{\mu \nu}, \qquad g^{\mu \nu} = \mbox{diag}(1,-1,-1).$$ Note that the Zeeman term, if necessary, can be added explicitly both to the original Hamiltonian (\[Hamiltonian.DDW\]) and the Lagrangian (\[Dirac.conjugated.Lagrangian\]). Here, however, this term is neglected. Green’s function in an external magnetic field {#sec:Green.magnetic} ============================================== To derive the transport coefficients we need an explicit representation for the fermionic Green’s function in an external field. The calculation, following the Schwinger (proper time) approach [@Schwinger:1951:PR], is sketched in Appendix \[sec:A\] and the result is (we set for convenience $v_F = v_D =1$ and restore them when it is needed according to the prescription given in Appendix \[sec:B\]) $$\label{Schwinger.representation} S(x-y) = \exp \left( i e \int \limits_y^x A_{\lambda}^{\mathrm{ext}} d z^{\lambda} \right) {\tilde S}(x-y),$$ where the translation invariant part ${\tilde S}$ in the Matsubara frequency-momentum representation is given by $$\label{Schwinger.representation.translation} \begin{split} {\tilde S}(i \omega, \mathbf{p}) = & - \int \limits_{0}^{\infty} d s \exp \left[-s \left(\Delta^2 -(i \omega)^2 + \mathbf{p}^2 \frac{\tanh (eBs)}{eBs}\right) \right] \\ & \times \left[ \left(i \omega \gamma^0 - p_1 \gamma^1 - p_2 \gamma^2 + \Delta + \frac{1}{i}(p_2 \gamma^1 - p_1 \gamma^2) \tanh (e B s) \right) \left(1 + \frac{1}{i} \gamma^1 \gamma^2 \tanh (e B s) \right) \right]. \end{split}$$ Here we replaced the Hubbard-Stratonovich field $\varphi(x)$ by its mean-field value $\Delta = \langle \varphi (x)\rangle$. The chemical potential $\mu$ has to be taken into account via the shift $i\omega \to i \omega + \mu$ \[see Eq. (\[G.translation.invariant\]) in Appendix \[sec:B\]\]. As already mentioned, the gap $\Delta$ can be generated by an external magnetic field [@Gusynin:1995:PRD]. The value of the gap has to be determined from the minimum condition of the corresponding effective potential, see e.g. [@Gusynin:1995:PRD; @Ferrer:2002; @Gorbar:2002:PRB]. Here, however, we are mainly interested in the case of $\Delta = 0$ and will concentrate on $\mu$ and $T$ dependences of the Lorenz ratio $L$. The propagator (\[Schwinger.representation.translation\]) can be decomposed over the Landau level poles [@Gusynin:1995:PRD; @Chodos:1990:PRD] (see also Appendix \[sec:A\] for the details): $$\label{Landau.levels} {\tilde S}(i \omega, \mathbf{p}) = \exp\left( - \frac{\mathbf{p}^2}{|e B|}\right) \sum_{n=0}^{\infty} (-1)^n \frac{S_n(B,i \omega, \mathbf{p})}{(i\omega)^2 - \Delta^2 - 2 |e B| n},$$ where $$\label{Sn} S_n(B,i \omega, \mathbf{p}) = 2 (\Delta + i \omega \gamma^0) \left[ P_{-} L_n \left(2\frac{\mathbf{p}^2}{|eB|}\right) - P_{+} L_{n-1}\left(2 \frac{\mathbf{p}^2}{|eB|} \right) \right] + 4 (p_1 \gamma^1 + p_2 \gamma^2) L_{n-1}^1 \left(2 \frac{\mathbf{p}^2}{|eB|} \right)$$ with $P_{\pm} = [1\pm \mbox{sgn} (e B)i \gamma^1 \gamma^2]/2 = (1 \mp \sigma_1)/2$ being projectors and $L_n$, $L_n^1$ Laguerre’s polynomials ($L_{-1}^1 \equiv 0$). In what follows for convenience we take $e B > 0$. The spectral function $A_D(\omega, \mathbf{p})$ associated with the translation-invariant part ${\tilde S}$ of the Green’s function is defined as $$\label{spectral.def} A_D(\omega, \mathbf{p}) = \frac{1}{2 \pi i} \left[ {\tilde S}^A(\omega - i0, \mathbf{p}) - {\tilde S}^R(\omega + i0, \mathbf{p}) \right],$$ where the retarded, ${\tilde S}^R$, and advanced, ${\tilde S}^A$, Greens’ functions are ${\tilde S}^R(\omega + i 0, \mathbf{p}) = {\tilde S}(i \omega \to \omega + i 0, \mathbf{p})$ and ${\tilde S}^A(\omega - i 0, \mathbf{p}) = {\tilde S}(i \omega \to \omega - i 0, \mathbf{p})$. Using the definition (\[spectral.def\]) one obtains the spectral function associated with the “Dirac” fermion Green’s function $\langle \chi \bar{\chi} \rangle$ (\[Landau.levels\]) for a clean system $$\label{AD.clean} A_D(\omega, \mathbf{p}) = \exp\left( - \frac{\mathbf{p}^2}{e B}\right) \sum_{n=0}^{\infty} (-1)^n \left[ \frac{(\gamma^0 M_n + \Delta) f_1(\mathbf{p}) + f_2(\mathbf{p})}{2 M_n} \delta(\omega - M_n) + \frac{(\gamma^0 M_n - \Delta) f_1(\mathbf{p}) - f_2(\mathbf{p})}{2 M_n} \delta(\omega + M_n) \right],$$ where $M_n = \sqrt{\Delta^2 + 2 e B n}$ and $$\label{f} f_1(\mathbf{p}) = 2\left[ P_{-} L_n \left(2\frac{\mathbf{p}^2}{eB}\right) - P_{+} L_{n-1}\left(2 \frac{\mathbf{p}^2}{eB} \right) \right], \qquad f_2(\mathbf{p}) = 4 (p_1 \gamma^1 + p_2 \gamma^2) L^{1}_{n-1}\left(2 \frac{\mathbf{p}^2}{eB} \right).$$ To consider transport phenomena for a more realistic case one should introduce into the theory the effect of scattering on impurities. In general this can be done by considering dressed fermion propagators that include the self-energy $\Sigma(\omega)$ due to the scattering on impurities. This self-energy in turn has to be found self-consistently by solving the corresponding Schwinger-Dyson equation. This scheme was in fact used in the Ref. [@Kim:2002:PRB], and the WF violation obtained by KC at finite temperatures in the unitary limit is essentially due to the nontrivial frequency dependence of the scattering rate $\Gamma(\omega) = -\mbox{Im} \Sigma^R(\omega)$. As mentioned in Introduction, here we choose the case of constant width $\Gamma = \Gamma(\omega = 0) = 1/(2 \tau)$, $\tau$ being a mean free time of quasiparticles, so that the $\delta$-like quasiparticle peaks in Eq. (\[AD.clean\]) acquire a Lorentzian shape: $$\label{Gamma} \delta(\omega \pm M_n) \to \frac{\Gamma}{\pi} \frac{1}{(\omega \pm M_n)^2 + \Gamma^2}.$$ Such a broadening of Landau levels was found to be a rather good approximation valid in not very strong magnetic fields [@Prange.book]. Finally, we establish a link between the Green’s function (\[Schwinger.representation\]) considered here and the Green’s function (\[Green.common\]) used in the preceding section. Taking into account that the average $\langle \chi \chi^{\dagger} \rangle$ is related to the “Dirac’s average” via $\langle \chi \chi^{\dagger} \rangle = \langle \chi \bar{\chi} \rangle \gamma^{0}$ one obtains for $B = 0$: $$\label{G.B=0} G(i\omega, \mathbf{p}) = {\tilde S}(i \omega_m,\mathbf{p}; B =0) \gamma^0 = \frac{i \omega + p_1 \sigma_3 + p_2 \sigma_2 + \Delta \sigma_1} {(i \omega)^2 - \mathbf{p}^2 - \Delta^2},$$ so that restoring $v_F$, $v_D$, replacing $i\omega_n \to i \omega_n + \mu$, and setting $\Delta =0$ we recover the Green’s function (\[Green.nodal\]). Electrical conductivity {#sec:electrical} ======================= General expression for electrical conductivity ---------------------------------------------- The frequency dependent longitudinal electrical conductivity can be calculated from the current-current correlation function $$\label{el.cur-cur.tensor} \Pi^{CC}_{\alpha \beta}(i \Omega) = - \int \limits_{0}^{\beta} d \tau e^{i \Omega \tau} \langle T_{\tau} j_{\alpha}^{\dagger}(\tau, \mathbf{0}) j_{\beta}(0, \mathbf{0}) \rangle$$ by means of the Kubo formula [@Mahan:book] $$\label{el.cur-cur.scalar} \sigma(\Omega) = - \frac{\mbox{Im} \Pi_{R}^{CC}(\Omega +i 0)}{\Omega},$$ where $\Pi_{R}^{CC}(\Omega + i0) = \Pi^{CC}(i \Omega \to \Omega + i0)$ is the longitudinal polarization \[see Eq. (\[tensor2scalar\])\]. The expression for the electrical current operator for the Hamiltonian (\[Hamiltonian.DDW\]) was already derived in Eq. (\[electric.current.DDW\]). Having the current, we are in a position to calculate the current-current correlation function (\[el.cur-cur.tensor\]), which is given by the bubble diagram. The influence of the impurity vertex corrections on the transport properties of another ($d$SC) nodal system was considered in Ref. [@Durst:2000:PRB]. The calculation of the bare bubble polarization function is in fact similar for all transport coefficients in $d$SC and DDW nodal liquids, so that it is convenient to define a generalized polarization function $\Pi^{g g^{\prime}}(\Omega)$ that depends on the coupling parameters $g$, $g^{\prime}$ and the generalized velocity $\mathbf{V}(\mathbf{k})$. The polarization function is calculated in Appendix \[sec:B\]. Using the general result (\[bubble.final.tensor\]) for the case of interest, $g = g^{\prime} = e$ in the limit $\Omega \to 0$, we find that the dc conductivity of isotropic system is given by $$\label{sigma.like.Langer} \sigma = \pi e^2 \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \int \limits_{-\infty}^{\infty} d\omega (-n_F^{\prime}(\omega - \mu)) \mbox{tr} \left[ A(\omega, {\bf k}) V_{\alpha}(\mathbf{k}) A(\omega, {\bf k}) V_{\alpha}(\mathbf{k}) \right] ,$$ where $-n_{F}^{\prime}(\omega -\mu) = (1/4T) \cosh^{-2}[(\omega -\mu)/2T]$ is the derivative of the Fermi distribution and summation over the dummy variable is implied. At this point it is instructive to compare Eq. (\[sigma.like.Langer\]) with the expression for the conductivity obtained by Langer \[Eq. (4.8) of Ref. [@Langer:1962a:PR]\] in the bubble approximation. Langer’s expression has the same form as Eq. (\[sigma.like.Langer\]) except the matrix-valued $e V_{\alpha}(\mathbf{k})$ is replaced by the vertex with coinciding fermion momenta and energies $$\label{Lambda.Langer} \Lambda_i(\mathbf{k}, \omega) = e \frac{k_i}{m} - e \frac{\partial}{\partial k_i} \Sigma(\mathbf{k}, \omega),$$ where $\Sigma(\mathbf{k}, \omega)$ is the usual self-energy \[he considered the quadratic dispersion law, so that $\partial \epsilon(\mathbf{k})/\partial k_i = k_i/m$\]. In fact Eq. (\[Lambda.Langer\]) is a direct consequence of the Ward identity. Thus knowing the self-energy one immediately obtains the dc conductivity in the bubble approximation. Furthermore, the bubble conductivity of Ref. [@Langer:1962a:PR] becomes exact in the zero-temperature limit. Since in the mean-field approximation $\Sigma(\mathbf{k}, \omega)$ is replaced by the DDW gap $\sigma_2 D(\mathbf{k})$, Langer’s expression reduces to Eq. (\[sigma.like.Langer\]). It is also evident from Eq. (\[Lambda.Langer\]) that for the momentum-independent self-energy the bubble conductivity coincides with the conductivity calculated in the [*bare bubble*]{} approximation, but clearly this is [*not the case*]{} of the DDW gap. Expression (\[sigma.like.Langer\]) can also be derived in the lowest order approximation from a general expression for the dc conductivity, with the vertex corrections taken into account, obtained by Eliashberg [@Eliashberg:1962:JETP]. Making the nodal approximation (see Eq. (\[bubble.final\])) we get $$\label{conductivity.general.final} \begin{split} \sigma(B,T) & = \frac{2 \pi e^2}{v_{F}v_{D}} \int \frac{d^{2}p}{(2\pi)^{2}} \int \limits_{-\infty}^{\infty} d\omega (-n_F^{\prime}(\omega-\mu)) \\ & \times \left( v_F^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{1} A_D(\omega, {\bf p}) \gamma^{1} \right] + v_D^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{2} A_D(\omega, {\bf p}) \gamma^{2} \right] \right). \end{split}$$ Finally we note that one could also calculate conductivity directly working with the nodal Lagrangian (\[Dirac.conjugated.Lagrangian\]) and expressing the bubble in terms of the corresponding electrical current operator $$j_x = e v_{F} \bar{\chi}_s \gamma^{1} \chi_s, \qquad j_y = e v_{D} \bar{\chi}_s \gamma^{2} \chi_s,$$ which is also valid in the presence of an external field. This way of calculation is exactly the same as in Ref. [@Yang:2002:PRB] and final result agrees with (\[conductivity.general.final\]). Calculation of conductivity {#sec:conductivity.calculation} --------------------------- Straightforward calculation of the trace in Eq. (\[conductivity.general.final\]), with $A_D(\omega, \mathbf{k})$ from Eqs. (\[AD.clean\]) and (\[f\]) but with the $\delta$ functions in Eq. (\[AD.clean\]) replaced by the Lorentzians (\[Gamma\]), gives $$\label{tr.same} \begin{split} & \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{1,2} A_D(\omega, {\bf p}) \gamma^{1,2} \right] = \frac{4 \Gamma^2}{\pi^2} \exp \left( - \frac{2 \mathbf{p}^2}{eB} \right) \\ & \times \sum_{n,m =0}^{\infty} (-1)^{n+m+1} \frac{[(\omega^2 + M_n^2 + \Gamma^2) (\omega^2 + M_m^2 + \Gamma^2) - 4 \omega^2 \Delta^2][L_n L_{m-1} + L_{n-1} L_{m}] \mp 32 (p_1^2 - p_2^2) L_{n-1}^1 L_{m-1}^1} {[(\omega^2 + M_n^2 + \Gamma^2)^2 - 4 \omega^2 M_n^2] [(\omega^2 + M_m^2 + \Gamma^2)^2 - 4 \omega^2 M_m^2]}, \end{split}$$ where all Laguerre’s polynomials depend on $\frac{2 \mathbf{p}^2}{eB}$, and the minus sign corresponds to $\gamma^1$ and the plus sign to $\gamma^2$, respectively. The integration over momentum $p$ can be easily done after extending the upper limit of integration to $\infty$, so that one can use the orthogonality of Laguerre’s polynomials $$\int \limits_{0}^{\infty} d xe^{-x} x^{\alpha} L_{m}^{\alpha} (x) L_{n}^{\alpha} (x) = \Gamma(1+\alpha) \frac{(n+\alpha)!}{n! \alpha!} \delta_{mn} \quad \mbox{with} \quad x = \frac{2 \mathbf{p}^2}{e B},$$ and obtain electrical conductivity in terms of the sum over the transitions between neighboring Landau levels $$\label{sigma.final1} \begin{split} \sigma(B,T) & = e^2 \frac{v_F^2 + v_D^2}{v_F v_D} \frac{e B \Gamma^2}{2 \pi^2 T} \\ & \times \sum_{n=0}^{\infty} \int \limits_{-\infty}^{\infty} d\omega \frac{1}{\cosh^2 \frac{\omega - \mu}{2T}} \frac{(\omega^2 + M_n^2 + \Gamma^2) (\omega^2 + M_{n+1}^2 + \Gamma^2) - 4 \omega^2 \Delta^2} {[(\omega^2 + M_n^2 + \Gamma^2)^2 - 4 \omega^2 M_n^2] [(\omega^2 + M_{n+1}^2 + \Gamma^2)^2 - 4 \omega^2 M_{n+1}^2]}. \end{split}$$ The sum over $n$ in Eq. (\[sigma.final1\]) can be expressed via the digamma function $\psi$ as described in Ref. [@Ferrer:2002], and the final expression for the electrical conductivity is $$\label{sigma.final2} \sigma = e^2 \alpha \int \limits_{-\infty}^{\infty} \frac{d \omega}{4T\cosh^2 \frac{\omega - \mu}{2T}} \mathcal{A}(\omega,B,\Gamma,\Delta(B)),$$ where we introduced the function $\mathcal{A}$ $$\label{A.def} \begin{split} \mathcal{A}(\omega,B,\Gamma,\Delta(B))= \frac{1}{\pi^2} \frac{\Gamma^2}{(e B)^2 + (2 \omega \Gamma)^2} & \left\{ 2 \omega^2 + \frac{(\omega^2 + \Delta^2 + \Gamma^2)(eB)^2 - 2 \omega^2 (\omega^2 - \Delta^2 + \Gamma^2) eB} {(\omega^2 - \Delta^2 - \Gamma^2)^2 + 4 \omega^2 \Gamma^2} \right. \\ &\left. - \frac{\omega(\omega^2 - \Delta^2 + \Gamma^2)}{\Gamma} \mbox{Im} \psi \left( \frac{\Delta^2 + \Gamma^2 - \omega^2 - 2 i \omega \Gamma} {2eB}\right) \right\} \end{split}$$ and $$\label{alpha} \alpha = \frac{v_F}{v_D} + \frac{v_D}{v_F}.$$ Another representation of Eq. (\[A.def\]) can be obtained using the series representation of $\psi$ function and writing the expression in curly brackets in fractions of $1/(\Gamma^2 + x^2)$, $$\label{A.series} \begin{split} \mathcal{A}(\omega,B,\Gamma,\Delta(B))= \frac{1}{\pi^2} \frac{\Gamma^2}{(e B)^2 + (2 \omega \Gamma)^2} & \left\{ 2 \omega^2 + \frac{\frac{(eB)^2}{2} + eB \omega(\omega + \Delta)} {(\omega+\Delta)^2 + \Gamma^2} + \frac{\frac{(eB)^2}{2} + eB \omega(\omega - \Delta)} {(\omega-\Delta)^2 + \Gamma^2} \right. \\ &\left. + eB\omega \sum_{n=1}^{\infty} \frac{1}{M_n} \left[\frac{\Delta^2 + M_n^2 + 2 \omega M_n}{(\omega + M_n)^2 + \Gamma^2} + \frac{2 \omega M_n - \Delta^2 - M_n^2}{(\omega - M_n)^2 + \Gamma^2}. \right] \right\}, \end{split}$$ The representation (\[A.series\]) is particularly convenient for studying the narrow width limit $\Gamma \ll T, \sqrt{eB}$ when we can replace the fractions $\Gamma/(\Gamma^2 + x^2)$ by $\pi \delta(x)$: $$\label{A.series.delta} \begin{split} \mathcal{A}(\omega,B,\Gamma,\Delta(B))= \frac{\Gamma}{\pi} & \left\{ \frac{1}{(e B)^2 + 4 \Delta^2 \Gamma^2} \left[\frac{(eB)^2}{2} \delta(\omega + \Delta) + \frac{(eB)^2}{2} \delta(\omega - \Delta) \right] \right. \\ & \left. + \sum_{n=1}^{\infty} \frac{2 (e B)^2 n}{(e B)^2 + 4 (\Delta^2 + 2 eB n) \Gamma^2} \left[ \delta(\omega + M_n) + \delta(\omega - M_n) \right] \right\}, \end{split}$$ where we kept $\Gamma^2$ in the denominators in order to be able to reproduce a smooth behavior of $\sigma(B)$ and $\kappa(B)$ in the limit $B \to 0$. We are now in a position to study different asymptotic regimes defined by different relations among the parameters $\Gamma$, $T$, $\mu$, $B$, and $\Delta$. Zero magnetic field {#sec:conductivity.B=0} ------------------- We begin with the limit of vanishing magnetic field ($B =0$). Using the large $z$ asymptote of the $\psi$ function $$\psi(z) = \ln z - \frac{1}{2z} - \frac{1}{12z^2} + \frac{1}{120z^4} + O\left(\frac{1}{z^6} \right)$$ we obtain $$\label{A.B=0} \mathcal{A}(\omega,B=0,\Gamma,\Delta)= \frac{1}{2 \pi^2} \left[ 1 + \frac{\omega^2 - \Delta^2 + \Gamma^2}{2 |\omega| \Gamma} \left(\frac{\pi}{2} - \arctan \frac{\Delta^2 + \Gamma^2 - \omega^2} {2|\omega| \Gamma} \right) \right].$$ Putting $\Delta = 0$ also in Eq. (\[A.B=0\]), we get $$\label{A.B=0.Delta=0} \mathcal{A}(\omega,B=0,\Gamma,\Delta =0)= \frac{1}{2 \pi^2} \left[ 1 + \frac{\omega^2 + \Gamma^2}{\omega \Gamma} \arctan \frac{\omega}{ \Gamma} \right].$$ ### Limit $T \to 0$ The limit $T \to 0$ is significantly simplified by the fact that $\cosh^{-2}$ term in Eq. (\[sigma.final2\]) can be replaced by the $\delta$ function $$\label{electric.B=0.final} \sigma = e^2 \alpha \mathcal{A}(\mu,B=0,\Gamma,\Delta) ,$$ where $\mathcal{A}$ is given by Eq. (\[A.B=0\]). For the case of zero gap $\Delta =0$ using Eq. (\[A.B=0.Delta=0\]) we obtain that for $\mu =0$ $$\label{sigma.mu=0} \sigma (\mu =0)= \frac{e^2 \alpha}{\pi^2}$$ and for $|\mu| \gg \Gamma$ $$\label{sigma.mu>Gamma} \sigma = \frac{e^2 \alpha}{4\pi} \frac{|\mu|}{\Gamma}.$$ Comparing Eq. (\[sigma.mu=0\]) with Eq. (57) of YN, one can see that our result is twice as large, but contains the same prefactor $\alpha$ that reflects the fact that the electrical current operator has the component $\sim \mathbf{v}_D$. On the other hand, the result of KC is $\sim v_F/v_D$ due to the fact that they took the current without the $\mathbf{v}_D$ component, and it is twice as large as Eq. (\[sigma.mu=0\]) because KC integrated over the full Brillouin zone instead of the reduced one. Expression (\[sigma.mu=0\]) can also be compared with its superconducting counterpart [@Durst:2000:PRB] $\sigma_{SC} = (e^2/\pi^2) v_{F}/v_{\Delta}$, where $v_{\Delta}$ is the $d$SC gap velocity. As one can see the numerical prefactor is [*exactly the same*]{}, but there is no $v_{\Delta}/v_{F}$ term in $\sigma_{SC}$ because the electrical current in the $d$SC state is $\sim\mathbf{v}_F$. Finally one can also compare our $\mu \neq 0$ expression (\[sigma.mu&gt;Gamma\]) with Eq. (13) of KC: $$\label{KC.mu} \sigma(T \ll D_0) \simeq \frac{e^2}{2 \pi} \frac{v_F}{v_D} \frac{D_0}{\gamma_0},$$ where in the Born limit $\gamma_0$ is related to $\Gamma(\omega)$ via $\Gamma (\omega) = \gamma_0 [(\omega + \mu)/D_0]$. Substituting the value of $\Gamma(0)$ in Eq. (\[sigma.mu&gt;Gamma\]) one can see that it reduces to Eq. (\[KC.mu\]) except for the abovementioned differences in the factors. ### Limit $T \ll \Gamma$ In order to obtain low-temperature corrections to conductivities it is convenient to apply the Sommerfeld expansion $$\label{Fermi.expansion} \begin{split} & \int \limits_{-\infty}^{\infty} \frac{(\omega - \mu)^n}{4T\cosh^2 \frac{\omega - \mu}{2T}} f(\omega) d \omega \simeq \int \limits_{-\infty}^{\infty} \frac{(\omega - \mu)^n}{4T\cosh^2 \frac{\omega - \mu}{2T}} d \omega f(\mu) \\ & + \int \limits_{-\infty}^{\infty} \frac{(\omega - \mu)^{n+1}}{4T\cosh^2 \frac{\omega - \mu}{2T}} d \omega f^{\prime}(\mu) + \frac{1}{2} \int \limits_{-\infty}^{\infty} \frac{(\omega - \mu)^{n+2}}{4T\cosh^2 \frac{\omega - \mu}{2T}} d \omega f^{\prime \prime}(\mu) + O \left( \left(\frac{T}{\mbox{max} (\mu, \Gamma)} \right)^{n+3} \right). \end{split}$$ Here the function $f$ should be nonsingular and not too rapidly varying in the vicinity of $\omega = \mu$ and the integrals on the right hand side of Eq. (\[Fermi.expansion\]) are $$\int \limits_{-\infty}^{\infty} \frac{d \omega}{4T\cosh^2 \frac{\omega - \mu}{2T}} =1, \qquad \int \limits_{-\infty}^{\infty} \frac{d \omega (\omega - \mu)^2}{4T\cosh^2 \frac{\omega - \mu}{2T}} = \frac{\pi^2 T^2}{3}, \qquad \int \limits_{-\infty}^{\infty} \frac{d \omega(\omega - \mu)^4}{4T\cosh^2 \frac{\omega - \mu}{2T}} = \frac{7 \pi^4 T^4}{15} .$$ It is easy to derive the next to the leading term in $T^2$ to the conductivity (\[electric.B=0.final\]) for $\Delta =0$, $$\label{sigma.B=0.final2} \sigma = e^2 \alpha \left[ \mathcal{A}(\omega = \mu, 0, \Gamma, 0) + \frac{\pi^2 T^2}{6} \mathcal{A}^{\prime \prime}_\omega(\omega= \mu, 0, \Gamma,0) \right] ,$$ where the derivatives of $\mathcal{A}$ with respect to $\omega$ are $$\label{A.derivatives} \begin{split} & \mathcal{A}^{\prime}_{\omega} (\omega, 0, \Gamma, 0) = \frac{1}{2 \pi^2 \omega} \left( 1 + \frac{\omega^2 - \Gamma^2}{\omega \Gamma} \arctan \frac{\omega}{\Gamma} \right), \\ & \mathcal{A}^{\prime \prime}_{\omega} (\omega, 0, \Gamma, 0) = \frac{1}{\pi^2 \omega^2} \left(\frac{\Gamma}{\omega} \arctan \frac{\omega}{\Gamma} - \frac{\Gamma^2}{\omega^2 + \Gamma^2} \right). \end{split}$$ For $\mu = 0$ case Eq. (\[sigma.B=0.final2\]) gives the $T^2$ correction to the expression (\[sigma.mu=0\]): $$\sigma (\mu =0)= \frac{e^2 \alpha}{\pi^2} \left[1 + \frac{\pi^2}{9} \frac{T^2}{\Gamma^2} \right], \qquad T \ll \Gamma.$$ ### Limit $T \gg \Gamma$ For $\Gamma \to 0$ from Eq. (\[A.B=0\]) we get $$\label{A.B=0.T>Gamma} \mathcal{A}(\omega,B=0,\Gamma,\Delta)= \frac{1}{4 \pi^2} \frac{\omega^2 - \Delta^2}{|\omega| \Gamma} \pi \theta( \omega^2 - \Delta^2).$$ For the $\Delta = 0$ case retaining more terms in the expansion of Eq. (\[A.B=0.Delta=0\]), we obtain $$\label{A.B=D=0.T>Gamma} \mathcal{A}(\omega,B=0,\Gamma,\Delta =0)= \frac{1}{4 \pi} \left( \frac{|\omega|}{\Gamma} + \frac{\Gamma}{|\omega|} \right).$$ Substituting Eq. (\[A.B=D=0.T&gt;Gamma\]) in the expression (\[sigma.final2\]) for $\sigma$ we arrive at $$\label{sigma.T>Gamma} \sigma = e^2 \alpha \frac{T}{2\pi \Gamma} \ln \left(2 \cosh \frac{\mu}{2T} \right), \quad T \gg \Gamma .$$ As one can easily see for $\mu \gg T$ Eq. (\[sigma.T&gt;Gamma\]) reduces to Eq. (\[sigma.mu&gt;Gamma\]) reflecting the fact that when $\mu$ is the largest parameter the value of conductivity is not sensitive to the relation between $T$ and $\Gamma$. As mentioned before, the magnetic catalysis phenomenon [@Gusynin:1995:PRD; @Ferrer:2002; @Gorbar:2002:PRB] implies that the gap $\Delta$ is generated [*only*]{} in the presence of an external field. Nevertheless, one can also study the consequences of the gap opening even for $B =0$ case to gain a deeper insight to a more complicated case $B, \Delta \neq 0$. In particular, considering the $\Delta \gg T$ case, we obtain from Eqs. (\[A.B=0.T&gt;Gamma\]) and (\[sigma.final2\]) that $$\label{sigma.Delta} \sigma = \frac{e^2 \alpha }{16\pi T \Gamma} \int \limits_{\Delta}^{\infty} d \omega \frac{\omega^2 - \Delta^2}{\omega} \left[\frac{1}{\cosh^2 \frac{\omega-\mu}{2T}} + (\mu \to - \mu) \right] \simeq \frac{e^2 \alpha}{4 \pi \Gamma} \begin{cases} & \dfrac{\mu^2 - \Delta^2}{|\mu|} - \dfrac{\pi^2 T^2 \Delta^2}{3 \mu^2 |\mu|}, \qquad \quad |\mu| > \Delta, \\ & 2 T \ln 2, \qquad \qquad \qquad \qquad \quad |\mu| = \Delta,\\ & 4 T \cosh \dfrac{\mu}{T} \exp\left( - \dfrac{\Delta}{T} \right), \qquad |\mu| < \Delta. \end{cases}$$ It is clear from Eq. (\[sigma.Delta\]) that the opening of the gap $\Delta$ results in the thermally activated behavior of conductivity only for $|\mu | < \Delta$. As discussed in Sec. \[sec:sigma.numerical\] this observation remains valid even in the presence of an external field. Nonzero magnetic field ---------------------- There are not so many cases available for analytical investigation for $B \neq 0$ and we have to integrate numerically Eq. (\[sigma.final2\]) with $\mathcal{A}$ given by Eq. (\[A.def\]). Nevertheless in a few cases analytical expressions for the conductivity can be obtained and here we begin with considering these results. ### Limit $T \to 0$ As one can notice, Eq. (\[electric.B=0.final\]) is in fact valid even for nonzero $B$ because only the first of term of the Sommerfeld expansion (\[Fermi.expansion\]) contributes, i.e., $$\label{electric.B.final} \sigma = e^2 \alpha \mathcal{A}(\mu,B,\Gamma,\Delta) , \qquad \forall \mu, \quad \Gamma \neq 0.$$ ### Narrow width case To study the narrow width limit $\Gamma \to 0$ we use the representation (\[A.series.delta\]) to arrive at $$\begin{split} \label{sigma.field.narrow} \sigma = e^2 \alpha \frac{\Gamma}{4 \pi T} & \left\{ \frac{(eB)^2}{2[(e B)^2 + 4 \Delta^2 \Gamma^2]} \left[\frac{1}{\cosh^2 \frac{\Delta + \mu}{2T}} + \frac{1}{\cosh^2 \frac{\Delta - \mu}{2T}} \right] \right. \\ & \left. + \sum_{n=1}^{\infty} \frac{2 (e B)^2 n}{(e B)^2 + 4 (\Delta^2 + 2 eB n) \Gamma^2} \left[ \frac{1}{\cosh^2 \frac{\sqrt{\Delta^2 + 2 eBn}+\mu}{2T}} + \frac{1}{\cosh^2 \frac{\sqrt{\Delta^2+ 2 eBn}-\mu}{2T}} \right] \right\}. \end{split}$$ The fact that the conductivity is proportional to the scattering rate $\Gamma$ in the limit $\Gamma \to 0$ means that in contrast to the zero-field case (see Sec. \[sec:conductivity.B=0\]) it results from transitions of quasiparticles between neighboring cyclotron orbits. Using Eq. (\[sigma.field.narrow\]) one easily gets the asymptotic expression for conductivity in the [*strong-field limit*]{}, $\sqrt{e B} \gtrsim 4 T$ for $\mu = \Delta =0$ $$\label{sigma.Bstrong} \sigma = e^2 \alpha \frac{\Gamma}{4 \pi T}$$ that shows that $\sigma$ becomes field independent in the strong field. ### Numerical calculation of electrical conductivity {#sec:sigma.numerical} To investigate numerically the behavior of electrical and thermal conductivities, and to make comparison with experiment, we need to restore all model parameters, such as $\hbar$, $c$, $k_B$, $v_{F}$, and $v_D$ in Eqs. (\[sigma.final2\]) and (\[A.def\]). As discussed after Eq. (\[scaling\]) the prefactor $\alpha$ in Eq. (\[sigma.final2\]) is already fixed and one should only substitute $T \to k_B T$, $\Gamma \to \hbar \Gamma$, and $e B \to (\hbar v_F v_D/c) e B$. It is convenient to measure all energetic quantities in K, which results in the following replacement: $$\label{units} \sqrt{2 e B} \to \sqrt{\frac{\hbar v_F v_D 2 e B}{c}} [\mbox{K}] = 63.9 \sqrt{\frac{v_{D}}{v_{F}}} v_{F} [\mbox{eV} \cdot \mbox{\AA}] \sqrt{B [\mbox{Tesla}]} = 4.206 \times 10^{-6} \sqrt{\frac{v_{D}}{v_{F}}} v_{F} [\mbox{cm}/\mbox{s}] \sqrt{B [\mbox{Tesla}]} ,$$ where in the first equality $v_{F}$ is given in $\mbox{eV} \cdot \mbox{\AA}$ and in $\mbox{cm}/\mbox{s}$ in the second one. In particular, for $v_F = 1.5 \mbox{eV} \cdot \mbox{\AA}$ \[this roughly agrees with the value of $t$ given after Eq. (\[Green.nodal\])\] and $v_F/v_D = 24$ using Eq. (\[units\]) we obtain that $e B \to 200 \cdot \mbox{K}^2 \cdot B [\mbox{Tesla}]$. In what follows we use this estimate to compute all numerical expressions. There is a larger uncertainty for the value of $\Gamma$ that we could choose for our computations. For example, for clean YBCO monocrystals the estimated [@Ando:2000:PRB] value of the scattering rate due to impurities $\Gamma_0 \sim 1-2 \mbox{K}$, so that we use the value $\Gamma = 2 \mbox{K}$. In Figs. \[fig:3\] and \[fig:4\] we show the temperature dependence of the conductivity for three different values of the applied field at half-filling (Fig. \[fig:3\]) and slightly away from it (Fig. \[fig:4\]). For $\mu = B= 0$ the conductivity increases as the temperature $T$ grows. However, when the magnetic field is nonzero, the dependence $\sigma(T)$ becomes non-monotonic and there is first a decrease in $\sigma(T)$ as the temperature increases. For $\mu, B \neq 0$ (see Fig. \[fig:4\]) the conductivity slowly increases as the temperature grows and this dependence becomes almost flat as the field increases. This tendency agrees with Eqs. (\[sigma.field.narrow\]) and (\[sigma.Bstrong\]). In zero field both figures reveal an “‘insulator’’ (i.e., increasing with temperature) behavior in agreement with our analytical result (\[sigma.T&gt;Gamma\]). This behavior is a consequence of using a constant value for the scattering rate $\Gamma$ in our model. The growth of the conductivity with increasing temperature is directly related to the increasing number of thermally excited quasiparticles. The choice of temperature-independent $\Gamma$ might be reasonable in the narrow low-temperature region. It is essential, however, that the observed decrease and flattening $\sigma(T,B)$ in the nonzero external field are due to the assumption $\Gamma(B) = \mbox{const}$ discussed after Eq. (\[Gamma\]) and are not related to the $\Gamma(T) = \mbox{const}$ approximation just mentioned. Comparing Figs. \[fig:3\] and \[fig:4\], one can see that in zero field the increase of $|\mu|$ (opening of the pockets on the Fermi surface) leads to the increase of the conductivity. On the contrary, in the presence of the external field, the conductivity [*decreases*]{} as the pockets on the Fermi surface open. One can notice that the value of conductivity $\sigma$ at $\mu = T = 0$ [*is field independent*]{}. This is due to the fact that for $\omega = 0$, the function $\mathcal{A}$ \[see Eq. (\[A.def\]) and Fig. \[fig:14\]\] becomes field independent. Using Eqs. (\[sigma.final2\]) and (\[A.def\]) it is also possible to investigate the case of $\Delta \neq 0$. The main results for $B \neq 0$ can be foreseen from Eq. (\[sigma.Delta\]) and are the following. When $T,|\mu| \lesssim \Delta$ the behavior of $\sigma$ becomes thermally activated, i.e., governed by the factor $\sim \exp(-\Delta/T)$. In contrast, for $\Delta \lesssim |\mu|$ the behavior of $\sigma(T)$ is rather similar to the case of $\Delta = 0$. This reflects the fact that the gap $\Delta$ induced by the magnetic catalysis, in contrast to a superconducting gap, [*is not tied*]{} to the Fermi surface and there are gapless excitations for $|\mu| \gtrsim \Delta$. In Figs. \[fig:5\] and \[fig:6\] we show, respectively, the dependence of $\sigma(B)$ for three different values of $T$ and fixed $\mu$ and for three different values of $\mu$ at fixed $T$. Both figures show that $\sigma$ decreases as a function of $B$. This decrease reflects the fact mentioned before Eq. (\[sigma.final1\]) that in the presence of magnetic field only transitions between neighboring Landau levels contribute into electrical (and thermal) conductivity. A further increase of the field and entering in the [*strong field regime*]{}, $\sqrt{e B} \gtrsim 4 T$ suppresses the transition between Landau levels, so that as one can see from Fig. \[fig:5\], the conductivity $\sigma(B)$ becomes field independent. Furthermore for $T =10 \mbox{K}$ the value of normalized conductivity is $\approx 0.15$, in agreement with Eq. (\[sigma.Bstrong\]), which is derived under the assumption $\Gamma \ll T$ and taking into account only the transition between the lowest levels. In Fig. \[fig:6\] we observe Shubnikov-de Haas oscillations of the conductivity that are due to the Landau-level crossing of the Fermi pockets (an analytical treatment of the magnetic oscillations in the conductivities will be presented elsewhere). Thermal conductivity {#sec:thermal} ==================== General expression for thermal conductivity ------------------------------------------- Thermal conductivity can be calculated from the the energy current-current correlation function $$\label{polarization.EE} \Pi^{EE}_{\alpha \beta}(i \Omega) = - \int \limits_{0}^{\beta} d \tau e^{i \Omega \tau} \langle T_{\tau} j_{\alpha}^{E \dagger}(\tau, \mathbf{0}) j_{\beta}^{E}(0, \mathbf{0}) \rangle,$$ and the correlation function of energy current with the electrical current $$\label{polarization.EC} \Pi_{\alpha \beta}^{EC}(i \Omega) = - \int \limits_{0}^{\beta} d \tau e^{i \Omega \tau} \langle T_{\tau} j_{\alpha}^{E \dagger}(\tau, \mathbf{0})j_{\beta}(0, \mathbf{0}) \rangle$$ using a thermal Kubo formula [@Mahan:book; @foot1] $$\label{thermal.def} \frac{\kappa(\Omega)}{T} = - \frac{1}{T^2} \frac{\mbox{Im} \Pi_{R}^{EE}(\Omega +i0)}{\Omega} - S^2(\Omega) \sigma(\Omega,B,T) = - \frac{1}{T^2} \frac{\mbox{Im} \Pi_{R}^{EE}(\Omega +i0)}{\Omega} - \frac{1}{T^2} \frac{[\mbox{Im} \Pi_{R}^{EC}(\Omega +i0)]^2} {\sigma(\Omega) \Omega^2}.$$ Here $\Pi_{R}^{EE}(\Omega +i0) = \Pi^{EE}(i \Omega \to \Omega + i0)$ is the longitudinal polarization \[see Eq. (\[tensor2scalar\])\] and $S(\Omega,B,T)$ is the thermopower $$S(\Omega) = - \frac{1}{T} \frac{\mbox{Im} \Pi_{R}^{EC} (\Omega)} {\mbox{Im} \Pi_{R}^{CC} (\Omega)},$$ with the longitudinal $\Pi_{R}^{EC}(\Omega +i0) = \Pi^{EC}(i \Omega \to \Omega + i0)$. The term with the thermal power $S$ ensures that the energy current is evaluated under the condition of vanishing electrical current [@Langer:1962:PR] (see also Ref. [@Mahan:book]). Usually, for $T \ll \mu $ this term is considered to be unimportant because it is $ \sim T^2/\mu^2$ times less than the first term of Eq. (\[thermal.def\]). It is also zero for $\mu =0$, but nevertheless its contribution is important for the case of interest, $|\mu| \lesssim T$. The structure of the energy current (\[heat.current2.B=0\]) is similar to that of the electrical current (\[electric.current.DDW\]). Hence, the calculation of the corresponding polarization functions (\[polarization.EE\]) and (\[polarization.EC\]) is almost identical to calculating the bubble (\[el.cur-cur.tensor\]). Using the general result (\[bubble.final\]) to the case of interest: $g = g^{\prime} = \omega+ \Omega/2 - \mu$, we arrive at $$\label{thermal.conductivity.general.final1} \begin{split} - \frac{\mbox{Im} \Pi_R^{EE}(\Omega)} {T^2 \Omega}& = \frac{\pi}{v_{F}v_{D}} \int \frac{d^{2}p}{(2\pi)^{2}} \int \limits_{-\infty}^{\infty} d\omega \frac{\tanh \frac{\omega - \mu + \Omega}{2T} - \tanh \frac{\omega - \mu}{2T}}{\Omega} \left( \frac{\omega - \mu + \Omega/2 }{T} \right)^2 \\ & \times \left[ v_F^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{1} A_D(\omega + \Omega, {\bf p}) \gamma^{1} \right] + v_D^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{2} A_D(\omega + \Omega, {\bf p}) \gamma^{2} \right] \right], \end{split}$$ and for $g = \omega+ \Omega/2 - \mu$, $g^{\prime} = e$ $$\label{thermal.conductivity.general.final2} \begin{split} - \frac{\mbox{Im} \Pi_R^{EC}(\Omega)} {T \Omega}& = \frac{e \pi}{v_{F}v_{D}} \int \frac{d^{2}p}{(2\pi)^{2}} \int \limits_{-\infty}^{\infty} d\omega \frac{\tanh \frac{\omega - \mu + \Omega}{2T} - \tanh \frac{\omega - \mu}{2T}}{\Omega} \frac{\omega - \mu + \Omega/2 }{T} \\ & \times \left[ v_F^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{1} A_D(\omega + \Omega, {\bf p}) \gamma^{1} \right] + v_D^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{2} A_D(\omega + \Omega, {\bf p}) \gamma^{2} \right] \right]. \end{split}$$ Similarly to the case of the electrical current, one could also obtain the energy current operator from the nodal Lagrangian (\[Dirac.conjugated.Lagrangian\]) in the external magnetic field as done in Ref. [@Ferrer:2002], $$\label{energy.current.nodal} j_x^E = \frac{i v_{F}}{2} ( \bar{\chi}_s \gamma^{1} \partial_t \chi_s - \partial_t \bar{\chi}_s \gamma^{1} \chi_s), \qquad j_y^E = \frac{i v_{D}}{2} ( \bar{\chi}_s \gamma^{2} \partial_t \chi_s - \partial_t \bar{\chi}_s \gamma^{2} \chi_s)$$ to derive the expression (\[thermal.conductivity.general.final2\]). The current (\[energy.current.nodal\]) corresponds to the current used by YN. The calculation of thermal conductivity from Eq. (\[thermal.conductivity.general.final2\]) follows exactly the same route as for the electrical conductivity in Sec. \[sec:conductivity.calculation\] and finally we arrive at $$\label{thermal.conductivity.general.final3} \begin{split} \frac{\kappa (B,T)}{T} = \alpha & \left\{ \int \limits_{-\infty}^{\infty} d \omega \left( \frac{\omega-\mu}{T} \right)^2 \frac{1}{4T\cosh^2 \frac{\omega - \mu}{2T}} \mathcal{A}(\omega,B,\Gamma,\Delta(B)) \right. \\ & - \left. \frac{e^2 \alpha}{\sigma (B,T)} \left[ \int \limits_{-\infty}^{\infty} d \omega \frac{\omega-\mu}{T} \frac{1}{4T\cosh^2 \frac{\omega - \mu}{2T}} \mathcal{A}(\omega,B,\Gamma,\Delta(B)) \right]^2 \right\} , \end{split}$$ where $\mathcal{A}$ is the same function (\[A.def\]) as for the electrical conductivity and $\alpha$ is given by Eq. (\[alpha\]). Thus we are ready to study the thermal conductivity as a function of $\Gamma$, $T$, $\mu$, $B$, and $\Delta$. Zero magnetic field {#zero-magnetic-field} ------------------- To study the $B=0$ case we substitute the expression for $\mathcal{A}(B=0)$ given by Eq. (\[A.B=0\]) into the general expression for thermal conductivity (\[thermal.conductivity.general.final2\]). ### Limit $T \to 0$ Corresponding to Eq. (\[electric.B=0.final\]) limit $T\to 0$ is given by $$\label{kappa.B=0.final} \frac{\kappa}{T} = \frac{\pi^2}{3} \alpha \mathcal{A}(\mu,B=0,\Gamma,\Delta) .$$ Then using Eq. (\[A.B=0.Delta=0\]) for $\mu = \Delta =0$, we obtain $$\label{kappa.mu=0} \frac{\kappa (\mu =0)}{T} = \frac{\alpha}{3} .$$ The factor $\alpha$ in Eq. (\[kappa.mu=0\]) is present also in the DDW [@Yang:2002:PRB; @Kim:2002:PRB] and $dSC$ [@Durst:2000:PRB] cases where the same expression for the thermal current has been used. The overall numerical factor $1/3$ in Eq. (\[kappa.mu=0\]) is the same as in Ref. [@Durst:2000:PRB] and it is different from Refs. [@Yang:2002:PRB; @Kim:2002:PRB] in the same way as the corresponding factor in expression (\[sigma.mu=0\]) for the electrical conductivity. ### Limit $T \ll \Gamma$ Using the Sommerfeld expansion (\[Fermi.expansion\]) we can also derive next to the leading term in $T^2$ to the thermal conductivity for $\Delta = 0$ $$\label{kappa.B=0.final2} \frac{\kappa}{T} = \frac{\alpha}{3}\left[ \mathcal{A}(\omega = \mu,0,\Gamma, 0) + \frac{7 \pi^2 T^2}{30} \mathcal{A}^{\prime \prime}_{\omega}(\omega = \mu,0,\Gamma, 0) - \frac{\pi^2 T^2}{9} \frac{(\mathcal{A}^{\prime}_{\omega}(\omega = \mu,0,\Gamma, 0))^2} {\mathcal{A}(\omega = \mu,0,\Gamma, 0)} \right],$$ where the derivatives of $\mathcal{A}$ are given by Eq. (\[A.derivatives\]). For $\mu = 0$, since there is no contribution from the thermal power term $\Pi^{EC}$, Eq. (\[kappa.B=0.final2\]) reduces to the expression $$\frac{\kappa (\mu =0)}{T} = \frac{\alpha}{3} \left[ 1 + \frac{7 \pi^2}{15} \frac{T^2}{\Gamma^2}\right] , \qquad T \ll \Gamma.$$ ### Limit $T \gg \Gamma$ For $\Gamma \to 0$ and $\Delta = 0$ substituting Eq. (\[A.B=D=0.T&gt;Gamma\]) into the expression (\[thermal.conductivity.general.final3\]) for $\kappa$ we obtain (see Appendix \[sec:C\]) $$\label{kappa.T>Gamma} \begin{split} \frac{\kappa}{T} \simeq \alpha \left[ \frac{\pi}{12} \frac{\mu}{\Gamma} - \frac{\pi^3}{36} \frac{T^2}{\Gamma \mu} \right], \qquad \mu \gg T \end{split}$$ where the first term $\sim \mu$ of $\kappa$ arises from the first term of Eq. (\[thermal.conductivity.general.final3\]) and the second term $\sim 1/\mu$ originates from the thermal power term $\Pi^{EC}$. Thus, as one would expect, the thermal power contribution into the thermal conductivity is important only for $|\mu| \lesssim T$. For the case $\mu = \Delta = 0$ using Eq. (\[A.B=D=0.T&gt;Gamma\]) we arrive at the expression $$\label{kappa.T>Gamma.mu=0} \begin{split} \frac{\kappa(\mu =0)}{T} = \frac{\alpha}{\pi} \left[ \frac{9 \zeta(3)}{4} \frac{T}{\Gamma} + \frac{\ln 2}{2} \frac{\Gamma}{T} \right], \qquad \Gamma \ll T \end{split}$$ where $\zeta (z)$ is the Riemann zeta function. Finally using the representation (\[A.B=0.T&gt;Gamma\]), which is valid for $\mu =0$ we can get the thermal conductivity for nonzero $\Delta$ [@Ferrer:2002]: $$\label{kappa.Delta} \frac{\kappa(\mu =0)}{T} = \frac{\alpha}{8\pi T^3 \Gamma} \int \limits_{\Delta}^{\infty} d \omega \frac{|\omega| (\omega^2 - \Delta^2)}{\cosh^2 \frac{\omega}{2T}} \simeq \alpha \frac{\Delta^2}{\pi \Gamma T} \exp \left( - \frac{\Delta}{T} \right), \qquad \Delta \gg T.$$ Nonzero magnetic field ---------------------- As for the electrical conductivity there are not so many cases available for analytical investigation for $B \neq 0$ and we have to integrate numerically Eq. (\[thermal.conductivity.general.final3\]) with $\mathcal{A}$ given by Eq. (\[A.def\]). There are still a few cases when the analytical treatment is possible and we begin by looking at them. ### Limit $T \to 0$ As one can check, Eq. (\[kappa.B=0.final\]) is valid even for nonzero $B$, so that $$\label{kappa.B.final} \frac{\kappa}{T} = \frac{\pi^2}{3} \alpha \mathcal{A}(\mu,B=0,\Gamma,\Delta) , \qquad \forall \mu, \quad \Gamma \neq 0.$$ ### Narrow width case Using the narrow width representation (\[A.series.delta\]) for $\mathcal{A}$, we arrive at the following result, $$\label{kappa.field.narrow} \begin{split} \frac{\kappa}{T} = \alpha \frac{\Gamma}{4 \pi T^3} & \left\{ \frac{(eB)^2}{2[(e B)^2 + 4 \Delta^2 \Gamma^2]} \left[\frac{(\Delta + \mu)^2}{\cosh^2 \frac{\Delta + \mu}{2T}} + \frac{(\Delta - \mu)^2} {\cosh^2 \frac{\Delta - \mu}{2T}} \right] \right. \\ & \left. + \sum_{n=1}^{\infty} \frac{2 (e B)^2 n}{(e B)^2 + 4 (\Delta^2 + 2 eB n) \Gamma^2} \left[ \frac{(\sqrt{\Delta^2 + 2 eBn}+\mu)^2} {\cosh^2 \frac{\sqrt{\Delta^2 + 2 eBn}+\mu}{2T}} + \frac{(\sqrt{\Delta^2+ 2 eBn}-\mu)^2} {\cosh^2 \frac{\sqrt{\Delta^2+ 2 eBn}-\mu}{2T}} \right] \right\} + \frac{\kappa_1}{T}, \end{split}$$ where $\kappa_1$ is a term originating from the condition of vanishing electrical current. As was discussed above, it is zero for $\mu =0$ and can be neglected for $|\mu| \gg T$. In particular, for $\mu = \Delta =0$ in the [*strong field limit*]{}, $\sqrt{e B} \gtrsim 4T$ we obtain from Eq. (\[kappa.field.narrow\]) the expression $$\label{kappa.Bstrong} \frac{\kappa}{T} = \alpha \frac{8 \Gamma e B}{\pi T^3} e^{-\frac{\sqrt{e B}}{T}}.$$ ### Numerical calculation of thermal conductivity {#sec:kappa.numerical} In Figs. \[fig:7\] and \[fig:8\] we present the results for the thermal conductivity that were obtained for the same values of the model parameters as the data for electrical conductivity shown in Figs. \[fig:3\] and \[fig:4\]. Comparing these figures, one can see that the behavior of the thermal conductivity is rather similar to the behavior of electrical conductivity and to see more subtle differences we have to consider the temperature dependence of the Lorenz number as done in Sec. \[sec:WF\]. It is also interesting to investigate numerically the contribution from the second term of Eq. (\[thermal.conductivity.general.final3\]) to the thermal conductivity. This is done in Fig. \[fig:9\], where one can clearly see that when the temperature grows and $T$ becomes $\lesssim |\mu|$ the second term of Eq. (\[thermal.conductivity.general.final3\]) makes an important negative contribution to the thermal conductivity. In Figs. \[fig:10\] and \[fig:11\] we present the results for the dependence of thermal conductivity on the magnetic field $B$ that were obtained for the same values of the model parameters as the data for electrical conductivity shown in Figs. \[fig:5\] and \[fig:6\]. Wiedemann-Franz law {#sec:WF} =================== After deriving in Secs. \[sec:electrical\] and \[sec:thermal\] the expressions for electrical and thermal conductivities in the various limits we are ready to consider their implications for the WF law. Limit $T \to 0$, arbitrary field, and chemical potential -------------------------------------------------------- For finite $\Gamma$, arbitrary $\mu$, $B$, and $\Delta$ in the limit $T \to 0$ one can see from Eq. (\[electric.B.final\]) \[see also Eq.  (\[electric.B=0.final\]) for $B =0$\] and (\[kappa.B.final\]) \[see also Eq. (\[kappa.B=0.final\]) for $B=0$\] that WF law is maintained: $$\label{L0} L_0 = \frac{\kappa}{\sigma T} = \frac{\pi^2}{3} \frac{k_B^2}{e^2} .$$ This result can be understood from the qualitative arguments given, for example, in Ref. [@Taylor.book]. The expressions for the electrical, Eq. (\[sigma.final2\]), and thermal, Eq. (\[thermal.conductivity.general.final3\]), conductivities calculated within the bare bubble approximation are very similar in the limit $T \to 0$ when the second term of Eq. (\[thermal.conductivity.general.final3\]) vanishes. Both expressions contain the function $\mathcal{A}(\omega)$ multiplied by the derivative of the Fermi distribution $f(\omega)= -n_F^{\prime}(\omega -\mu)$, with the only difference that the thermal conductivity is also multiplied by the factor $\sim (\omega -\mu)^2$. These prefactors $f(\omega)$ and $g(\omega) = - (\omega - \mu)^2 n_F^{\prime}(\omega -\mu)/T^2$ are shown in Figs. \[fig:12\] and \[fig:13\]. When the temperature $T$ goes to zero, the $\delta$-like spikes of $f(\omega)$ and $g(\omega)$ occur at the same value $\omega = \mu$ and the WF law is maintained. Zero magnetic field at $T \neq 0$ --------------------------------- For $\mu = 0$ and $T \gg \Gamma$ considering Eqs. (\[kappa.T&gt;Gamma.mu=0\]) and (\[sigma.T&gt;Gamma\]) we obtain that the WF law is violated, $$\label{Lorenz.mu=0} L = \frac{\kappa}{\sigma T} = \frac{9 \zeta(3)}{2\ln 2} \frac{k_B^2}{e^2} \simeq 2.37 L_0,$$ while for $\mu \gg T$ Eqs. (\[kappa.T&gt;Gamma\]) and (\[sigma.T&gt;Gamma\]) show that the Lorenz number has its usual value $$\label{Lorenz.mu>T} L = L_0.$$ These two results can also be understood qualitatively using Figs. \[fig:12\] and \[fig:13\]. For $|\mu| \gg T >0$ the electrical conductivity is determined by energies $\omega \approx \mu =0$, while the thermal conductivity is determined by the energies near $\omega \approx \mu \pm k_B T$. If the function $\mathcal{A}(\omega)$ does not vary appreciably over the energy range $\mu - k_B T$ to $\mu + k_B T$, the Lorenz number may still be $\sim 1$ even for finite $T$, $T \ll \mu$, as follows from Eq. (\[Lorenz.mu&gt;T\]). If $\mu$ becomes of the same order of $T$, the dimensionless values $\mu/ T$ and $\mu/T \pm 1$ are apart and the Lorenz number deviates from 1 as seen from Eq. (\[Lorenz.mu=0\]). For nonzero $\Delta \gg T$ and $\mu = 0$ we get from Eqs. (\[sigma.Delta\]) and (\[kappa.Delta\]) that $$\label{Lorenz.Delta.mu=0} L = \frac{k_B^2}{e^2} \frac{\Delta^2}{T^2},$$ i.e., $L > L_0$ for $\Delta > (\pi/\sqrt{3}) T$. This dominance of the thermal conductivity can be easily understood from the fact that when the gap opens, the electrical conductivity diminishes more strongly because it is determined by the energies $\omega \approx \mu$ near the Fermi surface. Nonzero magnetic field at $T \neq 0$ ------------------------------------ For $\mu = \Delta =0$ in the [*strong field limit*]{}, $\sqrt{e B} \gtrsim 4 T$ from Eqs. (\[sigma.Bstrong\]) and (\[kappa.Bstrong\]) we obtain that the Lorenz number becomes field dependent: $$L = \frac{32 eB}{T^2} e^{-\frac{\sqrt{e B}}{T}}.$$ In the presence of magnetic field the function $\mathcal{A}(\omega)$ (see Fig. \[fig:14\]) varies appreciably over the energy range $\omega - k_B T$ to $\omega + k_B T$, so that it is nonzero for $\omega = 0$ and very small at $\omega = \pm k_B T$. This feature of $\mathcal{A}(\omega)$, along with the small values of $\mu$, can produce rather strong violation of the WF law. Our qualitative arguments are indeed confirmed by Figs. \[fig:15\] and \[fig:16\] where we show the temperature dependence of $L(T)$ at half-filling (Fig. \[fig:15\]) and away from it (Fig. \[fig:16\]). Both these figures are computed on the basis of Eqs. (\[sigma.final2\]), (\[thermal.conductivity.general.final3\]), and (\[A.def\]). For $\mu =0$ the value of $L$ always goes down from $L_0$ as the temperature increases, then $L$ increases crossing $L_0$, and finally it goes to its zero-field value (\[Lorenz.mu=0\]). One can see from Fig. \[fig:16\] that for nonzero $\mu$ the deviations of $L(T)$ from $L_0$ become less pronounced (see also the discussion of Fig. \[fig:17\]). It is also instructive to investigate the importance of the second term of thermal conductivity, Eq. (\[thermal.conductivity.general.final3\]), for the Lorenz number calculated with and without this term (as already done in Fig. \[fig:9\] for the thermal conductivity itself). These results are given in Fig. \[fig:17\] and they show that the second term of thermal conductivity is indeed crucial to get the correct answer. Note that for the case of nonzero $\mu$ the situation with the WF law is more complicated because the value of $\mu$ may coincide with $\omega$, corresponding to the maximum of $\mathcal{A}(\omega)$. In this case $L(T)$ will firstly increase, as one can see in Fig. \[fig:17\] for $B= 12 \mbox{T}$. Conclusions {#sec:conclusions} =========== Let us compare the results derived in the present paper with the experimental results obtained in Ref. [@Hill:2001:Nature]. It is still unknown whether the DDW state exists and/or plays an important role in the electron doped compound used in Ref. [@Hill:2001:Nature], so that making this comparison would help to address these questions. First, we observe that while the results of [@Hill:2001:Nature] suggest that the WF law is violated at $T \to 0$, [*there is no*]{} violation of the WF law in this limit in the DDW scenario of the pseudogap. Since in Fig. 3 of Ref. [@Hill:2001:Nature] the electrical conductivity is a constant, the line $\kappa_e(T)/T$ directly represents the normalized Lorenz number $L(T)/L_0$. For finite temperatures there is then some similarity between Fig. \[fig:15\] (or Fig. \[fig:16\] for $\mu \neq 0$ case) and Fig. 3 of Ref. [@Hill:2001:Nature] where as $T$ increases the thermal conductivity crosses from the region with $\kappa_e/T < L_0/\rho_0$ to the region with $\kappa_e(T)/T > L_0/\rho_0 $ resembling the character of the WF law violation seen in the experiment. However, in the experiment, the electrical conductivity is flat while the thermal conductivity changes significantly in a subkelvin range. On the other hand, theoretical calculations show that the electrical (see Figs. \[fig:3\] and \[fig:4\]) and thermal (see Figs. \[fig:7\] and \[fig:8\]) conductivities vary simultaneously in a range of 10 K. Note that the wider range of temperatures can be probably attributed to the fact that the values of the model parameters we took for the numerical estimates are more appropriate for the hole-doped compounds. It is obvious from Fig. \[fig:15\] that such a behavior of $L(T)$ is due to the [*presence*]{} of the magnetic field. This confirms our claim that to interpret theoretically the experiment in Ref. [@Hill:2001:Nature] one should take into account the influence of the external field. Our main results can be summarized as follows. \(1) We have obtained analytical expressions for electrical conductivity (\[sigma.final2\]) and thermal conductivity (\[thermal.conductivity.general.final3\]) in the DDW state in presence of an external magnetic field. \(2) We have established a correspondence between the expression for the electrical conductivity (\[sigma.like.Langer\]) in the DDW state written in terms of the generalized velocity $\mathbf{V}(\mathbf{k})$, Eq. (\[velocity.general\]), and the dc conductivity derived by Langer [@Langer:1962a:PR] in the bubble approximation with the vertex (\[Lambda.Langer\]). \(3) We have shown that in the DDW system in the presence of impurities the WF law holds in $T \to 0$ limit for an arbitrary field $B$ and chemical potential $\mu$. This is checked within the bubble approximation, i.e., not including the impurity vertex. The influence of the impurity vertex on the properties of the DDW state in zero field at half-filling can be considered using the heuristical arguments of Durst and Lee [@Durst:2000:PRB] for $d$SC state. Since in the $d$SC state the thermal (and spin) currents are proportional to the group velocity, they can relax through either intranode scattering or scattering between nodes. As a result, the different types of scattering play nearly the same role and therefore vertex corrections do not contribute to the thermal (and spin) conductivity. This is not the case of the electrical current in the $d$SC state, because it depends only on the Fermi velocity. Thus, the electrical current can relax more effectively via scattering from node to node than it can via scattering within a single node. This difference is taken into account by considering vertex corrections which modify the bare bubble expression for the electrical conductivity. Since the electrical current in the DDW state is also proportional to the group velocity, the corresponding impurity vertex should not contribute to the electrical conductivity. Moreover, the thermal current in the DDW state is also proportional to the group velocity, so that one can rely on the same arguments about the impurity vertex as for the $d$SC case. Thus we do not expect that in the DDW state the impurity vertex can be a source of the WF violation at $T \to 0$. These observations should be confirmed by detailed calculations of the impurity vertex corrections, which, however, are beyond the scope of the present paper. \(4) For finite temperatures $T \lesssim |\mu|$, the WF law violation is possible and in zero field the thermal conductivity dominates over the electrical conductivity, i.e., $L(T)/L_0 > 1$. \(5) For $T \lesssim |\mu|$ in the nonzero field the WF law violation becomes even stronger than in zero field and depending on the temperature both regimes $L(T)/L_0 \ll 1$ and $L(T)/L_0 > 1$ are possible. \(6) For $T \ll |\mu|$ there is no WF violation even in the presence of magnetic field. Finally, we would like to stress that the results of the present paper should be applicable not only to the DDW state, but to a wider class of theories. Here we started from the DDW Hamiltonian (\[Hamiltonian.DDW\]) and, to simplify the problem, approximated it by the QED$_3$ Lagrangian (\[Dirac.conjugated.Lagrangian\]). There is, however, a number of systems, e.g., pyrolitic graphite [@Gorbar:2002:PRB], that can also be described by the Lagrangian (\[Dirac.conjugated.Lagrangian\]), so that our results would also be relevant for them. Acknowledgments =============== We gratefully acknowledge A. Gröger for a stimulating discussion. S.G.Sh would like to thank N. Andrenacci, L. Benfatto and L. Carlevaro for helpful discussions and W. Kim for useful correspondence. This work was supported by Research Project No. 20-65045.01 of the Swiss NSF. The work of V.P.G. was supported by the SCOPES Projects No. 7UKPJ062150.00/1 and No. 7 IP 062607 of the Swiss NSF and by Grant No. PHY-0070986 of NSF (USA). Nodal fermion Green’s function in an external magnetic field {#sec:A} ============================================================ In the coordinate space, the fermion propagator has the following form in the proper-time representation $$\label{Schwinger.1} S(x,y) = (i \hat{D} + \Delta)_x \langle x\left|\frac{-1}{\Delta^2 + \hat{D}^2} \right|y\rangle = -i (i \hat{D} + \Delta)_x \int \limits_{0}^{\infty} d s \langle x| \exp [ -is(\Delta^2 + \hat{D}^2)] |y\rangle,$$ where $\hat{D} = \gamma^{\nu} D_{\nu}$ and the ‘’long’’ derivative $D_{\nu}$ is given by Eq. (\[long.derivative\]). The matrix element $\langle x| \exp [ -is(\Delta^2 + \hat{D}^2)] |y\rangle$ can be calculated using the Schwinger (proper time) approach [@Schwinger:1951:PR] (see also a pedagogical overview in Ref. [@Dittrich.book]). The main idea of this method is based on the interpretation of $\langle x | \exp [ -is \hat{D}^2] |y\rangle$ as the coordinate representation of the proper-time “evolution operator” $U(s) = \exp[- i Hs]$, where we have introduced the “Hamiltonian” $$H = \hat{D}^2 = D^2 - \frac{e}{2} \sigma^{\rho \nu} F_{\rho \nu}, \qquad F_{\rho \nu} = \partial_{\rho} A_{\nu} - \partial_{\nu} A_{\rho}, \qquad \sigma^{\rho \nu} = \frac{i}{2} [\gamma^{\rho}, \gamma^{\nu}].$$ The matrix element of the evolution operator can be evaluated [@Dittrich.book], for example, using either operator or functional integral formalism: $$\label{U.final} \langle x | U(s) | y \rangle = \frac{e^{-i \pi/4}}{8 (\pi s)^{3/2}} \exp \left[ i e \int \limits_y^x d \xi_{\nu} A^{\nu} (\xi) + \frac{i}{4}(x-y) f(s) (x-y) - L(s) + i \frac{e}{2} \mathbb{\sigma} \mathbb{F} s \right],$$ with $$f(s) \equiv e \mathbb{F} \coth (e \mathbb{F} s), \qquad L(s) \equiv \frac{1}{2} \mbox{tr} \ln \left(\frac{\sinh e \mathbb{F}s}{e \mathbb{F} s}\right),$$ where we have used matrix notations, e.g., $F_{\rho \nu} \equiv (\mathbb{F})_{\rho\nu}$ and the integral with $A_{\nu}$ is calculated along the straight line. This operator differs from the four-dimensional version of $\langle x | U(s) | y \rangle$ [@Dittrich.book] in the power dependence of the proper time and in the numerical prefactors. Plugging (\[U.final\]) into (\[Schwinger.1\]) one obtains $$\label{Schwiner.2} \begin{split} S(x,y) &= \exp \left( i e \int \limits_y^x d \xi_{\nu} A^{\nu} (\xi) \right) \frac{e^{-i3 \pi/4}}{8 \pi^{3/2}} \int \limits_{0}^{\infty} \frac{ds}{s^{3/2}} \left[ \Delta - \frac{1}{2} \gamma^{\rho} [f(s) + e \mathbb{F}]_{\rho \nu} (x-y)^{\nu}\right] \\ & \times \exp \left[- i \Delta^2 s + \frac{i}{4}(x-y) f(s) (x-y) - L(s) + i \frac{e}{2} \mathbb{\sigma} \mathbb{F} s \right]. \end{split}$$ Finally, to perform calculations for the case of interest one should evaluate $f(s)$, $L(s)$ and $\exp [i (e/2) \mathbb{\sigma} \mathbb{F} s]$ for $\mathbb{F}$ containing purely magnetic background field $B$. The corresponding expressions for these values are $$\begin{split} & f(s)_{\rho \nu} = - \frac{1}{s} \left[ g_{\rho \nu} + \frac{(\mathbb{F}^2)_{\rho \nu}}{B^2} (1 - eBs \cot (eBs)) \right],\\ & \exp[-L(s)] = \frac{e B s}{\sin e Bs}, \\ & \exp \left[i \frac{e}{2} \mathbb{\sigma} \mathbb{F} s \right] = \cos e B s + \gamma^1 \gamma^2 \sin e B s. \end{split}$$ In the Matsubara frequency-momentum representation this leads to Eq. (\[Schwinger.representation.translation\]). It is convenient to write down an alterative form of the fermion propagator in a magnetic field as a sum over the Landau-level poles. In deriving it we follow Refs. [@Gusynin:1995:PRD; @Chodos:1990:PRD]. Introducing the shorthand notations we rewrite Eq. (\[Schwinger.representation.translation\]) as follows, $$\begin{split} {\tilde S}(i \omega, \mathbf{p}) = & - \int \limits_{0}^{\infty} d s \exp\left[- s\left(a + \mathbf{p}^2 \frac{\tanh(eBs)}{eBs}\right) \right] \times \left[ \left( b + c \tanh(|eBs|) \right) \left(1 + d \tanh (|eBs|) \right) \right], \end{split}$$ where $$\begin{split} & a = \Delta^2 - (i\omega)^2, \qquad b = i \omega \gamma^0 - p_1 \gamma^1 - p_2 \gamma^2 + \Delta,\\ & c = -i(p_2 \gamma^1- p_1 \gamma^2) \mbox{sgn}(e B), \qquad d = -i\gamma^1 \gamma^2 \mbox{sgn}(e B). \end{split}$$ Using the identity $\tanh x = 1 - 2 \exp(-2x) /[1 + \exp(-2x)]$, the relation [@Gradshtein.book] $$(1-z)^{- (\alpha+1)} \exp \left(\frac{xz}{z-1}\right) = \sum_{n=0}^{\infty} L_n^{\alpha} (x) z^n, \qquad |z| < 1,$$ with $z = - \exp(-2|eB|s)$, $x = 2 \frac{\mathbf{p}^2}{|eB|}$, we arrive at the expression $$\begin{split} &{\tilde S}(i \omega, \mathbf{p}) = - \exp\left( - \frac{\mathbf{p}^2}{|e B|}\right) \\ & \times \int \limits_0^{\infty} d s \exp\left[ -s a + \frac{2\mathbf{p}^2}{|e B|} \frac{z}{z-1} \right] \left[ \left(b + c\right) \left( 1 + d \right) + \left(4cd +2c + 2bd \right) \frac{z}{1-z} + 4 cd \frac{z^2}{(1-z)^2}\right]. \end{split}$$ Then using the identity $L_n^{\alpha-1}(x) = L_n^{\alpha}(x) - L_{n-1}^{\alpha}(x)$ and the definitions $L_n \equiv L_n^0$, $L_{-1}^{\alpha} =0$ we obtain $$\begin{split} {\tilde S} & (i \omega, \mathbf{p}) = -\exp\left( - \frac{\mathbf{p}^2}{|e B|}\right) \int \limits_{0}^{\infty} d s \exp (-sa) \left[ \left(b + c\right) \left( 1 + d \right) \sum_{n=0}^{\infty} L_n(x) z^n \right. \\ & \left. + \left(-\left(b + c\right) \left( 1 + d \right) + 2c + 2bd \right) \sum_{n=0}^{\infty} L_{n-1}(x)z^n + 4cd \sum_{n=0}^{\infty} L_{n-1}^1 (x) z^n \right] \\ = & - \exp\left( - \frac{\mathbf{p}^2}{|e B|}\right) \int \limits_{0}^{\infty} d s \exp (-sa) \left\{ (\Delta + i \omega \gamma^0) \left[ (1- \mbox{sgn}(e B) i \gamma^1 \gamma^2) \sum_{n=0}^{\infty} L_n(x) z^n - (1+ \mbox{sgn}(e B) i \gamma^1 \gamma^2) \sum_{n=0}^{\infty} L_{n-1}(x)z^n \right] \right. \\ & \left. + 4 (p_1 \gamma^1 + p_2 \gamma^2) \sum_{n=0}^{\infty} L_{n-1}^1 (x) z^n \right\}. \end{split}$$ Finally, integrating over $s$ we arrive at Eq. (\[Landau.levels\]). Generalized polarization bubble and nodal approximation {#sec:B} ======================================================= The calculations of electrical, thermal, and even spin conductivities are quite similar for both DDW and $d$SC [@Durst:2000:PRB] nodal systems. Thus instead of repeating the same calculation several times and to underline the similarities and differences between DDW and $d$SC cases, it is rather convenient to define a generalized polarization tensor $\Pi^{g g^{\prime}}$ that depends on the generalized coupling parameters $g$, $g^{\prime}$ $$\label{couplings} g , g^{\prime} = \left[e, \begin{cases} i \omega + i \Omega/2\\ \omega + \Omega/2 - \mu \end{cases} \right].$$ Two lines in the definition of $g, g^{\prime}$ correspond to the Matsubara and real frequencies, respectively, and the origin of the chemical potential $\mu$ in the second line will become clear later. The generalized bubble is $$\label{generalized.bubble} \Pi^{g g^{\prime}}_{\alpha \beta}(i\Omega) = 2 T \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \sum_{i \omega} g g^{\prime} \mbox{tr} [G(i \omega, \mathbf{k}) V_{\alpha}(\mathbf{k}) G(i \omega + i \Omega, \mathbf{k}) V_{\beta}(\mathbf{k})],$$ where $G(i\omega, \mathbf{k})$ is the Green’s function (\[Green.common\]), but evaluated in the external field, and $\mathbf{V}(\mathbf{k})$ is generalized velocity given by Eq. (\[velocity.general\]). The integral is over the reduced Brillouin zone (RBZ) and the factor 2 before the integral is due to the spin degree of freedom $s$. To compare this expression with $d$SC case \[see Eq. (A1) of Ref. [@Durst:2000:PRB]\] we note that the summation over the spin degree of freedom [*is already included*]{} in the Nambu formalism. Using the spectral representation for the fermion Green’s function $$\label{G.translation.invariant} G(i \omega + \mu, {\bf k}) = \int \limits_{-\infty}^{\infty} \frac{A(\omega_1, {\bf k})}{i\omega + \mu - \omega_{1}}d\omega_{1},$$ where $A(\omega_1, {\bf k})$ is the spectral density, we arrive at $$\Pi^{g g^{\prime}}_{\alpha \beta}(i \Omega) = 2 \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \int d\omega_{1} \int d\omega_{2} \mbox{tr} \left[ A(\omega_1, {\bf k}) V_{\alpha}(\mathbf{k}) A(\omega_2, {\bf k}) V_{\beta}(\mathbf{k}) \right] R$$ with $$\label{Matsubara.sum} R = T \sum_{i\omega} g g^{\prime} \frac{1}{i\omega + \mu - \omega_{1}} \frac{1}{i\omega + \mu + i \Omega - \omega_{2}}.$$ Since the intermediate results differ depending on the frequency dependence of the coupling parameters $g$, $g^{\prime}$, we consider the frequency-independent and frequency-dependent couplings separately. For $g=g^{\prime} = e$ (frequency independent coupling), evaluating the sum and then continuing $i \Omega \to \Omega+ i 0$ we get $$\label{R.independent} R = g^2 \frac{n_{F}(\omega_1 - \mu) - n_F(\omega_2 - \mu)}{\omega_1 - \omega_2 + \Omega + i0} = \frac{g^2}{2} \frac{\tanh \frac{\omega_2-\mu}{2T} - \tanh\frac{\omega_1- \mu}{2T}} {\omega_1 - \omega_2 + \Omega + i 0}.$$ For the frequency-dependent coupling $g = g^{\prime}= i \omega + i \Omega/2$ the evaluation of the Matsubara sum (\[Matsubara.sum\]) gives [@foot2] $$R = T \sum_{i\omega} \frac{(i \omega + i \Omega/2)^2} {(i\omega + \mu - \omega_{1})(i\omega + \mu + i \Omega - \omega_{2})} = \frac{(\omega_1 - \mu + i \Omega/2)^2 n_F(\omega_1 - \mu) - (\omega_2 - \mu - i \Omega/2)^2 n_F(\omega_2 - \mu)} {\omega_1 - \omega_2 + i \Omega}.$$ Then, continuing to real frequencies $i \Omega \to \Omega + i0$ we obtain $$\label{R.dependent} \begin{split} \mbox{Im} R = \frac{\pi}{2} (\omega_1 - \mu + \Omega/2)^2 \left[ \tanh \frac{\omega_1 - \mu}{2T} - \tanh \frac{\omega_1 -\mu + \Omega}{2T} \right] \delta (\omega_1 - \omega_2 + \Omega). \end{split}$$ Comparing the imaginary part of Eq. (\[R.independent\]) with Eq. (\[R.dependent\]) one can see that the latter equation can be obtained from the former by the direct substitution of $g = g^{\prime} = \omega + \Omega/2 - \mu$. This explains the difference between the upper and lower lines in Eq. (\[couplings\]). Similarly, evaluating the Matsubara sum (\[Matsubara.sum\]) with $g = e$, $g^{\prime}= i \omega + i \Omega/2$ we get $$R = T \sum_{i\omega} \frac{e (i \omega + i \Omega/2)} {(i\omega + \mu - \omega_{1})(i\omega + \mu + i \Omega - \omega_{2})} = \frac{e(\omega_1 - \mu + i \Omega/2) n_F(\omega_1 - \mu) - e(\omega_2 - \mu - i \Omega/2) n_F(\omega_2 - \mu)} {\omega_1 - \omega_2 + i \Omega},$$ so that $$\begin{split} \mbox{Im} R = \frac{\pi}{2} e(\omega_1 - \mu + \Omega/2) \left[ \tanh \frac{\omega_1 - \mu}{2T} - \tanh \frac{\omega_1 -\mu + \Omega}{2T} \right] \delta (\omega_1 - \omega_2 + \Omega). \end{split}$$ Finally, we obtain for the imaginary part of the tensor polarization $\Pi_{R}^{g g^{\prime}}(\Omega + i0)$ the following expression $$\label{bubble.final.tensor} \mbox{Im} \Pi^{g g^{\prime}}_{\alpha \beta}(\Omega + i 0) = \pi \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \int \limits_{-\infty}^{\infty} d\omega g g^{\prime} \left[ \tanh \frac{\omega - \mu}{2T} - \tanh \frac{\omega -\mu + \Omega}{2T} \right] \mbox{tr} \left[ A(\omega, {\bf k}) V_{\alpha}(\mathbf{k}) A(\omega + \Omega, {\bf k}) V_{\beta}(\mathbf{k}) \right] .$$ Looking at Eqs. (\[generalized.bubble\]) and (\[bubble.final.tensor\]) one can notice that the Green’s function $G$ and the associated spectral density $A$ in the external field are in fact unknown. Instead of considering these functions, we have constructed in Sec. \[sec:Green.magnetic\] the Green’s function (\[Landau.levels\]) for the linearized nodal Lagrangian (\[Dirac.conjugated.Lagrangian\]) that is valid in the vicinity of the four nodal points. Thus, in Eq. (\[generalized.bubble\]) we replace the integration over the reduced Brillouin zone by the integral over the $\mathbf{k}$-space surrounding each node and sum over the four nodal subzones: $$\label{scaling} \int_{\mathrm{RBZ}} \frac{d^2 k}{(2 \pi)^2} \rightarrow \frac{1}{2}\sum_{j=1}^{4}\int\frac{d k_x d k_x}{(2\pi)^{2}} \rightarrow \frac{1}{2}\sum_{j=1}^{4}\int\frac{d^{2}p}{(2\pi)^{2}v_{F}v_{D}} = \frac{1}{4 \pi v_F v_{D}} \sum_{j=1}^{4}\int \limits_0^{p_0} p dp \int \limits_{0}^{2\pi} \frac{d \theta}{2 \pi},$$ where $p_1 = v_{F} k_x = p \cos \theta$, $p_2 = v_{D} k_y = p \sin \theta$, $p = \sqrt{p_1^2 + p_2^2}$, $p_0 = \sqrt{\pi v_{F} v_{D}}/ (2 a)$ and the local nodal coordinate systems $(k_x, k_y)$ are shown in Fig. \[fig:2\]. Note again that, comparing with [@Durst:2000:PRB], an extra factor $1/2$ appears due to the fact that the original integral is over the reduced Brillouin zone. The advantage of the scaled variables $p_1$ and $p_2$ is that they correspond to the “relativistic” case $v_F = v_D = c =1$. Moreover, we can use the spectral function $A_D$ in Eq. (\[AD.clean\]) inside the integral over $p$, so that all necessary factors with $v_{F}$ and $v_{D}$ are already outside the integral (\[scaling\]). We should only provide the rule that allows us to restore the model parameters coming along with the magnetic field in the final result [@Liu:1999:NP]: $e B \to (\hbar v_F v_D/c)e B$. We replace the generalized velocities $\mathbf{V}(\mathbf{k})$ in Eq. (\[bubble.final.tensor\]) by their values $\mathbf{V}_{\mathbf{k} = \mathbf{N}}$ on the Fermi surface at half-filling ($\mu =0$). Note that for the DDW case this approximation is more severe than for the $d$SC case because for $\mu \neq 0$ the true Fermi surface does not coincide with the Fermi surface at half-filling. This approximation puts some restrictions on the values of $\mu$, so that we cannot move far away from half-filling. The bubble (\[bubble.final.tensor\]) takes the form $$\label{bubble.nodal} \mbox{Im} \Pi^{g g^{\prime}}_{\alpha \beta}(\Omega + i0) = \frac{\pi }{2 v_{F}v_{D}} \sum_{j=1}^{4} \int \frac{d^{2}p}{(2\pi)^{2}} \int \limits_{-\infty}^{\infty} d\omega g g^{\prime} \left[ \tanh \frac{\omega - \mu}{2T} - \tanh \frac{\omega -\mu + \Omega}{2T} \right] \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{0} V_{\alpha} A_D(\omega + \Omega, {\bf p}) \gamma^{0} V_{\beta} \right] ,$$ where the spectral density (\[AD.clean\]) includes the broadening of the spectral lines (\[Gamma\]) due to impurities and we inserted the $\gamma^0$ matrix \[see Eq. (\[G.B=0\])\]. Note that in Eq. (\[G.translation.invariant\]) we used only the translationary invariant part (\[Schwinger.representation.translation\]) of the Green’s function (\[Schwinger.representation\]) calculated in the external field, since its translation non-invariant part cancels out when substituted in the bubble $\Pi$. Since we are interested only in the longitudinal conductivities and because the system is isotropic, we can define the longitudinal polarization function $\Pi^{g g^{\prime}}$ as follows (the sum over dummy index is implied): $$\label{tensor2scalar} \Pi^{g g^{\prime}}(\Omega) \equiv \frac{1}{2} \Pi^{g g^{\prime}}_{\alpha \alpha}(\Omega).$$ Then evaluating the sum over nodes in Eq. (\[bubble.nodal\]) by using the identities $$\label{nodal.sum} \sum_{j=1}^{4} v_{l \alpha}^{(j)} v_{l \beta}^{(j)} = 2v_{l}^{2} \delta_{\alpha \beta} , \quad \sum_{j=1}^{4} v_{l \alpha}^{(j)} v_{l^{\prime} \beta}^{(j)} = 2 v_l v_{l^{\prime}} \epsilon_{l l^{\prime}} \epsilon_{\alpha \beta} \quad (l \neq l^{\prime}), \qquad \mathbf{v}_{l} \equiv \{\mathbf{v}_F, \mathbf{v}_D\},$$ with $\epsilon_{ij}$ being antisymmetric tensor, we arrive at the final result $$\label{bubble.final} \begin{split} & \mbox{Im} \Pi^{g g^{\prime}}(\Omega + i0) = \frac{\pi}{v_{F}v_{D}} \int \frac{d^{2}p}{(2\pi)^{2}} \int \limits_{-\infty}^{\infty} d\omega g g^{\prime} \left[ \tanh \frac{\omega - \mu}{2T} - \tanh \frac{\omega -\mu + \Omega}{2T} \right] \\ & \times \left( v_F^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{1} A_D(\omega + \Omega, {\bf p}) \gamma^{1} \right] + v_D^2 \mbox{tr} \left[ A_D(\omega, {\bf p}) \gamma^{2} A_D(\omega + \Omega, {\bf p}) \gamma^{2} \right] \right). \end{split}$$ Since both the electrical current (\[electric.current.DDW\]) and thermal current (\[heat.current2.B=0\]) have two terms, a “Fermi” term proportional to $\mathbf{v}_F$ and $\sigma_3$ and a “gap” term proportional to $\mathbf{v}_D$ and $\sigma_2$, our calculation of the current-current polarization function (\[generalized.bubble\]) is in fact similar to the calculation of the thermal current bubble for the $d$SC case done in Ref. [@Durst:2000:PRB]. Evaluating the polarization function (\[bubble.nodal\]) we obtain four bubbles: Fermi-Fermi, Fermi-gap, gap-Fermi, and gap-gap. However, since $\mathbf{v}_F \perp \mathbf{v}_D$ at each of the gap nodes, the two cross terms cancel, as reflected in the second identity in Eq. (\[nodal.sum\]). Therefore both electrical and thermal longitudinal conductivities have two terms: the Fermi term with the velocity $v^2_F$ and the gap term with the velocity $v^2_D$. Thus, finally we approximated the polarization bubble for the Hamiltonian (\[Hamiltonian.DDW\]) by the polarization bubble for the nodal Lagrangian (\[Dirac.conjugated.Lagrangian\]). Some integrals {#sec:C} ============== Here we derive the expression (\[kappa.T&gt;Gamma\]) for $\kappa$. It follows from the general expression (\[thermal.conductivity.general.final3\]) with $\mathcal{A}$ given by Eq. (\[A.B=D=0.T&gt;Gamma\]). The corresponding integrals are easily evaluated in terms of polylogarithmic functions, $\mbox{Li}_{n}(z)$, [@Wolfram]: $$\label{J} \begin{split} J & =\int \limits_{-\infty}^\infty\frac{d\omega(\omega-\mu)^2|\omega|} {\cosh^2((\omega-\mu) /2T)} \\ & = 8T^2\mu^2\ln(2\cosh(\mu/2T)) + 16T^3 \mu[ \mbox{Li}_2[-e^{\mu/T}]- \mbox{Li}_2[-e^{-\mu/T}]] - 24T^4 [ \mbox{Li}_3[-e^{\mu/T}]+ \mbox{Li}_3[-e^{-\mu/T}]] \end{split}$$ and $$\label{I} \begin{split} I& =\int \limits_{-\infty}^\infty\frac{d\omega(\omega-\mu)|\omega|} {\cosh^2((\omega-\mu) /2T)} = 8T^3[ \mbox{Li}_2[-e^{-\mu/T}]- \mbox{Li}_2[-e^{\mu/T}]] - 8T^2\mu\ln(2\cosh(\mu/2T)) . \end{split}$$ These rather complicated expressions for $I$ and $J$ can be expressed via elementary functions in the limit $\mu \gg T$. We begin with the more simple case of $I$. Using the identity [@Wolfram] $$\label{Li2} \mbox{Li}_2(z)=-\mbox{Li}_2(1/z)-\frac{1}{2}\ln^2(-z)-\frac{\pi^2}{6}, \quad z \notin (0,1),\qquad \mbox{Li}_2(0) = 0,$$ we obtain that in the limit $\mu \gg T$ $$\mbox{Li}_2[-e^{-\mu/T}]- \mbox{Li}_2[-e^{\mu/T}] = 2 \mbox{Li}_2(-e^{-\mu/T})+ \frac{1}{2}\ln^2(e^{\mu/T})+ \frac{\pi^2}{6} \simeq \frac{\mu^2}{2 T^2} + \frac{\pi^2}{6}.$$ Thus the final expression for $I$ reads as $$\label{I1} I \simeq \frac{4 \pi^2}{3} T^3, \qquad \mu \gg T.$$ To simplify $J$, in addition to Eq. (\[Li2\]), we use the corresponding identity for $\mbox{Li}_3(z)$ [@Wolfram]: $$\mbox{Li}_3(z)= \mbox{Li}_3(1/z)-\frac{1}{6}\ln^3(-z)-\frac{\pi^2}{6} \ln(-z), \quad z \notin (0,1), \qquad \mbox{Li}_3(0) = 0,$$ and for $\mu \gg T$ we obtain that $$\label{J1} \begin{split} J \simeq \frac{4 \pi^2}{3} T^3 \mu, \qquad \mu \gg T. \end{split}$$ Putting Eqs. (\[J1\]) and (\[I1\]) along with Eq. (\[sigma.T&gt;Gamma\]) into Eq. (\[thermal.conductivity.general.final3\]), we arrive at the final result (\[kappa.T&gt;Gamma\]). [99]{} R.W. Hill, C. Proust, L. Taillefer, P. Fournier,and R.L. Greene, Nature [**414**]{}, 711 (2001). C. Proust, E. Boaknin, R.W. Hill, L. Taillefer, and A.P. Mackenzie, Phys. Rev. Lett. [**89**]{}, 147003 (2002). J.S. Langer, Phys. Rev. [**128**]{}, 110 (1962). P.L. Taylor, [*A Quantum Approach to the Solid State,*]{} (Prentice-Nall, New Jersey, 1970). L. Smr[č]{}ka and P. St[ř]{}eda, J. Phys. C: Solid State Phys. [**10**]{}, 2153 (1977). A. Houghton, S. Lee, J.B. Marston, Phys. Rev. B [**65**]{}, 220503 (2002). X. Yang and C. Nayak, Phys. Rev. B [**65**]{}, 064523 (2002). W. Kim and J.P. Carbotte, Phys. Rev. B [**66**]{}, 033104 (2002). S. Chakravarty, R.B. Laughlin, D.K. Morr, and C. Nayak, Phys. Rev. B [**63**]{}, 094503 (2001). T. Timusk and B. Statt, Rep. Progr. Phys. [**62**]{}, 61 (1999). C. Nayak, Phys. Rev. B [**62**]{}, 4880 (2000). J.L. Tallon and J.W. Loram, Physica C [**349**]{}, 53 (2001). O. Vafek, A. Melikyan, Z. Te[š]{}anovi[ć]{}, Phys. Rev. B [**64**]{}, 224508 (2001). A.S. Mel’nikov in [*Vortices in Unconventional Superconductors and Superfluids*]{}, edited by R.P. Huebener, N. Schopohl and G.E. Volovik (Springer, Berlin, 2002), Chap. 12. A.A. Nersesyan and G.E. Vachanadze, J. Low Temp. Phys. [**77**]{}, 293 (1989). C.P. Enz, [*A Course on Many-Body Theory Applied to Solid-State Physics,*]{} (World Scientific, Singapore, 1992). A.C. Durst and P.A. Lee, Phys. Rev. B [**62**]{}, 1270 (2000). D. Morr, Phys. Rev. Lett. [**89**]{}, 106401 (2002). W. Dittrich, H. Geis, [*Probing the Quantum Vacuum: Perturbative Effective action approach in Quantum Electrodynamics and its Application,*]{} Springer, 2000. R. Jackiw and S. Templeton, Phys. Rev. D [**23**]{}, 2291 (1981). V.P. Gusynin, V.A. Miransky, and I.A. Shovkovy, Phys. Rev. Lett. [**73**]{}, 3499 (1994); Phys. Rev. D [**52**]{}, 4718 (1995). J. Schwinger, Phys. Rev. [**82**]{}, 664 (1951). E.J. Ferrer, V.P. Gusynin, V. de la Incera, preprint cond-mat/0203217. E.V. Gorbar, V.P. Gusynin, V.A. Miransky, I.A. Shovkovy, Phys. Rev. B [**66**]{}, 045108 (2002). A. Chodos, K.Everding, and D.A. Owen, Phys. Rev. D [**42**]{}, 2881 (1990). R.E. Prange, [*The Quantum Hall Effect*]{}, edited by R.E. Prange and S.M. Girvin, (Springer Verlag, New York, 1987). G. D. Mahan, [*Many-Particle Physics*]{}, Plenum Press, New York, 1990. J.S. Langer, Phys. Rev. [**127**]{}, 5 (1962). G.M. Eliashberg, Sov. Phys. JETP [**14**]{}, 886 (1962). Y. Ando, J. Takeya, Y. Abe, K. Nakamura, and A. Kapitulnik, Phys. Rev. B [**62**]{}, 626 (2000). The original Kubo formula requires modification in the case of Hall thermal conductivity \[Ref.13 [@Vafek:2001:PRB] and M. Jonson, S.M. Girvin, Phys. Rev. B [**29**]{}, 1939 (1984)\] in the presence of magnetic field while it remains intact for the longitudinal thermal conductivity studied here. I.S. Gradshtein and I.M. Ryzhik, [*Table of Integrals, Series and Products,*]{} (Academic Press, Orlando, 1980). The sum over frequencies appears to be divergent, however, as was shown by V. Ambegaokar and A. Griffin \[Phys. Rev. A [**137**]{}, 1151 (1965)\], this divergence results from an improper treatment of time derivatives inside the time-ordered product of heat currents Eq. (\[polarization.EE\]). This divergence disappears when the problem is treated more carefully. The prescription is simply to ignore it. W.V. Liu, Nucl. Phys. [**B 556**]{}, 553 (1999). http://functions.wolfram.com/ZetaFunctionsandPolylogarithms/

--- abstract: 'This paper presents a new computer code to solve the general relativistic magnetohydrodynamics (GRMHD) equations using distributed parallel adaptive mesh refinement (AMR). The fluid equations are solved using a finite difference Convex ENO method (CENO) in $3+1$ dimensions, and the AMR is Berger-Oliger. Hyperbolic divergence cleaning is used to control the $\nabla\cdot {\bf B}=0$ constraint. We present results from three flat space tests, and examine the accretion of a fluid onto a Schwarzschild black hole, reproducing the Michel solution. The AMR simulations substantially improve performance while reproducing the resolution equivalent unigrid simulation results. Finally, we discuss strong scaling results for parallel unigrid and AMR runs.' address: | $1$ Department of Physics and Astronomy, Louisiana State University, Baton Rouge, LA 70803-4001, USA\ $2$ Department of Physics and Astronomy, Brigham Young University, Provo, UT 84602, USA\ $3$ Department of Physics, Long Island University – C.W. Post Campus, Brookville, NY 11548, USA author: - 'Matthew Anderson$^{1}$, Eric W. Hirschmann$^{2}$, Steven L. Liebling$^{3}$, David Neilsen$^{2}$' bibliography: - './mhd.bib' title: Relativistic MHD with Adaptive Mesh Refinement --- Introduction {#sec:introduction} ============ The interaction of gravitational and electromagnetic fields together with rotation is believed to power the central engines of many astrophysical phenomena including relativistic jets in active galactic nuclei (AGN), other forms of black hole accretion, gamma-ray bursts (GRB), and core collapse supernovae. In addition, interactions between strong gravitational and electromagnetic fields are believed to result in the transport of angular momentum in accretion disks via the magnetorotational instability (MRI) and in the extraction of black hole energy via the Blandford-Znajek mechanism. The expected ubiquity of magnetic fields in the vicinity of strongly gravitating compact objects has spurred increased theoretical efforts to understand these astrophysical systems. The difficulties, however, are formidable as the physical laws describing these phenomena are nonlinear, evolutionary, and, in general, without simplifying symmetries. As a consequence, numerical simulation of these systems becomes crucial for better understanding them. Even then, there are likely considerable aspects of the physics that remain out of reach of current computational resources. For instance, it is not unrealistic to imagine that dissipative and radiative aspects of these problems will be important for accurately modeling certain types of phenomena. However, with the hope of capturing some of the relevant physics, a number of groups have begun developing methods and codes for evolving some of the relativistic components of these complicated scenarios. In this large and growing body of work, a significant effort is directed at special relativistic magnetohydrodynamics (RMHD). However, with increasing interest in the interaction of gravitational and electromagnetic fields, one must also couple the equations of relativistic MHD to general relativity either through a curved space background or the dynamical field equations themselves. The earliest attack on this general problem was Wilson’s pioneering work evolving a rotating, axisymmetric star with a poloidal magnetic field [@Wilson1975]. Subsequently, little numerical work was done until very recently with several groups developing codes for evolving the general relativistic MHD (GRMHD) equations on fixed backgrounds  [@Koide:2000; @DeVilliers:2002ab; @Gammie:2003rj; @Baumgarte:2002b; @Komissarov:2004; @Anton:2005gi; @Anninos:2005] and in dynamical spacetimes  [@Duez:2005sf; @Shibata:2005gp]. One of the difficulties alluded to above in the numerical simulation of these sorts of systems is that one must solve the GRMHD equations over a large range of time and length scales. The computational requirements necessary to adequately resolve multiscale phenomena using only a single resolution mesh are often too high for available resources. This is particularly true when the full, dynamical GRMHD equations are being considered. Adaptive mesh refinement (AMR) therefore becomes crucial in order to reduce the computational requirements necessary for modeling such systems. There are currently MHD codes in use which incorporate AMR, but most are focused on the nonrelativistic problem [@Balsara2001; @Gombosi2000; @Gombosi2001]. Notable among these are the work of Balsara which extended a version of constrained transport to AMR. For relativistic MHD, only the work by Anninos  [@Anninos:2005] incorporates adaptive mesh refinement. Theirs is a finite volume approach with divergence cleaning. This paper describes our algorithm for solving the GRMHD equations with AMR. Building on the work presented in [@Neilsen2005], where the RMHD equations were solved on overlapping grids, some key elements of our algorithm are: (1) The Convex Essentially Non-Oscillatory (CENO) method for the MHD equations, (2) Hyperbolic divergence cleaning for controlling the solenoidal constraint, (3) Berger-Oliger AMR, (4) Weighted Essentially Non-Oscillatory (WENO) interpolation for communications from coarse to fine grids, and (5) discretization in time via method of lines. The CENO scheme is robust, and has three advantages for our code. First, the method does not require the spectral decomposition of the Jacobian matrix, making it relatively efficient. Central and central-upwind schemes are known to give results nearly identical to those of more complicated methods for many problems [@Lucas-Serrano:2004aq; @Shibata:2005jv]. With the added capability of AMR, we find that we are able to efficiently resolve very fine solution features. A second advantage of CENO for our purposes is that it is a finite difference, or vertex centered, method. As we will solve the Einstein equations with finite differences, using a finite difference fluid scheme simplifies coupling the two sets of equations with AMR. The simplification arises because fluid and geometric variables are always defined at the same point as grids are refined. Finite volume and finite difference grids become staggered with respect to each other as they are refined. While we find this simplification advantageous in our present work, we note that our AMR code [had]{} can combine finite difference and finite volume schemes [@motl2006]. A third advantage is that ENO schemes are easily extended to higher order accuracy. Our CENO scheme can reconstruct fluid variables to both first and second order, resulting in second and third order evolution schemes, respectively. We choose hyperbolic divergence cleaning [@Dedner2002] to limit growth in the solenoidal constraint on the magnetic field, $\nabla\cdot {\bf B}=0$. Hyperbolic divergence cleaning is simple to implement. A single hyperbolic field is added to the system and coupled to the evolution equations for $\bf B$. Divergence cleaning gave good results in earlier tests [@Neilsen2005] and allows us to freely choose prolongation methods for AMR. While $\nabla\cdot {\bf B}=0$ is not satisfied to machine precision for any particular discrete divergence operator, we find that $||\nabla\cdot {\bf B}||$ does converge to zero, and, in the tests presented here, $||\nabla\cdot {\bf B}||$ is roughly the same order of magnitude as the expected truncation error. Mesh refinement is necessary for obtaining accurate numerical solutions for three dimensional systems in general relativity, and AMR is essential for complicated problems where the refinement regions can not be guessed [*a priori.*]{} We use the infrastructure to provide Berger-Oliger style AMR [@Berger]. has a modular design allowing one to easily implement many different sets of evolution equations, and different modules can be combined, for example, to solve both the Einstein and MHD equation simultaneously. supports higher order differencing schemes, and our implementation of the MHD equations is fully third order accurate [@Lehner2006]. Refinement regions can be specified in a variety of ways. provides a shadow hierarchy for specifying refinement criteria using truncation error estimates, or the user may specify problem specific criteria, such as refining on gradients or other solution features. supports different interpolation schemes (we choose WENO interpolation for this work) and supports both finite volume and finite difference equations or combinations of both. Finally, as discussed below, scales well in strong scaling tests in both unigrid and AMR tests. Finally, we discretize the continuum equations first in space (creating a semi-discrete system) and then discretize in time using the method of lines. This gives us considerable flexibility in choosing discretization schemes appropriate for very different types of equations. For example, the MHD equations are solved here with high-resolution shock-capturing methods, while we might solve the Einstein equations using methods that preserve a discrete energy norm. Using the method of lines, we can easily and consistently combine these two sets of semi-discrete equations in a uniform time integration. Time integrators can be independently chosen for their desired properties or order of integration. For example, we choose for this work a third-order Runge–Kutta scheme that preserves the TVD condition [@ShuOsherI]. The remaining sections of this paper give further details to the algorithm sketched above and present code tests in both flat space and on a Schwarzschild black hole background. We first present the GRMHD equations used in this work. The MHD equations in general relativity {#subsec:equations} ======================================= A number of derivations of the GRMHD equations have appeared in the literature, e.g., [@Baumgarte:2002a; @Sloan:1985; @Evans:1988; @Zhang:1989; @Gammie:2003rj; @DeVilliers:2002ab; @Neilsen2005], and thus we simply present the equations to be solved here. The numerical methods that we use are very similar to those we have used in our previous MHD work using overlapping grids. Thus our presentation here is short, and the reader may refer to that work for more detail [@Neilsen2005]. The spacetime metric is written in terms of the conventional ADM 3+1 variables, namely $$\rmd s^2 = -\alpha^2 \,\rmd t^2 + h_{ij}(\rmd x^i + \beta^i \,\rmd t)(\rmd x^j + \beta^j \,\rmd t),$$ where $\alpha$ is the lapse, $\beta^i$ is the shift, and $h_{ij}$ is the $3$-metric on the spacelike hypersurfaces. Units are chosen such that $c=1$ and $G=1$. We denote the extrinsic curvature as $K_{ab}$ and the Christoffel coefficients with respect to the $3$-metric as ${^{3}{\Gamma}}_{ab}^i$. As our focus in this paper is fixed background geometries, we will omit here a discussion of our approach to evolving the Einstein equations and only present the relevant matter equations. Future papers will address dynamical spacetimes. The equations for MHD on a curved background, as in flat spacetime, can be written, for the most part, in balance law form, namely $$\partial_t\bbu + \partial_k\bbf\,^k(\bbu) = \bbs(\bbu). \label{eq:balance}$$ where $\bbu$ is a state vector, $\bbf\,^k$ are flux functions, and $\bbs$ are source terms. For the current case, these are $$\begin{aligned} % % \fl \partial_t \left( \sqrt{h} \, D \right) + \partial_i \left[ \sqrt{-g} \, D \left( v^i - {\beta^i \over \alpha} \right) \right] = 0,\label{eq:ev_D} \\ % % \fl \partial_t \left( \sqrt{h} \, S_b \right) + \partial_i \left[ \sqrt{-g} \left( S_b \left( v^i - {\beta^i \over \alpha} \right) + P \, h^i{}_b - \frac{1}{W^2}\left(B^iB_b - \frac{1}{2} \, h^i{}_b \, B^jB_j \right) \right.\right.\nonumber\\ \qquad\qquad \left.\left. - \frac{1}{2} \, B^j v_j \left(B^i v_b - \frac{1}{2} \, h^i{}_b \, B^j v_j\right) \right) \right] \nonumber\\ % \qquad\qquad\qquad = \sqrt{-g} \, \left[ \, {^{3}{\Gamma}}_{ab}^i \left( \perp\! T \right)^a{}_i + {1\over\alpha} S_a \partial_b \beta^a - {1\over\alpha} \partial_b \alpha \, E \right], \\ % % \fl \partial_t \left( \sqrt{h} \, \tau \right) + \partial_i \left[ \sqrt{-g} \left( S^i - \frac{\beta^i}{\alpha} \, \tau - v^i D \right) \right] = \sqrt{-g} \, \left[ \left( \perp\! T \right)^{ab} \, K_{ab} - \frac{1}{\alpha} \, S^a \partial_a \alpha \right],\\ % % \fl \partial_t \left( \sqrt{h} \, B^b \right) + \partial_i \left[ \sqrt{-g} \, \left( B^b \left( v^i - \frac{\beta^i}{\alpha} \right) - B^i \left( v^b - \frac{\beta^b}{\alpha} \right) \right) \right] = 0 , \label{eq:ev_B} \\ % % \fl\frac{1}{\sqrt{h}} \, \partial_i \left(\sqrt{h} \, B^i \right) = 0 \label{eq:sol_const} , % %\end{aligned}$$ where the quantity $\bigl( \perp\! T \bigr)^i{}_b$ is the spatial projection of the stress tensor given in terms of the matter fields by $$\begin{aligned} \bigl( \perp\! T \bigr)^i{}_b &=& v^i S_b + P \cdot h^i{}_b - {1\over W^2} \, \Bigl[ B^i B_b - {1 \over 2} h^i{}_b \cdot B^2 \Bigr] \nonumber\\ &&\qquad - \bigl( B^jv_j \bigr) \Bigl[ B^i v_b - {1\over2} h^i{}_b \cdot \bigl( B^jv_j \bigr) \Bigr] . \end{aligned}$$ In the above, we work from a set of “primitive" variables ${\bbw} = (\rho_0 , v^i , P , B^j)^{\rm T}$ consisting of the energy density, $\rho_0$, the components of the coordinate velocity of the fluid, $v^i$, the fluid pressure, $P$, and the magnetic field in the fluid frame, $B^j$. From these, we define a set of conservative variables, ${\bbu} = (D, S_b, \tau , B^j)^{\rm T}$ where the relativistic density $D$, momentum $S_b$, and energy $E = \tau + D$ are given in terms of the primitive variables by $$\begin{aligned} D &=& W \rho_0,\\ S_i &=& \left[h_e W^2 + B^2\right]v_i - \left(B^j v_j\right)B_i,\\ \tau &=& h_e W^2 + B^2 - P - \frac{1}{2}\left[\left(B^j v_j\right)^2 + \frac{B^2}{W^2}\right] - W\rho_0,\end{aligned}$$ and $B^j$ remains unchanged. We have also defined $B^2 = B_i B^i$, the fluid enthalpy $h_e = \rho_0 ( 1 + \epsilon ) + P$ with $\epsilon$ the fluid’s internal energy, and the Lorentz factor $W = (1-v_i v^i)^{-1/2}$. Note that spatial indices are lowered and raised by the 3-metric $h_{ij}$ and its inverse. Finally, to close the system, we assume a $\Gamma$-law equation of state $$P = (\Gamma-1)\, \rho_0\epsilon,$$ where $\Gamma$ is the usual adiabatic index. Note that while Eqns. – are indeed in balance law form, Eq.  is not. This last equation, of course, is the solenoidal constraint and must be dealt with separately. One of the benefits of our chosen numerical scheme is that it does not require the full spectral decomposition of the system of evolution equations. However, we do find it useful to have some information about the possible speeds of some of the waves in our system. This information comes by solving for the eigenvalues of the Jacobian matrix, ${\cal J\,}^k$, associated with the flux, ${\bbf\,}^k(\bbu)$, in the $k$ direction where $${\cal J}\,^k = \frac{\partial \bbf\,^k(\bbu)}{\partial \bbu} .$$ On doing this for our system for the $k$ direction, say, one gets the general relativistic generalization [@Anton:2005gi] of the seven wave speeds of flat space MHD [@Anile; @Komissarov:1999; @DelZanna2002rv; @Leismann:2005]. These include the entropy wave and two Alfv[é]{}n waves, $$\begin{aligned} {^{\rm ent}\lambda}^k & = & \alpha v^k - \beta^k \\ {^{\rm A}\lambda}^k_{\pm} & = & \alpha v^k - \beta^k - {B^k \over h_e W^2 + B^2} \left[ B^j v_j \pm \Bigl( h_e (B^j v_j)^2 + {B^2 \over W^2} \Bigr)^{1/2} \right] \end{aligned}$$ and the four (“fast" and “slow") magnetosonic waves, ${^{\rm f,s}\lambda}^k_{\pm}$, that are the zeros of the fourth order polynomial $$\begin{aligned} \fl 0 = h_e W^4 (1 - c_s^2) \bigl( \alpha v^k - \beta^k - \lambda^k \bigr)^4 \nonumber \\ \fl \qquad + \left[ \bigl( \beta^k + \lambda^k \bigr)^2 - \alpha^2 h^{kk} \right] \cdot \Bigl[ \bigl( \alpha v^k - \beta^k - \lambda^k \bigr)^2 \bigl( h_e W^2 c_s^2 + B^2 + W^2 ( B^j v_j )^2 \bigr) \Bigr. \nonumber \\ \Bigl. \qquad\qquad\qquad - \, c_s^2 \Bigl( W (B^j v_j) \bigl( \alpha v^k - \beta^k - \lambda^k \bigr) + \alpha {B^k \over W} \Bigr)^2 \Bigr] \label{eq:quartic}\end{aligned}$$ where $c_s$ is the local sound speed and is given by $$h_e c_s^2 = \rho_0 \, {\partial P \over \partial \rho_0} + {P \over \rho_0} \, {\partial P \over \partial \epsilon} .$$ We solve numerically using the DRTEQ4 routine from the publicly available CERN Program Library. The roots from DRTEQ4 are then refined using a Newton-Raphson solver. Finally, we note that the MHD equations are written in terms of both the conservative and primitive variables. The transformation from conservative variables to primitive variables is transcendental, requiring the solution of a single transcendental, nonlinear equation, and is outlined in [@Neilsen2005]. Numerical approach {#sec:numerical} ================== This section briefly summarizes the numerical scheme we use to solve the relativistic MHD equations. The fluid equations are solved with the Convex Essentially Non-Oscillatory (CENO) scheme. CENO is based on a finite difference discretization, which simplifies the coupling to the Einstein equations with AMR. Hyperbolic divergence cleaning controls growth of error in the solenoidal constraint, and gives some flexibility in choosing other components of the numerical algorithm with AMR. CENO {#subsec:methods} ---- The CENO scheme was developed by Liu and Osher [@LiuOsher] to efficiently solve equations in balance law form $$\partial_t\bbu + \partial_k\bbf\,^k(\bbu) = \bbs(\bbu), \label{eq:balance2}$$ where $\bbu$ is a state vector, $\bbf\,^k$ are flux functions, and $\bbs$ source terms. We use the modification of CENO for relativistic fluids of Del Zanna and Bucciantini [@DelZanna:2002qr]. is solved using the method of lines. The semi-discrete form in one dimension is $$\frac{d \bbu_i}{dt} = - \frac{\skew 8\hat\bbf_{i+1/2} - \skew 8\hat\bbf_{i-1/2}}{\triangle x} + \bbs(\bbu_i), \label{eq:sd_eq}$$ where $\skew 8\hat\bbf$ is a consistent numerical flux. We use both the Lax–Friedrichs flux and the HLL flux for the numerical flux. The Lax–Friedrichs flux is $${\bbf}\,^{\rm LF}_{i+1/2} = \frac{1}{2}\left[ \bbf(\bbu^L_{i+1/2}) + \bbf(\bbu^R_{i+1/2}) - (\bbu^R_{i+1/2} - \bbu^L_{i+1/2})\right],$$ where $\bbu^L_{i+1/2}$ and $\bbu^R_{i+1/2}$ are the left and right reconstructed states at $x_{i+1/2}$. The HLL flux [@Harten] is a central-upwind flux that uses the maximum characteristic velocities for both left- and right-moving waves, $\lambda_\ell$ and $\lambda_r$, respectively, $${\bbf}\,^{\rm HLL} = \frac{\lambda^+_r \bbf(\bbu^\ell) - \lambda^-_\ell\bbf(\bbu^r) + \lambda^+_r \lambda^-_\ell(\bbu^r - \bbu^\ell)}% {\lambda^+_r - \lambda^-_\ell},$$ where $$\begin{aligned} \lambda^-_\ell &=& \min(0,\lambda_\ell)\\ \lambda^+_r &=& \max(0,\lambda_r).\end{aligned}$$ For highly relativistic flows the Lax–Friedrichs flux gives results very similar to the HLL flux; the maximum characteristic velocities approach the speed of light. The point-valued fluxes ${\bbf}_{i+1/2}$ are then converted into consistent numerical fluxes, $\skew 8\hat\bbf_{i+1/2}$. For a second order scheme, $\skew 8\hat\bbf_{i+1/2}={\bbf}_{i+1/2}$, while the correction for the third-order scheme is [@ShuOsherI] $$\skew 8\hat \bbf_{i+1/2}= \left( 1 - \frac{1}{24}{\cal D}^{(2)}\right) \bbf_{i+1/2},$$ where ${\cal D}^{(2)}$ is a non-oscillatory second-order difference operator. The operator used in this work is specified in [@Neilsen2005]. The accuracy of the overall numerical scheme is determined by the reconstruction of the fluid states $\bbu^L_{i+1/2}$ and $\bbu^R_{i+1/2}$ from the solution known at grid points, i.e., the solution at $x_{i-p}, \ldots, x_i, \ldots, x_{i+q}$ for integer $p$ and $q$. Linear and quadratic reconstructions discussed below lead to second and third order methods, respectively, for smooth solutions. As is commonly done in relativistic fluid dynamics, we reconstruct the primitive variables rather than the conservative variables. This is because the conservative to primitive variable transformation is transcendental, and computationally rather expensive. The reconstruction is performed hierarchically, meaning that a reconstruction of order $n$ is calculated from the reconstruction of order $n-1$ using a general algorithm. This allows one to construct schemes of formally very high order. Thus, we first obtain a linear reconstruction, $L_i$, which is then used to create the second order reconstruction. $L_i$ is defined on the domain $[x_{i-1/2},x_{i+1/2}]$ as $$L_i(x) = u_i + u'_i(x-x_i),$$ where $u'_i$ is the limited slope $$u'_i = \mbox{minmod}(D_- u_i, D_+ u_i).$$ Here we have defined one-sided and centered difference operators as $$D_{\pm} u_i = \pm \frac{u_{i\pm 1} - u_i}{\triangle x},\qquad D_{0} u_i = \frac{u_{i+1} - u_{i-1}}{2\triangle x},$$ and the minmod limiter is $$\mbox{minmod}(a_1, a_2, \cdots) = \left\{ \begin{array}{ll} \min\{a_k\} & \hbox{if all $a_k > 0$,}\\ \max\{a_k\} & \hbox{if all $a_k < 0$,}\\ 0 & \hbox{otherwise.} \end{array}\right. \label{eq:minmod}$$ The first-order reconstruction, $L_i(x)$, is thus equivalent to the linear TVD reconstruction. A second order reconstruction proceeds by constructing three candidate quadratic functions, $Q^k_i(x)$, which are then compared to $L_i(x)$. When the solution is smooth, one of the quadratic functions is chosen for the reconstruction. Near discontinuities, however, the linear reconstruction is retained, thus giving solutions similar to TVD schemes for discontinuous solutions. The three candidate quadratic functions are $$Q^k_{i}(x) = u_{i+k} + D_0 u_{i+k} (x - x_{i+k}) + \frac{1}{2}D_+D_- u_{i+k} (x - x_{i+k})^2 ,$$ with $k=-1, 0, 1$. Weighted differences with respect to $L_i(x)$ are then calculated $$d^k(x) = \alpha^k \left( Q^k_i(x) - L_i(x)\right). \label{eq:eno_differences}$$ The weights $\alpha^k$ are chosen to bias the reconstruction towards the centered polynomial: $\alpha^0 = 0.7$, and $\alpha^{-1} = \alpha^1 = 1$. When the differences $d^k$ all have the same sign, we choose the $Q^k_i(x)$ for which $d^k$ has the smallest magnitude. When the $d^k(x)$ have differing signs, we revert to the first order reconstruction. Finally, the semi-discrete equations are integrated with the optimal third-order Runge–Kutta that preserves the TVD condition [@ShuOsherI] $$\begin{aligned} \bbu^{(1)} &=& \bbu^{n} + \triangle t L(\bbu^n),\nonumber\\ \bbu^{(2)} &=& \frac{3}{4}\bbu^{n} + \frac{1}{4}\bbu^{(1)} + \frac{1}{4}\triangle t L(\bbu^{(1)}),\\ \bbu^{n+1} &=& \frac{1}{3}\bbu^{n} + \frac{2}{3}\bbu^{(2)} + \frac{2}{3}\triangle t L(\bbu^{(2)}).\nonumber\end{aligned}$$ Hyperbolic Divergence Cleaning {#subsec:hdc} ------------------------------ The time evolution of the magnetic field is governed by above. However, $\bf B$ is also subject to the solenoidal constraint $\nabla \cdot {\bf B}=0$. The continuum evolution equations preserve this constraint, although it may be violated in numerical evolutions. These violations can lead to unphysical numerical solutions [@BrackbillBarnes; @Brackbill]. Some differencing schemes for the Maxwell equations and MHD are designed such that a particular discretization of the solenoidal constraint is satisfied to machine precision. These constrained transport methods, naturally, do not give solutions that exactly satisfy the continuum constraint, and the magnitude of the constraint error can be estimated by using an independent discrete divergence operator. Constrained transport methods for classical MHD have been used with AMR [@Balsara2001b; @TothRoe; @LiLi; @Balsara2004]. Divergence cleaning methods are an alternative approach to constrained transport, and allow some flexibility in designing the numerical algorithm. Elliptic divergence cleaning methods require the solution of a Poisson equation (either explicitly, or implicitly in Fourier space), and some common implementations have been reviewed for classical MHD by Tóth [@Toth] and Balsara and Kim [@BalsaraKim]. Tóth reports favorably on divergence cleaning, while Balsara and Kim argue that constrained transport performs better for a wider class of problems. Hyperbolic divergence cleaning is quite efficient, easy to implement, and usually gives good results [@Dedner2002]. A new field $\psi$ is added to the equations and coupled to the evolution equations for $\bf B$. The field $\psi$ acts as a generalized Lagrange multiplier, similar to the $\lambda$-system used in solving the Einstein equations [@Brodbeck1999]. Having some freedom in choosing the equation for $\psi$, we choose $$\begin{aligned} \label{eq:newB} \partial_t B^b + \partial_i\left( B^b v^i - B^iv^b\right) + h^{bj}\partial_j \psi &=& 0,\\ \qquad\qquad \frac{1}{c_h^2}\partial_t \psi + \frac{1}{c_p^2}\psi + {\bf \nabla} \cdot {\bf B} &=& 0. \label{eq:psi}\end{aligned}$$ It can be shown that $\psi$ satisfies the telegraph equation, whose solutions are damped, traveling waves. The parameters $c_h$ and $c_p$ control the speed and damping of the constraint propagation, respectively. We use $c_h = 1$ and $c_p \in \left[1,12\right]$ in the tests examined here. Using larger values of $c_h$ requires smaller Courant factors and did not change results significantly in [@Neilsen2005]. In contrast, the optimal damping factor, $c_p$, is related to the size of the initial shock discontinuity. Generally, the larger the shock, the larger the value necessary for $c_p$. Finally, work is underway to develop constraint preserving boundary conditions consistent with hyperbolic divergence cleaning for the MHD equations [@Palenzuela]. Adaptive Mesh Refinement {#subsec:amr_methods} ------------------------ AMR provides the ability to add grid refinement where and when needed. This need is determined by some refinement criterion. At any given level of resolution, points which meet this criterion are flagged and a new, finer level is created which includes all such flagged points. Similarly, when no points are flagged, the level is removed. For the tests presented here, the maximum number of levels of refinement was limited to two. In other words, our runs have grids with three different resolutions. The fine and coarse grids communicate in AMR through prolongation and restriction. Fine grids are created by interpolating the solution from a parent grid (prolongation), and the fine grid solution is communicated to coarser grids through restriction. In prolongation we interpolate the conservative variables onto finer grids using third-order WENO interpolation [@SebastianShu; @Neilsen2005]. This interpolation scheme is designed to work well with discontinuous functions by adjusting the interpolation stencil to the local smoothness of the function. This avoids oscillations near discontinuities, which often cause primitive variable solvers for relativistic fluids to fail. For restriction on vertex centered grids, the fine grid values are copied directly to the coarse grid (direct injection). If a point on a vertex centered coarse grid is also found on two or more finer grids, the restriction operation averages the values on the finer grids for the solution at the coarser grid point. The distributed AMR infrastructure that we employ is [had]{}. [had]{} is a F77 based Berger-Oliger [@Berger] type AMR package presented in [@Liebling] using the message passing interface (MPI) for distributed parallelism. [had]{} has a modular design, allowing one to solve different sets of equations with the same computational infrastructure. Unlike many other publicly available AMR toolkits, including [@SAMRAI; @Chombo; @Paramesh; @AMROC; @Boxlib], [had]{} is natively vertex centered. This can be advantageous in numerical relativity, as many difference schemes for the Einstein equations are vertex centered. Support for cell centered grids in [had]{} is also available. [had]{} supports subcycling of grids in time for full space-time AMR and can in principle accommodate arbitrary orders of accuracy in both space and time. An example has been shown using third order accurate AMR simulations [@Lehner2006]. The [had]{} clustering algorithm is Berger-Rigoutsos [@Rigoutsos]; the load balancing algorithm is the least loaded scheme [@Rendleman]. Considerable flexibility is provided in developing a refinement criterion for the [had]{} infrastructure, and many were explored in conjunction with the numerical tests presented here, such as refining on gradients in the density, pressure or the magnetic field. All criteria resulted in similar adaptive mesh hierarchies for the relativistic rotor and spherical blast wave tests; consequently, those tests center the refinement on the evolving shock front. A shadow hierarchy [@Pretorius] has recently been added to [had]{} for truncation error estimation and was used as the refinement criterion in the spherical accretion tests of a fluid falling onto a Schwarzschild black hole. Numerical Results {#sec:nr} ================= In this section we examine three relativistic MHD test problems and one accretion test problem using AMR. The MHD test problems are selected because their solutions have very sharp features that require high resolutions. The accretion test problem demonstrates the AMR capabilities with a curved space background and an excision region. The first problem is a one-dimensional blast wave problem introduced by Balsara [@Balsara2001]. We compare the unigrid shocktube results with the exact solution, and then compare unigrid results with the resolution-equivalent AMR results. The second and third problems are three-dimensional extensions of standard two-dimensional tests: a spherical blast wave and a spherical relativistic rotor [@DelZanna2002rv; @Shibata:2005gp; @Neilsen2005]. We present unigrid and AMR results for these test problems, and discuss the effects of hyperbolic divergence cleaning. The last problem is the accretion of a fluid onto a Schwarzschild black hole. We numerically recover the steady state solution using AMR with a shadow hierarchy as the refinement criterion. Finally, we present scaling results for parallel unigrid and AMR runs. In the tests below we use the discrete $L_2$ norm $$\begin{aligned} || \, u \, ||_2 = \left[ {1\over N-1} \, \sum_i^N \left(u_i\right)^2 \right]^{1/2}\end{aligned}$$ where $u$ is a discrete function defined at $N$ locations, $u_i$. Riemann problem test {#subsec:rpt} -------------------- Balsara introduced several test Riemann problems for relativistic MHD [@Balsara2001], and we choose here his third blast wave problem for its very narrow features as a test of our AMR. The initial parameters for this Riemann problem are given in . The exact solution for this problem is given by Giacomazzo and Rezzolla [@Giacomazzo2005jy], and is plotted in the figures below for comparison. [**Balsara blast wave parameters**]{} $\rho_0$ P $v^x$ $v^y$ $v^z$ $B^x$ $B^y$ $B^z$ ------- ---------- ------ ------- ------- ------- ------- ------- ------- Left 1.0 1000 0.0 0.0 0.0 10.0 7.0 7.0 Right 1.0 0.1 0.0 0.0 0.0 10.0 0.7 0.7 : The initial parameters for the Balsara blast wave. The discontinuity is initially placed at $x=0$ and the system is evolved along the $x$ axis. The adiabatic index is $\Gamma = 5/3$. The exact solution to this problem is given in [@Giacomazzo2005jy]. The Courant factor used in the numerical solution is $0.2$.[]{data-label="table:blastwave"} The blast wave problem is implemented using the three-dimensional [had]{} infrastructure for AMR and simulated along a coordinate axis. shows the blast wave with a strong initial pressure difference centered at $x=0$ evolved along the $x$-axis at several unigrid resolutions. The plots show $\rho_0$, $v^x$, $v^y$, and $B^y$ at time $t=0.4$. The unigrid simulations were performed in the $x$ direction on a domain of $\left[-0.5,0.5\right]$ with a Courant factor of $0.2$. ------------- ------------- [**(a)**]{} [**(b)**]{} [**(c)**]{} [**(d)**]{} ------------- ------------- A series of two-level AMR simulations of the blast wave test were conducted with refinement centered on the shock propagation. compares a unigrid simulation with a two-level AMR simulation where the resolution of the finest mesh in the AMR hierarchy is the same as that in the unigrid simulation. The AMR capability to reproduce the resolution-equivalent unigrid result depends on how well the refinement region tracks the shock throughout the time history of the evolution. In our tests we observe the AMR blast wave simulations reproducing the resolution-equivalent unigrid result to within $~0.1\%$. ------------- ------------- [**(a)**]{} [**(b)**]{} ------------- ------------- Spherical blast wave {#subsec:sbw} -------------------- The spherical blast wave consists of a uniform fluid background with a small spherical region where the pressure is $10^6$ times larger than the background. The background pressure is $10^{-2}$, and $P=10^4$ inside a central sphere of radius $0.08$. This is the three-dimensional extension of the cylindrical blast wave studied in [@DelZanna2002rv; @Shibata:2005gp; @Neilsen2005]. The parameters for the spherical blast wave are given in . We first calculate the solution on a single uniform grid, and then draw comparisons to the AMR results. shows $z=0$ cuts of the uniform grid solution at $t=0.4$, and shows line plots of the pressure, $P$, along the $x$- and $y$-axes at three different resolutions. Adaptive mesh refinement substantially improves performance for the spherical blast wave while returning results nearly identical to the unigrid result. shows the resulting mesh hierarchy and pressure at time $t=0.4$ for the spherical blast wave in an AMR simulation with two levels of refinement. This AMR simulation uses hyperbolic divergence cleaning and is the AMR equivalent of the divergence cleaning unigrid simulation in . The refinement criteria are set to center refinement on the shock. The relative simplicity of this solution—the outgoing shock is the dominant feature of the solution–allows many refinement criteria to produce similar mesh hierarchies. The AMR simulation in was performed on 32 processors and was a factor of eight times faster than the equivalent unigrid simulation. Like the Balsara blast wave case examined in \[subsec:rpt\], adaptivity significantly reduces the computational overhead required to adequately resolve the multiscale features that appear in the simulation. Finally, we monitor the violations of the solenoidal constraint during both free evolutions and evolutions with hyperbolic divergence cleaning. shows the magnitude of the $L_2$ norm of $\nabla\cdot {\bf B}$ at three different resolutions with and without divergence cleaning. There are some subtleties in interpreting $L_2$ norms of the constraint violation, which arise primarily because Richardson-like convergence can not be defined for discontinuous functions [@Neilsen2005]. However, it appears here that the constraint violations are propagated at roughly the same velocity as the out-going shock, and the difference between free and cleaned evolutions is small. [**Spherical blast wave parameters**]{} $\rho_0$ $P$ $v^x$ $v^y$ $v^z$ $B^x$ $B^y$ $B^z$ ---------------- ---------- -------- ------- ------- ------- ------- ------- ------- Inside sphere 1.0 $10^4$ 0.0 0.0 0.0 4.0 0.0 0.0 Outside sphere 1.0 0.01 0.0 0.0 0.0 4.0 0.0 0.0 : The initial parameters for the spherical blast wave. The data consists of a uniform fluid background with the pressure set to $10^4$ inside a sphere of radius $0.08$ centered at the origin. The adiabatic index is $\Gamma = 4/3$. The domain of simulation is $\{x,y,z\}\in\left[ -1,1\right]$. The Courant factor is $0.3$.[]{data-label="table:sphblastwave"} With Cleaning Difference ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ $P$ ![Unigrid simulation results for the spherical blast wave at $t=0.4$, showing a slice along the $z=0$ plane. The $x$ axis is the horizontal direction. The pressure found using hyperbolic divergence cleaning is shown on the left. The difference between the pressure found with and without hyperbolic divergence cleaning is shown on the right. This gives an estimation of the relative errors that arise in free evolutions. The simulations were performed using a resolution of $h=0.006410$ on a domain of $\{x,y,z\}\in\left[ -1,1\right]$.[]{data-label="fig:2dsphshock"}](sphShock_p.ps "fig:"){height="6.0cm"} $\Delta P$ ![Unigrid simulation results for the spherical blast wave at $t=0.4$, showing a slice along the $z=0$ plane. The $x$ axis is the horizontal direction. The pressure found using hyperbolic divergence cleaning is shown on the left. The difference between the pressure found with and without hyperbolic divergence cleaning is shown on the right. This gives an estimation of the relative errors that arise in free evolutions. The simulations were performed using a resolution of $h=0.006410$ on a domain of $\{x,y,z\}\in\left[ -1,1\right]$.[]{data-label="fig:2dsphshock"}](sphshock_diff_p.ps "fig:"){height="6.0cm"} ------------- ------------- [**(a)**]{} [**(b)**]{} ------------- ------------- \ [**(a)**]{}\ \ [**(b)**]{}\ Relativistic rotor {#subsec:rr} ------------------ The relativistic rotor test case starts with a rigidly rotating fluid and evolves it in the presence of a magnetic field. This problem is discussed and examined in $2+1$ dimensions in [@Neilsen2005; @DelZanna2002rv]. Here we examine the relativistic rotor in $3+1$ dimensions, confining the initially rigidly rotating fluid to a sphere of radius $0.1$ with the angular momentum vector pointing in the $+z$ direction. The fluid is initially rotating with an angular velocity of $9.95$. The initial data and relevant evolution parameters are given in . Results using a uniform computational grid form the standard against which we measure the AMR results. The $L_2$ norms of the ${\bf \nabla} \cdot {\bf B}$ constraint violation as a function of time using three different unigrid resolutions are shown in , comparing results obtained both with and without hyperbolic divergence cleaning. Hyperbolic divergence cleaning significantly improves the constraint preservation in the relativistic rotor case. 2-D slices of the pressure along the $z=0$ plane at time $t=0.4$ are shown in . Adaptive mesh refinement results are presented in figures \[fig:ramr\] and \[fig:ramrdiff\]. These figures present a two-level AMR simulation with refinement centered on the shock front. This AMR simulation required five times fewer CPU hours than the equivalent unigrid simulation. shows the resulting pressure and mesh hierarchy at time $t=0.4$. compares the difference along the $x$ and $y$ axes of the pressure between the AMR and equivalent unigrid simulation. The AMR simulation was found to reproduce the unigrid results to well within 0.1%. [**Relativistic rotor parameters**]{} $\rho$ $P$ $B^x$ $B^y$ $B^z$ ---------------- -------- ----- ------- ------- ------- Inside sphere 10.0 1.0 1.0 0.0 0.0 Outside sphere 1.0 1.0 1.0 0.0 0.0 : The initial parameters for the relativistic rotor. The data consists of an initially rigidly rotating fluid inside a sphere of radius $0.1$ centered at the origin with a magnetic field. The fluid is rotating with $\omega = 9.95$ around the $z$ axis. The adiabatic index is $\Gamma = 5/3$. The domain of simulation is $\left[ -1,1\right]$ in each of the $x$,$y$, and $z$ directions. The Courant factor used in the presented relativistic rotor simulations was $0.2$.[]{data-label="table:relrotor"} With Divergence Cleaning Difference ----- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $P$ ![Unigrid simulation results for the relativistic rotor at time 0.4, showing a slice along the $z=0$ plane. The $x$ axis is the horizontal direction. The pressure found using hyperbolic divergence cleaning is shown on the left. The difference between the pressure found with and without hyperbolic divergence cleaning is shown on the right. This gives an estimation of the relative errors that arise in free evolutions. The simulations were performed using a resolution of $h=0.00625$ on a domain of $\{x,y,z\}\in\left[ -1,1\right]$.[]{data-label="fig:2drelrotor"}](rotor_c_p.ps "fig:"){height="6.0cm"} $\Delta P$ ![Unigrid simulation results for the relativistic rotor at time 0.4, showing a slice along the $z=0$ plane. The $x$ axis is the horizontal direction. The pressure found using hyperbolic divergence cleaning is shown on the left. The difference between the pressure found with and without hyperbolic divergence cleaning is shown on the right. This gives an estimation of the relative errors that arise in free evolutions. The simulations were performed using a resolution of $h=0.00625$ on a domain of $\{x,y,z\}\in\left[ -1,1\right]$.[]{data-label="fig:2drelrotor"}](rotor_diff_p.ps "fig:"){height="6.0cm"} ------------- ------------- [**(a)**]{} [**(b)**]{} ------------- ------------- Accretion onto a Black Hole {#subsec:accretion} --------------------------- Numerical simulation of fluid accretion onto a Schwarzschild black hole provides a test of the AMR infrastructure using a curved space background and an excision region. The steady state solution is given by Michel [@Michel] and has been explored previously using ingoing Eddington-Finkelstein coordinates [@Papadopoulos]. This solution describes the continuous spherical accretion of a fluid onto a black hole. We set a fixed black hole metric using ingoing Eddington-Finkelstein coordinates. To avoid the singularity inside the black hole, we implement a cubic excision region. The excision cube is located at the center of the grid and has a half width of $0.3M$. The boundary condition at the excision region is a copy condition; points next to the excision region are simply copied into the excision region when necessary for reconstruction. The mass of the black hole, $M$, is set to one and the black hole is placed at the center of the grid. The sonic radius, $r_c$, is selected to be $400M$ with a density $\rho_c = 0.01$. The domain of simulation is $\{x,y,z\}\in\left[ -15M,15M\right]$. The Michel steady state solution is found following the procedure described in [@Papadopoulos] and the outer boundary is kept fixed at this solution, providing a continual influx of mass onto the black hole. For radius $r > 2.5M$ the Michel steady state solution is set as initial data; for $r \leq 2.5M$ the initial data are set to be $\rho_0 = 0.1$, $P=0.1$, and $v^i = 0$. The fluid falls onto the black hole and eventually reaches steady state. A comparison of the Michel steady state solution and numerical solution at $t=50M$ is given in figure \[fig:accretion\]. The AMR grid structures at times $t=0M$ and $t=50M$ are given in figure \[fig:accretion2\]. A convergence test for $\rho_0$ is presented in figure \[fig:accretion3\]. ------------------ ------------------- [**$t = 0M$**]{} [**$t = 50M$**]{} ------------------ ------------------- Scaling {#subsec:scaling} ------- Unigrid and mesh refinement parallel scaling tests for the spherical blast wave are given in . The results presented are the strong scaling results; the global problem size was kept constant while the number of processors varied. Strong scaling tests are problem dependent and vary according to the size of the global problem selected for investigation. However, they also give the most direct indication of performance speed-up across a wide range of processors for a particular problem. Weak scaling tests are where the global problem size is increased as the number of processors increases in order to keep the problem size local to each processor constant. Weak scaling tests were also performed on the spherical blast wave problem on 16–256 processors without showing any significant performance degradation. In the unigrid strong scaling tests of (a) a $121^3$ spherical blast wave problem was evolved for 80 iterations on 1–128 processors. Speed-up is defined as $$\begin{aligned} {\rm speedup}(n) = \frac{{\rm Run~time~on~one~processor}}{{\rm Run~time~on~{\it n}~processors}}. \end{aligned}$$ As the number of processors increases, the communication eventually overshadows the local processor computation. For the test problem size examined, this begins on $\geq 64$ processors. For comparison, strong scaling results for a different unigrid TVD MHD code are given in [@CITA], where communication overhead saturation occurs on $> 64$ processors using a $240^3$ mesh size and $8$ processors the base scaling value. In the mesh refinement strong scaling tests of (b) a $81^3$ spherical blast wave with a single level of mesh refinement was evolved for 30 iterations. The base number of processors for scaling measurement was $8$ on account of memory considerations. Results are shown for 8–80 processors. ------------- ------------- [**(a)**]{} [**(b)**]{} ------------- ------------- Conclusion {#sec:conclusion} ========== We have presented three flat space relativistic MHD tests and one fluid accretion test using vertex-centered distributed adaptive mesh refinement and the approximate Riemann solver algorithm for the GRMHD equations presented in [@Neilsen2005]. Each of the three relativistic MHD tests, including the Balsara black wave and the spherical shock and relativistic rotor, have sharp features requiring high resolutions. In each of these cases, substantial performance gains of AMR versus unigrid were observed. Two level AMR simulations required between 5–8 times fewer CPU hours than the equivalent unigrid cases. AMR results reproduced the unigrid results, often better than 0.1%. Hyperbolic divergence cleaning was examined in connection with the spherical blast wave and relativistic rotor cases. It had a positive impact on constraint control in both cases. The impact was especially pronounced in the relativistic rotor case, which has more features interior to the outgoing wave front than the spherical blast wave. Hyperbolic divergence cleaning worked equally well in both unigrid and AMR tests. Fluid accretion onto a Schwarzschild black hole tests the code using a curved space background and a black hole excision region as well as using AMR. In this test, a shadow hierarchy for truncation error estimation was used as the AMR refinement criterion in recovering the known steady state solution. Parallel performance measures were presented in connection with the spherical blast wave. Speed-ups achieved using the spherical blast wave were reported both with unigrid and mesh refinement simulations. Performance speed-ups were found on up to at least $128$ processors with unigrid and up to at least $80$ processors with mesh refinement. By its very nature, the strong scaling test is problem dependent: the problem size is fixed while the number of processors is varied. This is in contrast to the weak scaling tests often presented in numerical relativity, where the problem size per processor is fixed. Strong scaling tests, however, address the real-world questions of how long it takes to solve a particular problem, and how to do it most efficiently. Having presented these tests, we now turn to some questions of astrophysical interest mentioned in the introduction. In particular, we have included fully dynamical general relativity in our code using the Einstein equations specified in [@Tiglio:2003xm]. In future work we hope to present evolutions of TOV stars as well as rotating, magnetized neutron stars. Recent work suggests interesting effects of general relativity with rotation on supermassive polytropes and the bar mode instability [@Zink:2005rr]. The addition of magnetic fields to these systems may suggest new questions and provide new insight into magnetic astrophysical phenomena. One part of this question includes understanding not only the interior of a magnetized rotating neutron star, but its magnetosphere as well. Ideal MHD codes based on Godunov-type schemes frequently encounter difficulties when $B^2 \gg \rho_0$, as relatively small truncation errors in the evolution of the conserved variables lead to large fractional errors in computing the internal energy density and other primitive variables. While we have tried to create a robust primitive variable solver, this difficulty can not be avoided for high-resolution shock-capturing schemes. Therefore, a full numerical study of such a star and magnetosphere may require coupling the equations of ideal MHD for the interior solution with the equations of force-free electrodynamics for the exterior. We are actively pursuing this question. Another area to be targeted in future papers is constraint preserving boundary conditions for MHD. Currently we use the conventional outflow boundary conditions. This could be improved by using outer boundary conditions that are constraint preserving. Additionally, for some systems we wish to require that no incoming modes enter the domain. To this end it would be useful to construct the full spectral decomposition of our system in order to be able to determine ingoing and outgoing modes. Work in this direction has already begun and shows promise. We are pleased to thank Luis Lehner, Carlos Palenzuela, Ignacio Olabarrieta, Patrick Motl, Tanvir Rahman, Oscar Reula, and Joel Tohline for helpful discussions and comments during the course of this work. We also thank Bruno Giacomazzo and Luciano Rezzolla for sharing their computer code to solve the exact Riemann problem for relativistic MHD. This work was supported by the National Science Foundation under grants PHY-0326311 and PHY-0244699 to Louisiana State University, PHY-0326378 and PHY-0502218 to Brigham Young University, and PHY-0325224 to Long Island University. This research was also supported in part by the National Science Foundation through TeraGrid resources provided by SDSC under allocation award PHY-040027.

--- abstract: | Let $U:[0,\infty)^2 \to [0,\infty)$ be a measurable kernel satisfying: 1. $U(x,y)$ is nonincreasing in $x$ and nondecreasing in $y$; 2. there exists a constant ${\vartheta}>0$ such that $$U(x,z) \le {\vartheta}\left( U(x,y)+U(y,z) \right)$$ for all $0\le x<y<z<\infty$; 3. $U(0,y)>0$ for all $y>0$. Let $0<q<1< p <\infty$. We prove that the weighted inequality $$\left( \int_0^\infty \left( \int_0^t f(x)U(x,t){\mathrm{\,d}x}\right)^q w(t) {\mathrm{\,d}t}\right)^{\frac1q}\le C \left( \int_0^\infty f^p(t)v(t){\mathrm{\,d}t}\right)^{\frac1p}$$ holds for all nonnegative measurable functions $f$ on $(0,\infty)$ if and only if $${\left(}\int_0^\infty \left( \int_t^\infty w(x){\mathrm{\,d}x}\right)^{\frac rp}w(t) \left( \int_0^t U^{p'}(z,t){v^{1\!-p'}}(z){\mathrm{\,d}z}\right)^{\frac r{p'}}{\mathrm{\,d}t}{\right)}^{\frac1r}<\infty$$ and $${\left(}\int_0^\infty \left( \int_t^\infty w(x) {U^q}(t,x) {\mathrm{\,d}x}\right)^{\frac rp}w(t) \sup_{z\in(0,t)} {U^q}(z,t)\left( \int_0^z {v^{1\!-p'}}(s){\mathrm{\,d}s}\right)^{\frac r{p'}}{\mathrm{\,d}t}{\right)}^{\frac1r}<\infty,$$ where $p':=\frac{p}{p-1}$ and $r:=\frac{pq}{p-q}$. Analogous conditions for the case $p=1$ and for the dual version of the inequality are also presented. address: - 'Karlstad University, Faculty of Health, Science and Technology, Department of Mathematics and Computer Science, 651 88 Karlstad, Sweden' - 'Charles University, Faculty of Mathematics and Physics, Department of Mathematical Analysis, Sokolovská 83, 186 75 Praha 8, Czech Republic' author: - Martin Křepela title: | Boundedness of Hardy-type operators with a kernel:\ integral weighted conditions for the case $0<q<1\le p<\!\infty$ --- Introduction ============ Operators of the general form $$Tf(x) = \int_0^\infty f(y)U(x,y) {\mathrm{\,d}y},$$ where $U$ is a *kernel*, play an indispensable role in various areas of analysis. The means of their investigation, naturally, greatly depend on additional properties of the kernel $U$. In the present article, we study the so-called *Hardy-type* operators $$\label{700} Hf(x) = \int_0^x f(y)U(y,x) {\mathrm{\,d}y}, \qquad \textnormal{and}\qquad H^*f(x) = \int_x^\infty f(y)U(x,y) {\mathrm{\,d}y},$$ where the kernel $U:[0,\infty)^2\to [0,\infty)$ is a measurable function which has the following properties: 1. $U(x,y)$ is nonincreasing in $x$ and nondecreasing in $y$; 2. there exists a constant ${\vartheta}>0$ such that for all $0\le x<y<z<\infty$ it holds $$ U(x,z) \le {\vartheta}\left( U(x,y)+U(y,z) \right);$$ 3. $U(0,y)>0$ for all $y>0$. If ${\vartheta}>0$ and $U$ is a function satisfying the conditions above with the given parameter ${\vartheta}$ in point (ii), then we, for the sake of simplicity, call $U$ a *${\vartheta}$-regular kernel*. The simplest case of a ${\vartheta}$-regular kernel $U$ is the constant $U\equiv 1$, with which $H$ and $H^*$ become the ordinary Hardy and Copson (“dual Hardy”) operators, respectively. Other examples of ${\vartheta}$-regular kernels include the Riemann-Liouville kernel $$U(x,y)=(y-x)^\alpha, \quad \alpha>0,$$ the logarithmic kernel $$U(x,y)=\log^\alpha {\left(}\frac yx {\right)}, \quad \alpha>0,$$ and the kernels $$U(x,y)=\int_x^y u(t){\mathrm{\,d}t}\textnormal{\quad and\quad } U(x,y)={\operatornamewithlimits{ess\, sup\,}}_{t\in(x,y)}u(t),$$ where $u$ is a given nonnegative measurable function. These operators find applications, for instance, in the theory of differentiability of functions, interpolation theory and more topics involving function spaces. The two last-named examples of ${\vartheta}$- regular kernels prove to be particularly useful in a study of the so-called iterated Hardy operators [@GKPS; @GS], for example. The particular aspect we investigate in this paper is boundedness of the operators $H$ and $H^*$ with a ${\vartheta}$-regular kernel $U$ between weighted Lebesgue spaces. In order to define these spaces, we need to introduce several auxiliary terms first. Throughout the text, by a *measurable function* we always mean a Lebesgue measurable function (on an appropriate subset of ${\mathbb{R}}$). The symbol ${\mathscr M_+}$ denotes the cone of all nonnegative measurable functions on $(0,\infty)$. A *weight* is a function $w\in{\mathscr M_+}$ on $(0,\infty)$ such that $$0<\int_0^t w(s){\mathrm{\,d}s}<\infty \textnormal{\quad for\ all\quad} t>0.$$ Finally, if $v$ is a weight and $p\in(0,\infty]$, then the *weighted Lebesgue space* ${L^p(v)}={L^p(v)}(0,\infty)$ is defined as the set of all real-valued measurable functions $f$ on $(0,\infty)$ such that $$\begin{aligned} \|f\|_{{L^p(v)}} & :=\ \left(\int_0^\infty |f(t)|^p v(t){\mathrm{\,d}t}\right)^{\frac1p}<\infty & \textnormal{if\ }p<\infty, \\ \|f\|_{L^{\!\infty}(v)} & :=\ {\operatornamewithlimits{ess\, sup\,}}_{t\in(0,\infty)} |f(t)|v(t) <\infty & \textnormal{if\ }p=\infty. \end{aligned}$$ Note that if $p\in(0,1)$, then $({L^p(v)}, \|\cdot\|_{{L^p(v)}})$ is in general not a normed linear space because of the absence of the Minkowski inequality in this case. However, as we deal only with the case $1\le p<\infty$ anyway, this detail is not of our concern here. Throughout the text, if $p\in(0,1)\cup(1,\infty)$, then $p'$ is defined by $p'=\frac{p}{p-1}$. Analogous notation is used for $q'$. In the following, assume that ${\vartheta}\in(0,\infty)$, $U$ is a ${\vartheta}$-regular kernel, $H$ is the corresponding operator from and $v$, $w$ are weights. Boundedness of the operator $H$ between ${L^p(v)}$ and ${L^q(w)}$ was completely characterized for $p,q\in[1,\infty]$. The authors credited for this work are Bloom and Kerman [@BK], Oinarov [@Oi] and Stepanov [@S1]. The results of [@Oi], for instance, have the following form. Let $1<p\le q<\infty$. Then $H:{L^p(v)}\to{L^q(w)}$ is bounded if and only if $$E_1 := \sup_{t\in(0,\infty)} {\left(}\int_t^\infty {U^q}(t,x)w(x){\mathrm{\,d}x}{\right)}^{\frac1q}{\left(}\int_0^t {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac1{p'}}< \infty$$ and $$E_2 := \sup_{t\in(0,\infty)} {\left(}\int_t^\infty w(x){\mathrm{\,d}x}{\right)}^{\frac1q}{\left(}\int_0^t {U^{p'}}(x,t){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac1{p'}}<\infty.$$ Moreover, the least constant $C$ such that the inequality $$\label{701} \| Hf\|_{{L^q(w)}} \le C \|f\|_{{L^p(v)}}$$ holds for all $f\in{\mathscr M_+}$ satisfies $C\approx E_1 + E_2$. Let $1<q<p<\infty$ and $r:=\frac{pq}{p-q}$. Then $H:{L^p(v)}\to{L^q(w)}$ is bounded if and only if $$E_3 := {\left(}\int_0^\infty {\left(}\int_t^\infty {U^q}(t,x)w(x){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_0^t {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{q'}}{v^{1\!-p'}}(t) {\mathrm{\,d}t}{\right)}^{\frac1r}< \infty$$ and $$E_4 := {\left(}\int_0^\infty {\left(}\int_t^\infty w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_0^t {U^{p'}}(x,t){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}{\right)}^{\frac1r}<\infty.$$ Moreover, the least constant $C$ such that holds for all $f\in{L^p(v)}$ satisfies $C\approx E_3 + E_4$. The conditions obtained in [@BK; @S1] have a slightly different form, a more detailed comparison between them is found in [@S1]. As for the “limit cases”, conditions for the case $p=\infty$ and $q\in(0,\infty]$ are obtained very easily, the same applies to the case $q=1$ and $p\in[1,\infty)$ in which one simply uses the Fubini theorem. Yet another possible choice of parameters is $p=1$ and $q\in(1,\infty]$. It was (at least for $q<\infty$) included in [@Oi Theorem 1.2] and the conditions may be recovered from that article by correctly interpreting the expressions involving the symbol $p'$ in there. Another option is to follow the more general theorem [@KA Chapter XI, Theorem 4]. If $0<p<1$, then the operator $H$ can never be bounded (provided that $U$, $v$, $w$ are nontrivial, which is always assumed here). The problem in here lies in the fact that for each $t>0$ there exists $f_t\in L^p(v)$ which is not locally integrable at the point $t$. For more details, see e.g. [@L2]. No such difficulty arises if $0<q<1\le p<\infty$. In this case, $H$ may indeed be bounded between ${L^p(v)}$ and ${L^q(w)}$ and it is perfectly justified to ask for the conditions under which this occurs. As for the known answers to this question, the situation is however much worse than in the other cases. When assumed $U\equiv 1$, i.e. for the ordinary Hardy operator, the boundedness characterization was found by Sinnamon [@Si91] and it corresponds to the condition $E_3<\infty$ (with $U\equiv 1$, of course). In the general case, in [@S1] it was shown that the condition $E_3<\infty$ is sufficient but not necessary for $H:{L^p(v)}\to{L^q(w)}$ to be bounded, while the condition $$E_5 := {\left(}\int_0^\infty {\left(}\int_t^\infty {U^q}(t,x)w(x){\mathrm{\,d}x}{\right)}^\frac{p'}{q} {v^{1\!-p'}}(t) {\mathrm{\,d}t}{\right)}^{\frac1{p'}}<\infty$$ is necessary but not sufficient. For related counterexamples, see [@S2]. The fact that the two conditions do not meet is a significant drawback. An equivalent description of the optimal constant $C$ in is usually substantial for the result to be applicable in any way. Lai [@L] found equivalent conditions by proving that, with $0<q<1<p<\infty$, the operator $H$ is bounded from ${L^p(v)}$ to ${L^q(w)}$ if and only if $$\widetilde{D}_1 := \sup_{\{t_k\}} \sum_{k} {\left(}\int_{t_k}^{t_{(\!k\!+\!1\!)}} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} {U^{p'}}(x,{{t_k}}) {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}<\infty$$ as well as $$\widetilde{D}_2 := \sup_{\{t_k\}} \sum_{k} {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} w(t) {U^q}({{t_k}},t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}<\infty.$$ The suprema in here are taken over all covering sequences, i.e. partitions of $(0,\infty)$ (see [@L] or Section 2 for the definitions), and $r:=\frac{pq}{p-q}$, as usual. Moreover, these conditions satisfy $\widetilde{D}_1 + \widetilde{D}_2 \approx C^r$ with the least $C$ such that holds for all $f\in{\mathscr M_+}$. Corresponding variants for $p=1$ are also provided in [@L]. The earlier use of similar partitioning techniques in the paper [@MRS] of Martín-Reyes and Sawyer should be also credited. Unfortunately, even though the $\widetilde{D}$-conditions are both sufficient and necessary, they are only hardly verifiable due to their discrete form involving all possible covering sequences. This fact has hindered their use in various applications (see e.g. [@GS]). In contrast, in the case $1<q<p<\infty$ it is known (see [@S2; @L]) that $\widetilde{D}_1 + \widetilde{D}_2 \approx A_3^r + A_4^r$. This does not apply when $0<q<1\le p<\infty$, as shown by the results of [@S1] mentioned earlier. Rather recently, Prokhorov [@P] found conditions for $0<q<1\le p<\infty$ which have an integral form but involve a function $\zeta$ defined by $$\zeta(x):=\sup \left\{ y\in(0,\infty); \int_y^\infty \!w(t){\mathrm{\,d}t}\ge ({\vartheta}^q\!+1) \int_x^\infty \! w(t){\mathrm{\,d}t}\right\}, \quad x>0.$$ The conditions presented in [@P] even involve this function iterated three times. The presence of such an implicit expression involving the weight $w$ virtually prevents any use of these conditions in applications which require further manipulation $w$ (see Section 4 for an example). Finding explicit integral conditions for the case $0<q<1\le p<\infty$, which would have a form comparable e.g. to $E_3$ and $E_4$, hence remained an open problem.\ In this paper, we solve this problem and provide the missing integral conditions. No additional assumptions on the weights $v$, $w$ and the ${\vartheta}$-regular kernel $U$ are required here, neither are any implicit expressions. The results are presented in Theorems \[7\], \[144\] and Corollaries \[7\*\], \[144\*\]. The proofs are based on the well-known method of *dyadic discretization* (or *blocking technique*, see [@GE] for a basic introduction into this method). The particular variant of the technique employed here is essentially the same as the one used in [@K6]. Concerning the structure of this paper, this introduction is followed by Section 2 where additional definitions and various auxiliary results are presented. Section 3 consists of the main results, their proofs and some related remarks. In the final Section 4 we present certain examples of applications of the results. Definitions and preliminaries ============================= Let us first introduce the remaining notation and terminology used in the paper. We say that ${\mathbb{I}}\subseteq {\mathbb{Z}}$ is an *index set* if there exist $k_{\min},\, k_{\max}\in{\mathbb{Z}}$ such that $k_{\min}\le k_{\max}$ and $${\mathbb{I}}=\{ k\in{\mathbb{Z}},\,k_{\min}\le k \le k_{\max}\}.$$ Moreover, we denote $${\mathbb{I}}_0:={\mathbb{I}}\setminus \{k_{\min},k_{\max}\}.$$ Let ${\mathbb{I}}$ be an index set containing at least three indices. Then a sequence of points $\{t_k\}_{k\in{\mathbb{I}}}$ is called a *covering sequence* if $t_{k_{\min}}=0$, $t_{k_{\max}}=\infty$ and $t_k<t_{(\!k\!+\!1\!)}$ whenever $k\in{\mathbb{I}}\setminus\{k_{\max}\}$. Next, let $z\in{\mathbb{N}}\,\cup\{0\}$ and $n,k\in{\mathbb{N}}$ are such that $0\le k <n$. We write $z{\operatorname{\,mod}}n = k$ if there exists $j\in{\mathbb{N}}\,\cup\{0\}$ such that $z=jn + k$. In other words, $k$ is the remainder after division of the number $z$ by the number $n$.\ In the next part, we present various auxiliary results which will be needed later. The first two propositions are simple consequences of the Hölder inequality (for functions and sequences) and of the characterizations of the dual spaces to the spaces $L^p$ and $l^p$. For more details see e.g. [@K6]. \[107\] Let $v$ be a weight, $p\in(1,\infty)$ and $0\le x<y\le \infty$. Let $f$ and ${\varphi}$ be nonnegative measurable functions on $(x,y)$. Then the Hölder inequality gives $$\int_x^y f(s){\varphi}(s){\mathrm{\,d}s}\le \left(\int_x^y f^p(s)v(s){\mathrm{\,d}s}\right)^{\frac1p}\left(\int_x^y {\varphi}^{p'}\!(s){v^{1\!-p'}}(s){\mathrm{\,d}s}\right)^\frac1{p'}.$$ Moreover, if $\int_x^y {\varphi}^{p'}\!(s){v^{1\!-p'}}(s){\mathrm{\,d}s}<\infty$, then there exists a nonnegative measurable function $g$ supported in $[x,y]$ and such that $\int_x^y g^p(s)v(s){\mathrm{\,d}s}= 1$ and $$\left(\int_x^y {\varphi}^{p'}\!(s){v^{1\!-p'}}(s){\mathrm{\,d}s}\right)^\frac1{p'} \le 2 \int_x^y g(s){\varphi}(s){\mathrm{\,d}s}<\infty.$$ If $\int_x^y {\varphi}^{p'}\!(s){v^{1\!-p'}}(s){\mathrm{\,d}s}=\infty$, then for every $M\in{\mathbb{N}}$ there exists a nonnegative measurable function $h$ supported in $[x,y]$ and such that $\int_x^y h^p(s)v(s){\mathrm{\,d}s}= 1$ and $$\int_x^y h(s){\varphi}(s){\mathrm{\,d}s}> M.$$ \[108\] Let ${\mathbb{I}}$ be an index set. Let $\{a_k\}_{k\in{\mathbb{I}}}$ and $\{b_k\}_{k\in{\mathbb{I}}}$ be two nonnegative sequences. Assume that $0<q<p<\infty$. Then the Hölder inequality gives $$\left(\sum_{k\in{\mathbb{I}}} a_k^q b_k\right)^\frac1q \le \left(\sum_{k\in{\mathbb{I}}} a_k^p\right)^\frac1p \left( \sum_{k\in{\mathbb{I}}}b_k^\frac{p}{p-q}\right)^\frac{p-q}{pq}.$$ Moreover, if $\sum_{k\in{\mathbb{I}}}b_k^\frac{p}{p-q}\!<\infty$, then there exists a nonnegative sequence $\{c_k\}_{k\in{\mathbb{I}}}$ such that $\sum_{k\in{\mathbb{I}}} c_k^p=1$ and $$\left( \sum_{k\in{\mathbb{I}}}b_k^\frac{p}{p-q}\right)^\frac{p-q}{pq} \le 2 \left(\sum_{k\in{\mathbb{I}}} c_k^q b_k\right)^\frac1q <\infty.$$ Proposition \[107\] has also a variant for $p=1$, giving $$\int_x^y f(s){\varphi}(s){\mathrm{\,d}s}\le \int_x^y f(s)v(s){\mathrm{\,d}s}\ {\operatornamewithlimits{ess\, sup\,}}_{s\in(x,y)} {\varphi}(s)v^{-1}(s){\mathrm{\,d}s}.$$ The remaining part of Proposition \[107\] is then also true, provided that the expressions are modified accordingly. The next proposition was proved in [@GHS Proposition 2.1], more comments may be found e.g. in [@K6]. It is a fundamental part of the discretization method. \[3\] Let $0<\alpha<\infty$ and $1<D<\infty$. Then there exists a constant $C_{\alpha,D}\in(0,\infty)$ such that for any index set ${\mathbb{I}}$ and any two nonnegative sequences $\{b_k\}_{k\in{\mathbb{I}}}$ and $\{c_k\}_{k\in{\mathbb{I}}}$, satisfying $$b_{(\!k\!+\!1\!)}\ge D\, b_k \textit{\ for all\ } k\in{\mathbb{I}}\setminus\{k_{\max}\},$$ it holds $$\sum_{k=k_{\min}}^{k_{\max}} \left( \sum_{m=k}^{k_{\max}} c_m \right)^\alpha b_k \le C_{\alpha,D} \sum_{k=k_{\min}}^{k_{\max}} c_k^\alpha b_k$$ and $$\sum_{k=k_{\min}}^{k_{\max}} \left(\sup_{k\le m \le k_{\max}} \!\! c_m \right)^\alpha b_k \le C_{\alpha,D} \sum_{k=k_{\min}}^{k_{\max}} c_k^\alpha b_k .$$ The following result is an analogy to the previous proposition. We present a simple proof, although the result is also well known. \[89\] Let $0<\alpha<\infty$ and $1<D<\infty$. Then there exists a constant $C_{\alpha,D}\in(0,\infty)$ such that for any index set ${\mathbb{I}}$ and any two nonnegative sequences $\{b_k\}_{k\in{\mathbb{I}}}$ and $\{c_k\}_{k\in{\mathbb{I}}}$, satisfying $$b_{(\!k\!+\!1\!)} \ge D\, b_k \textit{\ for all\ } k\in{\mathbb{I}}\setminus\{k_{\max}\},$$ it holds $$\sup_{k_{\min}\le k \le k_{\max}} \left( \sum_{m=k}^{k_{\max}} c_m \right)^\alpha b_k \le C_{\alpha,D} \sup_{k_{\min}\le k \le k_{\max}} c_k^\alpha b_k.$$ It holds $$\begin{aligned} \sup_{k_{\min}\le k \le k_{\max}} {\left(}\sum_{m=k}^{k_{\max}} c_m {\right)}^\alpha b_k & = \sup_{k_{\min}\le k \le k_{\max}} {\left(}\sum_{m=k}^{k_{\max}} c_m b_m^{-\!\frac1\alpha} b_m^{\!\frac1\alpha} {\right)}^\alpha b_k \\ & \le \sup_{k_{\min}\le k \le k_{\max}} {\left(}\sum_{m=k}^{k_{\max}} b_m^{-\!\frac1\alpha} {\right)}^\alpha b_k \sup_{k\le i\le k_{\max}} c_i^\alpha b_i \\ & \le \sup_{k_{\min}\le k \le k_{\max}} {\left(}b_k^{-\!\frac1\alpha} \sum_{m=0}^{k_{\max}-k} D^{-\!\frac m\alpha} {\right)}^\alpha b_k \sup_{k\le i\le k_{\max}} c_i^\alpha b_i^\alpha \\ & \le {\left(}\sum_{m=0}^{\infty} D^{-\!\frac m\alpha} {\right)}^\alpha \sup_{k_{\min}\le i\le k_{\max}} c_i^\alpha b_i^\alpha \\ & = \frac D{(D^{\frac 1\alpha}\!-1)^\alpha} \sup_{k_{\min}\le i\le k_{\max}} c_i^\alpha b_i^\alpha. \end{aligned}$$ Applying Proposition \[3\], one obtains the next result. It is useful to handle inequalities involving a ${\vartheta}$-regular kernel. \[4\] Let $0<\alpha<\infty$ and ${\vartheta}\in [1,\infty)$. Let $U$ be a ${\vartheta}$-regular kernel. Then there exists a constant $C_{\alpha,{\vartheta}}\in(0,\infty)$ such that, for any index set ${\mathbb{I}}$, any increasing sequence $\{t_k\}_{k\in{\mathbb{I}}}$ of points from $(0,\infty]$ and any nonnegative sequence $\{a_k\}_{k\in{\mathbb{I}}\setminus\{k_{\max}\}}$ satisfying $$\label{6} a_{(\!k\!+\!1\!)}\ge 2{\vartheta}^\alpha a_k \textit{\ for all\ } k\in{\mathbb{I}}\setminus\{k_{\max},k_{\max}-1\},$$ it holds $$\sum_{k=k_{\min}}^{k_{\max}-1} a_k U^\alpha(t_k,t_{k_{\max}}) \le C_{\alpha,{\vartheta}} \sum_{k=k_{\min}}^{k_{\max}-1} a_k U^\alpha(t_k,t_{(\!k\!+\!1\!)}) .$$ Naturally, we may assume that ${\mathbb{I}}$ contains at least three indices. Let $k\in{\mathbb{I}}\setminus\{k_{\max}\}$. By iterating the inequality $$\label{61} U(x,z) \le {\vartheta}U(x,y) + {\vartheta}U(z,y) \qquad (x<y<z)$$ from the definition of the ${\vartheta}$-regular kernel, we get $$\begin{aligned} U(t_k, t_{k_{\max}}) & \le {\vartheta}U(t_k, t_{(\!k\!+\!1\!)}) + {\vartheta}U(t_{(\!k\!+\!1\!)}, t_{k_{\max}}) \\ & \le {\vartheta}U(t_k, t_{(\!k\!+\!1\!)}) + {\vartheta}^2 U(t_{(\!k\!+\!1\!)}, t_{(k+2)}) + {\vartheta}^2 U(t_{(k+2)}, t_{k_{\max}}) \\ & \ \vdots \\ & \le \sum_{m=k}^{k_{\max}-1} {\vartheta}^{m-k+1} U(t_m,t_{(m+1)}) \\ & = {\vartheta}^{-k} \sum_{m=k}^{k_{\max}-1} {\vartheta}^{m+1} U(t_m,t_{(m+1)}). \end{aligned}$$ Set $b_k := {\vartheta}^{-\alpha k}a_k$ for $k\in{\mathbb{I}}\setminus\{k_{\max}\}$. Then, by , for all $k\in{\mathbb{I}}\setminus\{k_{\max},k_{\max}-1\}$ it holds $b_{(\!k\!+\!1\!)}\ge 2 b_k$. We obtain $$\begin{aligned} \sum_{k=k_{\min}}^{k_{\max}-1} a_k U^\alpha(t_k,t_{k_{\max}}) & \le \sum_{k=k_{\min}}^{k_{\max}-1} {\vartheta}^{-\alpha k} a_k \left( \sum_{m=k}^{k_{\max}-1} {\vartheta}^{m+1} U(t_m,t_{(m+1)} \right)^\alpha \nonumber\\ & = \sum_{k=k_{\min}}^{k_{\max}-1} b_k \left( \sum_{m=k}^{k_{\max}-1} {\vartheta}^{m+1} U(t_m,t_{(m+1)}) \right)^\alpha \nonumber\\ & \le C_\alpha \sum_{k=k_{\min}}^{k_{\max}-1} b_k {\vartheta}^{\alpha(\!k\!+\!1\!)} U^\alpha(t_k,t_{(\!k\!+\!1\!)}) \label{60}\\ & = C_\alpha {\vartheta}^\alpha \sum_{k=k_{\min}}^{k_{\max}-1} a_k U^\alpha(t_k,t_{(\!k\!+\!1\!)}). \nonumber \end{aligned}$$ To get the inequality , we used Proposition \[3\], setting $D:=2$ and $c_m:=U(t_m,t_{(m+1)})$ for the relevant indices $m$. This proves the statement. Notice that, by the definitions at the beginning of this section, we consider only finite index sets (and therefore also finite covering sequences later on). However, all the results of this section hold for infinite sequences as well. This may be easily shown by using a limit argument. We will nevertheless continue working with finite index sets and covering sequences only. The notion of supremum is used regularly even where it relates to a finite set and where it therefore could be replaced by a maximum. For further remarks see the last part of Section 3. The final basic result concerns ${\vartheta}$-regular kernels and reads as follows. \[59\] Let $0\le a<b<c\le\infty$, $0<\alpha<\infty$ and $1\le {\vartheta}<\infty$. Let $U$ be a ${\vartheta}$-regular kernel and $\psi$ be a nonincreasing nonnegative function defined on $(0,\infty)$. Then $$\sup_{z\in[a,c)} U^\alpha(a,z)\psi(z) \le (1+{\vartheta}) {\left(}\sup_{z\in[a,b]} U^\alpha(a,z)\psi(z) + \sup_{z\in[b,c)} U^\alpha(b,z)\psi(z) {\right)}.$$ If $c<\infty$, the result is unchanged if the intervals $[a,c)$ and $[b,c)$ in the suprema are replaced by $[a,c]$ and $[b,c]$, respectively. The result is a consequence to the following simple observation. $$\begin{aligned} \sup_{z\in[a,c)} U^\alpha(a,z)\psi(z) & \le \sup_{z\in[a,b]} U^\alpha(a,z)\psi(z) + \sup_{z\in[b,c)} U^\alpha(a,z)\psi(z) \\ & \le \sup_{z\in[a,b]} U^\alpha(a,z)\psi(z) + {\vartheta}U^\alpha(a,b) \sup_{z\in[b,c)}\psi(z) + {\vartheta}\sup_{z\in[b,c)} U^\alpha(b,z)\psi(z)\\ & = \sup_{z\in[a,b]} U^\alpha(a,z)\psi(z) + {\vartheta}U^\alpha(a,b) \psi(b) + {\vartheta}\sup_{z\in[b,c)} U^\alpha(b,z)\psi(z)\\ & \le (1+{\vartheta}) {\left(}\sup_{z\in[a,b]} U^\alpha(a,z)\psi(z) + \sup_{z\in[b,c)} U^\alpha(b,z)\psi(z) {\right)}. \end{aligned}$$ Main results ============ This section contains the main theorems and their proofs. Remarks to the results and proof techniques can be found at the end of the section. The notation $A\lesssim B$ means that $A\le C B$, where the constant $C$ may depend *only* on the exponents $p$, $q$ and the parameter ${\vartheta}$. In particular, this $C$ is always independent on the weights $w$, $v$, on certain indices (such as $k$, $n$, $j$, $K$, $N$, $J$, $\mu,\dots$), on the number of summands involved in sums etc. We write $A\approx B$ if both $A\lesssim B$ and $B\lesssim A$. \[7\] Let $0<q<1<p<\infty$, $r:=\frac{pq}{p-q}$ and $0<{\vartheta}<\infty$. Let $v$, $w$ be weights. Let $U$ be a ${\vartheta}$-regular kernel. Then the following assertions are equivalent: 1. There exists a constant $C\in(0,\infty)$ such that the inequality $$\label{8} \left( \int_0^\infty {\left(}\int_t^\infty f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}\right)^{\frac1q}\le C {\left(}\int_0^\infty f^p(t)v(t){\mathrm{\,d}t}{\right)}^{\frac1p}$$ holds for all functions $f\in{\mathscr M_+}$. 2. Both the conditions $$D_1 := \sup_{\substack{\{t_k\}_{k\in{\mathbb{I}}} \\ \textnormal{covering}\\ \textnormal{sequence}}} \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x) {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}<\infty$$ and $$D_2 := \sup_{\substack{\{t_k\}_{k\in{\mathbb{I}}} \\ \textnormal{covering}\\ \textnormal{sequence}}} \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}<\infty$$ are satisfied. 3. Both the conditions $$A_1 := \int_0^\infty \left( \int_0^t w(x){\mathrm{\,d}x}\right)^{\frac rp}w(t) \left( \int_t^\infty U^{p'}(t,z){v^{1\!-p'}}(z){\mathrm{\,d}z}\right)^{\frac r{p'}}{\mathrm{\,d}t}<\infty$$ and $$A_2 := \int_0^\infty \left( \int_0^t w(x) {U^q}(x,t) {\mathrm{\,d}x}\right)^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z)\left( \int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}\right)^{\frac r{p'}}{\mathrm{\,d}t}<\infty$$ are satisfied. Moreover, if $C$ is the least constant such that holds for all functions $f\in{\mathscr M_+}$, then $$C^r \approx D_1 + D_2 \approx A_1 + A_2.$$ The variant of the previous theorem for $p=1$ reads as follows. \[144\] Let $0<q<1=p$ and $0<{\vartheta}<\infty$. Let $v$, $w$ be weights. Let $U$ be a ${\vartheta}$-regular kernel. Then the following assertions are equivalent: 1. There exists a constant $C\in(0,\infty)$ such that the inequality holds for all functions $f\in{\mathscr M_+}$. 2. Both the conditions $$D_3 := \sup_{\substack{\{t_k\}_{k\in{\mathbb{I}}} \\ \textnormal{covering}\\ \textnormal{sequence}}} \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(t){\mathrm{\,d}t}{\right)}^{1-q'} \!\!\! {\operatornamewithlimits{ess\, sup\,}}_{x\in({{t_k}},t_{(\!k\!+\!1\!)})} U^{-q'}({{t_k}},x)\, v^{q'}\!(x){\mathrm{\,d}x}<\infty$$ and $$D_4 := \sup_{\substack{\{t_k\}_{k\in{\mathbb{I}}} \\ \textnormal{covering}\\ \textnormal{sequence}}} \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\right)}^{1-q'} \!\!\! {\operatornamewithlimits{ess\, sup\,}}_{x\in({{t_k}},t_{(\!k\!+\!1\!)})} v^{q'}\!(x){\mathrm{\,d}x}<\infty$$ are satisfied. 3. Both the conditions $$A_3 := \int_0^\infty \left( \int_0^t w(x){\mathrm{\,d}x}\right)^{-q'} w(t)\ {\operatornamewithlimits{ess\, sup\,}}_{z\in(t,\infty)} U^{-q'}(t,z)\, v^{q'}\!(z)\, {\mathrm{\,d}t}<\infty$$ and $$A_4 := \int_0^\infty \left( \int_0^t w(x) {U^q}(x,t) {\mathrm{\,d}x}\right)^{-q'} w(t)\ {\operatornamewithlimits{ess\, sup\,}}_{z\in(t,\infty)} {U^q}(t,z)\, v^{q'}\!(z) \, {\mathrm{\,d}t}<\infty$$ are satisfied. Moreover, if $C$ is the least constant such that holds for all functions $f\in{\mathscr M_+}$, then $$C^{-q'} \approx D_3 + D_4 \approx A_3 + A_4.$$ By performing a simple change of variables $t \to \frac1t$, one gets the two corollaries below. They are formulated without the discrete conditions, those corresponding to Corollary \[7\*\] were presented in Section 1. An interested reader may also derive all the discrete conditions easily from their respective counterparts in Theorems \[7\] and \[144\]. \[7\*\] Let $0<q<1<p<\infty$, $r:=\frac{pq}{p-q}$ and $0<{\vartheta}<\infty$. Let $v$, $w$ be weights. Let $U$ be a ${\vartheta}$-regular kernel. Then the following assertions are equivalent: 1. There exists a constant $C\in(0,\infty)$ such that the inequality $$\label{8*} \left( \int_0^\infty {\left(}\int_0^t f(x)U(x,t){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}\right)^{\frac1q}\le C {\left(}\int_0^\infty f^p(t)v(t){\mathrm{\,d}t}{\right)}^{\frac1p}$$ holds for all functions $f\in{\mathscr M_+}$. 2. Both the conditions $$A^*_1 := \int_0^\infty \left( \int_t^\infty w(x){\mathrm{\,d}x}\right)^{\frac rp}w(t) \left( \int_0^t U^{p'}(z,t){v^{1\!-p'}}(z){\mathrm{\,d}z}\right)^{\frac r{p'}}{\mathrm{\,d}t}<\infty$$ and $$A^*_2 := \int_0^\infty \left( \int_t^\infty w(x) {U^q}(t,x) {\mathrm{\,d}x}\right)^{\frac rp}w(t) \sup_{z\in(0,t]} {U^q}(z,t)\left( \int_0^z {v^{1\!-p'}}(s){\mathrm{\,d}s}\right)^{\frac r{p'}}{\mathrm{\,d}t}<\infty$$ are satisfied. Moreover, if $C$ is the least constant such that holds for all functions $f\in{\mathscr M_+}$, then $$C^r \approx A^*_1 + A^*_2.$$ \[144\*\] Let $0<q<1=p$ and $0<{\vartheta}<\infty$. Let $v$, $w$ be weights. Let $U$ be a ${\vartheta}$-regular kernel. Then the following assertions are equivalent: 1. There exists a constant $C\in(0,\infty)$ such that the inequality holds for all functions $f\in{\mathscr M_+}$. 2. Both the conditions $$A^*_3 := \int_0^\infty \left( \int_t^\infty w(x) {\mathrm{\,d}x}\right)^{-q'} w(t)\ {\operatornamewithlimits{ess\, sup\,}}_{z\in(0,t)} U^{-q'}(z,t)\, v^{q'}\!(z)\, {\mathrm{\,d}t}<\infty$$ and $$A^*_4 := \int_0^\infty \left( \int_t^\infty w(x) {U^q}(t,x) {\mathrm{\,d}x}\right)^{-q'} w(t)\ {\operatornamewithlimits{ess\, sup\,}}_{z\in(0,t)} {U^q}(z,t)\, v^{q'}\!(z) \, {\mathrm{\,d}t}<\infty$$ are satisfied. Moreover, if $C$ is the least constant such that holds for all functions $f\in{\mathscr M_+}$, then $$C^{-q'} \approx A^*_3 + A^*_4.$$ The next part contains the proofs. The core components of the discretization method used in this article are summarized in Theorem \[9\] below. It is presented separately for the purpose of possible future reference since this particular variant of discretization may be used even in other problems (cf. [@K6]). Throughout the text, parentheses are used in expressions that involve indices, producing symbols such as $t_{(\!k\!+\!1\!)},$ $t_{k_{(\!n\!+\!1\!)}}$, etc. The parentheses do not have a special meaning, i.e. $t_{(\!k\!+\!1\!)}$ simply means $t$ with the index $k+1$. They are used to make it easier to distinguish between objects as $t_{k_{(\!n\!+\!1\!)}}$ and $t_{(k_n+1)}$, which, in general, are different and both of them appear frequently in the formulas. \[9\] Let $0<q<\infty$ and $1\le{\vartheta}<\infty$. Define $$\Theta := 2 {\vartheta}^{q}.$$ Let $U$ be a ${\vartheta}$-regular kernel. Let $K\in{\mathbb{Z}}$ and $\mu\in{\mathbb{Z}}$ be such that $\mu\le K-2$. Define the index set $$\label{23} {{\mathbb{Z}_\mu}}:= \{k\in{\mathbb{Z}};\ \mu\le k \le K-1 \}.$$ Let $w$ be a weight such that $\int_0^\infty w = \Theta^K$. Let $\{t_k\}_{k=-\infty}^{K}\subset (0,\infty]$ be a sequence of points such that $$\label{17} \int_0^{t_k} w(x){\mathrm{\,d}x}= {\Theta^k}$$ for all $k\in{\mathbb{Z}}$ such that $k\le K$ and $t_K=\infty$. For all $k\in{\mathbb{Z}}$ such that $k\le K-1$, denote $${{\Delta_k}}:= [t_k, t_{(\!k\!+\!1\!)})$$ and $$U({{\Delta_k}}) := U\left(t_k, t_{(\!k\!+\!1\!)}\right).$$ Then there exist a number $N\in{\mathbb{N}}$ and an index set $\{k_n\}_{n=0}^N\subset {{\mathbb{Z}_\mu}}$ with the following properties. - It holds $k_0=\mu$ and $k_{(\!n\!+\!1\!)}=K$. Whenever $n\in\{0,\ldots,N\}$, then $k_n + 1 \le k_{(\!n\!+\!1\!)}$ and therefore also $$\label{10} t_{(k_n + 1)} \le t_{k_{(\!n\!+\!1\!)}}.$$ If we define $$\label{20} {\mathbb{A}}:= \{ n\in{\mathbb{N}};\ n\le N,\ k_n + 1 < k_{(\!n\!+\!1\!)} \},$$ then $$\label{11} {{\mathbb{Z}_\mu}}= \{ k_{(\!n\!+\!1\!)}\!-1;\ n\in{\mathbb{N}}\cup\{0\},\ n\le N \} \cup \{ k;\ k\in{\mathbb{Z}},\ n\in{\mathbb{A}},\ k_n\le k \le k_{(\!n\!+\!1\!)}\!-2 \}.$$ - For every $n\in{\mathbb{N}}$ such that $n\le N-1$ it holds $$\label{12} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{U^q(\Delta_k)}\ge \Theta \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^k}{U^q(\Delta_k)}$$ and $$\label{13} \sum_{k=\mu}^{k_n-1} {\Theta^k}{U^q(\Delta_k)}\le \frac{\Theta}{\Theta-1} \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^k}{U^q(\Delta_k)}.$$ - For every $n\in{\mathbb{A}}$ it holds $$\label{14} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^k}{U^q(\Delta_k)}< \Theta \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^k}{U^q(\Delta_k)}.$$ - For every $n\in{\mathbb{N}}$, $k\in{{\mathbb{Z}_\mu}}$ and $t\in (0,\infty]$ such that $n\le N$, $k\le k_{(\!n\!+\!1\!)}\!-1$ and $t\in(t_k,t_{(\!k\!+\!1\!)}]$ it holds $$\label{15} \int_{t_\mu}^t w(x) {U^q(x,t)}{\mathrm{\,d}x}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}+ {\Theta^k}{U^q(t_k,t)}.$$ If the same conditions hold and it is even satisfied that $k\le k_{(\!n\!+\!1\!)}\!-2$, then $$\label{57} \int_{t_\mu}^t w(x) {U^q(x,t)}{\mathrm{\,d}x}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}.$$ - Define $k_{(-1)}:=\mu-1$. Then for every $n\in{\mathbb{N}}$ such that $n\le N$ it holds $$\label{16} \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}\lesssim \int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}.$$ At first, observe that it is indeed possible to choose the sequence $\{t_k\}$ with the required properties because the weight $w$ is locally integrable. Since $w$ may take zero values, the sequence $\{t_k\}$ need not be unique. In that case, we choose one fixed $\{t_k\}$ satisfying the requirements. From we deduce that $$\label{18} {\Theta^k}= \int_0^{t_k}w(s){\mathrm{\,d}s}= \frac1{\Theta-1} \int_{{\Delta_k}}w(s){\mathrm{\,d}s}= \frac{\Theta}{\Theta-1} \int_{\Delta_{(k-1)}}\!\!w(s){\mathrm{\,d}s}$$ for all $k\in{\mathbb{Z}}$ such that $k\le K-1$. We proceed with the construction of the index subset $\{k_n\}$. Define $k_0:=\mu$ and $k_1:=\mu+1$ and continue inductively as follows.\ ($\ast$) Let $k_0,\ldots,k_n$ be already defined. Then - If $k_n=K$, define $N:=n-1$ and stop the procedure. - If $k_n<K$ and there exists an index $j$ such that $k_n<j\le K$ and $$\label{19} \sum_{k=k_n}^{j-1} {\Theta^k}{U^q(\Delta_k)}\ge \Theta \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^k}{U^q(\Delta_k)},$$ then define $k_{(\!n\!+\!1\!)}$ as the smallest index $j$ for which holds. Then proceed again with step ($\ast$) with $n+1$ in place of $n$. - If $k_n<K$ and and holds for no index $j$ such that $k_n<j\le K$, then define $N:=n$, $k_{(\!n\!+\!1\!)}:=K$ and stop the procedure. In this manner, one obtains a finite sequence of indices $\{k_0,\ldots,k_N\}\subseteq{{\mathbb{Z}_\mu}}$ and the final index $k_{(\!n\!+\!1\!)}=K$. We will call each interval ${{\Delta_k}}$ the *$k$-th segment*, and each interval $[t_{k_n},t_{(k_n+1)})$ the *$n$-th block*. If $n\in{\mathbb{N}}$ is such that $n\le N$, then the $n$-th block either consists of the single $k_n$-th segment, in which case it holds $$k_{(\!n\!+\!1\!)}=k_{n}+1,$$ or the $n$-th segment contains more than one segment and then $$k_{(\!n\!+\!1\!)}>k_{n}+1,$$ If the $n$-th block is of the second type, then $n\in{\mathbb{A}}$, according to the definition . Hence, is satisfied, even though the set ${\mathbb{A}}$ may be empty. The relation in plain words says that each segment is either the last one (i.e., with the highest index $k$) in a block, or it belongs to a block consisting of more than one segment and the investigated segment is not the last one of those. We have now proved (i). The property follows directly from the construction. If $n\in{\mathbb{N}}$ is such that $n\le N$, then by iterating one gets $$\sum_{k=\mu}^{k_n-1} {\Theta^k}{U^q(\Delta_k)}= \sum_{i=0}^{n-1} \sum_{k=k_i}^{k_{(\!i\!+\!1\!)}-1} \!\! {\Theta^k}{U^q(\Delta_k)}\le \sum_{i=0}^{n-1} \Theta^{i-n+1} \!\!\!\! \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} \!\! {\Theta^k}{U^q(\Delta_k)}\le \frac{\Theta}{\Theta\!-\!1} \sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} \!\! {\Theta^k}{U^q(\Delta_k)}.$$ Hence, holds and (ii) is then proved. Property (iii) is again a direct consequence of the way the blocks were constructed. We proceed with proving (iv). Let $n\in{\mathbb{N}}$, $k\in{{\mathbb{Z}_\mu}}$ and $t\in (0,\infty]$ be such that $n\le N$, $k\le k_{(\!n\!+\!1\!)}\!-1$ and $t\in(t_k,t_{(\!k\!+\!1\!)}]$. Then the following sequence of inequalities is valid: $$\begin{aligned} \int_{t_\mu}^t w(x) {U^q(x,t)}{\mathrm{\,d}x}& = \int_{t_\mu}^{t_k} w(x) {U^q(x,t)}{\mathrm{\,d}x}+ \int_{t_k}^t w(x) {U^q(x,t)}{\mathrm{\,d}x}\nonumber\\ & \lesssim \int_{t_\mu}^{t_k} w(x) {U^q(x,t_k)}{\mathrm{\,d}x}+ \int_{t_\mu}^{t_k} w(x){\mathrm{\,d}x}\ {U^q(t_k,t)}+ \int_{t_k}^t w(x) {U^q(x,t)}{\mathrm{\,d}x}\nonumber\\ & \le \sum_{j=\mu}^{k-1} \int_{{{\Delta_j}}} w(x){\mathrm{\,d}x}\ U^q(t_j,t_k) + \int_{t_\mu}^{t_{(\!k\!+\!1\!)}} w(x){\mathrm{\,d}x}\ {U^q(t_k,t)}\nonumber\\ & \lesssim \sum_{j=\mu}^{k-1} {\Theta^j}U^q(t_j,t_k) + {\Theta^k}{U^q(t_k,t)}\label{21}\\ & \lesssim \sum_{j=\mu}^{k-1} {\Theta^j}{U^q(\Delta_j)}+ {\Theta^k}{U^q(t_k,t)}. \label{22} \end{aligned}$$ In here, step follows by , and step by Proposition \[4\]. If $k\le k_n$, then $$\sum_{j=\mu}^{k-1} {\Theta^j}{U^q(\Delta_j)}\le \sum_{j=\mu}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}.$$ The second inequality here follows by . If $k>k_n$, then $n\in{\mathbb{A}}$, $k_n+1\le k \le k_{(\!n\!+\!1\!)}\!-1$ and it holds $$\sum_{j=\mu}^{k-1} {\Theta^j}{U^q(\Delta_j)}\le \sum_{j=\mu}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^j}{U^q(\Delta_j)}= \sum_{j=\mu}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}+ \sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^j}{U^q(\Delta_j)}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}.$$ The last inequality is granted by and . We have proved that $$\sum_{j=\mu}^{k-1} {\Theta^j}{U^q(\Delta_j)}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}.$$ Applying this in the inequality obtained at , we get the estimate . If we now add the assumption $k\le k_{(\!n\!+\!1\!)}\!-2$, then still holds and, in addition to that, we get $${\Theta^k}{U^q(t_k,t)}\le {\Theta^k}{U^q(\Delta_k)}\le \sum_{j=\mu}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^j}{U^q(\Delta_j)}\lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}.$$ In here, the last inequality follows from and . Applying this result to , we obtain and (iv) is thus proved. To prove (v), let $n\in{\mathbb{N}}$ be such that $n\le N$ and observe the following: $$\begin{aligned} \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^j}{U^q(\Delta_j)}& \lesssim \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1}\ \int_{\Delta_{j-1}} w(t){\mathrm{\,d}t}\ {U^q(\Delta_j)}\le \sum_{j=k_{(\!n\!-\!1\!)}}^{k_n-1}\ \int_{\Delta_{j-1}} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}\\ & = \int_{t_{(\!k_{(\!n\!-\!1\!)}-1\!)}}^{t_{(k_n-1)}} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}\le \int_{t_{ k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}. \end{aligned}$$ In the first step, was used. In the last one, we used the inequality $t_{ k_{(\!n\!-\!2\!)}} \le t_{(\! k_{(\!n\!-\!1\!)}-1 \!)}$ which follows from . Without loss of generality, we may assume that ${\vartheta}\in[1,\infty)$. Indeed, if the kernel $U$ is ${\vartheta}$-regular with ${\vartheta}\in(0,1)$, then $U$ is obviously also $1$-regular.\ “(ii)$\Rightarrow$(i)”. Assume that $D_1<\infty$ and $D_2<\infty$. Let us prove that holds for all $f\in{\mathscr M_+}$ with the least constant $C$ satisfying $C^r \lesssim D_1 + D_2$. At first, let us assume that there exists $K\in{\mathbb{Z}}$ such that $\int_0^\infty w = 2^K$. Let $\mu\in{\mathbb{Z}}$ be such that $\mu\le K-2$ and define ${{\mathbb{Z}_\mu}}$ by . Let $\{t_k\}_{k=-\infty}^{K}\subset (0,\infty]$ be a sequence of points such that $t_K=\infty$ and holds for all $k\in{\mathbb{Z}}$ such that $k\le K$. Let $\{k_n\}_{n=0}^N\subset{{\mathbb{Z}_\mu}}$ be the subsequence of indices granted by Theorem \[9\]. Related notation from Theorem \[9\] will be used in what follows as well. Suppose that $f\in {\mathscr M_+}\cap L^p(v)$. Then [\[bcounter51\]]{}[\[bcounter52\]]{}[\[bcounter53\]]{} $$\begin{aligned} \int_{{{t_\mu}}}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}& = \sum_{k\in{{\mathbb{Z}_\mu}}} \int_{{{\Delta_k}}} {\left(}\int_t^\infty f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}\nonumber\\ & \lesssim \sum_{k\in{{\mathbb{Z}_\mu}}} {\Theta^k}{\left(}\int_{{{t_k}}}^\infty f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \label{1080}\\ & \lesssim \sum_{n=0}^N \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \quad + \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^N \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \quad + \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{t_k}},t_{k_{(\!n\!+\!1\!)}}){\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \quad + \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x)U(t_{k_{(\!n\!+\!1\!)}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & =: {B_{\ref{bcounter51}}} + {B_{\ref{bcounter52}}} + {B_{\ref{bcounter53}}}.\nonumber \end{aligned}$$ Inequality follows from . Furthermore, we have [\[bcounter54\]]{}[\[bcounter55\]]{} $$\begin{aligned} {B_{\ref{bcounter51}}} & = \sum_{n=0}^N \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}\!-1} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \! {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}\!-1} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \! {\Theta^k}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-10pt} f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \quad + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \!\! f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}\!-1} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-18pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q + \sum_{n\in{\mathbb{A}}} \! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \! {\Theta^k}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-18pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \quad + \sum_{n\in{\mathbb{A}}} \! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}){\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-20pt} f(x){\mathrm{\,d}x}{\right)}^q + \sum_{n\in{\mathbb{A}}} \! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{\left(}\int_{{{t_k}}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \!\!\! f(x)U({{t_k}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-20pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q + \sum_{n\in{\mathbb{A}}} \! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} \!\!\! f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & =: {B_{\ref{bcounter54}}} + {B_{\ref{bcounter55}}}.\nonumber \end{aligned}$$ For the role of the symbol ${\mathbb{A}}$, see . In the next step, for formal reasons define $t_{(k_{(\!N+2)}-1)}:=\infty$. Then we get $$\begin{aligned} {B_{\ref{bcounter54}}} & = \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-12pt} f(x)U(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \le \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-12pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x) {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\frac{q}{p'}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} f^p(x)v(x){\mathrm{\,d}x}{\right)}^{\frac qp} \label{109}\\ & \le {\left(}\sum_{n=0}^N \Theta^{{\frac rq}k_{(\!n\!+\!1\!)}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-17pt} {U^{p'}}\!(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x) {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac{q}r} \! {\left(}\sum_{n=0}^N \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-17pt} f^p(x)v(x){\mathrm{\,d}x}{\right)}^{\frac qp} \label{110}\\ & \le {\left(}\sum_{n=0}^N \Theta^{{\frac rq}k_{(\!n\!+\!1\!)}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x) {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac{r}{p'}} {\right)}^{\frac{q}r} \|f\|_{L^p(v)}^q \nonumber\\ & \lesssim {\left(}\sum_{n=0}^N {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}-2\!)}} \hspace{-10pt} w(x){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x) {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac{q}r} \|f\|_{L^p(v)}^q \label{111}\\ & \le {\left(}\sum_{n=0}^N {\left(}\int_{t_{(\!k_n\!-\!1\!)}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-5pt} w(x){\mathrm{\,d}x}{\right)}^{\!{\frac rq}} {\left(}\int_{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}^{t_{(\!k_{(\!n\!+\!2\!)}-\!1\!)}} {U^{p'}}\!\!(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},x) {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac{q}r} \|f\|_{L^p(v)}^q \label{112}\\ & \le D_1^{\frac qr} \|f\|^q_{L^p(v)}.\nonumber \end{aligned}$$ The Hölder inequality for functions was used in , and its discrete version (see Proposition \[108\]) was used in . Step follows from . In we used the inequalities $t_{(k_n\!-1)} \le t_{(k_{(\!n\!+\!1\!)}-2)}$ and $t_{k_{(\!n\!+\!1\!)}} \le t_{(k_{(\!n\!+\!2\!)}-1)}$ which hold for all $n\in\{0,\ldots,N\}$ and both follow from or the additional formal definition in the case $n=N$. Step ensures that the sequence $\{t_{(k_n-1)}\}_{n=0}^N$ can be extended into a covering sequence (formally, $\{t_{(k_n-1)}\}_{n=0}^N$ itself is not a covering sequence since $t_{(k_0\!-1)}=t_{(\mu\!-1\!)}>0$). Regarding the term ${B_{\ref{bcounter55}}}$, one has $$\begin{aligned} {B_{\ref{bcounter55}}} & = \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \le \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{\Delta_k}}) {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! f(x){\mathrm{\,d}x}{\right)}^q \label{113}\\ & \le \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{\Delta_k}}) {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^\frac{q}{p'} {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! f^p(x)v(x){\mathrm{\,d}x}{\right)}^\frac{q}{p} \label{115}\\ & \le {\left(}\sum_{n\in{\mathbb{A}}} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \!\!\! {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\!{\frac rq}} \!{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac qr} {\left(}\sum_{n\in{\mathbb{A}}} \int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! f^p(x)v(x){\mathrm{\,d}x}{\right)}^\frac{q}{p} \label{116}\\ & \le {\left(}\sum_{n\in{\mathbb{A}}} {\left(}\int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}{\right)}^{\frac rq}\!{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac qr} \|f\|_{L^p(v)}^q \label{117}\\ & \le D_2^{\frac qr} \|f\|_{L^p(v)}^q. \nonumber \end{aligned}$$ Inequality follows from Proposition \[4\]. In steps and we used the appropriate versions of the Hölder inequality, cf. Propositions \[107\] and \[108\]. Inequalities and give the estimate . We proved $${B_{\ref{bcounter51}}} \lesssim {B_{\ref{bcounter54}}} + {B_{\ref{bcounter55}}} \lesssim (D_1 + D_2)^\frac qr \|f\|_{L^p(v)}^q.$$ We continue with the term ${B_{\ref{bcounter52}}}$. $$\begin{aligned} {B_{\ref{bcounter52}}} & = \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{t_k}},t_{k_{(\!n\!+\!1\!)}}){\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{\Delta_k}}){\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x){\mathrm{\,d}x}{\right)}^q \label{118}\\ & = \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{\Delta_k}}){\left(}\sum_{i=n+1}^N \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-2pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{\Delta_k}}){\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} f(x){\mathrm{\,d}x}{\right)}^q \label{119}\\ & \le \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{\Delta_k}}){\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^\frac{q}{p'} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} f^p(x)v(x) {\mathrm{\,d}x}{\right)}^\frac{q}{p} \label{120}\\ & \le {\left(}\sum_{n=0}^{N-1} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} \!\! {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\!{\frac rq}} \! {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac qr} \! {\left(}\sum_{n=0}^{N-1} \int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} f^p(x)v(x) {\mathrm{\,d}x}{\right)}^{\!\frac{q}{p}} \label{121}\\ & \lesssim {\left(}\sum_{n=0}^{N-1} {\left(}\int_{t_{k_{(\!n\!-\!1\!)}}}^{t_{k_{(\!n\!+\!1\!)}}} w(t) U^q(t,t_{k_{(\!n\!+\!1\!)}}) {\mathrm{\,d}t}{\right)}^{\!{\frac rq}} \! {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{r}{p'}} {\right)}^{\!\frac qr} \|f\|_{L^p(v)}^q \label{122}\\ & \le D_2^{\frac qr} \|f\|_{L^p(v)}^q. \nonumber \end{aligned}$$ Step follows from Proposition \[4\]. Proposition \[3\] supplied with gives . In and we used the Hölder inequality (see Propositions \[107\] and \[108\]). To get , one uses . We obtained $${B_{\ref{bcounter52}}} \lesssim D_2^\frac qr \|f\|_{L^p(v)}^q.$$ In what follows, without loss of generality we will assume that $N\ge 2$. If $N=1$, then the terms involving $\sum_{j=0}^{N-2}$ (or similar) are simply not present in the calculations below. The term ${B_{\ref{bcounter53}}}$ is treated as follows. [\[bcounter56\]]{}[\[bcounter57\]]{} $$\begin{aligned} {B_{\ref{bcounter53}}} & = \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x)U(t_{k_{(\!n\!+\!1\!)}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \hspace{-8pt} f(x)U(t_{k_{(\!n\!+\!1\!)}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & = \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\sum_{i=n\!+\!1}^{N} \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-2pt} f(x)U(t_{k_{(\!n\!+\!1\!)}},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\sum_{i=n\!+\!1}^{N} \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x)U(t_{k_i},x){\mathrm{\,d}x}{\right)}^q \! + \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\sum_{i=n\!+\!2}^{N} U(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) \! \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & =: {B_{\ref{bcounter56}}} + {B_{\ref{bcounter57}}}.\nonumber \end{aligned}$$ Furthermore, it holds $$\begin{aligned} {B_{\ref{bcounter56}}} & = \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\sum_{i=n\!+\!1}^{N} \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x)U(t_{k_i},x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} f(x)U(t_{k_{(\!n\!+\!1\!)}},x){\mathrm{\,d}x}{\right)}^q \label{123}\\ & \le \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} U(t_{k_{(\!n\!+\!1\!)}},x){v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\frac q{p'}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} f^p(x)v(x) {\mathrm{\,d}x}{\right)}^\frac{q}{p} \label{124}\\ & \le {\left(}\sum_{n=0}^{N-1} \Theta^{{\frac rq}k_{(\!n\!+\!1\!)}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} U(t_{k_{(\!n\!+\!1\!)}},x) {v^{1\!-p'}}\!(x) {\mathrm{\,d}x}{\right)}^{\frac r{p'}} {\right)}^{\frac qr} {\left(}\sum_{n=0}^{N-1} \int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} f^p(x)v(x) {\mathrm{\,d}x}{\right)}^\frac{q}{p} \label{125}\\ & \lesssim {\left(}\sum_{n=0}^{N-1} {\left(}\int_{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} w(x) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} U(t_{k_{(\!n\!+\!1\!)}},x) {v^{1\!-p'}}\!(x) {\mathrm{\,d}x}{\right)}^{\frac r{p'}} {\right)}^{\frac qr} \|f\|_{L^p(v)}^q \label{126}\\ & \le {\left(}\sum_{n=0}^{N-1} {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} w(x) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-3pt} U(t_{k_{(\!n\!+\!1\!)}},x) {v^{1\!-p'}}\!(x) {\mathrm{\,d}x}{\right)}^{\frac r{p'}} {\right)}^{\frac qr} \|f\|_{L^p(v)}^q \label{127}\\ & \le D_1^{\frac qr} \|f\|_{L^p(v)}^q.\nonumber \end{aligned}$$ Step follows by Proposition \[3\]. As usual, in and we used the Hölder inequality. The inequality is granted by , and is a consequence of . Next, for the term ${B_{\ref{bcounter57}}}$ we have $$\begin{aligned} {B_{\ref{bcounter57}}} & = \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}} {\left(}\sum_{i=n\!+\!2}^{N} U(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) \! \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \le \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}} \sum_{i=n\!+\!2}^{N} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \label{128}\\ & \le \sum_{i=2}^{N} \sum_{n=0}^{i-2} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \le \sum_{i=2}^{N} \sum_{k=\mu}^{k_i-1} {\Theta^k}{U^q}({{t_k}},t_{k_i}) {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \nonumber\\ & \lesssim \sum_{i=2}^{N} \sum_{k=\mu}^{k_i-1} {\Theta^k}{U^q(\Delta_k)}{\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f(x){\mathrm{\,d}x}{\right)}^q \label{1280}\\ & \le \sum_{i=2}^{N} \sum_{k=\mu}^{k_i-1} {\Theta^k}{U^q(\Delta_k)}{\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\frac q{p'}} {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f^p(x)v(x){\mathrm{\,d}x}{\right)}^{\frac qp} \label{129}\\ & \le {\left(}\sum_{i=2}^{N} {\left(}\sum_{k=\mu}^{k_i-1} {\Theta^k}{U^q(\Delta_k)}{\right)}^{{\frac rq}} {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}} {\right)}^{\frac qr} {\left(}\sum_{n=0}^{N-1} \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} f^p(x)v(x){\mathrm{\,d}x}{\right)}^{\frac qp} \label{130}\\ & \lesssim {\left(}\sum_{i=2}^{N} {\left(}\int_{t_{k_{(\!i\!-\!2\!)}}}^{t_{k_i}} w(t) U^q(t,t_{k_i}) {\mathrm{\,d}t}{\right)}^{\!{\frac rq}} {\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \hspace{-3pt} {v^{1\!-p'}}\!(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}} {\right)}^{\frac qr} \|f\|_{L^p(v)}^q \label{131}\\ & \le D_2^{\frac qr} \|f\|_{L^p(v)}^q.\nonumber \end{aligned}$$ Inequality follows from concavity of the $q$-th power for $q<1$. In one uses Proposition \[4\]. The Hölder inequality gives and . Estimate follows from and . We proved $${B_{\ref{bcounter53}}} \lesssim {B_{\ref{bcounter56}}} + {B_{\ref{bcounter57}}} \lesssim (D_1 + D_2)^\frac qr \|f\|_{L^p(v)}^q.$$ Combined with the other estimates of ${B_{\ref{bcounter51}}}$ and ${B_{\ref{bcounter52}}}$, this yields $$\int_{{{t_\mu}}}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}\lesssim (D_1 + D_2)^\frac qr \|f\|_{L^p(v)}^q.$$ Observe that the constant related to the symbol “$\lesssim$” in here does not depend on the choice of $\mu$. The reader may nevertheless notice that the construction of the $n$-blocks in fact depends on $\mu$. However, the constants in the “$\lesssim$”-estimates proved with help of that construction are indeed independent of $\mu$. Hence, we may perform the limit pass $\mu\to -\infty$. Since ${{t_\mu}}\to 0$ as $\mu\to -\infty$, the monotone convergence theorem (and taking the $q$-th root) yields $${\left(}\int_{0}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\lesssim (D_1 + D_2)^\frac 1r \|f\|_{L^p(v)}$$ for the fixed function $f\in {\mathscr M_+}\cap L^p(v)$. Since the function $f$ was chosen arbitrarily and the constant represented in “$\lesssim$” does not depend on $f$, the inequality holds with $C=(D_1 + D_2)^\frac 1r$ for all functions $f\in {\mathscr M_+}$. Clearly, if $C$ is the least constant such that holds for all $f\in{\mathscr M_+}$, then $$\label{132} C^r \lesssim D_1 + D_2.$$ At this point, recall that so far we have assumed that $\int_0^\infty w(x){\mathrm{\,d}x}= \Theta^K$ for a $K\in{\mathbb{Z}}$. Let us now complete the proof of this part for a general weight $w$. At first, if $\int_0^\infty w(x){\mathrm{\,d}x}$ is finite but not equal to any integer power of $\Theta$, the result is simply obtained by multiplying $w$ by a constant $c\in(1,2)$ such that $\int_0^\infty cw(x){\mathrm{\,d}x}= \Theta^K$ for a $K\in{\mathbb{Z}}$, and then using homogeneity of the expressions $\int_{0}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}$, $D^{\frac qr}_1$ and $D^{\frac qr}_2$ with respect to $w$. Finally, let us suppose that $\int_0^\infty w(x){\mathrm{\,d}x}=\infty$. Choose an arbitrary function $f\in {\mathscr M_+}\cap L^p(v)$. For each $m\in{\mathbb{N}}$ define $w_m:=w\chi_{[0,m]}$ and denote by $D_{1,m}$ the expression $D_1$ with $w$ replaced by $w_m$. Similarly we define $D_{2,m}$. Since the weight $w$ is locally integrable, for each $m\in{\mathbb{N}}$ it holds $\int_0^\infty w_m(x){\mathrm{\,d}x}<\infty$. Hence, by the previous part of the proof we get $${\left(}\int_{0}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w_m(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\lesssim (D_{1,m} + D_{2,m})^\frac 1r \|f\|_{L^p(v)}.$$ Obviously, for all $m\in{\mathbb{N}}$ it holds $w_m \le w$ pointwise, hence $D_{1,m} \le D_1$ and $D_{2,m} \le D_2$. Thus, we get $${\left(}\int_{0}^\infty \! {\left(}\int_t^\infty \!\!\! f(x)U(t,x){\mathrm{\,d}x}{\right)}^q w_m(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\lesssim (D_1 + D_2)^\frac 1r \|f\|_{L^p(v)}.$$ The constant in “$\lesssim$” does not depend on $m$ or $f$ and the latter was arbitrarily chosen. Since $w_m\uparrow w$ pointwise as $m\to\infty$, the monotone convergence theorem (for $m\to\infty$) yields that holds for all functions $f\in {\mathscr M_+}$ and the best constant $C$ in satisfies . The proof of this part is now complete.\ “(i)$\Rightarrow$(ii)”. Suppose that holds for all $f\in {\mathscr M_+}$ and $C\in(0,\infty)$ is the least constant such that this is true. We need to show that $D_1 + D_2 \lesssim C^r$. Let $\{t_k\}_{k\in{\mathbb{I}}}$ be a covering sequence indexed by a set ${\mathbb{I}}=\{k_{\min},\ldots,k_{\max}\}\subset{\mathbb{Z}}$. At first, let us show that $$\label{133} \int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}< \infty \quad \textnormal{for\ all\ }k\in{\mathbb{I}}_0.$$ Suppose, for a contradiction, that $k\in{\mathbb{I}}_0$ and $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}t}= \infty$. Then, by Proposition \[107\], for every $M\in{\mathbb{N}}$ there exists a function $g_M$ supported in $[{{{t_k}}},{t_{(\!k\!+\!1\!)}}]$ and such that $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_M^p(t)v(t){\mathrm{\,d}t}= 1$ and $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_M(x)U({{t_k}},x){\mathrm{\,d}x}> M$. Since ${{t_k}}>0$, by definition of a weight it holds $\int_0^{{{t_k}}} w(t){\mathrm{\,d}t}>0$. Thus, for every $M\in{\mathbb{N}}$ one gets $$\begin{aligned} {\left(}\int_0^\infty {\left(}\int_t^\infty g_M(x)U(t,x){\mathrm{\,d}x}{\right)}^q w(t){\mathrm{\,d}t}{\right)}^{\frac1q}& \ge {\left(}\int_0^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac1q}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_M(x)U({{t_k}},x){\mathrm{\,d}x}\\ & > M {\left(}\int_0^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac1q}\\ & = M {\left(}\int_0^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac1q}\|g_M\|_{L^p(v)}, \end{aligned}$$ which contradicts . Hence, must be satisfied. Since $\{t_k\}$ was chosen arbitrarily, together with local integrability of $w$ is in fact sufficient to prove that $D_1<\infty$. However, we aim to prove a stronger assertion, namely that $D_1\lesssim C^r$. To do so, we proceed as follows. Having verified , for each $k\in{\mathbb{I}}_0$ we may use Proposition \[107\] to find a measurable function $g_k$ supported in $[{{t_k}}, t_{(\!k\!+\!1\!)}]$ and such that $\|g_k\|_{L^p(v)}=1$ as well as $$\label{134} {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac 1{p'}} \le 2 \int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_k(x) U(t_k,x){\mathrm{\,d}x}.$$ Furthermore, it holds $$\sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{{\frac r{p'}}} <\infty$$ since $w$ is locally integrable, holds and ${\mathbb{I}}_0$ consists of a finite number of indices. Hence, by Proposition \[108\] we can find a nonnegative sequence $\{c_k\}_{k\in{\mathbb{I}}_0}$ such that $\sum_{k\in{\mathbb{I}}_0} c_k^p = 1$ and $$\begin{gathered} {\left(}\sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{{\frac r{p'}}} {\right)}^{{\frac1r}} \\ \le 2 {\left(}\sum_{k\in{\mathbb{I}}_0} c_k^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac{q}{p'}} {\right)}^{\frac1q}. \label{135} \end{gathered}$$ Define a function $g:=\sum_{k\in{\mathbb{I}}_0} c_k g_k$ and recall that each $g_k$ is supported in $[{{t_k}}, t_{(\!k\!+\!1\!)}]$. Hence, $$\label{136} \|g\|_{L^p(v)} = {\left(}\sum_{k\in{\mathbb{I}}_0} c_k^p \|g_k\|_{L^p(v)}^p {\right)}^{\frac1p}= {\left(}\sum_{k\in{\mathbb{I}}_0} c_k^p {\right)}^{\frac1p}= 1.$$ Finally, we get the following estimate. $$\begin{aligned} & \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{{\frac r{p'}}} \nonumber\\ & \lesssim {\left(}\sum_{k\in{\mathbb{I}}_0} c_k^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},x){v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac{q}{p'}} {\right)}^{\frac rq}\label{137}\\ & \lesssim {\left(}\sum_{k\in{\mathbb{I}}_0} c_k^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} U({{t_k}},x) g_k(x){\mathrm{\,d}x}{\right)}^q {\right)}^{\frac rq}\label{138}\\ & = {\left(}\sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} U({{t_k}},x) g(x){\mathrm{\,d}x}{\right)}^q {\right)}^{\frac rq}\nonumber\\ & \le {\left(}\sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {\left(}\int_{t}^{t_{(\!k\!+\!1\!)}} U(t,x) g(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}{\right)}^{\frac rq}\nonumber\\ & \le {\left(}\int_0^\infty w(t) {\left(}\int_{t}^{\infty} U(t,x) g(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}{\right)}^{\frac rq}\nonumber\\ & \le C^r \|g\|_{L^p(v)}^r \label{139}\\ & = C^r. \label{140} \end{aligned}$$ In steps , , and we used , , and , respectively. Since the covering sequence $\{t_k\}_{k\in{\mathbb{I}}}$ was chosen arbitrarily, by taking supremum over all covering sequences we obtain $$D_1 \lesssim C^r.$$ In what follows, we are going to prove a similar estimate for $D_2$. Again, let $\{t_k\}_{k\in{\mathbb{I}}}$ be a covering sequence indexed by a set ${\mathbb{I}}=\{k_{\min},\ldots,k_{\max}\}\subset{\mathbb{Z}}$. Then it holds $$\label{142} \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}<\infty \quad \textnormal{for\ all\ }k\in{\mathbb{I}}_0,$$ and $$\label{141} \int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}<\infty \quad \textnormal{for\ all\ }k\in{\mathbb{I}}_0.$$ Let us prove these claims. At first, suppose that there exists $k\in{\mathbb{I}}_0$ such that $\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}=\infty$. By definition, the weight $v$ is locally integrable, thus the function $\chi_{[{{t_k}},{{t_k}}+1]}$ belongs to $L^p(v)$. Then $$\begin{aligned} \infty & = \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}\\ & = \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^\infty \chi_{[{{t_k}},{{t_k}}+1]}(x){\mathrm{\,d}x}{\right)}^q \\ & = \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {\left(}\int_{{{t_k}}}^\infty U(t,x) \chi_{[{{t_k}},{{t_k}}+1]}(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}\\ & \le \int_0^\infty w(t) {\left(}\int_t^\infty U(t,x) \chi_{[{{t_k}},{{t_k}}+1]}(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}, \end{aligned}$$ whereas $\|\chi_{[{{t_k}},{{t_k}}+1]}\|_{L^p(v)}<\infty$. This contradicts , hence holds. Next, assume that there exists $k\in{\mathbb{I}}_0$ such that $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}=\infty.$ Then, by Proposition \[107\], for every $M\in{\mathbb{N}}$ there exists a function $g_M$ such that $\|g_M\|_{L^p(v)}=1$ and $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_m(x) {\mathrm{\,d}x}> M$. By the definition of a weight and a ${\vartheta}$-regular kernel, the term $\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}) {\mathrm{\,d}t}$ is strictly positive. We get $$\begin{aligned} \int_0^\infty w(t) {\left(}\int_t^\infty U(t,x) g_M(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}& \ge \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} U(t,x) g_M(x){\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}\\ & \ge \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}) {\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} g_M(x){\mathrm{\,d}x}{\right)}^q \\ & \ge M^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}) {\mathrm{\,d}t}\\ & = M^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}) {\mathrm{\,d}t}\ \|g_M\|_{L^p(v)} \end{aligned}$$ for all $M\in{\mathbb{N}}$. This is a contradiction with . Hence, must be true. Thanks to , Proposition \[107\] yields that for every $k\in{\mathbb{I}}_0$ we can find a function $h_k$ supported in $[{{t_k}}, t_{(\!k\!+\!1\!)}]$ and such that $\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} h_k^p(x)v(x){\mathrm{\,d}x}= 1$ and $${\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}\le 2 \int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} h_k(x){\mathrm{\,d}x}.$$ Furthermore, it holds $$\sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}<\infty,$$ since the sum involves a finite number of terms and each of them is finite due to and . By Proposition \[108\], we may find a nonnegative sequence $\{d_k\}_{k\in{\mathbb{I}}_0}$ such that $\sum_{k\in{\mathbb{I}}_0} d_k^p = 1$ and $$\begin{gathered} {\left(}\sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac r{p'}}{\right)}^{\frac1r}\\ \le 2 {\left(}\sum_{k\in{\mathbb{I}}_0} d_k^q \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^\frac{q}{p'} {\right)}^{\frac1q}. \end{gathered}$$ Define the function $h:=\sum_{k\in{\mathbb{I}}_0} d_k h_k$. Then it is easy to verify that $\|h\|_{L^p(v)}=1$. Moreover, we get the following estimate. $$\begin{aligned} \sum_{k\in{\mathbb{I}}_0} \! {\left(}\int_{t_{\!(\!k\!-\!1\!)}}^{{{t_k}}} \hspace{-8pt} w(t) {U^q}(t,{{t_k}}\!){\mathrm{\,d}t}{\right)}^{\!{\frac rq}} \!\! {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}}\hspace{-8pt} {v^{1\!-p'}}\!\!(x){\mathrm{\,d}x}{\right)}^{\!{\frac r{p'}}} \hspace{-6pt} &\lesssim {\left(}\sum_{k\in{\mathbb{I}}_0} \! d_k^q \!\! \int_{t_{\!(\!k\!-\!1\!)}}^{{{t_k}}} \hspace{-8pt} w(t) {U^q}(t,{{t_k}}\!){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{\!(\!k\!+\!1\!)}} \hspace{-7pt} {v^{1\!-p'}}\!\!(x){\mathrm{\,d}x}{\right)}^{\!\frac{q}{p'}} {\right)}^{\!{\frac rq}} \\ & \lesssim {\left(}\sum_{k\in{\mathbb{I}}_0} \! d_k^q \!\! \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} \hspace{-6pt} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{\!(\!k\!+\!1\!)}}\!\! h_k(x){\mathrm{\,d}x}{\right)}^{\!q} {\right)}^{\!{\frac rq}} \\ & = {\left(}\sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} h(x){\mathrm{\,d}x}{\right)}^q {\right)}^{\frac rq}\\ & \le {\left(}\sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{{{t_k}}} w(t) {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} h(x) U(t,x) {\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}{\right)}^{\frac rq}\\ & \le {\left(}\sum_{k\in{\mathbb{I}}_0} \int_0^\infty w(t) {\left(}\int_{t}^{\infty} h(x) U(t,x) {\mathrm{\,d}x}{\right)}^q {\mathrm{\,d}t}{\right)}^{\frac rq}\\ & \le C^r \|h\|_{L^p(v)}\\ & = C^r. \end{aligned}$$ The covering sequence $\{t_k\}_{k\in{\mathbb{I}}}$ was arbitrarily chosen in the beginning, hence we may take the supremum over all covering sequences, obtaining the relation $$D_2 \lesssim C^r.$$ The proof of the implication “(i)$\Rightarrow$(ii)” and of the related estimates is then finished.\ “(iii)$\Rightarrow$(ii)”. Assume that $A_1<\infty$ and $A_2<\infty$. We will prove the inequality $D_1+D_2\lesssim A_1+A_2$. Let $\{t_k\}_{k\in{\mathbb{I}}}$ be an arbitrary covering sequence indexed by a set ${\mathbb{I}}$. Then it holds $$\begin{aligned} & \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(x){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},t) {v^{1\!-p'}}(t){\mathrm{\,d}t}{\right)}^{\frac r{p'}}\\ &\approx \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{t_{(\!k\!-\!1\!)}}^x w(s){\mathrm{\,d}s}{\right)}^{\frac rp}w(x){\mathrm{\,d}x}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {U^{p'}}({{t_k}},t) {v^{1\!-p'}}(t){\mathrm{\,d}t}{\right)}^{\frac r{p'}}\\ &\le \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{0}^x w(s){\mathrm{\,d}s}{\right)}^{\frac rp}w(x){\mathrm{\,d}x}{\left(}\int_{x}^{\infty} {U^{p'}}(x,t) {v^{1\!-p'}}(t){\mathrm{\,d}t}{\right)}^{\frac r{p'}}\\ &= A_1. \end{aligned}$$ Taking the supremum over all covering sequences, we obtain $D_1\lesssim A_1$. Similarly, for any fixed covering sequence $\{t_k\}_{k\in{\mathbb{I}}}$ we get $$\begin{aligned} & \sum_{k\in{\mathbb{I}}_0} {\left(}\int_{t_{(\!k\!-\!1\!)}}^{t_k} w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\approx \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{t_{(\!k\!-\!1\!)}}^t w(x) {U^q}(x,{{t_k}}){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}\, {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\lesssim \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{t_{(\!k\!-\!1\!)}}^t w(x) {U^q}(x,t){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}\, {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\qquad + \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{t_{(\!k\!-\!1\!)}}^t w(x) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) U^r(t,{{t_k}}) {\mathrm{\,d}t}\, {\left(}\int_{{{t_k}}}^{t_{(\!k\!+\!1\!)}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\le \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{0}^t w(x) {U^q}(x,t){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {U^q}(t,{{t_k}}){\mathrm{\,d}t}\, {\left(}\int_{{{t_k}}}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\qquad + \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{0}^t w(x) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) U^r(t,{{t_k}}) {\mathrm{\,d}t}\, {\left(}\int_{{{t_k}}}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\\ &\le \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{0}^t w(x) {U^q}(x,t){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_{z}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\\ &\qquad + \sum_{k\in{\mathbb{I}}_0} \int_{t_{(\!k\!-\!1\!)}}^{t_k} {\left(}\int_{0}^t w(x) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_t^\infty {U^{p'}}(t,s) {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\\ &= A_2 + A_1. \end{aligned}$$ Once again, taking the supremum over all covering sequences, we get $D_2\lesssim A_2+A_1$. Hence, we have shown that $D_1+D_2\lesssim A_1+A_2$ and the implication “(iii)$\Rightarrow$(ii)” is proved.\ “(ii) $\Rightarrow$ (iii)”. Suppose that $D_1<\infty$ and $D_2<\infty$ and let us show that $A_1+A_2\lesssim D_1+D_2$ then. Similarly as in the proof of “(ii) $\Rightarrow$ (i)”, let us first assume that $\int_0^\infty w = 2^K$ for some $K\in{\mathbb{Z}}$ . Let $\mu\in{\mathbb{Z}}$ be such that $\mu\le K-2$ and define ${{\mathbb{Z}_\mu}}$ by . Let $\{t_k\}_{k=-\infty}^{K}\subset (0,\infty]$ be the sequence of points from Theorem \[9\] and $\{k_n\}_{n=0}^N\subset{{\mathbb{Z}_\mu}}$ be the subsequence of indices granted by the same theorem. Then [\[bcounter1\]]{}[\[bcounter2\]]{}[\[bcounter3\]]{} $$\begin{aligned} & \int_{{{t_\mu}}}^\infty {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_t^\infty {U^{p'}}(t,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & = \sum_{k\in{{\mathbb{Z}_\mu}}} \int_{{{\Delta_k}}} {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_t^\infty {U^{p'}}(t,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & \le \sum_{k\in{{\mathbb{Z}_\mu}}} \int_0^{t_{(\!k\!+\!1\!)}} {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\mathrm{\,d}t}{\left(}\int_{t_k}^\infty {U^{p'}}(t_k,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{k\in{{\mathbb{Z}_\mu}}} {\Theta^\frac{kr}{q}}{\left(}\int_{t_k}^\infty {U^{p'}}(t_k,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{24}\\ & \approx \sum_{n=0}^N \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{\left(}\int_{t_k}^{t_{k_{(\!n\!+\!1\!)}}} {U^{p'}}(t_k,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \qquad + \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{U^r}(t_k,t_{k_{(\!n\!+\!1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \qquad + \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & =: {B_{\ref{bcounter1}}}+{B_{\ref{bcounter2}}}+{B_{\ref{bcounter3}}}\nonumber. \end{aligned}$$ In step we used . We continue by estimating each of the separate terms. [\[bcounter4\]]{}[\[bcounter5\]]{} $$\begin{aligned} {B_{\ref{bcounter1}}}\! & = \sum_{n=0}^N \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{\left(}\int_{t_k}^{t_{k_{(\!n\!+\!1\!)}}} {U^{p'}}(t_k,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\\ & = \sum_{n=0}^N \! \Theta^{(\! k_{(\!n\!+\!1\!)}\!-\! 1 \!){\frac rq}} \! {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-26pt} {U^{p'}}\! (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \hspace{-8pt} + \! \sum_{n\in{\mathbb{A}}} \!\! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \hspace{-8pt} {\Theta^\frac{kr}{q}}\! {\left(}\int_{{{t_k}}}^{t_{k_{(\!n\!+\!1\!)}}} \! {U^{p'}}\!({{t_k}},z){v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \\ & \lesssim \sum_{n=0}^N \! \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} \! {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-26pt} {U^{p'}}\! (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \hspace{-8pt} + \!\sum_{n\in{\mathbb{A}}} \!\! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \hspace{-8pt} {\Theta^\frac{kr}{q}}\! {\left(}\int_{{{t_k}}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \!\! {U^{p'}}\!({{t_k}},z){v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \\ & \qquad + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^\frac{kr}{q}}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-24pt} {U^{p'}}({{t_k}},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\\ & \lesssim \sum_{n=0}^N \! \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} \! {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-26pt} {U^{p'}}\! (t_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)},z){v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \hspace{-8pt} + \! \sum_{n\in{\mathbb{A}}} \!\! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \hspace{-8pt} {\Theta^\frac{kr}{q}}{U^r}\! ({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}) {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-24pt} {v^{1\!-p'}}\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \\ & \qquad + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^\frac{kr}{q}}{U^r}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{{{t_k}}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\\ & \qquad + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^\frac{kr}{q}}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-20pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\\ & \lesssim \sum_{n=0}^N \! \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} \! {\left(}\int_{\Delta_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-27pt} {U^{p'}}\!\! (t_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)},z){v^{1\!-p'}}\!\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \hspace{-8pt} + \! \sum_{n\in{\mathbb{A}}} \!\! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \hspace{-8pt} {\Theta^\frac{kr}{q}}{U^r}\!({{t_k}},t_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}\!) {\left(}\int_{{{t_k}}}^{t_{k_{\!(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \\ & \qquad + \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^\frac{kr}{q}}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-20pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\\ & \lesssim \sum_{n=0}^N \! \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} \! {\left(}\int_{\Delta_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-27pt} {U^{p'}}\!\! (t_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)},z){v^{1\!-p'}}\!\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \hspace{-8pt} + \! \sum_{n\in{\mathbb{A}}} \!\! \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} \hspace{-8pt} {\Theta^\frac{kr}{q}}{U^r}\!({{t_k}},t_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}\!) {\left(}\int_{{{t_k}}}^{t_{k_{\!(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!\!(z){\mathrm{\,d}z}{\right)}^{\!{\frac r{p'}}} \\ & := {B_{\ref{bcounter4}}} + {B_{\ref{bcounter5}}}. \end{aligned}$$ For ${B_{\ref{bcounter4}}}$ we have $$\begin{aligned} {B_{\ref{bcounter4}}} & = \sum_{n=0}^N \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=0}^N {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}} w(x) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{25}\\ & \le \sum_{k\in{{\mathbb{Z}_\mu}}} {\left(}\int_{\Delta_{(\!k-\!1\!)}} w(x) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{{{\Delta_k}}} {U^{p'}}(t_{(\!k-1\!)},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \le D_1. \nonumber \end{aligned}$$ In step we used . Let us formally define $k_{(-1)}:=\mu-1$ and proceed with estimating ${B_{\ref{bcounter5}}}$. $$\begin{aligned} {B_{\ref{bcounter5}}} & = \sum_{n\in{\mathbb{A}}} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^\frac{kr}{q}}{U^r}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \le \sum_{n\in{\mathbb{A}}} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^k}{U^q}({{t_k}},t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{26}\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-2} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{27}\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\sum_{k=k_{(\!n\!-\!1\!)}}^{k_n-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{28}\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(x) {U^q}(x,t_{k_n}) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{29}\\ & \le \sum_{n=1}^{N} {\left(}\int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(x) {U^q}(x,t_{k_n}) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & = \sum_{i=0}^1 \sum_{\substack{{1\le n\le N}\\ {n{\operatorname{\,mod}}2=i}}} {\left(}\int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} w(x) {U^q}(x,t_{k_n}) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{31}\\ & \lesssim D_2.\nonumber \end{aligned}$$ Since ${\frac rq}>1$, the estimate follows by convexity of the ${\frac rq}$-th power. Step is due to Proposition \[4\]. Step then follows by , and step by . Finally, in we split the even and odd indices $n$, so that the intervals $(t_{k_{(\!n\!-\!2\!)}},t_{k_n})$ involved in each $n$-indexed sum do not overlap. This standard step will be also used in other estimates further on. So far we have proved $${B_{\ref{bcounter1}}} \lesssim {B_{\ref{bcounter4}}} + {B_{\ref{bcounter5}}} \lesssim D_1 + D_2.$$ The term ${B_{\ref{bcounter2}}}$ is estimated as follows. $$\begin{aligned} {B_{\ref{bcounter2}}} & = \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{U^r}(t_k,t_{k_{(\!n\!+\!1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \le \sum_{n=0}^{N-1} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{U^q}(t_k,t_{k_{(\!n\!+\!1\!)}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{32}\\ & \lesssim \sum_{n=0}^{N-1} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{33}\\ & = \sum_{n=0}^{N-1} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\sum_{j=n+1}^N \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=0}^{N-1} {\left(}\sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{34}\\ & \lesssim \sum_{n=0}^{N-1} {\left(}\int_{t_{k_{(\!n\!-\!1\!)}}}^{t_{k_{(\!n\!+\!1\!)}}} w(x) {U^q}(x,t_{k_{(\!n\!+\!1\!)}}) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{35}\\ & = \sum_{i=0}^1 \sum_{\substack{{1\le n\le N}\\ {n{\operatorname{\,mod}}2=i}}} {\left(}\int_{t_{k_{(\!n\!-\!1\!)}}}^{t_{k_{(\!n\!+\!1\!)}}} w(x) {U^q}(x,t_{k_{(\!n\!+\!1\!)}}) {\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim D_2. \nonumber \end{aligned}$$ We used convexity of the ${\frac rq}$-th power to get . Step follows by Proposition \[4\]. Inequality is granted by Proposition \[3\] equipped with . Step follows by . We proved $${B_{\ref{bcounter2}}} \lesssim D_2.$$ The term ${B_{\ref{bcounter3}}}$ is first handled in the following way. [\[bcounter6\]]{}[\[bcounter7\]]{} $$\begin{aligned} {B_{\ref{bcounter3}}} & = \sum_{n=0}^{N-1} \sum_{k=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^\frac{kr}{q}}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & = \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\sum_{j=n+1}^N \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\sum_{j=n+2}^N {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},t_{k_j}) \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \quad + \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\sum_{j=n+1}^N \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {U^{p'}}(t_{k_j},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & =: {B_{\ref{bcounter6}}}+{B_{\ref{bcounter7}}}.\nonumber \end{aligned}$$ Then, for ${B_{\ref{bcounter6}}}$ we have $$\begin{aligned} {B_{\ref{bcounter6}}} & = \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\sum_{j=n+2}^N {U^{p'}}(t_{k_{(\!n\!+\!1\!)}},t_{k_j}) \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \le \sum_{n=0}^{N-2} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} \sum_{j=n+2}^N {U^r}(t_{k_{(\!n\!+\!1\!)}},t_{k_j}) {\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{36}\\ & = \sum_{j=2}^N \sum_{n=0}^{j-2} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {U^r}(t_{k_{(\!n\!+\!1\!)}},t_{k_j}) {\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \le \sum_{j=2}^N {\left(}\sum_{n=0}^{j-2} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_j}) {\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{38}\\ & \le \sum_{j=2}^N {\left(}\sum_{k=\mu}^{k_{(\!j\!-\!1\!)}} {\Theta^k}{U^q}(t_k,t_{k_j}) {\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{j=2}^N {\left(}\sum_{k=\mu}^{k_{(\!j\!-\!1\!)}} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{39}\\ & \le \sum_{j=2}^N {\left(}\sum_{k=\mu}^{k_j-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{j=2}^N {\left(}\sum_{k=k_{(\!j\!-\!1\!)}}^{k_j-1} {\Theta^k}{U^q}({{\Delta_k}}) {\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{58}\\ & \lesssim \sum_{j=2}^N {\left(}\int_{t_{k_{(\!j\!-\!2\!)}}}^{t_{k_j}} w(x) {U^q}(x,t_{k_j}){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{40}\\ & = \sum_{i=0}^1 \sum_{\substack{{2\le j\le N}\\ {j{\operatorname{\,mod}}2=i}}} {\left(}\int_{t_{k_{(\!j\!-\!2\!)}}}^{t_{k_j}} w(x) {U^q}(x,t_{k_j}){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim D_2. \nonumber \end{aligned}$$ Inequality follows from concavity of the ${\frac r{p'}}$-th power since ${\frac r{p'}}<1$. Similarly, convexity of the ${\frac rq}$-th power yields . Step is due to Proposition \[4\], step follows by , and in step we used . We continue as follows. $$\begin{aligned} {B_{\ref{bcounter7}}} & = \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\sum_{j=n+1}^N \int_{t_{k_j}}^{t_{k_{(\!j\!+\!1\!)}}} {U^{p'}}(t_{k_j},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=0}^{N-1} \Theta^{k_{(\!n\!+\!1\!)}{\frac rq}} {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} {U^{p'}}(t_{k_j},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{41}\\ & \lesssim \sum_{n=0}^{N-1} {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!1\!)}}} w(x){\mathrm{\,d}x}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} {U^{p'}}(t_{k_j},z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}\label{42}\\ & \le D_1. \nonumber \end{aligned}$$ To get , we used Proposition \[3\], and in we applied . We have proved $${B_{\ref{bcounter3}}} \lesssim {B_{\ref{bcounter6}}}+ {B_{\ref{bcounter7}}} \lesssim D_1 + D_2.$$ Combining all the estimates we have obtained so far, we get $$\label{43} \int_{{{t_\mu}}}^\infty {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_t^\infty {U^{p'}}(t,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\, \lesssim\, D_1 + D_2.$$ In the following part, we are going to perform estimates related to the term $A_2$. We have [\[bcounter8\]]{}[\[bcounter9\]]{} $$\begin{aligned} & \int_{{{t_\mu}}}^\infty {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & = \sum_{n=0}^{N} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & \quad + \sum_{n\in{\mathbb{A}}} \int_{t_{k_n}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & =: {B_{\ref{bcounter8}}}+{B_{\ref{bcounter9}}}. \nonumber \end{aligned}$$ By , the term ${B_{\ref{bcounter8}}}$ is further estimated as follows. [\[bcounter10\]]{}[\[bcounter11\]]{} $$\begin{aligned} {B_{\ref{bcounter8}}} & = \sum_{n=0}^{N} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-10pt} w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\\ & \quad + \sum_{n=0}^{N} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-10pt} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\\ & =: {B_{\ref{bcounter10}}}+{B_{\ref{bcounter11}}}. \end{aligned}$$ Notice that that, in ${B_{\ref{bcounter10}}}$, the term corresponding to $n=0$ is indeed omitted, since for any $t\in\Delta_{\mu}$ it holds $\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}\lesssim \Theta^\mu {U^q}({{t_\mu}},t)$ and the right-hand side is thus already represented by the $0$-th term in ${B_{\ref{bcounter11}}}$. Let us note that in what follows, expressions such as $\sup_{x\in(y,\infty]} {\varphi}(x)$ appear even where the argument ${\varphi}(x)$ is undefined for $x=\infty$. To fix this formal detail, suppose that, in such cases, $\sup_{x\in(y,\infty]} {\varphi}(x)$ is simply redefined as $\sup_{x\in(y,\infty)} {\varphi}(x)$. This will make expressions such as $\sum_{n=1}^{N} \sup_{x\in[t_{k_n},t_{k_{(\!n\!+\!1\!)}}]} {\varphi}(x)$ formally correct without need of treating the $(N\!+\!1)$-st summand separately. Besides this, the standard notation $\overline{\Delta}_k$ is used to denote the closure of ${{\Delta_k}}$, i.e. the interval $[{{t_k}},t_{(\!k\!+\!1\!)}]$. We then estimate ${B_{\ref{bcounter10}}}$. [\[bcounter12\]]{}[\[bcounter13\]]{} $$\begin{aligned} {B_{\ref{bcounter10}}} & = \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-10pt} w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in[t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},\infty)} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{200} \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{201} \\ & \quad + \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-7pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{004} \\ & \quad + \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(n\!+\!1\!)}-\!1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-7pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & =: {B_{\ref{bcounter12}}}+{B_{\ref{bcounter13}}}. \nonumber \end{aligned}$$ Inequality holds by , and is due to Proposition \[59\]. In we used . Next, we have [\[bcounter14\]]{}[\[bcounter15\]]{} $$\begin{aligned} {B_{\ref{bcounter12}}} & = \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \quad + \sum_{n=1}^{N\!-1} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} {U^q}(\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{207} \\ & \quad + \sum_{n=1}^{N\!-1} {\left(}\sum_{j=k_n}^{k_{(n\!+\!1\!)}-\!1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & =: {B_{\ref{bcounter14}}}+{B_{\ref{bcounter15}}}. \nonumber \end{aligned}$$ Step is based on . For each $n\in\{1,\ldots,N\}$ there exists a point $z_{(\!n\!+\!1\!)}\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}$ such that $$\label{202} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!\!(s){\mathrm{\,d}s}{\right)}^{\!{\frac r{p'}}} \!\!\! \le 2 {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-\!1\!)},z_{(\!n\!+\!1\!)}) {\left(}\int_{z_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} \!\! {v^{1\!-p'}}\!\!(s){\mathrm{\,d}s}{\right)}^{\!{\frac r{p'}}}\!\!.$$ Define also $z_{(-1)}:=0$ and $z_{(\!N\!+2\!)}:=\infty$. One then gets $$\begin{aligned} {B_{\ref{bcounter14}}} & = \sum_{n=1}^{N} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-10pt} \sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-7pt} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\Theta^{k_{(\!n\!+\!1\!)}\!-1} {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z_{(\!n\!+\!1\!)}) {\left(}\int_{z_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{203} \\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\!\!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}} \hspace{-15pt} w(t) {\mathrm{\,d}t}\ {U^q}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z_{(\!n\!+\!1\!)}) {\left(}\int_{z_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{204} \\ & \le \sum_{n=1}^{N} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\int_{t_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}}^{z_{(\!n\!+\!1\!)}} w(t) {U^q}(t,z_{(\!n\!+\!1\!)}) {\mathrm{\,d}t}\, {\left(}\int_{z_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \le \sum_{n=1}^{N} {\left(}\int_{t_{k_{(\!n\!-\!2\!)}}}^{t_{k_n}} \hspace{-10pt} w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}{\right)}^{\frac rp}\int_{t_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}}^{z_{(\!n\!+\!1\!)}} \hspace{-10pt} w(t) {U^q}(t,z_{(\!n\!+\!1\!)}) {\mathrm{\,d}t}\, {\left(}\int_{z_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{205} \\ & \le \sum_{n=1}^{N} {\left(}\int_{z_{(\!n\!-\!2\!)}}^{z_{(\!n\!+\!1\!)}} \hspace{-2pt} w(t) U^q(t,{z_{(\!n\!+\!1\!)}}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{z_{(\!n\!+\!1\!)}}^{z_{(\!n\!+\!2\!)}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{206} \\ & = \sum_{i=0}^3 \sum_{\substack{1\le n\le N\\ n {\operatorname{\,mod}}4 = i}} {\left(}\int_{z_{(\!n\!-\!2\!)}}^{z_{(\!n\!+\!1\!)}} \hspace{-2pt} w(t) U^q(t,{z_{(\!n\!+\!1\!)}}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{z_{(\!n\!+\!1\!)}}^{z_{(\!n\!+\!2\!)}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim D_2. \nonumber \end{aligned}$$ We used in , and in . Estimate follows from . To get , we used the relation $z_{(\!n\!-\!1\!)} \le t_{k_{(\!n\!-\!1\!)}} \le t_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}$ which holds for all relevant indices $n$. The second inequality $t_{k_{(\!n\!-\!1\!)}} \le t_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}$ follows from . Concerning ${B_{\ref{bcounter15}}}$, we obtain $$\begin{aligned} {B_{\ref{bcounter15}}} & = \sum_{n=1}^{N\!-1} {\left(}\sum_{j=k_n}^{k_{(n\!+\!1\!)}-\!1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & = \sum_{n=1}^{N\!-1} {\left(}\sum_{j=k_n}^{k_{(n\!+\!1\!)}-\!1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\sum_{i=n\!+\!1}^{N\!-1} \int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-1} {\left(}\sum_{j=k_n}^{k_{(n\!+\!1\!)}-\!1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{208} \\ & \lesssim \sum_{n=1}^{N\!-1} {\left(}\int_{t_{k_{(\!n\!-\!1\!)}}}^{t_{k_{(\!n\!+\!1\!)}}} \! w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{209} \\ & = \sum_{i=0}^2 \sum_{\substack{1\le n\le N\!-\!1\\ n {\operatorname{\,mod}}3 = i}} {\left(}\int_{t_{k_{(\!n\!-\!1\!)}}}^{t_{k_{(\!n\!+\!1\!)}}} \! w(t) U^q(t,t_{k_n}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim D_2. \nonumber \end{aligned}$$ Proposition \[3\] together with yields . Estimate follows from . We have proved $${B_{\ref{bcounter12}}} \lesssim {B_{\ref{bcounter14}}}+{B_{\ref{bcounter15}}} \lesssim D_2.$$ We proceed with the term ${B_{\ref{bcounter13}}}$. [\[bcounter16\]]{}[\[bcounter17\]]{} $$\begin{aligned} {B_{\ref{bcounter13}}} & = \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\!\! \Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \le \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{i\in\{n\!+\!1,\ldots,N\}} \sup_{z\in[t_{k_i}\!,\, t_{k_{(\!i\!+\!1\!)}}]} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{i\in\{n\!+\!1,\ldots,N\}} \sup_{z\in[t_{k_i}\!,\, t_{k_{(\!i\!+\!1\!)}}]} \hspace{-5pt} {U^q}(t_{k_i},z) {\left(}\int_z^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \quad + \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) {\left(}\int_{t_{k_i}}^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & =: {B_{\ref{bcounter16}}}+{B_{\ref{bcounter17}}}. \nonumber \end{aligned}$$ For ${B_{\ref{bcounter16}}}$ we have [\[bcounter18\]]{}[\[bcounter19\]]{} $$\begin{aligned} {B_{\ref{bcounter16}}} & = \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{i\in\{n\!+\!1,\ldots,N\}} \sup_{z\in[t_{k_i}\!,\, t_{k_{(\!i\!+\!1\!)}}]} \hspace{-5pt} {U^q}(t_{k_i},z) {\left(}\int_z^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}} ]} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{210} \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}} )} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^{t_{k_{(\!n\!+\!2\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \quad + \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & =: {B_{\ref{bcounter18}}}+{B_{\ref{bcounter19}}}. \nonumber \end{aligned}$$ In step we used Proposition \[3\], considering also . For each $n\in\{0,\ldots,N\!-\!1\}$ there exists a point $y_{(\!n\!+\!1\!)}\in[t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}} ]$ such that $$\label{211} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}} ]} \hspace{-12pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^{t_{k_{(\!n\!+\!2\!)}}} \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\!\! \le 2 {U^q}(t_{k_{(\!n\!+\!1\!)}},y_{(\!n\!+\!1\!)}) {\left(}\int_{y_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!2\!)}}} \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\!\!\!.$$ Define also $y_{(-1)}:=0$ and $y_{(\!N\!+2\!)}:=\infty$. $$\begin{aligned} {B_{\ref{bcounter18}}} & = \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}} ]} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^{t_{k_{(\!n\!+\!2\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},y_{(\!n\!+\!1\!)}) {\left(}\int_{y_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{212} \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \int_{\Delta_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}} \hspace{-12pt} w(t){\mathrm{\,d}t}\ {U^q}(t_{k_{(\!n\!+\!1\!)}},y_{(\!n\!+\!1\!)}) {\left(}\int_{y_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{213} \\ & \le \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \int_{y_{(\!n\!-\!2\!)}}^{y_{(\!n\!+\!1\!)}} \hspace{-2pt} w(t) {U^q}(t,y_{(\!n\!+\!1\!)}) {\mathrm{\,d}t}\, {\left(}\int_{y_{(\!n\!+\!1\!)}}^{t_{k_{(\!n\!+\!2\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{214} \\ & \lesssim \sum_{n=1}^{N\!-\!1} {\left(}\int_{y_{(\!n\!-\!2\!)}}^{y_{(\!n\!+\!1\!)}} \hspace{-2pt} w(t) {U^q}(t,y_{(\!n\!+\!1\!)}) {\mathrm{\,d}t}{\right)}^{{\frac rq}} {\left(}\int_{y_{(\!n\!+\!1\!)}}^{y_{(\!n\!+\!2\!)}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{215} \\ & = \sum_{i=0}^3 \sum_{\substack{1\le n\le N\!-\!1\\ n {\operatorname{\,mod}}4 = i}} {\left(}\int_{y_{(\!n\!-\!2\!)}}^{y_{(\!n\!+\!1\!)}} \hspace{-2pt} w(t) {U^q}(t,y_{(\!n\!+\!1\!)}) {\mathrm{\,d}t}{\right)}^{{\frac rq}} {\left(}\int_{y_{(\!n\!+\!1\!)}}^{y_{(\!n\!+\!2\!)}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim D_2. \nonumber \end{aligned}$$ In we used . Inequality follows from . To get , we used the inequality $y_{(\!n\!-\!2\!)} \le t_{k_{(\!n\!-\!1\!)}} \le t_{(\!k_{\!(\!n\!+\!1\!)}\!-\!1\!)}$ (cf. ) satisfied for all relevant indices $n$. This inequality, together with , also yields . Next, the term ${B_{\ref{bcounter19}}}$ is treated as follows. $$\begin{aligned} {B_{\ref{bcounter19}}} & = \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-2pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_{(\!n\!+\!1\!)}}^{k_{(\!n\!+\!2\!)}\!-\!1} \hspace{-2pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \Theta^{k_{(\!n\!+\!1\!)}} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_{(\!n\!+\!2\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{216} \\ & \le \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_{(\!n\!+\!1\!)}}^{k_{(\!n\!+\!2\!)}\!-\!1} \hspace{-2pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \sum_{j=k_{(\!n\!+\!1\!)}}^{k_{(\!n\!+\!2\!)}\!-\!1} \hspace{-2pt} {\Theta^j}{U^q}(t_j,t_{k_{(\!n\!+\!2\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_{(\!n\!+\!1\!)}}^{k_{(\!n\!+\!2\!)}\!-\!1} \hspace{-2pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rq}} {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{217} \\ & \le {B_{\ref{bcounter15}}} \nonumber\\ & \lesssim D_2. \nonumber \end{aligned}$$ Inequality is obtained by using , and inequality by Proposition \[4\]. The final estimate ${B_{\ref{bcounter15}}}\lesssim D_2$ was already proved before. We have obtained $${B_{\ref{bcounter16}}} \lesssim {B_{\ref{bcounter18}}}+{B_{\ref{bcounter19}}} \lesssim D_2.$$ Let us now return to the term ${B_{\ref{bcounter17}}}$. It holds $$\begin{aligned} {B_{\ref{bcounter17}}} & = \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \!\! \Theta^{k_{(\!n\!+\!1\!)}} \hspace{-6pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},t_{k_i}) {\left(}\int_{t_{k_i}}^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \le \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \hspace{-4pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \sum_{j=\mu}^{k_i\!-\!1} {\Theta^j}{U^q}(t_j,t_{k_i}) {\left(}\int_{t_{k_i}}^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \hspace{-4pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \sum_{j=\mu}^{k_i\!-\!1} {\Theta^j}{U^q(\Delta_j)}{\left(}\int_{t_{k_i}}^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{218} \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \hspace{-4pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \sum_{j=k_{\!(\!i\!-\!1\!)}}^{k_i\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\left(}\int_{t_{k_i}}^\infty \hspace{-5pt} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{219} \\ & = \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \hspace{-4pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \sum_{j=k_{\!(\!i\!-\!1\!)}}^{k_i\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\left(}\sum_{m=i}^{N} \int_{t_{k_m}}^{t_{k_{(\!m\!+\!1\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \hspace{-4pt} \sup_{i\in\{n\!+\!2,\ldots,N\}} \sum_{j=k_{\!(\!i\!-\!1\!)}}^{k_i\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\left(}\int_{t_{k_i}}^{t_{k_{(\!i\!+\!1\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{220} \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\!{\frac rp}} \sum_{j=k_{\!(\!n\!+\!1\!)}}^{k_{\!(\!n\!+\!2\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^{t_{k_{(\!n\!+\!3\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{221} \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\sum_{j=k_{\!(\!n\!+\!1\!)}}^{k_{\!(\!n\!+\!2\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^{t_{k_{(\!n\!+\!3\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{222} \\ & \lesssim \sum_{n=1}^{N\!-\!2} {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} w(t) {U^q}(t,{t_{k_{(\!n\!+\!2\!)}}}) {\mathrm{\,d}t}{\right)}^{{\frac rq}} {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^{t_{k_{(\!n\!+\!3\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{223} \\ & = \sum_{i=0}^2 \sum_{\substack{1\le n\le N\!-\!2\\ n {\operatorname{\,mod}}3 = i}} {\left(}\int_{t_{k_n}}^{t_{k_{(\!n\!+\!2\!)}}} \hspace{-2pt} w(t) {U^q}(t,{t_{k_{(\!n\!+\!2\!)}}}) {\mathrm{\,d}t}{\right)}^{{\frac rq}} {\left(}\int_{t_{k_{(\!n\!+\!2\!)}}}^{t_{k_{(\!n\!+\!3\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim D_2. \nonumber \end{aligned}$$ Inequality follows from Proposition \[4\], and inequality from . To get , one uses Proposition \[89\], considering also . Proposition \[3\], again with , yields . Step follows from . In we applied . Having proved ${B_{\ref{bcounter17}}} \lesssim D_2$, we may now complete several more estimates, namely $${B_{\ref{bcounter13}}} \lesssim {B_{\ref{bcounter16}}} + {B_{\ref{bcounter17}}} \lesssim D_2,$$ which, combined with the earlier results, gives $${B_{\ref{bcounter10}}} \lesssim {B_{\ref{bcounter12}}} + {B_{\ref{bcounter13}}} \lesssim D_2.$$ The next untreated expression is ${B_{\ref{bcounter11}}}$. It is estimated in the following way. [\[bcounter20\]]{}[\[bcounter21\]]{}[\[bcounter22\]]{} $$\begin{aligned} {B_{\ref{bcounter11}}} & = \sum_{n=0}^{N} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) \sup_{z\in[t,\infty)} \! {U^q}(t,z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber \\ & \lesssim \sum_{n=0}^{N} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) \hspace{-6pt} \sup_{z\in[t,t_{k_{(\!n\!+\!1\!)}}]} \hspace{-8pt} {U^q}(t,z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\label{001} \\ & \quad + \sum_{n=0}^{N\!-\!1} \! \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \hspace{-6pt} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p}\!(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) {\mathrm{\,d}t}\hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-12pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\!\! {v^{1\!-p'}}\!\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=0}^{N} \! \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \hspace{-6pt} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) \hspace{-6pt} \sup_{z\in[t,t_{k_{(\!n\!+\!1\!)}}]} \hspace{-8pt} {U^q}(t,z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\!{\mathrm{\,d}t}\nonumber \\ & \quad + \sum_{n=0}^{N} \! \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \hspace{-6pt} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) {U^q}(t,t_{k_{(\!n\!+\!1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\nonumber \\ & \quad + \sum_{n=0}^{N\!-\!1} \! \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \hspace{-6pt} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p}\! (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) {\mathrm{\,d}t}\hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-12pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\!\! {v^{1\!-p'}}\!\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & =: {B_{\ref{bcounter20}}} + {B_{\ref{bcounter21}}} + {B_{\ref{bcounter22}}}. \nonumber \end{aligned}$$ Inequality follows from Proposition \[59\]. Define $t_{(\!k_{(\!N\!+\!2\!)}-\!1\!)} := \infty$. Then we have $$\begin{aligned} {B_{\ref{bcounter20}}} & = \sum_{n=0}^{N} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \!\!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) \hspace{-6pt} \sup_{z\in[t,t_{k_{(\!n\!+\!1\!)}}]} \hspace{-8pt} {U^q}(t,z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\!\!{\mathrm{\,d}t}\nonumber \\ & \le \sum_{n=0}^{N} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} w(t) {\mathrm{\,d}t}\sup_{z\in\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-4pt} {U^r}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=0}^{N} \Theta^{{\frac rq}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \hspace{-5pt} \sup_{z\in{\overline{\Delta}_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}} \hspace{-5pt} {U^r}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},z) {\left(}\int_z^{t_{k_{(\!n\!+\!1\!)}}} {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{002} \\ & \le \sum_{n=0}^{N} \Theta^{{\frac rq}(\!k_{(\!n\!+\!1\!)}\!-1\!)} {\left(}\int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-10pt} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},s) {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=0}^{N} {\left(}\int_{t_{(\!k_n\!-\!1\!)}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} w(t) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}^{t_{(\!k_{(\!n\!+\!2\!)}-\!1\!)}} {U^{p'}}(t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},s) {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{226} \\ & \le D_1. \nonumber \end{aligned}$$ Step follows from . In we used and the inequalities $t_{(\!k_n\!-\!1\!)} \le t_{(\!k_{(\!n\!+\!1\!)}\!-2\!)}$ and $t_{k_{(\!n\!+\!1\!)}}\le t_{(\!k_{(\!n\!+\!2\!)}\!-1\!)}$ which hold for all $n\in\{0,\ldots,N\}$ thanks to and the definition of $t_{(\!k_{(\!N\!+\!2\!)}-\!1\!)}$. We continue with the term ${B_{\ref{bcounter21}}}$, for which we get $$\begin{aligned} {B_{\ref{bcounter21}}} & = \sum_{n=0}^{N} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) {U^q}(t,t_{k_{(\!n\!+\!1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\nonumber \\ & \le \sum_{n=0}^{N\!-\!1} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-15pt} w(t) {\mathrm{\,d}t}\ {U^r}({\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=0}^{N\!-\!1} \Theta^{{\frac rq}(\!k_{(\!n\!+\!1\!)}\!-1\!)} {U^r}({\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}}) {\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{003} \\ & \le \sum_{n=0}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-1} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}{\left(}\int_{t_{k_{(\!n\!+\!1\!)}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & = {\left(}\Theta^\mu {U^q}(\Delta_\mu) {\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}} \label{227} \\ & \lesssim {\left(}\int_0^{{{t_\mu}}} w(t){\mathrm{\,d}t}\, {U^q}(\Delta_\mu) {\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}} \label{006} \\ & \lesssim {\left(}\int_0^{t_{(\!\mu\!+\!1\!)}} w(t) {U^q}(t,t_{(\!\mu\!+\!1\!)}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}} \nonumber \\ & \lesssim D_2. \label{228} \end{aligned}$$ To get , we made use of . In we used the fact $$\label{007} \sum_{j=k_0}^{k_{1}\!-1} {\Theta^j}{U^q(\Delta_j)}= \Theta^\mu {U^q}(\Delta_\mu)$$ (recall that $k_0=\mu$ and $k_1=\mu+1$). Inequality is a consequence of. The final estimate follows from the relation ${B_{\ref{bcounter15}}}\lesssim D_2$ which was proved earlier. Concerning ${B_{\ref{bcounter22}}}$, we may write $$\begin{aligned} {B_{\ref{bcounter22}}} & = \sum_{n=0}^{N\!-\!1} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} \!\! \int_{\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \hspace{-25pt} U^{\frac{rq}p} (t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)},t) \, w(t) {\mathrm{\,d}t}\hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-12pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & \lesssim \sum_{n=0}^{N\!-\!1} \Theta^{{\frac rp}(\!k_{(\!n\!+\!1\!)}\!-1\!)} U^{\frac{rq}p}(\Delta_{(\!k_{(\!n\!+\!1\!)}\!-1\!)})\ \Theta^{k_{(\!n\!+\!1\!)}\!-1}\hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-12pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{005} \\ & \le \sum_{n=0}^{N\!-\!1} {\left(}\sum_{j=k_n}^{k_{(\!n\!+\!1\!)}\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\!\! \Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-5pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber \\ & = \Theta^{\frac{r\mu}{q}} U^\frac{rq}{p}(\Delta_\mu) \sup_{z\in[t_{k_1},\infty)} \hspace{-5pt} {U^q}(t_{k_1},z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter13}}} \label{008} \\ & \le \Theta^{\frac{r\mu}{q}} {\left(}\int_{{{t_\mu}}}^\infty {U^{p'}}({{t_\mu}},s) {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter13}}} \nonumber \\ & \lesssim {\left(}\int_0^{{{t_\mu}}} w(t){\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{{{t_\mu}}}^\infty {U^{p'}}({{t_\mu}},s) {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter13}}} \label{009} \\ & \lesssim D_1 + D_2. \label{010} \end{aligned}$$ Step follows from , step from , and step from . To obtain , we used the estimate ${B_{\ref{bcounter13}}}\lesssim D_2$ which was proved earlier. We have proved $${B_{\ref{bcounter11}}} \lesssim {B_{\ref{bcounter20}}} + {B_{\ref{bcounter21}}} + {B_{\ref{bcounter22}}} \lesssim D_1 + D_2.$$ Together with the estimate of ${B_{\ref{bcounter10}}}$ we obtained earlier, this also yields $${B_{\ref{bcounter8}}} \lesssim {B_{\ref{bcounter10}}} + {B_{\ref{bcounter11}}} \lesssim D_2.$$ In the next part, we return to the expression ${B_{\ref{bcounter9}}}$. It holds $$\begin{aligned} {B_{\ref{bcounter9}}} & = \sum_{n\in{\mathbb{A}}} \int_{t_{k_n}}^{t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}} \! {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}\! w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\nonumber\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-4pt} \int_{t_{k_n}}^{t_{(\!k_{\!(\!n\!+\!1\!)}-\!1\!)}} \! w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\label{250}\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-4pt} \int_{t_{k_n}}^{t_{(\!k_{\!(\!n\!+\!1\!)}-\!1\!)}} \! w(t) \sup_{z\in[t,t_{k_{(\!n\!+\!1\!)}}]} {U^q}(t,z) {\left(}\int_z^\infty \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\label{251}\\ & \quad + \sum_{n\in{\mathbb{A}}} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-6pt} \int_{t_{k_n}}^{t_{(\!k_{\!(\!n\!+\!1\!)}-\!1\!)}} \!\!\! w(t){\mathrm{\,d}t}\hspace{-6pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-9pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n\in{\mathbb{A}}} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-4pt} \int_{t_{k_n}}^{t_{(\!k_{\!(\!n\!+\!1\!)}-\!1\!)}} \! w(t) {U^q}(t,t_{(\!k_{(\!n\!+\!1\!)}\!-1\!)}) {\mathrm{\,d}t}\, {\left(}\int_{t_{k_n}}^\infty \!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{252}\\ & \quad + \sum_{n\in{\mathbb{A}}} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-2pt} \Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-8pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber\\ & \lesssim \sum_{n=1}^{N} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-2pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rq}\! {\left(}\int_{t_{k_n}}^\infty \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\label{253}\\ & \quad + \sum_{n=1}^{N\!-\!1} {\left(}\sum_{j=k_{(\!n\!-\!1\!)}}^{k_n\!-\!1} \hspace{-6pt} {\Theta^j}{U^q(\Delta_j)}{\right)}^{\frac rp}\hspace{-2pt} \Theta^{k_{(\!n\!+\!1\!)}\!-1} \hspace{-8pt} \sup_{z\in[t_{k_{(\!n\!+\!1\!)}},\infty)} \hspace{-8pt} {U^q}(t_{k_{(\!n\!+\!1\!)}},z) {\left(}\int_z^\infty \!\!\! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\nonumber\\ & = {\left(}\Theta^\mu {U^q}(\Delta_\mu) {\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}}+ {B_{\ref{bcounter13}}} \label{254} \\ & \lesssim {\left(}\int_0^{{{t_\mu}}} w(t){\mathrm{\,d}t}\, {U^q}(\Delta_\mu) {\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}} + {B_{\ref{bcounter13}}} \label{255} \\ & \lesssim {\left(}\int_0^{t_{(\!\mu\!+\!1\!)}} w(t) {U^q}(t,t_{(\!\mu\!+\!1\!)}) {\mathrm{\,d}t}{\right)}^{\frac rq}{\left(}\int_{t_{k_{1}}}^\infty {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}+ {B_{\ref{bcounter15}}} + {B_{\ref{bcounter13}}}\nonumber \\ & \lesssim D_2. \label{256} \end{aligned}$$ Estimate is granted by , and estimate by Proposition \[59\]. Step is based on . In we again applied . To get the relations and , we used and , respectively. The final inequality follows from the already known relations ${B_{\ref{bcounter15}}} \lesssim D_2$ and ${B_{\ref{bcounter13}}} \lesssim D_2$. We have shown $${B_{\ref{bcounter9}}} \lesssim D_2,$$ and thus also $$\int_{{{t_\mu}}}^\infty {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty \! {v^{1\!-p'}}\!(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}\!\! {\mathrm{\,d}t}\ \lesssim {B_{\ref{bcounter8}}} + {B_{\ref{bcounter9}}} \lesssim D_1 + D_2.$$ If we combine this inequality with , we reach $$\begin{aligned} & \int_{{{t_\mu}}}^\infty {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) {\left(}\int_t^\infty {U^{p'}}(t,z){v^{1\!-p'}}(z){\mathrm{\,d}z}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & \quad + \int_{{{t_\mu}}}^\infty {\left(}\int_{{{t_\mu}}}^t w(x) {U^q}(x,t) {\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in[t,\infty)} {U^q}(t,z) {\left(}\int_z^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}{\mathrm{\,d}t}\nonumber\\ & \lesssim D_1 + D_2. \nonumber \end{aligned}$$ The constant related to the symbol “$\lesssim$” in here does not depend on the choice of $\mu$, thus passing $\mu\to -\infty$ (notice ${{t_\mu}}\to 0$ as $\mu\to -\infty$) and applying the monotone convergence theorem yields $$A_1 + A_2 \lesssim D_1 + D_2.$$ We have so far assumed that $\int_0^\infty w(x){\mathrm{\,d}x}= \Theta^K$ for a $K\in{\mathbb{Z}}$. The result is extended to general weights $w$ by the same procedure as the one used at the end of the proof of the implication “(ii)$\Rightarrow$(i)”. The proof of the whole theorem is now complete. Theorem \[144\] is proved in almost exactly the same way as Theorem \[7\]. The difference is just in the use of appropriate “limit variants” of certain expressions for $p=1$. Namely, $${\left(}\int_y^z {U^{p'}}(y,x) {v^{1\!-p'}}(x) {\mathrm{\,d}x}{\right)}^{\frac1{p'}}\textnormal{\quad is\ replaced\ by\quad} {\operatornamewithlimits{ess\, sup\,}}_{x\in(y,z)} U(y,x)v^{-1}(x)$$ and $${\left(}\int_y^z {v^{1\!-p'}}(x){\mathrm{\,d}x}{\right)}^{\frac1{p'}}\textnormal{\quad is\ replaced\ by\quad} {\operatornamewithlimits{ess\, sup\,}}_{x\in(y,z)} v^{-1}(x),$$ whenever these expressions appear with some $0\le y<z\le\infty$. To clarify the correspondence between $A_2$ and $A_4$, let us note that $$\sup_{z\in[t,\infty)} {U^q}(t,z) {\operatornamewithlimits{ess\, sup\,}}_{s\in(z,\infty)} v^{q'}(s) = {\operatornamewithlimits{ess\, sup\,}}_{s\in(t,\infty)} v^{q'}(s) \sup_{z\in[t,s)} {U^q}(t,z) = {\operatornamewithlimits{ess\, sup\,}}_{s\in(t,\infty)} {U^q}(t,s) v^{q'}(s)$$ is true for all $t>0$. Naturally, the limit variant of Proposition \[107\] for $p=1$ is used in the proof as well. All the estimates are then analogous to their counterparts in the proof of Theorem \[7\]. Therefore, we do not repeat them in here. \(i) Theorem \[7\], which relates to the inequality , i.e. to the operator $H^*$, is the one proved here, while the result for $H$ (i.e. for ) is presented as Corollary \[7\*\]. Of course, the opposite order could have been chosen, since the version with $H$ instead of $H^*$ can be proved in an exactly analogous way. As mentioned before, the variants for $H$ and $H^*$ are equivalent by a change of variables in the integrals. The reason why the proof of the “dual” version is shown here is that the discretization-related notation is then the same as in [@K6]. \(ii) Discretization based on finite covering sequences is used here, although the double-infinite (indexed by ${\mathbb{Z}}$) variant is far more usual in the literature (cf. [@GS; @L; @S2]). The advantage of the finite version is that the proof works for $L^1$-weights $w$ and then it is easily extrapolated for the non-$L^1$ weights by the final approximation argument. In order to work with infinite partitions, one needs to assume $w\notin L^1$. The pass to the $L^1$-weights then cannot be done in such an easy way as in the opposite order. The authors usually omit the case $w\in L^1$ (see e.g. [@GS]). Besides that, there is no essential difference between in the techniques based on finite and infinite partitions. \(iii) In Theorems \[7\] and \[144\], the equivalence “(i)$\Leftrightarrow$(ii)” was known before [@L] and it is reproved here using another method than in [@L]. The main achievement is the equivalence “(i)$\Leftrightarrow$(iii)” which can also be proved directly, by the same technique and without need for the discrete $D$-conditions (cf. [@K6]). Doing so would however require constructing more different special functions (such as $g$ and $h$ in the “(i)$\Rightarrow$(ii)” part of Theorem \[7\]) and therefore also introducing additional notation. \(iv) The kernel $U$ is not assumed to be continuous. However, for every $t>0$ the function $U(t,\cdot)$ is nondecreasing, hence continuous almost everywhere on $(0,\infty)$. Thus, so is the function ${U^q}(t,\cdot){\left(}\int_\cdot^\infty {v^{1\!-p'}}(s){\mathrm{\,d}s}{\right)}^{\frac r{p'}}$. Therefore, the value of the expression $A_2$ remains unchanged if “$\sup_{z\in[t,\infty)}\!$” in there is replaced by “${\operatornamewithlimits{ess\, sup\,}}_{z\in[t,\infty)}\!$”. Although the latter variant may seem to be the “proper” one, both are correct in this case. Besides that, the range $z\in[t,\infty)$ in the supremum or essential supremum may obviously be replaced by $z\in(t,\infty)$ without changing the value of $A_2$. Applications ============ The integral conditions for the boundedness $H:L^p(v)\to L^q(w)$ with $0<q<1\le p<\infty$ may be used to complete [@GS Theorem 5.1] with two missing cases. (These cases are in fact included in [@GS] but covered there only by discrete conditions.) Denote by ${\mathscr M^\downarrow}$ the cone of all nonnegative nonincreasing functions on $(0,\infty)$. The result then reads as follows. \[800\] Let $u$, $v$, $w$ be weights, $0<q<p<\infty$, $q<1$ and $r=\frac{pq}{p-q}$. - Let $0<p\le 1$. Then the inequality $$\label{I1} {\left(}\int_0^\infty {\left(}\int_t^\infty f(s)u(s){\mathrm{\,d}s}{\right)}^q w(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\le C {\left(}\int_0^\infty f^p(t) v(t) {\mathrm{\,d}t}{\right)}^{\frac1p}$$ holds for all $f\in{\mathscr M^\downarrow}$ if and only if $$A_5 := {\left(}\int_0^\infty {\left(}\int_0^t w(x){\mathrm{\,d}x}{\right)}^{\frac rp}w(t) \sup_{z\in(t,\infty)} {\left(}\int_t^z u(s){\mathrm{\,d}s}{\right)}^r {\left(}\int_0^z v(y){\mathrm{\,d}y}{\right)}^{-{\frac rp}}{\mathrm{\,d}t}{\right)}^{\frac1r}<\infty$$ and $$A_6 := {\left(}\int_0^\infty \! {\left(}\int_0^t \!\! w(x) {\left(}\int_x^t \!\!u(s){\mathrm{\,d}s}{\right)}^{\!q} \!\! {\mathrm{\,d}x}{\right)}^{\!{\frac rp}} \!\! w(t) \sup_{z\in(t,\infty)} {\left(}\int_t^z \!\! u(s){\mathrm{\,d}s}{\right)}^{\!q} \! {\left(}\int_0^z \!\! v(y){\mathrm{\,d}y}{\right)}^{\!\!-{\frac rp}} \!\!\!\! {\mathrm{\,d}t}{\right)}^{\!{\frac1r}} \!\! < \infty.$$ Moreover, the least constant $C$ such that holds for all $f\in{\mathscr M^\downarrow}$ satisfies $C\approx A_5 + A_6$. - Let $p>1$. Then holds for all $f\in{\mathscr M^\downarrow}$ if and only if $A_6<\infty$, $$A_7 := {\left(}\int_0^\infty \! {\left(}\int_0^t \! w(x){\mathrm{\,d}x}{\right)}^{\frac rp}\! w(t) {\left(}\int_t^\infty \! {\left(}\int_t^z \! u(s){\mathrm{\,d}s}{\right)}^{p'} \! {\left(}\int_0^z \!\! v(y){\mathrm{\,d}y}{\right)}^{-p'} \!\!\! v(z) {\mathrm{\,d}z}{\right)}^{{\frac r{p'}}}\!\!\! {\mathrm{\,d}t}{\right)}^{\frac1r}\!\! <\infty$$ and $A_8<\infty$, where $$A_8 := \begin{cases} \displaystyle\ \, {\left(}\int_0^\infty \!\! w(t) {\left(}\int_0^t u(s) {\mathrm{\,d}s}{\right)}^q {\mathrm{\,d}t}{\right)}^{\frac1q}{\left(}\int_0^\infty \!\! v(y){\mathrm{\,d}y}{\right)}^{\!-{\frac1p}} \!\! <\infty & \textit{if\ } \displaystyle \int_0^\infty v(y){\mathrm{\,d}y}< \infty,\\ \displaystyle\ \, 0 & \textit{if\ } \displaystyle \int_0^\infty \!\! v(y){\mathrm{\,d}y}= \infty. \end{cases}$$ Moreover, the least constant $C$ such that holds for all $f\in{\mathscr M^\downarrow}$ satisfies $C\!\approx\! A_6 \!+ \!A_7\! + \!A_8$. \(i) By [@GS Theorem 4.1], holds for all $f\in{\mathscr M^\downarrow}$ if and only if $$\label{987} {\left(}\int_0^\infty {\left(}\int_t^\infty {\left(}\int_t^x u(s){\mathrm{\,d}s}{\right)}^p h(x) {\mathrm{\,d}x}{\right)}^\frac qp w(t) {\mathrm{\,d}t}{\right)}^\frac pq \le C^p \int_0^\infty h(s) \int_0^s v(y){\mathrm{\,d}y}{\mathrm{\,d}s}$$ holds for all $h\in{\mathscr M_+}$. In fact, [@GS Theorem 4.1] is stated with the assumption $\int_0^\infty v(y){\mathrm{\,d}y}= \infty$ which is, however, not used in the proof in [@GS]. Validity of for all $h\in{\mathscr M_+}$ is equivalent to the condition $A_5 + A_6 <\infty$ by Theorem \[144\], since $U(x,y)={\left(}\int_x^y u(s) {\mathrm{\,d}s}{\right)}^p$ is a ${\vartheta}$-regular kernel (with ${\vartheta}=2^p$). \(ii) By [@GS Theorem 2.1], holds for all $f\in{\mathscr M^\downarrow}$ if and only if $A_8 \le \infty$ and $${\left(}\int_0^\infty {\left(}\int_t^\infty \int_t^x u(s){\mathrm{\,d}s}\ h(x) {\mathrm{\,d}x}{\right)}^q w(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\le C {\left(}\int_0^\infty h^p(s) {\left(}\int_0^s v(y){\mathrm{\,d}y}{\right)}^{p} v^{1-p}(s) {\mathrm{\,d}s}{\right)}^{\frac1p}$$ holds for all $h\in{\mathscr M_+}$. The latter is, by Theorem \[7\], equivalent to the condition $A_6^* + A_7 <\infty$, where $$A_6^* := \! {\left(}\! \int_0^{\!\infty} \!\! {\left(}\! \int_0^t \!\!\! w(x) {\left(}\! \int_x^t \!\!\!u(s){\mathrm{\,d}s}{\right)}^{\!q} \!\! {\mathrm{\,d}x}{\right)}^{\!{\frac rp}} \!\! w(t) \!\! \sup_{z\in(t,\infty)} \! {\left(}\int_t^z \!\!\! u(s){\mathrm{\,d}s}{\right)}^{\!q} \! {\left(}\! \int_z^{\!\infty} \!\! {\left(}\int_0^x \! \!\! v(y){\mathrm{\,d}y}{\right)}^{\!\!-p'} \!\!\!\!\! v(x) {\mathrm{\,d}x}{\right)}^{\!\!{\frac r{p'}}} \!\!\!\! {\mathrm{\,d}t}{\right)}^{\!\!{\frac1r}}\!\!\!.$$ Since $$\int_z^\infty {\left(}\int_0^s \!\! v(y){\mathrm{\,d}y}{\right)}^{-p'}\!\!\! v(s){\mathrm{\,d}s}+ {\left(}\int_0^\infty \!\!\! v(y){\mathrm{\,d}y}{\right)}^{1-p'} \!\!\!\! \approx\, {\left(}\int_0^z \!\! v(y){\mathrm{\,d}y}{\right)}^{1-p'}$$ is satisfied for all $z>0$, it is easy to verify that $A_6^* \lesssim A_6$ and $A_6 \lesssim A_6^* + A_8$. In both cases (i) and (ii), the estimates on the optimal constant $C$ also follow from [@GS Theorem 2.1, Theorem 4.1] and Theorems \[7\] and \[144\]. In the case $0<q<p\le 1$, in [@GS Theorem 4.1] it was shown that holds for all $f\in{\mathscr M^\downarrow}$ if and only if $${\left(}\int_0^\infty {\left(}\sup_{x\in[t,\infty)} f(x) \int_t^x u(s){\mathrm{\,d}s}{\right)}^q w(t) {\mathrm{\,d}t}{\right)}^{\frac1q}\le C {\left(}\int_0^\infty f^p(t)v(t){\mathrm{\,d}t}{\right)}^{\frac1p}$$ holds for all $f\in{\mathscr M^\downarrow}$. Theorem \[800\] hence applies to this supremal operator inequality as well. Theorem \[800\] may be further applied to prove certain weighted Young-type convolution inequalities (cf. [@K1]) in parameter settings which could not be reached so far. For this particular application, it is important that the weight $w$ is not involved in any implicit conditions. For more details see [@K1]. As shown e.g. in [@SS Theorem 4.4], certain weighted inequalities restricted to convex functions are equivalently represented by weighted inequalities involving a Hardy-type operator with the $1$-regular Riemann-Liouville kernel $U(x,y)=(y-x)$. Hence, the results of this paper also provide characterizations of validity of those convex-function inequalities in the case $0<q<1\le p<\infty$. [99]{}

--- abstract: 'We study the automorphism groups of countable homogeneous directed graphs (and some additional homogeneous structures) from the point of view of topological dynamics. We determine precisely which of these automorphism groups are amenable (in their natural topologies). For those which are amenable, we determine whether they are uniquely ergodic, leaving unsettled precisely one case (the “semi-generic" complete multipartite directed graph). We also consider the Hrushovski property. For most of our results we use the various techniques of [@AKL12], suitably generalized to a context in which the universal minimal flow is not necessarily the space of all orders. Negative results concerning amenability rely on constructions of the type considered in [@ZUC13]. An additional class of structures (compositions) may be handled directly on the basis of very general principles. The starting point in all cases is the determination of the universal minimal flow for the automorphism group, which in the context of countable homogeneous directed graphs is given in [@JLNW14] and the papers cited therein.' address: ' University of Toronto)\' author: - Micheal Pawliuk and Miodrag Sokić bibliography: - 'references.bib' title: Amenability and unique ergodicity of automorphism groups of countable homogeneous directed graphs --- Introduction ============ The Kechris-Pestov-Todorcevic [@KPT05] correspondence relates the topological dynamics of automorphism groups of countable homogeneous structures to combinatorial problems related to structural Ramsey theory. Our aim here is to apply this correspondence and the methods of [@AKL12; @ZUC13], suitably generalized, to the determination of the amenable automorphism groups associated with countable homogeneous directed graphs, and also to determine which of these are uniquely ergodic; here one specific case remains open. The KPT correspondence relies on the identification of automorphism groups of countable homogeneous structures with closed subgroups of the full symmetric group on a countable set, with respect to its natural topology, and the more subtle connection due to Fraïssé between countable homogeneous structures in relational languages and classes of finite structures with certain closure properties, to be described in more detail in § \[sec:TechPrelim\]. The main points to be taken from [@KPT05] relate to the properties of *extreme amenability* and the determination of the *universal minimal flow*. Here a topological group $G$ is said to be extremely amenable if every continuous action on a compact set has a fixed point; a flow is a continuous action on a compact set; and a universal minimal flow is a minimal flow which covers any other minimal flow by a continuous $G$-invariant map. If one has a countable homogeneous structure $\Gamma$ associated on the one hand with the family $\Age(\Gamma)$ of all finite structures embedding into it, and on the other hand with the group $G = \Aut(\Gamma)$, then according to the KPT correspondence the property of extreme amenability for $G$ is equivalent to the Ramsey property for $\Age(\Gamma)$ (a structural Ramsey theorem). For example, the automorphism group of the randomly ordered random graph is extremely amenable; this is a reformulation of the structural Ramsey theorem for ordered graphs [@NR77; @NR83; @NR89]. A sharper interpretation of the structural Ramsey theorem for ordered graphs is as a determination of the universal minimal flow for the automorphism group of the (unordered) random graph. Under suitable hypotheses to be reviewed below, the universal minimal flow for an automorphism group can be identified with the set of expansions of the given structure to a category in which the structural Ramsey theorem holds; e.g., to the space of orderings of the random graph in the case at hand. Amenability and Unique Ergodicity. ---------------------------------- Using this determination of the universal minimal flow, [@AKL12] investigated problems of amenability and unique ergodicity for automorphism groups of countable homogeneous graphs. Here amenability requires, not a fixed point, but a $G$-invariant probability measure, while unique ergodicity requires amenability, but with a *unique* $G$-invariant probability measure. Thus amenability, unique ergodicity, and extreme amenability form a hierarchy of successively stronger properties. In [@ZUC13], Zucker gave examples of countable homogeneous directed graphs (including one tournament) with non-amenable automorphism group. We will work systematically through the classification of countable homogeneous directed graphs as given by [@Cher98]. In [@JLNW14] this classification was used to work out the universal minimal flows explicitly. In a number of cases (some already exploited by [@ZUC13]) the appropriate Ramsey class is not obtained from expansions by orders, so we will need to reformulate the methods of [@AKL12] in a somewhat broader setting. Imprimitive structures that is, structures carrying nontrivial equivalence relations require some specific attention when forming Ramsey expansions. In the simplest case, where the equivalence relation is a congruence, the analysis may be given in completely general terms. All other cases require individual attention (sometimes in large groups: the main family of examples to be considered is uncountable, but may be treated in a uniform manner). Evidently the reader will need to know something of the structure of each type of homogeneous directed graph, and the specifics of the expansions, whether by orderings or other additional structure, to classes with the Ramsey property. We will give this in an introductory section which may be used for reference. The specifics of the proof of the Ramsey property are irrelevant here, and even the meaning of the Ramsey property is not very germane, as [@AKL12] comes very close to giving a characterization of the properties of amenability and unique ergodicity that we can take as our starting point, once we have reformulated it at the appropriate level of generality. Summary of results. ------------------- It will be helpful at the outset to give a chart showing the various results to be obtained. Of course, this chart makes use of notation for specific families of countable homogeneous directed graphs to be discussed in detail a little farther on. In the classification of homogeneous structures, one usually places the imprimitive (or otherwise degenerate) examples ahead of the primitive ones. We find a different ordering more suitable here. We list a few degenerate examples, then the imprimitive cases involving a congruence (one might say, the highly imprimitive cases), then a few primitive examples that turn out to have non-amenable automorphism groups, and then the more typical cases in which the automorphism groups are amenable and the structures are not particularly degenerate, more or less in order of their structure-theoretic complexity. One of the more exotic structures from our point of view (the semi-generic complete multipartite directed graph) falls somewhere toward the middle of the classification, from that point of view. We also remark that we include in our analysis a property that we have neglected in this introduction, one of several that arise naturally in consideration of the problem of unique ergodicity, and one that is certainly of independent interest. The following summarizes what is now known, and what remains open, in regards to the amenability of the automorphism groups of structures from Cherlin’s classification. For readability, in the table we suppress $\Aut(\Gamma)$ and simply write the structure $\Gamma$. We also suppress the finite structures. \[table:results\] [lllll]{} Results\ Type & Notation & Amenable? & Hrushovski? & Uniquely ergodic?\ Composition & $\TT[\II_n], \II_n [\TT]$ & $\Leftrightarrow \TT$ & $\Leftrightarrow \TT$ & $\Leftrightarrow \TT$, § 2\ Some weak local orders & $\SS(2), \SS(3)$ & , [@ZUC13] & &\ ------------------------------- Generic p.o., variant, and 2-cover of $\QQ$ ------------------------------- & --------------------- $\PP$, $\PP(3), \hat{\QQ}$ --------------------- & ------------- , [@KSOK12] , § 5 ------------- & [@l@]{}\ & [@l@]{}\ \ Linear tournament & $\QQ$ & , [@pes98] & & , [@pes98]\ Generic tournament & $\TT^\omega$ & & $?$ & , [@AKL12]\ 2-cover of generic & $\hat{\TT^\omega}$ & , § 4.1 & $?$ & , § 8.1\ Generic multipartite & $\DD_n$ & , § 4.2 & $?$ & , § 8.2\ Semi-generic multipartite & $\SS$ & , § 4.3 & $?$ & $?$, § 8.3\ Generic omitter & $\GG_n, \FF(\T)$ & , § 4.4 & $?$ & , § 9 Amenability. ------------ To conclude this introduction we will give some indication of the methods used in the analysis of various cases. As the analysis is simpler in the case of amenability, we begin with that. Whether we deal with amenability or unique ergodicity, our starting point is a prior understanding of the universal minimal flow in concrete combinatorial terms, supplied in our case by [@JLNW14]. This can be taken largely as a black box: we begin with a countable homogeneous structure $\Gamma$, and a certain associated countable homogeneous structure $\Gamma^*$ (with the Ramsey property), and rather than considering $\Aut(\Gamma)$ and $\Aut(\Gamma^*)$, one uses the ideas of [@KPT05] as developed in [@AKL12] to characterize amenability and universal ergodicity in terms of the combinatorics of the two classes $\K := \Age(\Gamma), \K^* := \Age(\Gamma^*)$ consisting respectively of finite structures embedding into $\Gamma$, and finite structures embedding into $\Gamma^*$. Namely, amenability is equivalent to existence of a “random $\K^*$-expansion”, which is a function picking out for each $\K$-structure, a probability measure on the set of its expansions to a $\K^*$-structure, with a coherence condition corresponding to embeddings between structures in $\K$. And then, of course, unique ergodicity is equivalent to the uniqueness of this notion of random expansion. The treatment of amenability is relatively straightforward. To prove amenability, one may first try the uniform measure, according to which all expansions are equally likely. We may mention two of the standard examples. Example 1. 1.1. If the expansion from $\K$ to $\K^*$ involves adjunction of an additional ordering, then the coherence condition states that the restriction of a random ordering from a given structure to a smaller structure is random, which is probably clear as it stands and at the combinatorial level is a consequence of the fact that there is only one way to order the small structure (ignoring whatever structure is present beyond the order). 1.2. If the structure carries an equivalence relation $E$, then frequently the expansion consists of an ordering in which $E$-classes are convex. This may be thought of as an ordering of the $E$-classes followed by an ordering of each class, and is an example of what we will call a composition. In this case, if we have two such structures $\AA \leq \BB$, then a random ordering of $\BB$ clearly induces a random ordering of the classes of $\AA$, and then a random order of each class, independently. This proves the amenability of the automorphism group (with the uniform distribution as witness). There are a number of cases in which the expansion from $\Gamma$ to $\Gamma^*$, or from $\K$ to $\K^*$, involves the introduction of an arbitrary order, and the argument of (1.1) proves amenability. There are also a number of cases where we have a congruence $E$ on our structure in which the both amenability and unique ergodicity can be reduced to a treatment of the quotient structure and the structure on each class, in the manner of (1.2). When we dispose of cases that can be treated by these methods, we are left with several cases which are in fact non-amenable and for which a contradiction is achieved rapidly by examining the meaning of the coherence condition, and some other cases where the appropriate notion of random expansion must be written down, and the corresponding coherence condition checked. Unique Ergodicity. ------------------ Now we should say something about the methods used to prove, or disprove, unique ergodicity, which are more sophisticated than those used to check amenability, and frequently involve some computation. The first known example of a uniquely ergodic group that is neither extremely amenable nor compact was $S_\infty$ as shown in [@GW02]. In [@AKL12] it was shown that this is a general phenomenon; the automorphism groups of the Fraïssé limits of (1) the class of $K_n$-free graphs, (2) the class of $r$-uniform hypergraphs, (3) the class of finite metric spaces with distances in a given additive subsemigroup of $\RR^+$, (4) some classes of hypergaphs with forbidden configurations, are all uniquely ergodic. The approach in [@AKL12] was to examine consistent random expansions of these Fraïssé classes. First, amenability was established by verifying that the uniform measure was indeed a consistent random expansion. In their cases the reasonable expansions $\mathcal{K}^*$ were usually arbitrary linear orders, which automatically ensures that the uniform measure works, see their Proposition 9.3. Following that, they establish so called quantitative expansion properties, which amount to asymptotic bounds relating to how often (small) ordered structures embed into other (large) ordered structures. These bounds are established by counting the number of order expansions of a structure, and applying the probabilistic McDiarmid inequality. More concrete details will be provided in Section \[sec:UE\_McD\]. The cases dealt with here are more subtle since the expansions are usually more complicated than arbitrary linear orders. For example, in the case of $\mathbb{D}_n$ the expansions are linear orders that are convex with respect to the equivalence classes of vertices without edges. Quantifiably more complicated is the case of the semi-generic multipartite digraph $\mathbb{S}$, whose precompact expansion is more than just a collection of linear orders. In this case we were unable to establish unique ergodicity, but were able to establish it for slightly tamer expansions. For the most part, our approach is to use the tools established in [@AKL12] suitably generalized to handle precompact expansions. We streamline the methods into a black box theorem, Lemma \[lem:QOP\_strategy\], which is purely finitary and combinatorial. \[table:contents\] [ll]{} Contents\ § 1 Introduction & § 7 Unique ergodicity and\ & McDiarmid’s inequality\ § 2 Preliminaries, including composition & § 8 The random method\ -------------------------- § 3 Ramsey expansions of countable homogeneous -------------------------- & § 9 The hypergraph method\ § 4 Amenability & § 10 Conclusions and open questions\ § 5 Failures of amenability & § 11 Appendix\ § 6 The Hrushovski property Preliminaries, including composition {#sec:TechPrelim} ==================================== Now we will describe the mathematical objects and notions that we will be using. Sections 2.1-2.6 are intended to be used as a reference, and Sections 2.7-2.10 are a discussion of compositions. Amenability. ------------ Let $G$ be a topological group. A continuous action of $G$ on a compact Hausdorff space is called a **$G$-flow**. A $G$-flow is **minimal** if the orbit of every point is dense. If every $G$-flow has a $G$-invariant Borel probability measure, then we say that $G$ is **amenable**. We say that $G$ is **uniquely ergodic** if every minimal $G$-flow has a unique $G$-invariant Borel probability measure. We will go into more depth about various equivalent versions of amenability in Chapter 2. Throughout, we consider amenability and unique ergodicity for a collection of automorphism groups of countable structures related to directed graphs. These groups are not locally compact and they are not discrete, but they are non-Archimedian Polish groups, see [@BK96] for more details. Fraïssé Classes and Structures. ------------------------------- Let $\AA$ and $\BB$ be given structures. If there is an embedding from $\AA$ into $\BB$ then we write $\AA\into \BB$, if $\AA$ is a substructure of $\BB$ then we write $\AA \leq \BB$, and if $\AA$ and $\BB$ are isomorphic, then we write $\AA \cong \BB$. We write $\binom{\BB}{\AA} = \{\CC \leq \BB : \CC \cong \AA\}$. We say that a given structure is **locally finite** if each of its finitely generated substructures are finite. We denote by $\Age(\AA)$ the collection of all finite substructures of $\AA$. A structure $\AA$ is **ultrahomogeneous** if every isomorphism between two finite substructures can be extended to an automorphism of $\AA$. We say that $\AA$ is a **Fraïssé structure** if it is countably infinite, locally finite and ultrahomogeneous. Let $L$ be a signature and let $\K$ be a class of finite structures in $L$. Then $\K$ satisfies the: ----------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- (**HP**) **Hereditary Property**, if whenever $\AA \into \BB$ and $\BB \in \K$, then $\AA \in \K$. (**JEP**) **Joint Embedding Property**, if for all $ \AA, \BB \in \K$ there is a $\CC \in \K$ such that $\AA~\into~\CC$ and $\BB~\into~\CC$. (**AP**) **Amalgamation Property**, if for all $\AA,\BB,\CC \in \K$ and all embeddings $f: \AA \rightarrow \BB$ and $g: \AA \rightarrow \CC$ there is a $\DD \in \KK$ and embeddings $\overline{f}: \BB \rightarrow \DD$ and $\overline{g}: \CC \rightarrow \DD$ with $\overline{f} \circ f = \overline{g} \circ g$. (**SAP**) **Strong Amalgamation Property**, if in addition to **AP** we have $\overline{f}(\BB) \cap \overline{g}(\CC) = \overline{f} \circ f (\AA)$. ----------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- We say that $\K$ is a **Fraïssé class** if it satisfies **HP**, **JEP**, **AP**, contains finite structures of arbitrarily large finite cardinality, and only countably many different isomorphism types. If $\KK$ is a Fraïssé structure then its $\Age(\KK)$ is a Fraïssé class. Given a Fraïssé class $\K$ we have its Fraïssé limit $\text{Flim}(\K)$, which is a Fraïssé structure and is unique up to isomorphism. In this way there is a 1-1 correspondence between Fraïssé classes and Fraïssé structures. For more details, see [@HO93]. We consider a structure as a tuple $\AA = (A, \{R_i^A\}_{i \in I}, \{f_j^A\}_{j \in J})$ where $A$ is the underlying set of the structure, $R_i^A$ is the interpretation of the relational symbol in $\AA$ and $f_j^A$ is the interpretation of the functional symbol in $\AA$ for all $i \in I$ and all $j \in J$. If $J = \emptyset$ then we say that the structure is **relational** (or that the signature is relational). All of the structures studied within are relational, so for ease of notation we will appropriate $J$ to also serve as an index set for a collection of relations. In particular, we consider a directed graph (digraph) as a structure in the binary relational signature $\{\rightarrow\}$. The symbol $\rightarrow$ is always interpreted as an irreflexive and asymmetric relation. For a directed graph $(A, \rightarrow^A)$ we sometimes use the symbol $\perp^A$ to denote $\neg(x \rightarrow^A y \vee y \rightarrow^A x)$. A **tournament** is a digraph $(A, \rightarrow^A)$ with the property that for every $x \neq y \in A$ we have either $x \rightarrow^A y$ or $y \rightarrow^A x$ (but not both). In general, we will use the following typefaces: $\L, \K$ for classes, $\AA, \BB, \CC, \KK$ for structures, $A,B,C$ for universes (or underlying sets) of structures (with $L$ being reserved for the signature of a class) and $a,b,c$ for elements of underlying sets. Occasionally we will use $a,b$ for natural numbers that index the number of equivalence classes in $\AA$ and $\BB$. Reducts and the expansion property. ----------------------------------- Let $L \sse L^*$ be given signatures. Let $\K$ be a class of structures in $L$ and let $\K^*$ be a class of structures in $L^*$. If $\AA^* \in \K^*$ then we denote by $\AA^* \vert L$ the structure in $\K$ obtained by dropping the interpretations of the symbols in $L^* \setminus L$ in $\AA^*$, and define $\K^* \vert L := \{\AA^* \vert L : \AA^* \in \K^*\}$. We say that $\K^*$ is a **precompact expansion** of $\K$ provided that $\forall \AA \in \K$ there are only finitely many $\AA^* \in \K^*$ such that $\AA^* \vert L = \AA$. We say that $\K^*$ satisfies **the expansion property (EP)** (with respect to $\K$) if $\K^* \vert L = \K$ and for every $\AA \in \K$ there is a $\BB \in \K$ such that for every $\AA^*, \BB^* \in \K^*$ with $\AA^* \vert L = \AA$ and $\BB^* \vert L = \BB$ we have $\AA^* \into \BB^*$. We say that $\K^*$ is a **reasonable expansion** of $\K$ provided that it is a precompact expansion and $\forall \AA, \BB \in \K$, for every embedding $\pi : \AA \longrightarrow \BB$, $\forall \AA^* \in \K^*$ with $\AA^* \vert L = \AA$, there is a $\BB^* \in \K^*$ such that $\BB^* \vert L = \BB$ and $\pi$ is also an embedding of $\AA^*$ into $\BB^*$. Ramsey property. ---------------- We say that the class $\K$ satisfies the **Ramsey Property (RP)** (or is a **Ramsey class**) if for every (small) $\AA \in \K$ and every (medium) $\BB \in \K$ and every $r \in \N$ there is a (large) $\CC \in \K$ such that for every colouring $c : \binom{\CC}{\AA} \rightarrow \{1, \ldots, r\}$ there is a $\overline{\BB} \in \binom{\CC}{\BB}$ such that $c \restrict \binom{\overline{\BB}}{\AA}$ is a constant. We denote this using the arrow notation: $$\CC \longrightarrow (\BB)_r^\AA.$$ Another notion that will always appear with **RP** is rigidity. We say that a given structure is **rigid** if it has no nontrivial automorphisms. Consistent random expansions. {#sec:cons_rand_exp} ----------------------------- Let $L \sse L^*$ be given signatures. Let $\K$ and $\K^*$ be classes of structures in $L$ and $L^*$ respectively such that $\K^*\vert L = \K$. If $\BB \in \K^*$ then we write $\BB^* = \expand{\AA}$ where $\AA = \BB^* \vert L$ and $\AA^* = \BB^* \vert (L^* \setminus L)$. Colloquially, “$\AA$ is the old stuff, and $\AA^*$ is the new stuff” when we use the representation $\expand{\AA}$. For $\AA \in \K$ we denote by $\mu_\AA$ a measure on the set $$\K^*(\AA) := \{\AA^* : \expand{\AA} \in \K^*\}.$$ We will also have need for the related quantity $$\#(\AA) := \vert \K^*(\AA) \vert,$$ which is the number of expansions of $\AA$ in $\K^*$. Let $\AA \leq \BB$ be structures in $\K$ and let $\expand{\AA} \in \K^*$. Then we write: $$\#_{\K^*} (\AA^*, \BB) := \left\vert \{ \BB^* : \expand{\AA} \leq \expand{\BB} \in \K^* \}\right\vert,$$ which is the number of expansions of $\BB$ in $\K^*$ that extend $\expand{\AA}$. If there is no confusion then we write $\#(\AA^*,\BB)$, or occasionally we will write $\#_{\AA,\BB}(\AA^*)$. We say that the collection $\{\mu_\AA : \AA \in \K\}$ is a **consistent random $\K^*$-expansion on $\K$** (when it is clear from context we suppress the reference to $\K^*$) if we have: - Each $\mu_\AA$ is a probability measure on $\K^*(\AA)$. - Whenever $\varphi: \AA \longrightarrow \BB$ is an embedding, and $\expand{\AA} \in \K^*$, we have $\displaystyle{\mu_\AA (\{\AA^*\}) = \sum \{\mu_\BB (\{\BB^*\}) : \varphi \text{ embeds } \expand{\AA} \text{ into } \expand{\BB}\}.}$ We have **P**robability measures and **E**xtension properties. When it is clear from context we shall refer to a consistent random expansion as $(\mu_\AA)$, with no reference to $\K$. In the special case that $\varphi$ in (**E**) is an isomorphism we get that $\varphi_* \mu_\AA = \mu_\BB$, where $\varphi_* \mu_\AA$ is the push forward measure; call this (**I**) for **I**somorphism invariance. We will reference it explicitly later on. We assume that for $\AA \cong \BB$ in $\K$ and $\expand{\AA} \in \K^*$ we have $\BB^*$ such that $\expand{\AA} \cong \expand{\BB}$. Let $L^* \setminus L$ be a relational signature, and let $\K$ and $\K^*$ be Fraïssé classes such that $\K^*$ is a reasonable expansion of $\K$. Then we say that $(\K, \K^*)$ is an **excellent pair** if: 1. $\K^*$ is a Ramsey class of rigid structures, and 2. $\K^*$ satisfies the expansion property relative to $\K$. Amenability via expansions. --------------------------- The following is the key equivalence used to show amenability and non-amenability of the automorphism group of a Fraïssé structure. The version that appears as Proposition 9.2 in [@AKL12] is a special case of what we state, and the proof of this version is analogous. The arguments in the proof are standard and the proof uses the Carathéodory Extension Theorem, so the proof is omitted. \[threeone\] Let $(\K, \K^*)$ be an excellent pair. Then: 1. $\Aut(\Flim(\K))$ is amenable iff $\K$ has a consistent random $\K^*$-expansion. 2. $\Aut(\Flim(\K))$ is uniquely ergodic iff $\K$ has a unique consistent random $\K^*$-expansion. We remark that a consistent random expansion $(\mu_\AA)$ cannot be degenerate, which means that when $\expand{\AA} \in \K^*$, we have $\mu_\AA(\{\AA^*\}) \neq 0$. Otherwise, since $\Flim(\K)$ is separable, a degenerate measure for $\AA^*$ would give us a countable cover of the universal minimal flow by open sets each with measure 0. For more details, see the proof of [@KSOK12 Proposition 2.1]. Compositions - $\EK$ and $\LEK$. {#sec:EK_LEK_def} -------------------------------- Now we introduce the composition class $\LEK$. We mainly focus on the quotient structure $\LEK$, but to introduce it we first mention the class $\EK$, which is used to define $\LEK$. Intuitively, a structure in $\LEK$ is a (horizontal) structure $\KK \in \K$ and associated to each point $k \in \KK$ is a (vertical, possibly different) $\LL_k \in \L$. The $\K$-relations of elements in different $\LL_k$ “columns” are given by looking at the $\K$-relations of the corresponding $k \in \KK$. In this way, if you “quotient out” by the equivalence relation of being in the same $\LL_k$ column, then you get $\KK$. Alternatively, one can think of a structure in $\LEK$ as taking a structure $\KK \in \K$ then “blowing-up” each of its points $k \in \KK$ to a structure $\LL_k \in \L$. Let $\K$ be a class of structures in $L_I := \{R_i : i \in I\}$, a relational signature where each $R_i$ has arity $n_i$, and let $\sim$ be a binary relational symbol such that $\sim \, \notin L_I$. We denote by $\EK$ the class of relational structures of the form $$\AA = (A, \{R^A_i\}_{i \in I}, \sim^A)$$ with the properties: 1. $\sim^A$ is an equivalence relation on $A$ with equivalence classes denoted by $[a]_{\sim^A}$. 2. For $i \in I$ and $x_1, \ldots, x_{n_i}, y_1, \ldots, y_{n_i} \in A$ with $x_j \sim^A y_j$ (for all $j \leq n_i$) we have $$R^A_i (x_1, \ldots, x_{n_i}) \Leftrightarrow R^A_i (y_1, \ldots, y_{n_i}).$$ 3. Let $\bigslant{A}{\sim^A} := \{[a]_{\sim^A} : a \in A\}$ be the set of equivalence classes. Let $R_i^\bigslant{A}{\sim^A}$, for $i \in I$, be the relation defined on the set $\bigslant{A}{\sim^A}$ according to (2) with $$R_i^\bigslant{A}{\sim^A}([a_1]_{\sim^A}, \ldots, [a_{n_i}]_{\sim^A}) \Leftrightarrow R_i^A(x_1, \ldots, x_{n_i})$$ where $a_j \sim^A x_j$ for all $j \leq n_i$. Then we have $$\bigslant{\AA}{\sim^A} := (\bigslant{A}{\sim^A}, \{R_i^\bigslant{A}{\sim^A}\}_{i \in I}) \in \K.$$ Let $L_I := \{R_i : i \in I\}$ and $L_J := \{R_j : j \in J\}$ be disjoint relational signatures and let $\sim$ be a binary relational symbol such that $\sim \, \notin L_I \cup L_J$. Let $\K$ and $\L$ be classes of relational structures in $L_I$ and $L_J$ respectively. We denote by $\LEK$ the class of relational structures of the form $$\AA = (A, \{R^A_i\}_{i \in I}, \{R^A_j\}_{j \in J}, \sim^A)$$ with the properties: 1. $\AA \vert (L_I \cup \{\sim\}) \in \EK$. 2. For $j \in J$ and $x_1, \ldots, x_{n_j} \in A$ we have $$R^A_j (x_1, \ldots, x_{n_j}) \Rightarrow [x_1]_{\sim^A} = \ldots = [x_{n_j}]_{\sim^A}.$$ 3. For $a \in A$ we have $$([a]_{\sim^A}, \{R_j^A \cap ([a]_{\sim^A})^{n_j}\}_{j \in J}) \in \L.$$ Expansions of $\LEK$. {#sec:LEK_exp} --------------------- Let $L_I^* \supset L_I$ and $L_J^* \supset L_J$ be relational signatures such that $L_I^* \cap L_J^* = \emptyset$ and $\sim \, \notin L_I^* \cup L_J^*$. If $\K^*$ and $\L^*$ are expansions of the classes $\K$ and $\L$ such that $\K^* \vert L_I = \K$ and $\L^* \vert L_J = \L$ then we have that $$\left(\LEKstar\right) \vert ( L_I \cup L_J \cup \{\sim\} ) = \LEK.$$ Let $\AA \in \LEK$ be the finite structure which has $A_1, A_2, \ldots, A_a$ as its $\sim^A$-equivalence classes. Let $\AA_1, \AA_2, \ldots, \AA_a$ be structures in $\L$ which are placed on $A_1, \ldots, A_a$ respectively and let $\BB \in \K$ be the structure given by representatives of the equivalence classes. Then we write $\AA = (\BB : \AA_1, \ldots, \AA_a)$. Similarly, an expansion $\expand{\AA} \in \LEKstar$ of $\AA \in \LEK$ is given by the structures $\expand{\AA_1}, \ldots, \expand{\AA_a} \in \L^*$ and $\expand{\BB} \in \K^*$. So we write $$\expand{\AA} = (\expand{\BB} : \expand{\AA_1}, \ldots, \expand{\AA_a})$$ or, if there is no confusion $$\AA^* = (\BB^* : \AA_1^*, \ldots, \AA_a^*).$$ Excellent pair proposition. {#sec:LEK_excellent} --------------------------- The following technical proposition ensures that $(\LEK, \LEKstar)$ is an excellent pair, thus we may apply Proposition \[threeone\] to verify amenability of $\Aut(\Flim(\LEK))$. \[ExPair\_LEK\] Let $L_I^* \supset L_I$ and $L_J^* \supset L_J$ be relational signatures such that $L_I^* \cap L_J^* = \emptyset$ and let $\sim$ be a binary relational symbol such that $\sim \, \notin L_I^* \cup L_J^*$. Let $\K,\K^*, \L$ and $\L^*$ be classes of finite relational structures in $L_I, L_I^*, L_J$ and $L_J^*$ respectively. Let $\K^* \vert L_I = \K$ and $\L^* \vert L_J = \L$. Then we have: 1. If $\L$ and $\K$ are Ramsey classes of rigid structures then $\LEK$ is a Ramsey class of rigid structures. 2. If $\L^*$ satisfies **EP** with respect to $\L$ and $\K^*$ satisfies **EP** with respect to $\K$, then $\LEKstar$ satisfies **EP** with respect to $\LEK$. This follows by simple modifications of the proofs of Theorem 4.4 and Proposition 5.2 in [@SOK13]. $\LEK$. {#sec:LEK} ------- The following theorem is the main result of this section. Establishing this theorem was the genesis of this larger project, and after it was established we expanded our aims to the other digraphs on Cherlin’s classification. The forward implication in each of the parts is a straightforward, if tedious combinatorial verification. It can also be derived at the level of topological groups. The converse is more subtle so we include its proof. \[A\_LEK\] Let $(\L,\L^*)$ and $(\K,\K^*)$ be excellent pairs of classes of finite structures in distinct signatures. Then we have: 1. $\Aut(\Flim(\LEK))$ is amenable iff $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are amenable. \[A\_LEK\_one\] 2. $\Aut(\Flim(\LEK))$ is uniquely ergodic iff $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are uniquely ergodic. As an (almost) immediate corollary we get the unique ergodicity of $\Aut(\TT[\II_n])$ and $\Aut(\II_n[\TT])$, which are both part of Cherlin’s classification. The proof of this is broken up into five not entirely independent parts. The consistent random expansions presented in the amenability proofs will be used in the unique ergodicity proofs. Moreover, in order to not overly repeat ourselves, detailed proofs that some maps are actually consistent random expansions will only appear in the amenability proofs. These proofs are “direct” in the sense that they do not rely on heavy machinery. The essential claim in these proofs is that an expansion in the composition class is a composition of expansions, but there are details that need to be checked. In what follows the summations will always range over $\K^*$-expansions of fixed structures in $\K$. Assume that $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are amenable. Then by Proposition \[threeone\] there are consistent random expansions $\nu$ and $\mu$ on $\L$ and $\K$ respectively. We will define a consistent random expansion $\nu \otimes \mu$ on $\LEK$. Let $\SS = (\AA : \SS_1, \ldots, \SS_a) \in \LEK$ and $\expand{\SS} \in \LEKstar$ such that $\SS^* = (\AA^* : \SS_1^*, \ldots, \SS_a^*)$. Define $$(\nu \otimes \mu)_\SS (\{\SS^*\}) := \mu_\AA (\{\AA^*\}) \cdot \prod_{i=1}^a \nu_{\SS_i} (\{\SS_i^*\}),$$ and we check the conditions for being a consistent random expansion. $(\textbf{P})$ Observe that $$\begin{aligned} &\space \sum \left\{(\nu \otimes \mu)_\SS (\{\SS^*\}) : \expand{\SS} \in \LEKstar\right\} \\ &= \sum \left\{\mu_\AA (\{\AA^*\}) \cdot \prod_{i=1}^a \nu_{\SS_i} (\{\SS_i^*\}) : \expand{\AA}\in\K^*, \expand{\SS_i} \in \L^*, \forall i \leq a\right\} \\ &= \left(\sum_{\expand{\AA}\in\K^*} \mu_\AA (\{\AA^*\}) \right) \cdot \sum \left\{\prod_{i=1}^a \nu_{\SS_i} (\{\SS_i^*\}) : \expand{\SS_i} \in \L^*, \forall i \leq a\right\} \\ &= 1 \cdot \prod_{i=1}^a \left\{\sum \nu_{\SS_i} (\{\SS_i^*\}) : \expand{\SS_i} \in \L^* \right\} \\ &= 1 \cdot \prod_{i=1}^a 1 = 1, \end{aligned}$$ where the third equality follows from (**P**) on $\mu$, and the fourth equality follows from (**P**) on $\nu$. The second equality follows from a basic fact about sums and products. (**E**) Let $\SS = (\AA : \SS_1, \ldots, \SS_a)$ and $\TT = (\BB : \TT_1, \ldots, \TT_b)$ be structures in $\LEK$ such that $\SS \leq \TT$. Let $\SS^* = (\AA^* : \SS_1^*, \ldots, \SS_a^*)$ be such that $\expand{\SS}\in \LEKstar$. Since $\SS \leq \TT$ there is an $I \sse \{1, \ldots, b\}$ such that $\SS_i \leq \TT_i$ if and only if $i \in I$. We take $J = \{1, \ldots, b\}\setminus I$. Then $$\begin{aligned} & \sum \{(\nu \otimes \mu)_\TT (\{\TT^*\}) : \expand{\SS} \leq \expand{\TT} \in \LEKstar \} \\ &= \sum \left\{ \mu_\BB(\{\BB^*\}) \cdot \prod_{i=1}^b \nu_{\TT_i}(\{\TT_i^*\}) : \begin{aligned} &\scriptstyle{\expand{\AA} \leq \expand{\BB} \in\K^* }\\ &\scriptstyle{\expand{\SS_i} \leq \expand{\TT_i}\in\L^*, \text{ for } i \in I} \\ &\scriptstyle{\expand{\TT_i} \in \L^*, \text{ for } i \in J} \end{aligned} \right\} \\ &= \left( \sum \{\mu_\BB (\{\BB^*\}) : \expand{\AA} \leq \expand{\BB} \in \K^*\} \right)\\ &\hspace{0.7cm}\cdot \sum \left\{ \prod_{i=1}^b \nu_{\TT_i} (\{\TT_i^*\}) : \begin{aligned} &\scriptstyle{\expand{\SS_i} \leq \expand{\TT_i} \in \L^*, \text{ for } i \in I} \\ &\scriptstyle{\expand{\TT_i}\in\L^*, \text{ for } i \in J} \end{aligned} \right\} \\ &= \mu_\AA (\{\AA^*\}) \cdot \prod_{i\in I} \left\{ \sum \nu_{\TT_i}(\{\TT_i^*\}) : \expand{\SS_i} \leq \expand{\TT_i} \in \L^* \right\} \\ &\hspace{0.7cm}\cdot \prod_{j\in J} \left\{ \sum \nu_{\TT_j}(\{\TT_j^*\}) : \expand{\TT_j} \in \L^* \right\} \\ &= \mu_\AA (\{\AA^*\}) \cdot \prod_{i \in I} \nu_{\SS_i}(\{\SS_i^*\}) \cdot \prod_{j \in J} 1 = \mu_\AA (\{\AA^*\}) \cdot \prod_{i =1}^a \nu_{\SS_i}(\{\SS_i^*\}) = (\nu \otimes \mu)_\SS (\{\SS^*\}),\end{aligned}$$ where the third equality comes from (**E**) of $\mu$, the fourth equality comes from (**E**) of $\nu$ and the second equality uses the same basic fact about sums and products. This completes the verification that $\nu \otimes \mu$ is a consistent random expansion, and by Proposition \[threeone\] we have that $\Aut(\Flim(\LEK))$ is amenable. Assume that $\Aut(\Flim(\LEK))$ is amenable. By Proposition (\[threeone\]) there is a consistent random expansion $\rho$ on $\LEK$. We will show that $\Aut(\Flim(\K))$ is amenable. Let $\KK \in \K$ and $\expand{\KK}\in\K^*$. Let $\SS = (\KK : \SS_1, \ldots, \SS_a) \in \LEK$. Consider $$\mu_{\KK,\SS}(\{\KK^*\}) := \sum \{\rho_\SS(\{\SS^*\}) : \SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*), \expand{\SS}\in \LEKstar\}.$$ First, we show that $\mu_{\KK,\SS}$ is independent of our choice of $\SS$. Let $\TT = (\KK : \TT_1, \ldots, \TT_a)$ also be a structure in $\LEK$. If $\SS \leq \TT$, then $$\begin{aligned} \mu_{\KK, \TT}(\{\KK^*\}) &= \sum \{\rho_\TT (\{\TT^*\}) : \TT^* = (\TT_1^*, \ldots, \TT_a^*: \KK), \expand{\TT}\in\LEKstar\} \\ &= \sum_{\substack{\SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*) \\ \expand{\SS}\in \LEKstar }} \sum\{\rho_\TT (\{\TT^*\}) : \expand{\SS} \leq \expand{\TT} \in \LEKstar\} \\ &= \sum_{\substack{\SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*) \\ \expand{\SS}\in \LEKstar }} \rho_\SS (\{\SS^*\}) \\ &= \mu_{\KK,\SS}(\{\KK^*\})\end{aligned}$$ Where the third equality follows from (**E**) of $\rho$. If $\SS$ is not a substructure of $\TT$, then by **JEP** for $\LEK$ there is an $\RR \in \LEK$ such that $\SS \leq \RR$ and $\TT \leq \RR$. So by the above we have: $$\mu_{\KK, \SS}(\{\KK^*\}) = \mu_{\KK, \RR}(\{\KK^*\}) = \mu_{\KK, \TT}(\{\KK^*\})$$ Therefore $\mu_{\KK, \SS}(\{\KK^*\})$ is independent of the choice of structure $\SS$, so without ambiguity, we write $$\mu_\KK (\{\KK^*\}) := \mu_{\KK, \SS}(\KK^*)$$ where $\SS = (\KK : \SS_1, \ldots, \SS_a) \in \LEK$. Now we check that $\mu_\KK$ is a consistent random expansion for $\K$. (**P**) Fix any $\SS = (\SS_1, \ldots, \SS_a: \KK) \in \LEK$. Observe that $$\begin{aligned} & \sum \{\mu_\KK (\{\KK^*\}) : \expand{\KK} \in \K^*\} \\ &= \sum \{\mu_{\KK,\SS}(\{\KK^*\}) : \expand{\KK} \in \K^*\} \\ &= \sum_{\expand{\KK} \in \K^*} \sum \{\rho_\SS(\{\SS^*\}) : \expand{\SS} \in \LEKstar, \SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*)\} \\ &= \sum \{\rho_\SS(\{\SS^*\}) : \expand{\SS} \in \LEKstar \} = 1,\end{aligned}$$ where the third equality follows from (**P**) for $\rho$. (**E**) Let $\KK \leq \LL$ be structures in $\K$ and let $\expand{\KK} \in \K^*$. Let $\SS \leq \TT$ be structures in $\LEK$, with $\SS = (\KK : \SS_1, \ldots, \SS_a)$ and $\TT = (\LL : \TT_1, \ldots, \TT_b)$. Let $I := \{1 \leq i \leq b : \SS_i \leq \TT_i\}$ and $J := \{1, \ldots, b\} \setminus I$. Then we have $$\begin{aligned} & \sum \{\mu_\LL (\{\LL^*\}) : \expand{\KK} \leq \expand{\LL} \in \K^*\} \\ &= \sum \{\mu_{\LL,\TT} (\{\TT^*\}) : \expand{\KK} \leq \expand{\LL} \in \K^*\} \\ &= \sum_{\expand{\KK} \leq \expand{\LL} \in \K^*} \sum \{\rho_\TT (\{\TT^*\}) : \expand{\TT} \in \LEKstar, \TT^* = (\LL^* : \TT_1^*, \ldots, \TT_b^*)\} \\ &= \sum_{\substack{\expand{\SS} \in \LEKstar \\ \SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*)}} \sum \left\{ \rho_\TT (\{\TT^*\}) : \begin{aligned} &\scriptstyle{\expand{\SS} \leq \expand{\TT} \in \LEKstar }\\ &\scriptstyle{\TT^* = (\LL^* : \TT_1^*, \ldots, \TT_b^*)} \end{aligned} \right\} \\ &= \sum_{\substack{\expand{\SS} \in \LEKstar \\ \SS^* = (\KK^* : \SS_1^*, \ldots, \SS_a^*)}} \rho_\SS (\{\SS^*\}) = \mu_{\KK, \SS} (\{\KK^*\}) = \mu_\KK (\{\KK^*\}),\end{aligned}$$ where the fourth equality is by (**E**) of $\rho$. So we have shown that $\Aut(\Flim(\K))$ is amenable by Proposition \[threeone\]. Assume that $\Aut(\Flim(\LEK))$ is amenable. By Proposition (\[threeone\]) there is a consistent random expansion $\rho$ on $\LEK$. We will show that $\Aut(\Flim(\L))$ is amenable. Let $\PP$ be a one point structure in $\K$ and let $\expand{\PP} \in \K^*$. For $\LL \in \L$ there is an $\SS \in \LEK$ such that $\SS = (\PP : \LL)$, and every $\expand{\LL}\in\L^*$ gives us an $\SS^* = (\PP^* : \LL^*)$ such that $\expand{\SS} \in \LEKstar$. Using the consistent random expansion $\mu$ we previously defined, we introduce $$\gamma_\LL (\{\LL^*\}) := \rho_\SS (\{\SS^*\}) \cdot \frac{1}{\mu_{\PP} (\{\PP^*\})}.$$ Note that we must have $\mu_\PP (\{\PP^*\}) \neq 0$ since $\rho_\SS (\{\SS^*\}) \neq 0$ for all $\SS \in \LEK$ and $\expand{\SS} \in \LEKstar$. We prove that $(\gamma_\LL)$ is a consistent random expansion of $\L$ by checking (**P**) and (**E**). (**P**) Observe that $$\begin{aligned} & \sum \{\gamma_\LL (\{\LL^*\}) : \expand{\LL}\in\L^*\} \\ &= \sum_{\expand{\LL}\in\L^*} \{\rho_\SS (\{\SS^*\}) \cdot \frac{1}{\mu_{\PP} (\{\PP^*\})} : \expand{\SS} \in \LEKstar, \SS^* = (\LL^* : \PP^*)\} \\ &= \frac{1}{\mu_{\PP} (\{\PP^*\})} \cdot \sum_{\expand{\LL}\in\L^*} \{\rho_\SS (\{\SS^*\}) : \expand{\SS} \in \LEKstar, \SS^* = (\PP^* : \LL^*)\} \\ &= \frac{1}{\mu_{\PP} (\{\PP^*\})} \cdot \mu_\PP (\{\PP^*\}) = 1,\end{aligned}$$ where the third equality follows from the definition of $\mu$. (**E**) Let $\LL \leq \KK$ be structures in $\L$. Then there are $\SS \leq \TT \in \LEK$ such that $\SS = (\PP : \LL)$ and $\TT = (\PP : \KK)$. Let $\expand{\LL} \in \L^*$. Then we have the following $$\begin{aligned} & \sum \{\gamma_\KK (\{\KK^*\}) : \expand{\LL} \leq \expand{\KK} \in \L^*\} \\ &= \sum \left \{\frac{1}{\mu_\PP (\{\PP^*\})} \cdot \rho_\TT (\{\TT^*\}) : \expand{\LL} \leq \expand{\KK} \in \L^*, \TT^* = (\PP^*:\KK^*)\right\} \\ &= \frac{1}{\mu_\PP (\{\PP^*\})} \cdot \sum \{\rho_\TT (\{\TT^*\}) : \expand{\SS} \leq \expand{\TT}, \SS^* = (\PP^*:\LL^*)\} \\ &= \frac{1}{\mu_\PP (\{\PP^*\})} \cdot \rho_\SS(\{\SS^*\}) = \frac{1}{\mu_\PP (\{\PP^*\})} \cdot \mu_\PP (\{\PP^*\}) \cdot \gamma_\LL (\{\LL^*\}) = \gamma_\LL (\{\LL^*\}),\end{aligned}$$ where the third equality follows from (**E**) for $\rho$. This finishes the verification that $\Aut(\Flim(\LEK))$ is amenable. Assume that $\Aut(\Flim(\LEK))$ is uniquely ergodic. So we have that $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are amenable, and by Proposition \[threeone\] there are consistent random expansions $\mu$ and $\gamma$ on $\K$ and $\L$ respectively. Suppose that one of $\Aut(\Flim(\L))$ or $\Aut(\Flim(\K))$ is not uniquely ergodic. Then there is a consistent random expansion $\mu^\prime$ on $\K$ such that $\mu \neq \mu^\prime$ or there is a consistent random expansion $\gamma^\prime$ on $\L$ such that $\gamma \neq \gamma^\prime$. Then there is a structure $\KK \in \K$ and an expansion $\expand{\KK}\in\K^*$ such that $$\mu_\KK (\{\KK^*\}) \neq \mu_\KK^\prime(\{\KK^*\})$$ or there is a structure $\LL \in \L$ and an expansion $\expand{\LL}\in\L^*$ such that $$\gamma_\LL(\{\LL^*\}) \neq \gamma_\LL^\prime(\{\LL^*\}).$$ Now consider the structure $\SS = (\KK : \LL, \ldots, \LL) \in \LEK$ with expansion $\expand{\SS}\in\LEKstar$ where $\SS^* = (\KK^* : \LL^*, \ldots, \LL^*)$, with $a := \vert \KK \vert$ many $\LL$. Using similar arguments to the proof of $[(1), \Leftarrow]$ of this theorem, we have that $\gamma\otimes\mu, \gamma\otimes\mu^\prime$ and $\gamma^\prime \otimes\mu$ are consistent random expansions on $\LEK$. In particular if $\mu \neq \mu^\prime$, we have $$\begin{aligned} (\gamma\otimes\mu)_\SS (\{\SS^*\}) &= \mu_\KK (\{\KK^*\}) \cdot \prod_{i=1}^a \gamma_{\LL} (\{\LL^*\}) \\ &\neq \mu^\prime_\KK (\{\KK^*\}) \cdot \prod_{i=1}^a \gamma_{\LL} (\{\LL^*\}) = (\gamma\otimes \mu^\prime)_\SS (\{\SS^*\}),\end{aligned}$$ and if $\gamma \neq \gamma^\prime$, then we have $$\begin{aligned} (\gamma\otimes\mu)_\SS (\{\SS^*\}) &= \mu_\KK (\{\KK^*\}) \cdot \prod_{i=1}^a \gamma_{\LL} (\{\LL^*\}) \\ &\neq \mu_\KK (\{\KK^*\}) \cdot \prod_{i=1}^a \gamma^\prime_{\LL} (\{\LL^*\}) = (\gamma^\prime \otimes\mu)_\SS (\{\SS^*\}).\end{aligned}$$ Therefore we have two distinct consistent random expansions on $\LEK$. This is in contradiction to the unique ergodicity of $\LEK$, according to Theorem \[threeone\], so $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ must be uniquely ergodic. Now assume that $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are uniquely ergodic, and let $\mu$ and $\gamma$ be the unique consistent random expansions on $\K$ and $\L$ respectively. According to the first part of this theorem there is a consistent random expansion on $\LEK$. We will show that $\rho := \gamma\otimes \mu$ is the unique consistent random expansion on $\LEK$, as defined in the previous part of the proof. Let $\SS = (\KK : \SS_1, \ldots, \SS_a)$ be a structure in $\LEK$ with expansion $\expand{\SS} \in \LEKstar$ given by $\SS^* = (\KK^*:\SS_1^*, \ldots, \SS_a^*)$. In the previous proof of \[(1), $\Rightarrow \Aut(\Flim(\K))$ is amenable\] we described a consistent random expansion on $\K$ given by $\rho$, so we may notice that by unique ergodicity, this is exactly $\mu$. Fix any $\TT_2, \ldots, \TT_a \in \L$ which will be used to define measures on $\L$. They can be $\SS_2, \ldots, \SS_a$ if you like, but for purposes of clarity we use $\TT_2, \ldots, \TT_a$. Define $$p_0 := \mu_\KK (\{\KK^*\}).$$ Let $\LL \in \L$ with $\expand{\LL}\in\L^*$ be given. Consider the map $$\begin{aligned} \gamma_{\LL}^1 (\{\LL^*\}) &:= \frac{1}{p_0} \cdot \sum \left\{\rho_{\XX_1} (\XX_1^*) : \begin{aligned} &\scriptstyle{\XX_1 = (\KK : \LL, \TT_2, \ldots, \TT_a) \in \LEK}\\ &\scriptstyle{\XX_1^* = (\KK^*:\LL^*, \TT_2^*, \ldots, \TT_a^*) \in \LEKstar} \end{aligned} \right\} \end{aligned}$$ Notice that the sum does not run over $\LL^*$ and $\KK^*$, which are fixed. In a similar way as the proof of \[(1), $\Rightarrow \Aut(\Flim(\L))$ is amenable\] we may conclude that $\gamma^1$ is a consistent random expansion on $\L$ that does not depend on the choice of $\TT_2, \ldots, \TT_a$. Since $\Aut(\Flim(\L))$ is uniquely ergodic we must have that $\gamma^1 = \gamma$. In particular, for $\LL = \SS_1$ we can define $$p_1 := p_0 \cdot \gamma_{\SS_1} (\{\SS_1^*\}).$$ Now consider the map $\gamma^2$, for $\LL \in \L$ with $\expand{\LL} \in \L^*$ given by: $$\begin{aligned} \gamma_{\LL}^2 (\{\LL^*\}) &:= \frac{1}{p_1} \cdot \sum \left\{\rho_{\XX_2} (\{\XX_2^*\}) : \begin{aligned} &\scriptstyle{\XX_2 = (\KK:\SS_1, \LL, \TT_3, \ldots, \TT_a) \in \LEK}\\ &\scriptstyle{\XX_2^* = (\KK^*:\SS_1^*, \LL^*, \TT_3^*, \ldots, \TT_a^*) \in \LEKstar} \end{aligned} \right\} \end{aligned}$$ Notice that the sum does not run over $\SS_1^*, \LL^*$ and $\KK^*$, which are fixed. Again similar arguments as in the previous proof show that $\gamma^2$ is a consistent random expansion on $\L$, and by unique ergodicity we have that $\gamma_2 = \gamma$. In particular we define: $$p_2 := p_1 \cdot \gamma_{\SS_2} (\{\SS_2^*\}).$$ Continuing on in this way we obtain $$p_a := \mu_\KK (\{\KK^*\}) \cdot \gamma_{\SS_1}(\{\SS_1^*\}) \cdot \ldots \cdot \gamma_{\SS_a}(\{\SS_a^*\})$$ with $p_a = \rho_\SS (\{\SS^*\})$. Therefore we have proved that $\rho = \gamma \otimes \mu$, so it must be unique. The corollaries. ---------------- For $n \leq \omega$, define $[n] := \{i \in \omega : i < n\}$. Let $\TT$ be one of the tournaments $\QQ, \SS(2), \TT^\omega$ or $\CC_3,$ and let $T$ be the underlying set of $\TT$. For $n \leq \omega$ we denote by $\TT[\II_n]$ the directed graph with the underlying set $T \times [n]$ and the edge relation given by $$(x,i) \rightarrow (y,j) \text{ iff } x \rightarrow y$$ and $\II_n[\TT]$ the tournament with the underlying set $[n]\times T$ and edge relation given by $$(i,x) \rightarrow (j,y) \text{ iff } (i=j, x \rightarrow y)$$ Consider $\II_n$ as a structure on the empty signature. Therefore we have the following: \[cor:LEK\_cor\] Let $\TT$ be one of the tournaments $\QQ, \SS(2), \TT^\omega$ or $\CC_3$, and let $n \leq \omega$. Then, 1. $\Aut(\SS(2)[\II_n])$ is not amenable. For $\TT \neq \SS(2)$, $\Aut(\TT[\II_n])$ is uniquely ergodic. 2. $\Aut(\II_n [\SS(2)])$ is not amenable. For $\TT \neq \SS(2)$, $\Aut(\II_n [\TT])$ is uniquely ergodic. Non-amenability of $\Aut(\SS(2))$ comes from unpublished communication with A.Kechris, and the argument is given in Theorem \[thm:Sn\_amen\]. In the case that $\TT[\II_n]$ or $\II_n[\TT]$ is a finite structure then its automorphism group is finite and therefore uniquely ergodic. So let us assume that they are infinite. Thus we may view the structures as Fraïssé limits of the form $\LLEKK := \Flim(\LEK)$ where both $\L$ and $\K$ are Fraïssé structures, or one of them (e.g. $\II_n$) is not a Fraïssé structure simply because it is not an infinite structure. This can be rectified by going through the proof of Theorem \[A\_LEK\] and noticing that the assumption that $\L$ and $\K$ are infinite is not used. Ramsey expansions of countable homogeneous directed graphs ========================================================== We now introduce the directed graphs that appear in Cherlin’s classification. More detailed descriptions will be given in the relevant sections. Summary and known results. -------------------------- Denote by $\II_n$ the edgeless directed graph on $n$ vertices, where $n \leq \omega$. Denote by $\CC_3$ the directed $3$-cycle. Specifically, $\CC_3 = (C_3, \rightarrow^{C_3})$ is the directed graph such that $C_3 = \{a,b,c\}$ with $a \rightarrow^{C_3} b, b \rightarrow^{C_3} c$ and $c \rightarrow^{C_3} a$. The following is the classification of countable homogeneous directed graphs (see [@Cher98]). The infinite structures here are Fraïssé structures. 1. The finite digraphs $\CC_3$ and $\II_n$ for $n < \omega$. 2. $\II_\omega$ is the edgeless directed graph on $\omega$ vertices. 3. $\QQ, \SS(2)$ and $\TT^\omega$ are tournaments. 1. $\QQ$ is the set of rational numbers where we take $x \rightarrow^\QQ y$ iff $x < y$. 2. $\SS(2)$ is the **dense local ordering** which may be seen as the set of points on the unit circle with rational arguments such that $e^{i \theta} \rightarrow^{S(2)} e^{i \phi}$ iff $0 < \phi - \theta < \pi$. See Section \[sec:def\_S(n)\]. 3. $\TT^\omega$ is the generic tournament, i.e. the Fraïssé limit of the class of all finite tournaments. 4. $\TT[\II_n], \II_n[\TT]$, where $n \leq \omega$, and $\TT$ is one of the tournaments in (3) or $\CC_3$. This is a type of “$n$-point cover (or blowup)" of the nodes of the tournament $\TT$. See Section \[sec:LEK\]. 5. $\hat{\TT}$, for $\TT = \II_1, \CC_3, \QQ$ or $\TT^\omega$, is a type of “two point cover (or blowup)" of the vertices of a tournament. See Section \[sec:def\_blowups\]. 6. $\DD_n$, for $1 < n \leq \omega$, the complete $n$-partite directed graph with countably many nodes. See Section \[sec:def\_Dn\]. 7. $\SS$ is the semi-generic graph. See Section \[sec:def\_semigeneric\]. 8. $\SS(3)$ is a directed graph which may be seen as the set of points on the unit circle with rational arguments such that $e^{i \theta} \rightarrow^{S(3)} e^{i \phi}$ iff $0 < \phi - \theta < \frac{2 \pi}{3}$. This is not a tournament. See Section \[sec:def\_S(n)\]. 9. $\PP$ is the generic poset, i.e. the Fraïssé limit of the class of all finite posets such that $x \rightarrow^P y$ iff $x <^P y$. 10. $\PP(3)$ is the “twisted” generic poset in three parts. See Section \[sec:def\_P(3)\]. 11. $\GG_n$, for $n > 1$, is the generic directed graph with the property that $\II_{n+1}$ can’t be embedded in $\GG_n$. See Section \[sec:hypergraph\]. 12. $\FF(\T)$ is the Fraïssé limit of the class of finite directed graphs which do not embed any member of $\T$, where $\T$ is a fixed set of finite tournaments each of which has at least three vertices; they are **F**orbidden. See Section \[sec:hypergraph\]. In this paper we go through Cherlin’s classification and examine the automorphism group of each of these structures with respect to amenability and unique ergodicity. We consider each automorphism group as a topological group with the pointwise convergence topology, see [@BK96] for more details. First we give a list of known facts: 1. $\Aut(\II_\omega) \cong S_\infty$ is amenable and uniquely ergodic, see [@GW02], [@AKL12 Proposition 10.1]. 2. $\Aut(\QQ)$ and $\Aut(\TT^\omega)$ are amenable and uniquely ergodic, see [@pes98], [@AKL12 Theorem 6.1, Theorem 2.2]. 3. $\Aut(\SS(2))$ and $\Aut(\SS(3))$ are not amenable, see [@ZUC13 Theorem 3.1] and private communication with Kechris. 4. $\Aut(\PP)$ is not amenable, see [@KSOK12 Section 3]. In addition, the essential arguments for the amenability and unique ergodicity of $\Aut(\GG_n)$ and $\Aut(\FF(\T))$ are contained in [@AKL12 Theorem 5.1] which shows a similar theorem for undirected graphs (with some combinatorial conditions). Amenability and unique ergodicity of $\Aut(\GG_n)$ and $\Aut(\FF(\T))$ are not direct corollaries of their theorem or their proof, modifications had to be made. Weak local orders, $\SS(2), \SS(3)$ and $\SS(n)$. {#sec:def_S(n)} ------------------------------------------------- Fix $2 \leq n < \omega$. The directed graph $\SS(n)$ may be seen as the set of points on the unit circle with rational arguments such that $e^{i \theta} \rightarrow^{S(n)} e^{i \phi}$ iff $0 < \phi - \theta < \frac{2 \pi}{n}$. This is a tournament iff $n=2$, and is homogeneous iff $n=2,3$. Let $I_k$, for $0 \leq k \leq n-1$ be unary relational symbols. We consider the structure $\SS(n)^*$ in the signature $\{\rightarrow, I_0, \ldots, I_{n-1}\}$ where: - $\SS(n)^* \vert \{\rightarrow\} = \SS(n)$, - $I_k^{\SS(n)}(x) \Leftrightarrow x \in e^{i\theta}$ and $\theta \in (\frac{k \cdot 2\pi}{n}, \frac{(k+1) \cdot 2\pi}{n})$, Set $\mathcal{S}(n) := \Age(\SS(n))$ and $\mathcal{S}(n)^* := \Age(\SS(n)^*)$. For $n \geq 4$, $\SS(n)$ is not a Fraïssé structure. For $n=2,3$, $(\mathcal{S}(n), \mathcal{S}(n)^*)$ is an excellent pair, see [@LNS10; @NVT13]. Partial order $\PP$ and a variant $\PP(3)$. {#sec:def_P(3)} ------------------------------------------- Let $\P$ be the class of finite posets in signature $\{\leq\}$. Let $P_0, P_1, P_2$ be unary relational symbols and let $\P_3$ be the class of finite structures of the form $(A, \leq^A, P_0^A, P_1^A, P_2^A )$ where $(A, \leq^A) \in \P$ and $A = \bigsqcup_{i=0}^2 \{x : P_i^A(x)\}$. It is easy to see that $\P_3$ is a Fraïssé class with limit $\PP_3 = (R, \leq^R, P_0^R, P_1^R, P_2^R)$. Using $\PP_3$ we define the structure $\PP(3) = (R, \rightarrow^R)$ such that for $x,y \in R$ with $P_i^R(x)$ and $P_j^R(y)$ we have $x \rightarrow^R y$ iff one of the following conditions is satisfied: 1. $j = i$ and $x <^R y$; or 2. $j = i+1 \mod 3$ and $y <^R x$; or 3. $j = i+2 \mod 3$ and $x$ is $<^R$-incompatible with $y$. {width="40.00000%"} The structure $\PP(3)$ is a Fraïssé structure, see [@Cher98], which is called the generic twisted poset. Note that each $P_i$ induces a copy of $\PP$ that is cofinal in $\PP(3)$. We also consider $\PP(3)^* = (R, \rightarrow^R, P_0^R, P_1^R, P_2^R, \preceq^R)$, which is also a Fraïssé structure, where $\preceq^R$ is a linear order on $R$ that extends the partial order $(R, \leq^R)$ given by untwisting $(R, \rightarrow^R, P_0^R, P_1^R, P_2^R)$. In what follows we will use $\leq$ for untwisted partial orders, $\rightarrow$ for the corresponding twisted directed graph, and $\preceq$ for the linear order that extends $\leq$. We will not refer to an untwisted partial order’s natural directed graph. Let $\P(3) := \Age(\PP(3))$ and $\P(3)^* := \Age(\PP(3)^*)$. The pair $(\P(3), \P(3)^*)$ is an excellent pair, see [@JLNW14]. 2-covers of tournaments, $\hat{\QQ}$ and $\hat{\TT^\omega}$. {#sec:def_blowups} ------------------------------------------------------------ Here we discuss a way of 2-covering (or blowing up) points of a tournament so that it has much of the same structure, but it is no longer a tournament. In the case of covering $\TT^\omega$, the Ramsey expansion is straightforward, being essentially convex linear orders. In the case of covering $\QQ$, the Ramsey expansion is more subtle and must interact suitably with the linear order of $\QQ$. We explain these expansions in some depth because their discussion was unexpectedly absent from [@JLNW14]; in particular we show that these expansions are indeed Ramsey expansions. ### The 2-cover $\hat{\TT}$. Let $\TT$ be one of the following tournaments: $\II_1, \CC_3, \QQ$ or $\TT^\omega$, and let $T$ be its underlying set, and let $\T := \Age(\TT)$. We consider the structure $\hat{\TT}$ with underlying set[^1] $\hat{T} = T \times \{C,P\}$ and edge relation $\rightarrow^{\hat{T}}$ given by: $$(x,i) \rightarrow^{\hat{T}} (y,j) \Leftrightarrow \left((x \rightarrow^T y, i\neq j) \vee (y \rightarrow^T x, i=j)\right).$$  ### Description of $\hat{\T}$ and $\T^*$. For $\TT \in \{\II_1, \CC_3\}$ the structure $\hat{\TT}$ is finite so $\Aut(\hat{\TT})$ is finite, hence uniquely ergodic. For $\TT \in \{\QQ, \TT^\omega\}$, $\hat{\TT}$ is a Fraïssé structure with corresponding Fraïssé class $\hat{\T} := \Age(\hat{\TT})$. Let $\T^*$ be the collection of structures of the form $(A, \rightarrow^A, \leq^A)$ where $(A, \rightarrow^A) \in \T$ and $\leq^A$ is a linear order on $A$. So $(\T, \T^*)$ is an excellent pair, by [@AH78] and [@NR77; @NR83; @NR89]. ### Description of $\hat{\T}^*$. For each structure $(A, \rightarrow^A) \in \hat{\T}$, the relation $\perp^A$ is an equivalence relation which gives the partition $A = A_1 \sqcup \ldots \sqcup A_k$ where each class has at most two elements. In the following we describe a Fraïssé class $\hat{\T}^*$ such that $(\hat{\T},\hat{\T}^*)$ is an excellent pair. Let $\hat{\T}^*$ contain structures of the form $(A, \rightarrow^A, \leq^A, I_0^A, I_1^A)$ where: - $(A, \rightarrow^A) \in \hat{\T}$, - $\leq^A$ is a linear order on $A$, - $I_0^A$ and $I_1^A$ are unary relations on $A$ which partition $A$, and - each $A_i$ is an interval with respect to $\leq^A$. Though a slight abuse of notation, we denote by $\leq^A$ the linear ordering induced by $\leq^A$ on the set $\{A_1, \ldots, A_k\}$. The correct expansion of $\hat{\T}$ for $\T = \Age(\TT^\omega)$ allows arbitrary partitions by $I_0$ and $I_1$ in each $\perp$-equivalence class, so long as the smallest element in each equivalence class, with respect to the linear ordering, belongs to $I_0$. For $\T = \Age(\QQ)$ there is an additional subtlety which we will explain in the following section; essentially the expansion must be given by a transversal that coheres with the inherent linear order of $\QQ$, but for $\TT^\omega$ there is no such linear order to cohere with. This is similar, although not identical, to how the correct expansions for the generic partial orders are the linear orders that extend the partial orders. The proofs of the following two facts are relegated to the appendix. $\hat{\T}^*$ is a Ramsey class. $\hat{\T}^*$ satisfies the **EP** with respect to $\hat{\T}$. ### Description of $\hat{\T}^*$ for $\T = \Age(\QQ)$. Let $\AA = (A, \rightarrow^A) \in \hat{\T}$ be a structure with $k$ many $\perp^A$-equivalence classes $A_1, \ldots, A_k$ such that $\vert A_i \vert =2$ for all $i \leq k$. Suppose that $A_i = \{(i,C), (i,P)\}$ and that for $i \neq j$ we have: - $(i,C) \rightarrow^A (j,C) \Leftrightarrow i > j$; - $(i,P) \rightarrow^A (j,P) \Leftrightarrow i > j$; - $(i,P) \rightarrow^A (j,C) \Leftrightarrow i < j$; - $(i,C) \rightarrow^A (j,P) \Leftrightarrow i < j$.  Now we examine when a transversal $T \sse \AA$ forms a linear ordering, that is, a linear order on $T$ that also gives rise to the induced subgraph on $T$. The following lemma says that $T$ forms a linear order so long as there is at most one change of levels, and no zigzags. It also establishes that if a structure $\AA$ has $k$ equivalence classes each with two points, then $\AA$ has exactly $2k$ expansions. \[lem:l\]A sequence of vertices $(a_i)_{i=1}^k$ with $a_i \in A_i$ forms a linear ordering iff 1. All $a_i$ have the same second coordinate; or 2. There is $l < k$ such that for all $i < l$, the $a_i$ have the same second coordinate $m$, and for all $i \geq l$ the $a_i$ have the same second coordinate $n \neq m$. (See Figure \[fig:l\].) It is enough to consider the following. Let $i < j < k$ and $m \neq n$. Then we have: $$(i,m) \rightarrow^A (j,n) \rightarrow^A (k,m) \rightarrow^A (i,m)$$ so the directed graph induced by $(i,m), (j,n), (k,m)$ is not a linear ordering.  ![A linear order where $a_{l-1} < \ldots < a_1 < a_k < \ldots < a_l$.[]{data-label="fig:l"}](blowup_line_parts) We define an expansion class $\hat{\T}^*$ using such a sequence $\vec{a} = (a_i)_{i=1}^k$ in $A$, which forms a linear ordering $\leq^{\vec{a}}$. Then we introduce indicators $I_0^A$ and $I_1^A$ such that for $x \in A$ we have: $$\begin{aligned} I_1^A (x) &\Leftrightarrow x \in \{a_1, \ldots, a_k\} \\ I_0^A (x) &\Leftrightarrow x \notin \{a_1, \ldots, a_k\} \\\end{aligned}$$ We define a linear ordering $\leq^A$ on $A$ such that for $x \in A_i$ and $y \in A_j$ we have: $$x <^A y \Leftrightarrow \left((i=j,I_1^A(x)) \vee (a_i <^{\vec{a}} a_j)\right).$$ In the case where some columns of $\AA$ do not have two elements, we are a little more careful. For every structure $\AA \in \hat{\T}$ there is a unique structure $\BB \in \hat{\T}$, up to isomorphism, which contains the same number of $\perp^B$-equivalence classes of $\AA$ each of which has exactly two elements. Thus we may define an expansion of $\BB$, then by taking the restriction to $\AA$, we get an expansion of $\AA$. \[lem:hatT\_iso\] For all $\AA \in \hat{\T}$ and all $\expand[a]{\AA}, \expand[b]{\AA} \in \hat{\T}^*$, the structures $\expand[0]{\AA}$ and $\expand[1]{\AA}$ are isomorphic. Therefore $(\hat{\T}, \hat{\T}^*)$ satisfies the **EP**. Let $\vec{b} = (b_i)_{i=1}^k$ be a sequence in $A$ given by $b_i = (i,C)$ for each $i \leq k$. Then $\vec{b}$ forms a linear ordering and it induces an expansion of $\AA$, call it $\AA^b = (A, I_0^b, I_1^b, \leq^b)$ (where the vector notation is dropped for readability). Let $\vec{a}$ be another sequence in $A$ which induces the expansion $\AA^a$. Consider the map $\pi_{\vec{a}} : A \longrightarrow A$ given by: $$\pi_{\vec{a}}(i,m) = \begin{cases} (i+l, m) &: 1 \leq i \leq l \\ (i-l,1-m) &: l+1 \leq i \leq k \end{cases}$$ where $l$ is given by Lemma \[lem:l\]. It is easy to see that $\pi_{\vec{a}}$ is an automorphism of $\AA$ and moreover that $\pi_{\vec{a}}$ is an isomorphism between $\expand[a]{\AA}$ and $\expand[b]{\AA}$. Therefore any two expansions of $\AA$ are isomorphic. Complete $n$-partite directed graph, $\DD_n$. {#sec:def_Dn} --------------------------------------------- For $n\in\NN$ let $\D_n$ be the class of finite digraphs $(A, \rightarrow^A)$ in which $\perp^A$ is an equivalence relation with at most $n$ many equivalence classes. We will also consider $\D_\omega := \bigcup_{n < \omega} \D_n$. In this way we obtain Fraïssé classes $\D_n$, for $n \leq \omega$, with corresponding Fraïssé limits $\DD_n$, for $n \leq \omega$. We denote by $\D_\omega^*$ the class of finite structures of the form $(A, \rightarrow^A, \leq^A)$ where $(A, \rightarrow^A) \in \D_\omega$ and $\leq^A$ is a linear order on $A$ such that $\forall x,y,z \in A$ we have $$x <^A y <^A z, x \perp^A z \Rightarrow x \perp^A y \perp^A z,$$ which is a type of **convexity**. For a finite $n$, we let $\D_n^*$ be the class of finite structures of the form $(A, \rightarrow^A, \leq^A, \{I_i^A\}_{i=1}^n)$ where $(A, \rightarrow^A, \leq^A) \in \D_\omega^*$ and each $I_i^A$ is a unary relation on $A$ such that $\forall x,y \in A$ we have: - $(\exists i \leq n)(I_i^A(x))$, - $I_i^A(x), x \perp^A y \Rightarrow I_i^A(y)$, - $I_i^A(x), I_j^A(y), i < j \Rightarrow x<^A y$. The $I_i^A$ indicate the $n$-parts, and the ordering is convex with respect to the parts. Semi-generic multipartite digraph $\SS$ and variants $\SS_\leq, \SS_R$. {#sec:def_semigeneric} ----------------------------------------------------------------------- ### Description of $\SS$. Let $\mathcal{S}$ be the class of finite directed graphs of the form $(S, \rightarrow^S)$ with the following properties: 1. The binary relation $\perp^S$ defined on $S$ by $x \perp^S y \Leftrightarrow \neg(x \rightarrow^S y \vee y \rightarrow^S x)$ is an equivalence relation on $S$. (We call the equivalence classes **columns**.) 2. For $x,y,t_x,t_y \in S$, where $x \perp^S t_x$ and $y \perp^S t_y$, we have that the number of edges directed from $\{x,t_x\}$ to $\{y,t_y\}$ is even. {width="80.00000%"} Condition (1) ensures that the digraphs are complete $n$-partite, for some $n$. The parity condition (2) might seem artificial, but it has the following nice property which says “If you know three edges, then you know the fourth edge”. Let $\AA = (\{x,y,t_x,t_y\}, \rightarrow^S)$, with $x \perp^S t_x$, $y \perp^S t_y$, $\neg(t_x \perp^S t_y)$. If 3 of the directed edges between $\{x,t_x\}$ and $\{y,t_y\}$ are specified, then there is a unique directed edge between $\{x,t_x\}$ and $\{y,t_y\}$ that, when added, satisfies the parity condition.  Let $\mathcal{S}^*$ be the class of finite structures of the form $(A, \rightarrow^A, R^A, \leq^A)$ where $(A, \rightarrow^A) \in \mathcal{S}$, $R^A$ is a binary relation on $A$ and $\leq^A$ is a linear ordering on $A$ with the property that: 1. If $a$ is the number of $\perp^A$-equivalence classes, then there is a linear ordering $T = \{t_1 \leq t_2 \leq \ldots \leq t_a\}$, called a **transversal**, which we consider as a directed graph $\TT = (T, \rightarrow^T) \in \mathcal{S}$ given by $t_i \rightarrow t_j \Leftrightarrow t_i < t_j$. Then there is a $\BB = (B, \rightarrow^B) \in \mathcal{S}$ such that $\AA \leq \BB, \TT \leq \BB$ and $\BB$ also has $a$ many $\perp^B$-equivalence classes. Also, $\TT$ must be defined on each of the columns of $\AA$ (that is, $\forall x \in \AA, \exists t_i$ such that $x \perp t_i$). See Lemma \[lem:semi\_amalgam\] for further discussion. 2. If $R^A(x,y)$, then $\neg(x \perp^A y)$. If $t_i \perp^B x$, then we have $$R^A (x,y) \Leftrightarrow t_i \rightarrow^B y.$$ 3. If $x \perp^A z$ and $x <^A y <^A z$ then $x \perp^A y \perp^A z$. If $t_i \perp^B x$ and $t_j \perp^B y$, then $x <^A y \Leftrightarrow t_i < t_j$. (This is a type of **convexity**.) The condition (2), and the three of four lemma, ensures that the digraph structure of $\TT$ amalgamated with $\AA$ can be reconstructed from the transversal and the relations $R^A(x,y)$ (and *vice versa*). The condition (3) says that the linear order is **convex** with respect to the $\perp^A$-equivalence classes.  The classes $\mathcal{S}$ and $\mathcal{S}^*$ are Fraïssé classes with limits $\SS$ and $\SS^*$ respectively, see [@Cher98; @JLNW14]. \[lem:semi\_amalgam\] Every structure $\AA \in \mathcal{S}$ can be amalgamated with a transversal with the same columns. Moreover, the transversal can be chosen to respect an arbitrary linear order of the columns of $\AA$. Let $C_i$ for $i \leq a$ be an enumeration of the columns of $\AA$. From each column choose a $c_i \in C_i$. We will amalgamate a transversal $\TT = (T, \rightarrow^T) \in \mathcal{S}$, where $T = \{t_1 \leq t_2 \leq \ldots \leq t_a\}$, and $t_i \perp c_i$ for each $i\leq a$. First we describe a digraph structure on the nodes $X = \{c_i : i \leq a\} \cup \{t_i : i \leq a\}$. Between $c_i$ and $c_j$ maintain the same edge direction as in $\AA$, and add an edge from $t_i$ to $t_j$ iff $i < j$. For each $i \neq j$ there are many possible choices for the edges between $\{c_i, t_i\}$ and $\{c_j, t_j\}$ so that it respects the parity condition. Note that, in terms of the parity condition, the edges between $\{c_i, t_i\}$ and $\{c_j, t_j\}$ don’t interact with the edges to any other columns. Denote by $\XX$ this digraph structure on $X$. By the Strong Amalgamation Property for $\mathcal{S}$, $\AA$ and $\XX$ can be amalgamated along $\AA \restrict \{c_i : i \leq a\}$, which yields the desired result. ### The Relation $R$. Fixing a point $x \in \SS$, the relation $R_x$ induces an equivalence relation with two classes on each other column $A$ which does not contain $x$, where $$R_x (y) \Leftrightarrow R(x,y).$$ The three of four lemma ensures that if $x \perp x^\prime$, then $R_x$ and $R_{x^\prime}$ induce the same partition on $A$. Thus we refer to the **partition of a column** $A_i$ given by another column $A_j$. In fact, any $N$ columns $A_1, \ldots, A_N$ induce an equivalence relation on each other column $A$, which contains $2^N$ equivalence classes (some of which may be empty). These column partitions are equivalence relations that are finer than the $\perp$-equivalence relation, but to avoid confusion we shall refer to *column partitions* and $\perp$-*equivalence relations*. ![The column partitions two columns induce on each other. Notice the relative positions of $R_x$ and $R_y$.[]{data-label="fig:semi_R"}](semi_column_partition) For more discussion about the relation $R$ and column partitions see [@JLNW14 Section 10]. ### Variants. Now we provide some related expansions of $\mathcal{S} := \Age(\SS)$. These variants will allow us to provide partial results towards the unique ergodicity of $\Aut(\SS)$. We have a good understanding of each of the variants separately, but when combined they form $\SS$, which we do not fully understand. Let $\mathcal{S}_R := \mathcal{S}^* \vert \{\rightarrow, R\}$ and $\mathcal{S}_\leq := \mathcal{S}^* \vert \{\rightarrow, \leq\}$. It is not hard to see that $\mathcal{S}_R$ and $\mathcal{S}_\leq$ are Fraïssé classes, and that $(\mathcal{S}_R, \mathcal{S}^*)$ and $(\mathcal{S}_\leq, \mathcal{S}^*)$ are excellent pairs. Let $\SS_R := \Flim(\mathcal{S}_R)$ and $\SS_\leq := \Flim(\mathcal{S}_\leq)$. ### Generic omitters. {#sec:def_omitters} Here we introduce two Fraïssé structures, $\GG_n$ and $\FF(\T)$, that are generic subject to the condition that they omit a specific class of finite directed graphs. Let $\G_n$ ($n \geq 2$) be the class of finite directed graphs $\GG$ with the property that $\II_{n+1} \not\into \GG$. This is a Fra[ï]{}ss[é]{} class with limit $\GG_n := \Flim(\G_n)$. Let $\G_n^*$ be the class of structures of the form $(A, \rightarrow^A, \leq^A)$ with the property that $(A, \rightarrow^A) \in \G_n$ and $\leq^A$ is a linear order on $A$. Then $\G_n^*$ is a Fra[ï]{}ss[é]{} class and moreover $(\G_n, \G_n^*)$ is an excellent pair. This can be seen by using a partite method construction as in [@NR77; @NR83; @NR89], or by introducing new relation for “not edge". Let $\T$ be a collection of finite tournaments with $\vert T \vert \geq 3$ for all $T \in \T$, and let $\F(\T)$ denote the class of finite directed graphs $\AA$ with the property that $\BB \not\into \AA$ for all $\BB \in \T$. Then $\F(\T)$ is a Fra[ï]{}ss[é]{} class with limit $\FF(\T) := \Flim(\F(\T))$. Let $\F^*(\T)$ denote the class of finite structures of the form $(A, \rightarrow^A, \leq^A)$ with the property that $(A, \rightarrow^A) \in \F(\T)$ and $\leq^A$ is a linear order on $A$. Then $\F^*(\T)$ is a Fra[ï]{}ss[é]{} class and moreover $(\F(\T), \F^*(\T))$ is an excellent pair, again see [@NR77; @NR83; @NR89] and [@AH78]. Amenability {#sec:Amen} =========== Now we are in a position to verify that many of the previously mentioned structures have amenable automorphism groups. After establishing a sufficient density condition for amenability and unique ergodicity, we verify that $\DD_n, \hat{\TT^\omega}, \SS, \GG_n$ and $\FF(\T)$ all have amenable automorphism groups. After establishing Theorem \[dense\] verifying that these automorphism groups are amenable will amount to a routine counting of expansions of a structure. Density result about amenability. {#sec:dense} --------------------------------- Let $\F$ be a class of finite structures and let $\D \sse \F$. We say that $\D$ is **cofinal** (or dense) in $\F$ if for every $\AA \in \F$ there is a $\DD \in \D$ such that $\AA \leq \DD$. See section \[sec:cons\_rand\_exp\] for notation. \[dense\] Let $(\K,\K^*)$ be an excellent pair of Fraïssé classes. If for every $\AA \leq \BB$ in $\K$ and every $\expand[\prime]{\AA}, \expand[\prime\prime]{\AA} \in \K^*$ we have $$\begin{aligned} \#(\AA^\prime, \BB) = \#(\AA^{\prime\prime}, \BB) \label{ass:samenumberofexp}\end{aligned}$$ then $\Aut(\Flim(\K))$ is amenable. Moreover, if there is a cofinal subclass $\D \sse \K$ with the property that for every $\DD \in \D$ and every $\expand[\prime]{\DD}, \expand[\prime\prime]{\DD} \in \K^*$ we have $\expand[\prime]{\DD} \cong \expand[\prime\prime]{\DD}$, then $\Aut(\Flim(\K))$ is uniquely ergodic. For $\AA \in \K$ and $\expand[0]{\AA}\in \K^*$ define $$\mu_\AA (\{\AA^0\}) := \frac{1}{\#(\AA)}.$$ We check that $(\mu_\AA)_{\AA \in \K}$ defines a consistent random expansion. Condition (**P**) is clear so it remains to check (**E**). For $\AA \leq \BB \in \K$ and $\expand[0]{\AA} \in \K^*$ we have $$\begin{aligned} \#(\AA) \cdot \mu_\AA (\{\AA^0\}) &= 1 = \sum \{\mu_\BB (\{\BB^*\}) : \expand{\BB}\in\K^*\}\\ &= \sum_{\substack{\AA^* \\ \expand{\AA}\in\K^*}} \sum \{ \mu_\BB (\{\BB^*\}) : \expand{\AA} \leq \expand{\BB} \in \K^*\}\\ &= \#(\AA) \cdot \sum \{\mu_\BB (\{\BB^*\}) : \expand[0]{\AA} \leq\expand{\BB}\in\K^*\}\\ \Rightarrow \mu_\AA (\{\AA^0\}) &= \sum \{\mu_\BB (\{\BB^*\}) : \expand[0]{\AA} \leq \expand{\BB}\in\K^*\}.\end{aligned}$$ Here, (\[ass:samenumberofexp\]) is used for the fourth equality. By Proposition \[threeone\] we obtain that $\Aut(\text{Flim}(\K))$ is an amenable group. In order to show unique ergodicity, it is enough to show uniqueness of a consistent random $\K^*$ expansion by Proposition \[threeone\]. Let $(\mu_\AA)_{\AA \in \K}$ and $(\gamma_\AA)_{\AA\in\K}$ be two consistent random $\K^*$ expansions. From our assumptions and (**I**) we obtain that $\mu_\DD \equiv \gamma_\DD$ for all $\DD \in \D$. Fix $\AA \in \K$, and find $\DD \in \D$ such that $\AA \leq \DD$. For any fixed $\expand[0]{\AA}\in\K^*$ we have: $$\begin{aligned} \mu_\AA (\{\AA^0\}) &\overeasy \sum\{\mu_\DD (\{\DD^*\}) : \expand[0]{\AA} \leq \expand{\DD} \in \K^*\} \\ &= \sum\{\gamma_\DD (\{\DD^*\}) : \expand[0]{\AA} \leq \expand{\DD} \in \K^*\} \\ &\overeasy \gamma_\AA (\{\AA^0\})\end{aligned}$$ Therefore $\mu_\AA \equiv \gamma_\AA$ for all $\AA \in \K$ and uniqueness is verified. Let us immediately show the usefulness of this result. The class of finite rooted binary trees is not on Cherlin’s classification (as it is not a directed graph), but this density result gives amenability very quickly. ### Binary trees. {#subsec:binary_trees} Let $\B$ be the class of finite rooted binary trees. For $\BB \in \B$ we define $T(\BB)$ to be the set of terminal nodes of the tree $\BB$, and we define $\Delta(\BB)$ to be the structure $(T(\BB), C^\BB)$ where $C^\BB$ is a ternary relation on $T(\BB)$. We define $C^\BB$ such that for $x,y,z \in T(\BB)$ we have: $$\begin{aligned} C^\BB(x,y,z) \Leftrightarrow \space & x,y,z \text{ are distinct and the shortest path from } x \text{ to the root}\\ &\text{is disjoint from the shortest path from } y \text{ to } z.\end{aligned}$$ In this way we assign to each $\BB \in \B$ a unique $\Delta(\BB)$, but also each structure $\Delta(\BB)$ gives the unique binary tree $\TT$ with the fewest nodes such that $\Delta(\TT) = \Delta(\BB)$. Let $\H$ be the class of the structures of the form $\Delta(\BB)$ for $\BB \in \B$. Let $\O\H$ be the class of structures of the form $(A, C^A, \leq^A)$ where $(A,C^A) \in \H$ and $\leq^A$ is a linear ordering of $A$. We say that $\leq^A$ is **convex** on $(A,\leq^A)$ if $$C^A(x,y,z) \Rightarrow (x <^A y \wedge x<^A z) \vee (y <^A x \wedge z <^A x).$$ For $\BB \in \B$ and $b \in \BB$ we write $\lev(b) = n$ if the shortest path from $b$ to the root has $n$ edges. We also write $\BB(n) := \{b \in \BB : \lev(b) = n\}$ and $\BB \restrict n$ denotes the subtree given by $\{b \in \BB : \lev(b) \leq n\}$. Define $\BB[b]$ to be the subtree of $\BB$ with root $b$ which contains vertices in $\BB$ whose shortest path to the root of $\BB$ contains $b$. We say that $\BB \in \B$ is an **$n$-nice tree** if $T(\BB) = \BB(n)$ and $\vert \BB(n) \vert = 2^n$. We consider the cofinal subclass $\D \sse \H$ which is the collection of structures of the form $(A, C^A)$ for which there exists an $n$-nice tree $\BB \in \B$ such that $\Delta(\BB) = (A, C^A)$. We also have the subclass $\O\D \sse \O\H$, which contains structures in $\D$ augmented by a linear order. Let $\CH \sse \OH$ that contains the structures with convex linear orderings. Let $\COH$ be the class of structures of the form $(A, C^A, \leq^A, \preceq^A)$ where $(A,C^A,\leq^A) \in \OH$ and $(A,C^A, \preceq^A) \in \CH$. We have that $\H, \OH, \CH$ and $\COH$ are all Fraïssé classes, see [@AN98; @BOD12]. Moreover, we have the following. \[cor:OH\_amenable\] 1. $\Aut(\text{Flim} (\H))$ is uniquely ergodic. 2. $\Aut(\text{Flim} (\O\H))$ is amenable. \(i) We have that $(\H, \CH)$ is an excellent pair, see [@MIL79]. We will check the conditions in Theorem \[dense\] for the cofinal subclass $\D$. Fix $\AA \leq \BB$ structures in $\H$ and let $\expand[\prime]{\AA} \in\CH$. Let $\UU \in \B$ be the smallest tree such that $\Delta(\UU) = \BB$. Since this is the smallest tree, each non-terminal node has degree $3$ or $2$. Therefore $\#(\BB) = 2^b$, where $b$ is the number of non-terminal nodes in $\UU$. Similarly, if $\VV$ is the smallest tree such that $\Delta(\VV)=\AA$ then we have $\#(\AA) = 2^a$, where $a$ is the number of non-terminal nodes in $\VV$. Therefore we have that $\#(\AA^\prime,\BB) = 2^{b-a}$ only depends on $\AA$ and $\BB$, not $\AA^\prime$. \(ii) We have that $(\OH, \COH)$ is an excellent pair, see [@SOK15]. The conditions in Theorem \[dense\], part 1 follows as in the previous case, but because of rigidity, there is no cofinal class with the isomorphism condition as in part 2. We delay verifying unique ergodicity until Proposition \[prop:UE\_OH\]. Generic multipartite digraph $\DD_n$. ------------------------------------- \[thm:amen\_Dn\] For $n \leq \omega$, $\Aut(\DD_n)$ is amenable. Since $(\D_n, \D_n^*)$ is an excellent pair, see [@JLNW14 Theorem 8.6] for $n=\omega$ and [@JLNW14 Theorem 8.7] for $n < \omega$, it is enough to show that there is a consistent random $\D_n^*$ expansion of $\D_n$, by Proposition \[threeone\]. This will be done using Theorem \[dense\]. Let $\AA = (A, \rightarrow^A)$ and $\BB = (B, \rightarrow^B)$ be structures in $\D_\omega$ such that $\AA \leq \BB$, and let $\leq^A$ be a linear order on $A$ such that $(\AA, \leq^A) \in \D_\omega^*$. Let $A_1, \ldots, A_a$ be the $\perp^A$-equivalence classes of $\AA$ and let $B_1, \ldots, B_b$ be $\perp^B$-equivalence classes of $\BB$. Without loss of generality we may assume that $\leq^A$ induces a linear ordering on the $\perp^A$-equivalence classes such that $A_1 <^A \ldots <^A A_a$. There are $1 \leq i_1 < i_2 < \ldots < i_a \leq b$ such that $A_j \sse B_{i_j}$ for each $1 \leq j \leq a$. If $(\BB, \leq^B) \in \D_\omega^*$ is such that $(\AA, \leq^A) \leq (\BB, \leq^B)$, then $\leq^B$ and $\leq^A$ induce the same linear ordering on $\{B_{i_1}, \ldots, B_{i_a}\}$ and they agree on each $A_i$. Therefore we have: $$\begin{aligned} \#_{\AA,\BB} (\leq^A) &:= \left\vert\{ \leq^B : (\AA, \leq^A) \leq (\BB, \leq^B) \}\right\vert \\ &= \frac{b!}{a!} \cdot \prod_{k=1}^a \frac{\vert B_{i_k} \vert!}{\vert A_k \vert!} \cdot \prod_{k \notin \{i_1, \ldots, i_a\}} \vert B_k \vert ! \\ &= \frac{b!}{a!} \cdot \frac{\prod_{k=1}^b \vert B_{i_k} \vert!}{\prod_{k=1}^a \vert A_k \vert!}.\end{aligned}$$ Moreover, for $\AA_1, \AA_2 \in \binom{\BB}{\AA}$ we have $\#_{\AA_1,\BB} (\leq^{A_1}) = \#_{\AA_2,\BB} (\leq^{A_2})$, so we have that $\Aut(\DD_\omega)$ is amenable. Now suppose that $\AA$ and $\BB$ are structures in $\D_n$ for a fixed $n < \omega$. We have: $$\begin{aligned} \#_{\AA,\BB} (\leq^A, \{I_i^A\}_{i=1}^a) &:= \left\vert\{ (\leq^B, \{I_i^B\}_{i=1}^b ): (\AA, \leq^A, \{I_i^A\}_{i=1}^a) \leq (\BB, \leq^B, \{I_i^B\}_{i=1}^b) \in \D_n^* \}\right\vert \\ &= \frac{(n-a)!}{(n-a-b)!} \cdot \prod_{k=1}^a \frac{\vert B_{i_k} \vert!}{\vert A_k \vert!} \cdot \prod_{k \notin \{i_1, \ldots, i_a\}} \vert B_k \vert ! \\ &= \frac{(n-a)!}{(n-a-b)!} \cdot \frac{\prod_{k=1}^b \vert B_{i_k} \vert!}{\prod_{k=1}^a \vert A_k \vert!}.\end{aligned}$$ Again, for $\AA_1, \AA_2 \in \binom{\BB}{\AA}$ we have $\#_{\AA_1,\BB} (\leq^{A_1}, \{I_i^{A_1}\}_{i=1}^a) = \#_{\AA_2,\BB} (\leq^{A_2}, \{I_i^{A_2}\}_{i=1}^a)$, so we have that $\Aut(\DD_n)$ is amenable. 2-cover of the generic tournament $\hat{\TT^\omega}$. {#sec:That_amen} ----------------------------------------------------- \[thm:amen\_T\_hat\] The group $\Aut(\hat{\TT^\omega})$ is amenable. Since $(\hat\T, \hat\T^*)$ is an excellent pair, it is enough to show that there is a consistent random $\hat\T^*$ expansion of $\hat\T$, by Proposition \[threeone\]. For each $\AA \in \hat\T$ we define a measure $\mu_\AA$ by taking: $$\mu_\AA (\{\AA^*\}) := \frac{1}{\# (\AA)}.$$ We check, using Theorem \[dense\], that this is indeed a random consistent expansion of $\hat\T$. It is enough to show that for $\AA \leq \BB$ in $\hat\T$ and $\AA^*$ with $\expand{\AA} \in \hat\T^*$ the number $$\#_{\hat\T^*} (\AA^*, \BB) := \left\vert \{ \BB^* : \expand{\AA} \leq \expand{\BB} \in \T^* \} \right\vert$$ depends only on the isomorphism classes of $\AA,\BB$, and notably not on the particular expansion $\AA^*$. Let $\AA = (A, \rightarrow^A)$ and $\BB = (B, \rightarrow^B)$ be structures in $\hat\T$ such that $\AA \leq \BB$, and let $\leq^A$ be a linear order on $A$ such that $(\AA, \leq^A) \in \hat\T^*$. Let $A_1, \ldots, A_a$ be the $\perp^A$-equivalence classes of $\AA$ and let $B_1, \ldots, B_b$ be $\perp^B$-equivalence classes of $\BB$. Without loss of generality we may assume that $\leq^A$ induces a linear ordering on the $\perp^A$-equivalence classes such that $A_1 <^A \ldots <^A A_a$. Moreover we may assume that this linear order is induced by $l_i \in A_i$ where $1 \leq i \leq a$. We may also assume that $\vert A_i \vert = 2$ for all $1 \leq i \leq a$. There are $1 \leq i_1 < i_2 < \ldots < i_a \leq b$ such that $A_j = B_{i_j}$ for each $1 \leq j \leq a$. Define $I := \{i_1, i_2, \ldots, i_a\}$ and $J := \{1,2, \ldots, b\} \setminus I$. If $\expand{\BB} = (\BB, \leq^B, I_0^B, I_1^B) \in \hat\T^*$ is such that $\expand{\AA} = (\AA, \leq^A, I_0^A, I_1^A) \leq (\BB, \leq^B, I_0^B, I_1^B)$, then $\leq^B$ and $\leq^A$ induce the same linear ordering on $\{B_{i_1}, \ldots, B_{i_a}\}$ and they agree on each $A_i$. Furthermore $I_0^A = I_0^B$ and $I_1^A = I_1^B$ on $A_i$ (for $1 \leq i \leq a$). Therefore we have: $$\#_{\hat\T^*} (\AA^*,\BB) := \left\vert \{ \BB^* : \expand{\AA} \leq \expand{\BB} \in \hat\T^* \} \right\vert = \frac{b!}{a!} \cdot 2^{\vert J \vert}.$$ Clearly this does not depend on the particular expansion $\AA^*$, or the particular embedding of $\expand{\AA}$ into $\expand{\BB}$. Thus by Theorem \[dense\], we have the desired amenability. Interestingly, this argument does not work for $\Aut(\hat{\QQ})$, despite the superficial similarities between $\hat{\QQ}$ and $\hat{\TT^\omega}$. The rough idea is that $\QQ$ comes with its linear order, and the precompact expansions of $\hat{\QQ}$ must cohere with this linear order. This severely restricts the number of expansions of a finite substructure of $\hat{\QQ}$. With $\hat{\TT^\omega}$ there is no such linear order that must be cohered with. See Theorem \[thm:Qhat\] for more details. This can be extended to a rough heuristic: “If the expansions $\K^*$ reference a linear order on $\K$, then $\Aut(\Flim(\K))$ is not amenable". This heuristic can be seen in the examples studied in Section \[sec:non-amen\]. Semi-generic multipartite digraph $\SS$ and variants. {#sec:A_S} ----------------------------------------------------- \[thm:semi\_amen\] $\Aut(\SS)$ is amenable. Since $(\mathcal{S}, \mathcal{S}^*)$ is an excellent pair, see [@JLNW14 Lemma 10.7, Lemma 10.8], it is enough to show that there is a consistent random $\mathcal{S}^*$ expansion of $\mathcal{S}$, by Proposition \[threeone\]. For each $\AA \in \mathcal{S}$ we define a measure $\mu_\AA$ by taking: $$\mu_\AA (\{\AA^*\}) := \frac{1}{\# (\AA)}.$$ We check, using Theorem \[dense\], that this is indeed a random consistent expansion of $\mathcal{S}$. It is enough to show that for $\AA \leq \BB$ in $\mathcal{S}$ and $\AA^*$ with $\expand{\AA} \in \mathcal{S}$ the number $$\#_{\mathcal{S}^*} (\AA^*, \BB) := \left\vert \{ \BB^* : \expand{\AA} \leq \expand{\BB} \in \mathcal{S}^* \} \right\vert$$ depends only on the isomorphism classes of $\AA,\BB$ and $\expand{\AA}$, and notably not on the particular expansion of $\AA$. Let $A_1, \ldots, A_a$ be the list of $\perp^A$-equivalence classes in $\AA$, and let $\leq^A$ be a linear ordering on $\AA^*$ such that $A_1 \leq^A \ldots \leq^A A_a$, and let $R^A$ be the binary relation on $\AA^*$. Similarly, let $B_1, \ldots, B_b$ be the list of $\perp^B$-equivalence classes in $\BB$, and let $i_1 < i_2 < \ldots < i_a$ be such that $A_j \sse B_{i_j}$ for $1 \leq j \leq a$. Let $\BB^* = (B, R^B, \leq^B)$ be such that $\expand{\BB} \in \mathcal{S}^*$. If $\expand{\AA} \leq \expand{\BB}$ then $\leq^B$ extends $\leq^A$ and $R^B$ extends $R^A$. So we have: $$\begin{aligned} \#_{\mathcal{S}^*} (\AA^*, \BB) &= \frac{b!}{a!} \cdot \prod_{j=1}^a \frac{b_{i_j}!}{a_j !} \cdot \left(\prod_{\substack{1 \leq j \leq b, \\ j \notin \{i_1, \ldots, i_a\}}} b_j !\right) \cdot 2^{\binom{b}{2} - \binom{a}{2}} \\ &= \left( b! \cdot 2^{\binom{b}{2}} \cdot \prod_{j=1}^b b_j! \right) \left( a! \cdot 2^{\binom{a}{2}} \cdot \prod_{j=1}^a a_j! \right)^{-1}.\end{aligned}$$ Clearly this quotient depends only on the isomorphism classes of $\AA$ and $\BB$. Suppose that $B_k$ and $B_l$ are $\perp^B$-equivalence classes such that the linear ordering $\leq^B$ implies $B_k < B_l$. Then there are two ways to put $R^B$ between these two classes. Since $R^B$ extends $R^A$ we have to choose $R^B$ among $\binom{b}{2} - \binom{a}{2}$ pairs of column partitions, see Figure \[fig:semi\_R\] in section 3.6. This is why $2^{\binom{b}{2} - \binom{a}{2}}$ shows up. \[thm:SSexp\_amen\] $\Aut(\SS_R)$ and $\Aut(\SS_\leq)$ are amenable. The cases of $\Aut(\SS_R)$ and $\Aut(\SS_\leq)$ are similar to the case of $\Aut(\SS)$, which was just shown. Generic omitters $\GG_n$ and $\FF(\T)$. --------------------------------------- Since the expansions of these classes are arbitrary linear orders, $\Aut(\GG_n)$ and $\Aut(\FF(\T))$ are amenable by [@AKL12 Proposition 9.3]. Moreover, the uniform consistent random expansion is indeed a consistent random expansion. By Proposition \[threeone\] this ensures amenability of the automorphism groups. Failures of amenability {#sec:non-amen} ======================= As mentioned at the end of Section \[sec:That\_amen\], when the expansions $\K^*$ make reference to a canonical order of $\K$ we expect $\Aut(\Flim(\K))$ to be non-amenable. This heuristic works for the examples we investigate here: $\SS(n), \PP, \PP(3)$ and $\hat{\QQ}$. Weak local orders $\SS(n)$. {#sec:S(n)} --------------------------- The following argument comes from private communication with A. Kechris. The theorem for the case of $n=3$ was first shown in [@ZUC13], although the following argument has a distinct geometric crux. \[thm:Sn\_amen\] For $n = 2,3$, $\Aut(\SS(n))$ is not amenable. Fix $n=2$ or $3$. By Proposition \[threeone\] it is enough to show that there is no consistent random $\mathcal{S}(n)^*$ expansion of $\mathcal{S}(n)$. Suppose for the sake of contradiction that such an expansion $(\mu_\KK)$ exists. Consider the structures $\AA, \BB, \CC \in \mathcal{S}(n)$ where $\AA \leq \BB, \AA \leq \CC$ and $\BB \cap \CC = \AA$. Let: - $\AA = (A, \rightarrow^A)$, $A = \{x,y\}$ and $x \rightarrow^A y$; - $\BB = (B, \rightarrow^B)$, $B = A \cup \{b\}$, $x \rightarrow^B b$ and $b \rightarrow^B y$; - $\CC = (C, \rightarrow^C)$, $C = A \cup \{c\}$, $x \rightarrow^C c$ and $y \rightarrow^C c$. Consider also expansions (in $\mathcal{S}(n)^*$) of these structures. Let: - $\expand{\AA} = (\AA, I_0^A, I_1^A )$, where $I_0^A = \{x,y\}$; - $\expand{\BB} = (\BB, I_0^B, I_1^B )$, where $I_0^B = \{x,y,b\}$; - $\expand{\CC} = (\CC, I_0^C, I_1^C )$, where $I_0^C = \{x,y,c\}$; - $\expand[**]{\CC} = (\CC, I_0^{**}, I_1^{**})$, where $I_0^{**} = \{x,y\}, I_1^{**} =\{c\}$. ![$\expand{\AA}, \expand{\BB}, \expand{\CC}$ and $\expand[**]{\CC}$.](S2_good){width="80.00000%"} This gives us $$\begin{aligned} \mu_\AA (\{\AA^*\}) &\overeasy \sum\{\mu_\BB (\{\BB^\prime\}) :\expand{\AA} \leq \expand[\prime]{\BB} \in \mathcal{S}(n)^*\} \\ &= \mu_\BB (\{\BB^*\}) \end{aligned}$$ since there is only one expansion of $\BB$ in $\mathcal{S}(n)^*$ that extends $\expand{\AA}$, namely $\expand{\BB}$. Also, we have $$\begin{aligned} \mu_\AA (\{\AA^*\}) &\overeasy \sum\{\mu_\CC (\{\CC^\prime\}) : \expand{\AA} \leq \expand[\prime]{\CC} \in \mathcal{S}(n)^*\} \\ &= \mu_\CC (\{\CC^*\}) + \mu_\CC (\{\CC^{**}\}) \end{aligned}$$ since there are exactly two expansions of $\CC$ in $\mathcal{S}(n)^*$ which extend $\expand{\AA}$, namely $\expand{\CC}$ and $\expand[**]{\CC}$. Note $\BB \cong \CC$ and this isomorphism extends to all of $\SS(n)$, so $(\textbf{I})$ applies. Also $\expand{\BB} \cong \expand{\CC}$, so we have $$\mu_\CC (\{\CC^*\}) = \mu_\BB (\{\BB^*\}) = \mu_\CC (\{\CC^*\}) + \mu_\CC (\{\CC^{**}\})$$ So $\mu_\CC (\{\CC^{**}\}) = 0$, which is impossible for a consistent random expansion. We can also state a related result. Let $\O\mathcal{S}(n)$ be the class of structures of the form $(A, \rightarrow^A, \leq^A)$ where $(A, \rightarrow^A) \in \mathcal{S}(n)$ and $\leq^A$ is a linear order on $A$. Let $\O\mathcal{S}(n)^*$ be the class of the structures of the form $(A, \rightarrow^A, \leq^A, I_0^A, \ldots, I_{n-1}^A)$ where the structures $(A, \rightarrow^A, \leq^A) \in \O\mathcal{S}(n)$ and $(A, \rightarrow^A, I_0^A, \ldots, I_{n-1}^A) \in \mathcal{S}(n)^*$. Then we have that for $n=2,3$, $(\O\mathcal{S}(n), \O\mathcal{S}(n)^*)$ form an excellent pair of Fraïssé classes which both happen to satisfy the **SAP**, see [@SOK13]. Using the argument in the proof of Theorem \[thm:Sn\_amen\], we get the following corollary: \[cor:AutO2\] For $n = 2,3$, $\Aut(\Flim(\O\mathcal{S}(n)))$ is not amenable. In the case of $n \geq 4$ a similar argument will show that there is a non-trivial flow that is always assigned measure $0$. However, in this case there is no immediate contradiction since the flow need not be minimal; the space might support two disjoint non-minimal flows. A more subtle investigation would be necessary to establish the non-amenability of $\Aut(\SS(n))$, but it eludes the authors at this time. Generic partial order $\PP$. {#sec:P} ---------------------------- The main geometrical idea of the following proof is from [@KSOK12 Section 3]. $\Aut(\PP)$ is not amenable. Since $(\P, \P^*)$ is an excellent pair, it is enough to show that there is no consistent random $\P^*$ expansion of $\P$, by Proposition \[threeone\]. Suppose for the sake of contradiction that such an expansion $(\mu_\KK)$ exists. Consider the structures $\AA, \BB, \CC \in \P$ where $\CC= (C, <^C)$, $C = \{a,b,c\}$, $a <^C b$, with $\AA = \CC \restrict \{a,c\}$ and $\BB = \CC \restrict \{b,c\}$. Consider also expansions (in $\P^*$) of these structures. Let - $\expand{\AA} = (A, \prec^A)$, where $a \prec^A c$; - $\expand{\BB} = (B, \prec^B)$, where $b \prec^B c$. {width="80.00000%"} Now we consider the possible expansions $\expand{\CC} = (C, <^C, \prec^C)$ with $\expand{\BB} \leq \expand{\CC}$. We claim that there is only one such expansion possible. ![$\expand[0]{\CC}$ and $\expand[1]{\CC}$.](P_good_two){width="40.00000%"} We ask “Is $a \prec^C c$?”. A positive answer yields the expansion $\CC^0$, where $ a \prec^C b \prec^C c$ and $a \prec^C c$, a linear order. A negative answer would yield the cycle $a \prec^C b \prec^C c \prec^C a$, which cannot happen in a linear order. (This is the key geometric observation about $\P^*$.) Therefore: $$\begin{aligned} \mu_\BB (\{\BB^*\}) &\overeasy \sum\{\mu_\CC (\{\CC^\prime\}) : \expand{\BB} \leq \expand[\prime]{\CC} \in \P^*\} \\ &= \mu_\CC (\{\CC^0\}), \end{aligned}$$ Now we consider the possible expansions $\expand{\CC} = (C, <^C, \prec^C)$ with $\expand{\AA} \leq \expand{\CC}$. We claim that there are two such expansions possible. We ask “Is $b \prec^C c$?”. A positive answer yields the expansion $\CC^0$. A negative answer yields the expansion $\CC^1$, where $ a \prec^C c \prec^C b$ and $a \prec^C b$, a linear order. Therefore: $$\begin{aligned} \mu_\AA (\{\AA^*\}) &\overeasy \sum\{\mu_\CC (\{\CC^\prime\}) : \expand{\AA} \leq \expand[\prime]{\CC} \in \P^*\} \\ &= \mu_\CC (\{\CC^0\}) + \mu_\CC (\{\CC^1\}). \end{aligned}$$ Since $\AA \cong \BB$ and $\expand{\AA} = \expand{\BB}$ we have $$\mu_\CC (\{\CC^0\}) = \mu_\AA (\{\AA^*\}) = \mu_\BB (\{\BB^*\}) = \mu_\CC (\{\CC^0\}) + \mu_\CC (\{\CC^1\})$$ So $\mu_\CC (\{\CC^1\}) = 0$, which is impossible for a consistent random expansion. A similar argument will be used to show that $\PP(3)$, the so-called “twisted generic poset” has a non-amenable automorphism group. In that case there is an extra layer of notation which somewhat obscures the argument. “Twisted” generic partial order $\PP(3)$. {#sec:P(3)} ------------------------------------------ \[thm:AutP3\] $\Aut(\PP(3))$ is not amenable. Since $(\P(3), \P(3)^*)$ is an excellent pair, see [@JLNW14 Theorem 9.3], it is enough to show that there is no consistent random $\P(3)^*$ expansion of $\P(3)$, by Proposition \[threeone\]. Suppose for the sake of contradiction that such an expansion $(\mu_\KK)$ exists. Consider the structures $\AA, \BB, \CC \in \P(3)$ where $\CC= (C, \leq^C)$, $C = \{a,b,c\}$, $a \rightarrow^C b$, with $\AA = \CC \restrict \{a,c\}$ and $\BB = \CC \restrict \{b,c\}$. Consider also expansions (in $\P(3)^*$) of these structures. Let - $\expand{\AA} = (A, P_0^A, P_1^A, P_2^A, \preceq^A)$, where $P_0^A = \{a\}$, $P_1^A = \{c\}$ and $a \preceq^A c$; - $\expand{\BB} = (B, P_0^B, P_1^B, P_2^B, \preceq^B)$, where $P_0^B = \{b\}$, $P_1^B = \{c\}$ and $b \preceq^B c$. {width="80.00000%"} Now we consider the possible expansions $\expand{\CC} = (C, P_0^C, P_1^C, P_2^C, \preceq^C)$ with $\expand{\AA} \leq \expand{\CC}$. There are three options for the label of $b$, namely: $P_0^C(b), P_1^C(b)$ and $P_2^C(b)$. {width="80.00000%"} If $P_1^C (b)$, then we have $b \leq^C a$ and $a \leq^C c$ so $b \leq^C c$. This contradicts the fact that $P_1^C(b)$ and $P_1^C(c)$ guarantee that $b$ and $c$ are $\leq^C$ incomparable. If $P_2^C (b)$, then we have $a \leq^C c$ and $c \leq^C b$ so $a \leq^C b$. This contradicts the fact that $P_0^C(a), P_2^C(b)$ and $a \rightarrow^C b$ guarantee that $a$ and $b$ are $\leq^C$ incomparable. Therefore only $P_0^C(b)$ is possible, and so: $$\begin{aligned} \mu_\AA (\{\AA^*\}) &\overeasy \sum\{\mu_\CC (\{\CC^\prime\}) : \expand{\AA} \leq \expand[\prime]{\CC} \in \P(3)^*\} \\ &= \mu_\CC (\{\CC^0\}), \end{aligned}$$ where $\CC^0$ is given by $P_0^C (b), P_0^C(a), P_1^C (c)$, with $a \rightarrow^C c$ and $a \preceq^C b \preceq^C c$. On the other hand, there are many expansions $\expand{\CC}$ that respect $\expand{\BB}$, so $$\begin{aligned} \mu_\BB (\{\BB^*\}) &\overeasy \sum\{\mu_\CC (\{\CC^\prime\}) : \expand{\BB} \leq \expand[\prime]{\CC} \in \P(3)^*\} \\ &\geq \mu_\CC (\{\CC^0\}) + \mu_\CC (\{\CC^1\}), \end{aligned}$$ where $\CC^1$ is given by $P_1^C (a), P_0^C (b), P_1^C(c)$, with $a \rightarrow^C c$ and $b \preceq^C c \preceq^C a$. ![$\expand[0]{\CC}$ untwisted (1) and twisted (2); $\expand[1]{\CC}$ untwisted (3) and twisted (4).](P3_good_two){width="80.00000%"} Since $\AA \cong \BB$ and $\expand{\AA} \cong \expand{\BB}$ we have $$\mu_\CC (\{\CC^0\}) = \mu_\AA (\{\AA^*\}) = \mu_\BB (\{\BB^*\}) \geq \mu_\CC (\{\CC^0\}) + \mu_\CC (\{\CC^1\}).$$ So $\mu_\CC (\{\CC^1\}) = 0$, which is impossible for a consistent random expansion. 2-cover of the linear tournament $\hat{\QQ}$. {#sec:hat_Q} --------------------------------------------- \[thm:Qhat\] $\Aut(\hat{\QQ})$ is not amenable. Let $\T := \Age(\QQ)$. Suppose for the sake of contradiction that $\Aut(\Flim(\hat{\T}))$ is amenable. By Proposition \[threeone\], there is a consistent random expansion $(\mu_\AA)$. It is enough to work on the cofinal class of structures $\AA$ that have exactly two elements in each $\perp^A$-equivalence class. Moreover, since for all $\expand{\AA}, (\AA\oplus\BB^*) \in \hat{\T}^*$ we have $\expand{\AA} \cong (\AA\oplus\BB^*)$ we must have: $$\mu_\AA (\{\AA^*\}) = \frac{1}{2 \cdot k},$$ where $k$ is the number of $\perp^A$-equivalence classes, and $\AA = (A, \rightarrow^A)$. Let $\BB \in \hat{\T}$ be such that each $\perp^B$ equivalence class has two elements with the partition $B = \sqcup_{i=1}^3 B_i$ such that $B_i = \{(i,C), (i,P)\}$. Let the edges on $\{(1,j), (2,j), (3,j)\}$ be given by the natural linear order for $j= C,P$. Let $\AA$ be the substructure given by the initial segment $B_1 \sqcup B_2$. Let $\AA^*$ be obtained by the ordering $B_1 <^A B_2$ with $I_1 = \{(1,C),(2,P)\}$ and $I_0 = \{(1,P),(2,C)\}$.  Then there is only one $\BB^*$ such that $\expand{\AA} \leq \expand{\BB} \in \hat{\T}^*$, namely $I_1(3,P)$ and $I_0(3,C)$ thus: $$\begin{aligned} \frac{1}{4} = \mu_\AA (\{\AA^*\}) &\overeasy \sum \{\mu_\BB (\{\BB^\prime\}) : \expand{\AA} \leq \expand[\prime]{\BB} \in\hat{\T}^*\} \\ &= \mu_\BB (\{\BB^*\}) = \frac{1}{6},\end{aligned}$$ a contradiction. The Hrushovski property ======================= We make a brief mention of the Hrushovski property, which is closely related to amenability and unique ergodicity, as examined in [@AKL12]. In general, establishing that a Fraïssé class has the Hrushovski property is a challenging combinatorial problem. Our minor contribution in this area is to establish that the composition $\LEK$ interacts favourably with the Hrushovski property. A class $\K$ of finite structures is a **Hrushovski class** if for any $\KK \in \K$ and any finite sequence of partial isomorphisms $\phi_i : \AA_i \longrightarrow \BB_i$ (for $1 \leq i \leq k$) where $\AA_i, \BB_i \leq \KK$, there is a $\CC \in \K$, such that each $\phi_i$ (for $1 \leq i \leq k$) can be extended to an automorphism $\psi_i : \CC \longrightarrow \CC$. Recall the following proposition which appears as Proposition 13.1 in [@AKL12] for the special case of order expansions. \[prop:HruAmen\] Let $(\K, \K^*)$ be an excellent pair. If $\K$ is a Hrushovski class, then $\Aut(\Flim(\K))$ is amenable. We immediately get the following corollary. $\mathcal{S}(2)$, $\mathcal{S}(3)$, $\P$, $\P(3)$, $\Age(\SS(2)[\II_n])$, $\Age(\II_n [\SS(2)])$, and $\Age(\hat{\Q})$ are not Hrushovski classes. Now we present a proposition which says that the Hrushovski property behaves “exactly the way you’d want it to” with respect to the product class $\LEK$. \[prop:HruProd\] Let $\K$ and $\L$ be classes of finite relational structures, such that $\L$ satisfies the **JEP**. Then $\LEK$ is a Hrushovski class if and only if $\L$ and $\K$ are Hrushovski classes. \[$\Rightarrow$\] Suppose that $\LEK$ is a Hrushovski class. First we show that $\L$ is a Hrushovski class. Let $\LL \in \L$, and let $\PP \in \K$ be a one point structure. Fix a finite sequence of partial isomorphisms $\phi_i : \AA_i \longrightarrow \BB_i$ (for $1 \leq i \leq k$) where $\AA_i, \BB_i \leq \LL$. This gives related partial isomorphisms $\phi_i^\prime : (\PP:\AA_i) \longrightarrow (\PP:\BB_i)$. By the Hrushovski property of $\LEK$, there is a $\DD \in \LEK$, and automorphisms $\psi_i^\prime : \DD \longrightarrow \DD$, where $\psi^\prime_i$ extends $\phi_i^\prime$ for $1 \leq i \leq k$. Without loss of generality we may assume that $\DD = (\PP:\CC)$ and consequently, we must have automorphisms $\psi_i : \CC \longrightarrow \CC$ which extends $\psi_i$. So we have verified the Hrushovski property for $\L$. Now we show that $\K$ is a Hrushovski class. Let $\KK \in \K$, with $\vert \KK \vert = N$ and let $\phi_i : \AA_i \longrightarrow \BB_i$ (for $1 \leq i \leq k$) be a finite sequence of partial isomorphisms, where $\AA_i, \BB_i \leq \KK$. Let $\QQ \in \L$ be a one-point structure. We consider the structure $\DD := (\KK:\underbrace{\QQ, \ldots, \QQ}_{N\text{-times}}) \in \LEK$. Clearly every $\phi_i$ determines a unique $\phi_i^\prime : (\AA_i: \QQ, \ldots, \QQ) \longrightarrow (\BB_i:\QQ, \ldots, \QQ)$, so by the Hrushovski property for $\LEK$ there is a $\CC \in \LEK$ and automorphisms $\psi_i^\prime : \CC \longrightarrow \CC$ where $\psi_i^\prime$ extends $\phi_i^\prime$ for $1 \leq i \leq k$. If $\CC = (\RR:\PP_1, \ldots, \PP_N)$, then every $\phi_i^\prime$ determines an automorphism $\psi_i : \RR \rightarrow \RR$ which extends $\phi_i$. \[$\Leftarrow$\] Suppose that $\L$ and $\K$ are Hrushovski classes. Let $\AA = (\KK:\SS_1, \ldots, \SS_N) \in \LEK$ and let $\phi_i : \AA_i \longrightarrow \BB_i$ (for $1 \leq i \leq k$) be a finite sequence of partial isomorphisms in $\AA$. Since $\L$ satisfies the **JEP** we may assume that $\SS_1 = \ldots = \SS_k =: \SS$. Each $\phi_i$ is given by a partial isomorphism $\phi_i^\prime : \AA_i^\prime \longrightarrow \BB_i^\prime$ where $\AA_i^\prime, \BB_i^\prime \leq \KK$, and by a sequence of partial isomorphisms $(\phi_{i,s})_{s=1}^{n_i}$ inside $\SS$. Now, by the Hrushovski property for $\K$ there is a $\KK \leq \DD \in \K$ and automorphisms $\psi_i^\prime : \DD \longrightarrow \DD$ which extend the corresponding $\phi_i^\prime$. Moreover, there is an $\SS \leq \TT \in \L$ together with automorphisms $\phi_{i,s}^\prime : \TT \longrightarrow \TT$ which extend the corresponding $\phi_{i,s}^\prime$. The structure $\EE := (\DD:\TT, \ldots, \TT)$ contains $\AA$ and there are automorphisms $\psi_i : \EE \longrightarrow \EE$ given by $\psi_i^\prime$ and $\psi_{i,s}^\prime$ which extends $\phi_i$. This completes the verification that $\LEK$ is a Hrushovski class. We are now in a position to give a strengthening of Theorem \[A\_LEK\].1, with an alternate proof, in the special case that $\L$ and $\K$ are both Hrushovski Classes. Let $(\L, \L^*)$ and $(\K, \K^*)$ be excellent pairs of relational structures, where $\L$ and $\K$ are Hrushovski Classes. Then $\Aut(\Flim(\L)), \Aut(\Flim(\K))$ and $\Aut(\Flim(\LEK))$ are amenable. This follows immediately from Proposition \[prop:HruAmen\] and Proposition \[prop:HruProd\]. Unique ergodicity and McDiarmid’s inequality {#sec:UE_McD} ============================================ This section marks a shift, from establishing amenability to establishing unique ergodicity. Whereas the previous results were exact and finitary, we will now need to make use of asymptotics; we shift from finite combinatorics to analysis. To establish unique ergodicity we appeal to the probabilistic tools discussed in [@AKL12], suitably generalized to precompact expansions. First we examine the Quantitative Expansion Property, then we will see how this property, together with amenability, gives unique ergodicity. We will compress the probabilistic machinery into the black-box Lemma \[lem:QOP\_strategy\] which is combinatorial in nature. $\QOP$ and $\QOPstar$. {#sec:QOP} ---------------------- Here we look at two properties that allow us to push amenable automorphism groups up to uniquely ergodic. The following are **Q**uantitative **E**xpansion **P**roperties. The name comes from the property $\mathcal{QOP}$ in [@AKL12] which was concerned with expansions that are linear orderings. With suitable adaptions they apply to more general expansions, not just linear orderings. For fixed structures $\AA$ and $\BB$ with expansions $\expand{\AA}$ and $\expand{\BB}$, and $\E$ a set of embeddings of $\AA$ into $\BB$, define $$\Nexp(\E, \AA^*, \BB^*) := \left\vert\{\phi \in \E : \phi \text{ embeds } \expand{\AA} \text{ into } \expand{\BB}\}\right\vert.$$ If $\E$ is clear from context, we shall denote this set by $\Nexp(\AA^*,\BB^*)$. Also define $$\Nemb(\AA,\BB) := \left\vert\{\phi : \phi \text{ embeds } \AA \text{ into } \BB\}\right\vert.$$ Note that this is $\left\vert \Aut(\AA) \right\vert \cdot \left\vert \binom{\BB}{\AA} \right\vert$, and if $\AA$ is rigid, then this is just $\left\vert \binom{\BB}{\AA} \right\vert$. ### Definitions. Let $(\K,\K^*)$ be an excellent pair. We say that $\K^*$ satisfies the $\QOPstar$ if there is an isomorphism invariant map $\rho: \K^* \longrightarrow [0,1]$ such that for every $\expand{\AA}\in\K^*$ and every $\epsilon >0$, there is a $\BB \in \K$ and a non empty set of embeddings $\E$, from $\AA$ into $\BB$ with the property that for every $\expand{\BB}\in \K^*$ we have: $$\left\vert \frac{\Nexp(\E, \AA^*, \BB^*)}{\vert \E \vert} - \rho\expand{\AA} \right\vert < \epsilon.$$ Occasionally we will use the notation $a \overset{\epsilon}{\approx} b$ if $\vert a - b \vert < \epsilon$. Let $(\K,\K^*)$ be an excellent pair. We say that $\K^*$ satisfies the $\QOP$ if there is an isomorphism invariant map $\rho: \K^* \longrightarrow [0,1]$ such that for every $\AA\in\K$ and every $\epsilon >0$, there is a $\BB \in \K$ and a non empty set of embeddings $\E$, from $\AA$ into $\BB$ with the property that for every $\expand{\AA},\expand{\BB}\in \K^*$ we have: $$\left\vert \frac{\Nexp(\E, \AA^*, \BB^*)}{\vert \E \vert} - \rho\expand{\AA} \right\vert < \epsilon.$$ Note that in general $\QOP$ implies $\QOPstar$ (because $\QOP$ works for an arbitrary expansion, but in $\QOPstar$ you are working with a single expansion). Also, in a Hrushovski class, these are the same (see [@AKL12 Theorem 13.3]), and this is non-trivial. ### General Results and the Main Tool. The following theorem is one of the main reasons that we examine $\QOP$. It gives a method for ensuring that an amenable automorphism group is actually uniquely ergodic. \[QOP\_UE\] Let $(\K,\K^*)$ be an excellent pair, and suppose that $\Aut(\Flim(\K))$ is amenable and $\K^*$ satisfies the $\QOPstar$. Then $\Aut(\Flim(\K))$ is uniquely ergodic. With minor modifications, this follows from the Fubini-type argument presented in the proof of [@AKL12 Proposition 11.1]. ### Unique Ergodicity of $\Aut(\OH)$. Let us illustrate a direct verification of the $\QOPstar$ for a cofinal subclass of $\OH$. Recall the notation from Section \[subsec:binary\_trees\]. \[prop:UE\_OH\] $\Aut(\Flim(\OH))$ is uniquely ergodic. Amenability was proved in Corollary \[cor:OH\_amenable\].(ii). In order to prove unique ergodicity we will verify $\QOPstar$ for the class $\D$, which is enough by Theorem \[dense\]. Let $(A,C^A, \leq^A, \preceq^A) \in \COH$ be given where $(A, C^A, \leq^A) \in \D$, and let $\epsilon > 0$. Let $\BB$ be an $n+1$-nice tree such that $\Delta(\BB) = (A, C^A)$. In particular, we may assume that $A = T(\BB)$. Let $\preceq_1, \ldots, \preceq_l$ be the list of all linear orderings on $A$ such that $(A, C^A, \preceq_i) \in \CH$ for all $1 \leq i \leq l$. Let $\BB^\prime$ be an $n+l$-nice tree such that $\BB \restrict n = \TT$, and let $\BB_1, \ldots, \BB_l$ be the collection of subtrees of $\BB^\prime$ such that for $1 \leq i \leq l$ we have: - $\BB_i \restrict n = \BB$, - $(\forall x \in \BB(n))(\exists!x^\prime \in T(\BB_i))[x^\prime \in \BB^\prime[x]]$, - $T(\BB_i) \sse T(\BB^\prime)$. In this way each $\BB_i$ is a tree of height $n+l$; they are copies of $\BB$ with a branch of height $l$ appended to one of the terminal nodes of $\BB$. Now for $i \neq j$ we have $T(\BB_i) \cap T(\BB_j) = \emptyset$ and $\Delta(\BB_i) \cong \Delta(\BB_j)$. Moreover, every linear ordering $\preceq^\prime$ such that $(\Delta(\BB_i), \preceq^\prime) \in \CH$ is given by the unique linear ordering with the property that $(\Delta(\TT), \preceq) \in \CH$. More precisely, for $x,y \in T(\BB)$ and $x^\prime \in T(\BB^\prime[x]), y^\prime \in T(\BB^\prime[y])$ we have $x \preceq y \Leftrightarrow x^\prime \preceq^\prime y^\prime$. Therefore, $\preceq^\prime$ is given by one of the $\preceq_i$, and without loss of generality we denote such $\preceq^\prime$ by $\preceq_i$. On each $T(\BB_i)$ we put a linear ordering $\leq_i$ such that $$(T(\BB_i), \leq_i, \preceq_i) \cong (A, C^A, \leq^A, \preceq^A).$$ Let $\leq$ be a linear ordering on $T(\BB^\prime)$ which extends each $\leq_i$. Note that this is possible since $\BB_i \cap \BB_j = \emptyset$ for $i \neq j$. Let $\phi_i$ be the unique embedding from $(A, C^A, \leq^A)$ into $(\Delta(\BB^\prime), \leq)$ with image $(\Delta(\BB_i), \leq_i)$. Let $\E := \{\phi_i : 1 \leq i \leq l\}$. Now we may take $$\rho(A, C^A, \leq^A, \preceq^A) := \frac{1}{\#(A, C^A, \leq^A)}.$$ Now it is easy to see that in this way we can satisfy the condition of the $\QOPstar$, since for a given $\preceq$ with $(\Delta(\BB^\prime), \leq, \preceq)$ there is only one $\phi_i$ such that $\phi_i$ embeds $(A, C^A, \leq^A, \preceq^A)$ into $(\Delta(\BB^\prime), \leq, \preceq)$. ### $\QOP^*, \QOP$ and $\LEK$. We now show how the $\QOP$ and the $\QOP^*$ interact with $\LEK$. This will give us an alternate way to check unique ergodicity of $\Aut(\Flim(\LEK))$. \[prop:QOP\_LEK\] Let $(\K,\K^*)$ and $(\L,\L^*)$ be excellent pairs. If $\K^*$ and $\L^*$ satisfy $\QOPstar$, then $\LEK$ satisfies the $\QOPstar$. Let $\AA = (\SS_1, \ldots, \SS_k : \TT) \in \LEK$. There is an $\RR \in \K$ and an $\E_0$, a collection of embeddings from $\TT$ into $\RR$ which witnesses the $\QOPstar$ for $\K^*$. Also, there are $\LL_i \in \L$ for each $i$ and $\E_i$, a collection of embeddings from $\SS_i$ into $\LL_i$ which witnesses the $\QOPstar$ for $\L^*$. Consider the structure $\BB = (\LL_1, \ldots, \LL_k : \RR) \in \LEK$ with the collection $\E$ of all embeddings from $\AA$ into $\BB$. Each embedding from $\E$ is given by a member of $\E_0$ and a sequence of embeddings from $\E_1, \ldots, \E_k$. Let $\mu$ and $\nu$ be maps on $\K$ and $\L$ respectively that verifies the $\QOPstar$. We check that for $\AA^* = (\SS_1^*, \ldots, \SS_k^* : \TT^*)$ with $\expand{\AA} \in \LEKstar$ the following map verifies the $\QOPstar$ for $\LEKstar$: $$\rho\expand{\AA} := \mu\expand{\TT} \cdot \prod_{i=1}^k \nu (\expand{\SS_i})$$ Notice $$\begin{aligned} \frac{\Nexp(\E, \AA^*, \BB^*)}{\vert \E \vert} &= \frac{\Nexp(\E_0, \TT^*, \RR^*) \cdot \prod_{i=1}^k \Nexp(\E_i, \SS_i^*, \LL_i^*)}{\vert \E_0 \vert \cdot \vert \E_1 \vert \cdot \ldots \cdot \vert \E_k \vert}\\ &\approx (\mu\expand{\TT}\pm \epsilon) \cdot \prod_{i=1}^k (\nu(\expand{\SS_i})\pm \epsilon)\\ &\overset{(k+1)\epsilon}{\approx} \rho\expand{\AA}\end{aligned}$$ We trust that the reader can appropriately interpret the use of “$\approx$” in the second line. Since $\epsilon$ can be arbitrarily small, this completes the verification of the $\QOPstar$ for $\LEKstar$. \[QOP\_LEK\] Let $(\K,\K^*)$ and $(\L,\L^*)$ be excellent pairs. If $\K^*$ and $\L^*$ satisfy the $\QOP$, then $\LEK$ satisfies the $\QOP$. Let $(\K,\K^*)$ and $(\L,\L^*)$ be excellent pairs that satisfy the $\QOP$. If $\Aut(\Flim(\L))$ and $\Aut(\Flim(\K))$ are amenable then $\Aut(\Flim(\LEK))$ is uniquely ergodic. This follows from Proposition \[ExPair\_LEK\] (that $(\LEK, \LEKstar)$ is an excellent pair), Proposition \[QOP\_LEK\], Theorem \[A\_LEK\].i and Theorem \[QOP\_UE\]. Strategy for $\DD_n$ and $\hat{\TT^\omega}$. {#sec:strategyMcD} -------------------------------------------- Let $\KK$ be one of the directed graphs $\DD_n$ or $\hat{\TT^\omega}$. We will show that $\Aut(\KK)$ is uniquely ergodic using a method developed in [@AKL12 Section 3]. First we present a useful probabilistic inequality, and then we will discuss the general strategy. ### McDiarmid’s Inequality. The following theorem appears as Lemma 1.2 in [@McD89] and is a consequence of Azuma’s inequality. Let $\vec{X} = (X_1, \ldots, X_N)$ be a sequence of independent random variables and let $f(X_1, \ldots, X_N)$ be a real-valued function such that there are positive constants $a_i$, with $$\vert f(\vec{X}) - f(\vec{Y}) \vert \leq a_i,$$ whenever the vectors $\vec{X}$ and $\vec{Y}$ differ only in the $i^{\text{th}}$ coordinate. Then for $\zeta = \EE[f(\vec{X})]$ and all $\epsilon > 0$ we have: $$P[\vert f(\vec{X}) - \zeta \vert \geq \epsilon] \leq 2 \exp\left(\frac{-2\epsilon^2}{\sum_{i=1}^N a_i^2}\right).$$ Typically we will use $\vec{X} = (X_1, X_2, \ldots, X_{\binom{n}{2}})$, to talk about structures like the random directed tournament on $n$ vertices, and $\vec{X}$ will correspond to the direction of the $\binom{n}{2}$ edges. General Strategy. {#sec:QOP_strategy} ----------------- By Theorem \[QOP\_UE\], to show that $\Aut(\KK)$ is uniquely ergodic, it suffices to show that it is amenable and $\K := \Age(\KK)$ satisfies the $\QOPstar$. Showing amenability will usually be direct, and in the case of $\KK = \DD_n$ we have already shown amenability in Theorem \[thm:amen\_Dn\]. Showing that $\K$ satisfies the $\QOPstar$ is a more subtle affair. For the $\QOPstar$, for a (small) fixed $\HH \in \K$, (with around $k \cdot m$ vertices), we will find a (large, somewhat “random”) $\GG \in \K$, with $n$ vertices (or sometimes $n$ equivalence classes). To that end, let $\mathbf{G}$ be a uniformly random structure in $\K$ on $n$ fixed vertices (or sometimes $n$ fixed equivalence classes). In general, $\E := \{\phi : \phi \text{ embeds } \AA \text{ into } \BB\}$ will be the set of *all* embeddings from $\HH$ into $\mathbf G$, so $\vert \E \vert =: \Nemb(\HH, \mathbf G)$, and $\rho\expand{\HH} = \frac{1}{\# (\HH)}$, where $\# (\HH)$ is the number of expansions of $\HH$. We use the notation of $I(n,k,m) := \EE[\Nemb(\HH, \mathbf G)]$, the expected value of the number of embeddings of $\HH$ into $\mathbf G$. Note that $I(n,k,m)$ may also depend on other aspects of $\HH$ and $\mathbf G$, but in practice they won’t. In general only $n$ will be allowed to vary, and we will be concerned with large $n$. We will always establish two separate inequalities using the McDiarmid inequality. The first will be with the function $$f(\GG) := \frac{\Nemb(\HH, \mathbf G)}{I(n,k,m)},$$ and we will establish that changing $\mathbf G$ by a single edge changes $f(\mathbf G)$ by at most $O(\frac{1}{n^2})$. It is clear that $\EE[f(\mathbf G)] = 1$. McDiarmid’s inequality then yields $$\begin{aligned} P\left[\left\vert \frac{\Nemb(\HH,\mathbf G)}{I(n,m,k)} - 1 \right\vert \geq D \right] &\leq 2 \exp\left( \frac{-2D^2}{\binom{n}{2}\epsilon_1^2 n^{-4}}\right) \nn\\ &\leq 2 \exp(-\delta_1 D^2 n^2), \label{eqn:f_estimate}\end{aligned}$$ where $D = \frac{\epsilon}{2}$, fixed at the beginning, and $\delta_1$ does not depend on $n$, and the $\epsilon_1$ comes from $O(\frac{1}{n^2})$. The second inequality will be similar, applying McDiarmid’s inequality to the function $$f^*(\mathbf G) := \frac{\Nexp(\HH^*,\mathbf{G}^*)}{I(n,k,m)},$$ and we will establish that changing $\GG$ by a single edge changes $f^*(\mathbf G)$ by at most $O(\frac{1}{n^2})$. It is clear that $\EE[f^*(\mathbf G)] = \rho\expand{\HH}$. Thus McDiarmid’s Inequality yields: $$\begin{aligned} P\left[\left\vert \frac{\Nexp(\HH^*,\mathbf{G}^*)}{I(n,m,k)} - \rho\expand{\HH} \right\vert \geq D \right] &\leq 2 \exp\left( \frac{-2D^2}{\binom{n}{2}\epsilon_2^2 n^{-4}}\right) \nn\\ &\leq 2 \exp(-\delta_2 D^2 n^2) \label{eqn:f*_estimate}\end{aligned}$$ where $\rho\expand{\HH}$ and $\delta_2$ do not depend on $n$, and the $\epsilon_2$ comes from $O(\frac{1}{n^2})$. We then define the probability that $\mathbf G$ is not a suitable candidate: $$p = \frac{2 \cdot \#(\HH)\cdot \#(\mathbf G)}{e^{\delta_3 D^2 n^2}}$$ which will go to $0$ as $n$ gets large by Stirling’s approximation, since for us $\#(\HH) \cdot \#(\mathbf G) = O(n!)$, or $O((n!)^k)$, which corresponds to the number of pairs of expansions on $\HH$ and $\mathbf G$. So, except with probability $p$, by \[eqn:f\_estimate\] and \[eqn:f\*\_estimate\] we have, simultaneously: $$\left\vert \frac{\Nemb(\HH,\mathbf G)}{I(n,m,k)} - 1 \right\vert < D \text{ and }\left\vert \frac{\Nexp(\HH^*,\mathbf{G}^*)}{I(n,m,k)} - \rho\expand{\HH} \right\vert < D$$ for all $\expand{\HH}, \expand{\mathbf G} \in \K^*$. For large enough $n$, we have $p<1$, so a suitable $\GG$ will exist. The previous inequalities yield the following: $$\begin{aligned} \left\vert \frac{\Nexp(\HH^*,\GG^*)}{I(n,m,k)} - \frac{\Nexp(\HH^*,\GG^*)}{\Nemb(\HH,\GG)} \right\vert &= \frac{\Nexp(\HH^*,\GG^*)}{\Nemb(\HH,\GG)} \cdot \left\vert \frac{\Nemb(\HH,\GG)}{I(n,m,k)} - 1 \right\vert \\ &\leq 1 \cdot D = D.\end{aligned}$$ Finally, from the triangle inequality, we have that $\GG$ witnesses the $\QOP$: $$\begin{aligned} \left\vert \frac{\Nexp(\HH^*,\GG^*)}{\Nemb(\HH,\GG)} - \rho\expand{\HH} \right\vert \leq &\left\vert \frac{\Nexp(\HH^*,\GG^*)}{\Nemb(\HH,\GG)} - \frac{\Nexp(\HH^*,\GG^*)}{I(n,m,k)} \right\vert + \left\vert \frac{\Nexp(\HH^*,\GG^*)}{I(n,m,k)} - \rho\expand{\HH} \right\vert \\ \leq &2D \leq \epsilon.\end{aligned}$$ We summarize this in a lemma. \[lem:QOP\_strategy\] Using the notation defined above, suppose that $\Aut(\KK)$ is amenable, that changing $\mathbf G$ by a single edge changes $f$ and $f^*$ by no more than $O(\frac{1}{n^2})$, and that $\#(\mathbf G) \leq O((n!)^k)$. Then $\Aut(\KK)$ is uniquely ergodic. The random method ================= We are now in a position to apply Lemma \[lem:QOP\_strategy\] and establish the unique ergodicity of the automorphism groups of $\DD_\omega, \DD_n, \hat{\TT^\omega}, \SS_\leq$ and $\SS_R$. The structures $\GG_n$ and $\FF(\T)$ are more subtle and require more attention, so they will be addressed in the subsequent section. Unique ergodicity of $\Aut(\DD_n)$. {#sec:UEofDn} ----------------------------------- We will show unique ergodicity of $\Aut(\DD_n)$ in two steps. First we consider the special case of $n=\omega$, then we adapt the proof for $n < \omega$. Let $(n)_k$ be the number of injective maps from $\{1, \ldots, k\}$ into $\{1, \ldots, n\}$. Note that in general this is different from $\binom{n}{k}$. $\Aut(\DD_\omega)$ is uniquely ergodic. Let $\HH = (H, \rightarrow^H) \in \D_\omega$ have $k$ many $\perp^H$-equivalence classes with respective cardinalities $a_1, \ldots, a_k$. Let $G$ be the set with partition $G = \bigsqcup_{i=1}^k G_i$, with $\vert G_i \vert = m$ for $i \leq k$. We consider a sequence of independent uniformly random variables induced by a pair of elements $G(x,y)$ where $x \in G_i, y \in G_j$ and $i < j$. Each random variable indicates with probability $\frac{1}{2}$ that $x \rightarrow^G y$ and with probability $\frac{1}{2}$ that $y \rightarrow^G x$. In this way, the collection of random variables $G(x,y)$ gives a random directed graph $\mathbf G = (G, \rightarrow^G) \in \D_\omega$. Notice that $\# (\mathbf G) = (n!)(m!)^n = O(n!)$ and $\# (\HH) = k! \cdot a_1! \cdots a_k!$. We have: $$I(n,m,k,\vec{a}) := \EE[\Nemb(\HH,\mathbf G)] = (n)_k \prod_{i=1}^k (m)_{a_i} \cdot 2^{-\sum_{l<j} a_l a_j},$$ where $\vec{a} = (a_1, \ldots, a_k)$. For $$f(\mathbf G) := \frac{\Nemb(\HH, \mathbf G)}{I(n,m,k,\vec{a})}$$ we have $\EE[f(\mathbf G)] = 1$. If we change the direction of only one edge then we change $\Nemb(\HH,\mathbf G)$ by not more than $$(k)_2 \cdot (n-2)_{k-2} \cdot \prod_{i=1}^k (m)_{a_i},$$ and $f$ by not more than: $$\frac{(k)_2 \cdot (n-2)_{k-2} \cdot \prod_{i=1}^k (m)_{a_i}} {(n)_k \prod_{i=1}^k (m)_{a_i} \cdot 2^{-\sum_{l<j} a_l a_j}} = \frac{1}{n(n-1)} (k)_2 \cdot 2^{\sum_{l<j} a_l a_j} \leq \frac{\epsilon_1}{n^2}$$ for large enough $n$ and some positive constant $\epsilon_1$. Let $(\HH, \leq^H) \in \D_\omega^*$ and $(\mathbf G, \leq^G) \in \D_\omega^*$ be such that the $\perp$-equivalence classes are intervals with respect to $\leq^G$. A change in the direction of one edge will change the function: $$f^*(\mathbf G) := \frac{\Nexp(\leq^H,\leq^G)}{I(n,m,k,\vec{a})}$$ by not more than $$\frac{\binom{k}{2} \cdot (n-2)_{k-2} \cdot \prod_{i=1}^k (m)_{a_i}} {(n)_k \cdot \prod_{i=1}^k (m)_{a_i} \cdot 2^{-\sum_{l<j} a_l a_j}} = \frac{1}{n(n-1)}\binom{k}{2}\cdot 2^{\sum_{l<j} a_l a_j} \leq \frac{\epsilon_2}{n^2}$$ for large enough $n$ and some positive constant $\epsilon_2$. For the McDiarmid inequality we use $\rho\expand{\HH} = \frac{1}{k!a_1! \cdots a_k!}$. Thus we are finished by Lemma \[lem:QOP\_strategy\]. For $n <\omega, \Aut(\DD_n)$ is uniquely ergodic. Since it is more natural to let $n$ vary, we will show that $\Aut(\DD_N)$ is uniquely ergodic. For $N$ finite we will consider a similar random directed graph $\mathbf G$ given by the parameter $m$, the cardinality of the parts. In this case, $N$ will be fixed and we will adjust $m$. For the “small” structure $\HH$, we use the parameter $k$ for its number of parts. Notice that $\#(\mathbf G) = (N!)(m!)^N = O((m!)^N)$ and $\#(\HH) = (N)_k \cdot a_1! \cdots a_k!$, where $N$ is fixed, and $m$ can vary. So $\rho\expand{\HH} = \frac{1}{(N)_k\cdot a_1! \cdots a_k!}$. Let $$f(\mathbf G) := \frac{\Nemb(\HH,\GG)}{I(N,m,k,\vec{a})}.$$ A single change in the direction of one edge of parts $i,j$ in $\mathbf G$ will change $\Nemb(\HH,\mathbf G)$ by not more than $$(k)_2 \cdot (N-2)_{k-2} \cdot a_i \cdot (m-1)_{a_i-1} \cdot a_j \cdot (m-1)_{a_j-1} \cdot \prod_{l \neq i,j} \binom{m}{a_l},$$ and $f$ by not more than: $$\begin{aligned} & \frac{(k)_2 \cdot (N-2)_{k-2} \cdot a_i \cdot (m-1)_{a_i-1} \cdot a_j \cdot (m-1)_{a_j-1} \cdot \prod_{l \neq i,j} \binom{m}{a_l}} {(N)_k \cdot (m)_{a_1} \cdot \ldots \cdot (m)_{a_k} \cdot 2^{-\sum_{i^\prime<j^\prime} a_i^\prime a_j^\prime}} \\ &= \frac{(k)_2}{N(N-1)}\cdot a_i a_j \frac{1}{m^2}\cdot 2^{\sum_{i^\prime<j^\prime} a_i^\prime a_j^\prime} \\ &\leq \frac{\epsilon_1}{m^2}\end{aligned}$$ for a large enough $m$ and some positive constant $\epsilon_1$. For $\expand{\HH}, \expand{\mathbf G} \in \D_N^*$, for $$f^*(\mathbf G) := \frac{\Nexp(\HH^*,\mathbf G^*)}{I(N,m,k,\vec{a})}$$ a single change in direction of one edge between parts $i$ and $j$ will change $f^*(\mathbf G)$ by not more than $$\begin{aligned} & \frac{a_i \cdot (m-1)_{a_i-1} \cdot a_j \cdot (m-1)_{a_j-1} \cdot \prod_{l \neq i,j} \binom{m}{a_l}} {(N)_k \cdot (m)_{a_1} \cdot \ldots \cdot (m)_{a_k} \cdot 2^{\sum_{i^\prime<j^\prime} a_i^\prime a_j^\prime}} \\ &= \frac{1}{(k)_2}\cdot a_i a_j \frac{1}{m^2}\cdot 2^{\sum_{i^\prime<j^\prime} a_i^\prime a_j^\prime} \\ &\leq \frac{\epsilon_2}{m^2}\end{aligned}$$ for large enough $m$ and some positive constant $\epsilon_2$. Thus we are finished by Proposition \[lem:QOP\_strategy\]. The 2-cover $\hat{\TT^\omega}$ has a uniquely ergodic automorphism group. {#sec:hat_T_UE} ------------------------------------------------------------------------- Recall the notation from Section \[sec:def\_blowups\]. There we established that $\Aut(\hat{\TT^\omega})$ is amenable. The group $\Aut(\hat{\TT^\omega})$ is uniquely ergodic. Let $\T := \Age(\TT^\omega)$. Since $(\hat{\T},\hat{\T}^*)$ is an excellent pair, we may use Proposition \[threeone\] to establish unique ergodicity. We established that $\Aut(\hat{\TT^\omega})$ is amenable in Theorem \[thm:amen\_T\_hat\]. Let $\HH = (H, \rightarrow^H) \in \hat{\T}$ and let $H = H_1 \sqcup \ldots \sqcup H_k$ be the partition into $\perp^H$-equivalence classes, with $\vert H_i \vert = 2$ for all $i \leq k$. We consider a sequence of independent random variables $G(i,j)$ with $1 \leq i < j \leq n$. Each random variable indicates with probability $\frac{1}{2}$ that there is an edge between the equivalence classes $G_i$ and $G_j$, so that we obtain a graph in $\T$. Observe that there are only two options since, for a given vertex and equivalence class, there is exactly one in and one out vertex in this class. In this way, by doubling the points and making the canonical edge changes, the collection $G(i,j)$ of random variables gives a directed graph $\mathbf G = (G, \rightarrow^G) \in \hat{\T}$. Notice that $\# (\mathbf G) = n! \cdot 2^n = O(n!)$ and $\# (\HH) = k! \cdot 2^k$ so $\rho\expand{\HH} = \frac{1}{k! \cdot 2^k}$. In particular we have: $$I(n,k) := \EE[\Nemb(\HH,\mathbf G)] = (n)_k \cdot 2^k \cdot 2^{-\binom{k}{2}} = (n)_k \cdot 2^{-\binom{k}{2} +k}.$$ Define $$f(\mathbf G) := \frac{\Nemb(\HH,\mathbf G)}{I(n,k)}$$ so we have $\EE[f(\mathbf G)] = 1$. Changing a single value of a single $G(i,j)$ changes $\Nemb(\HH,\mathbf G)$ by not more than: $$(k)_2 \cdot (n-2)_{k-2} \cdot 2^k.$$ and $f(\mathbf G)$ by not more than: $$\frac{(k)_2 \cdot (n-2)_{k-2} \cdot 2^k}{(n)_k \cdot 2^{-\binom{k}{2}+k}} = \frac{1}{n(n-1)}\cdot \frac{(k)_2 \cdot 2}{2^{-\binom{k}{2}-k}} \leq \frac{\epsilon_1}{n^2}$$ for a large enough $n$ and some positive constant $\epsilon_1$. Now let $(\HH,\leq^H, I_1^H, I_2^H) \in \hat{\T}^*$ and let $\leq^G, I_1^G, I_2^G$ be given such that $(\mathbf G, \leq^G, I_1^G, I_2^G) \in \hat{\T}^*$. That is, each set $G_i$ comes with a partition given by $I_1^G$ and $I_2^G$, where $\leq^G$ is a linear ordering such that $G_1 \leq^G \ldots \leq^G G_k$. For $$f^*(\mathbf G) = \frac{\Nexp(\HH^*,\mathbf G^*)}{(n)_k \cdot 2^{-\binom{k}{2} +k}}$$ a change in a single $G(i,j)$ will change $f^*(\mathbf G)$ by not more than $$\frac{(k)_2 \cdot (n-2)_{k-2}}{(n)_k \cdot 2^{-\binom{k}{2}+k}} = \frac{1}{n(n-1)}\cdot \frac{(k)_2}{2^{-\binom{k}{2}+k}} \leq \frac{\epsilon_2}{n^2}$$ for a large enough $n$ and a fixed $k$. Thus we are finished by Proposition \[lem:QOP\_strategy\]. Expansions of the semi-generic digraph. {#sec:semi_exp} --------------------------------------- In Section \[sec:A\_S\] we established that the automorphism group of $\SS$, the semi-generic digraph, is amenable. In this section we provide some related expansions of $\mathcal{S} := \Age(\SS)$ and check that they satisfy the $\QOP$. This is not enough to get unique ergodicity of $\Aut(\SS)$, but provides a stepping stone to that result, and hones in on the difficulties it presents. Consider the classes $\mathcal{S}_R := \mathcal{S}^* \vert \{\rightarrow, R\}$ and $\mathcal{S}_\leq := \mathcal{S}^* \vert \{\rightarrow, \leq\}$. It is not hard to see that $\mathcal{S}_R$ and $\mathcal{S}_\leq$ are Fraïssé classes, and that $(\mathcal{S}_R, \mathcal{S}^*)$ and $(\mathcal{S}_\leq, \mathcal{S}^*)$ are excellent pairs. Let $\SS_R := \Flim(\mathcal{S}_R)$ and $\SS_\leq := \Flim(\mathcal{S}_\leq)$. \[thm:S\_exp\_UE\] $\Aut(\SS_R)$ is uniquely ergodic. Since we already have that $\Aut(\SS_R)$ is amenable, it is enough to prove the uniqueness of a consistent random expansion on a cofinal subclass of $\mathcal{S}_R$. Let $\HH := (H, \rightarrow^H, R^H) \in \mathcal{S}_R$ and let $H = H_1 \sqcup \ldots \sqcup H_k$ be the partition into $\perp^H$-equivalence classes, with $M =\vert H_i \vert = 2^{k-1}\cdot m$ for all $i \leq k$, for some natural number $m \geq 1$. Moreover, assume that $(H, \rightarrow^H)$ is in the cofinal subclass of $\mathcal{S}$ where all parts in each column have the same size. By part in each column, we mean each element of the partition of a column given by the other columns. Let $G = \bigsqcup_{i=1}^n G_i$ be a partition with $\vert G_i \vert = M$ for $i \leq n$. We consider a sequence of independent random variables $G(i,j)$ with $1 \leq i < j \leq n$. Each random variable $G(i,j)$ gives a pair of sets $(A,B)$, which is also given by $R$, such that: - $A \sse G_i$, - $B \sse G_j$, and - $\vert A \vert = \vert B \vert = \frac{M}{2}$. The partition given by $(A,B)$ is the same as the partition given by $(G_i \setminus A, G_j \setminus B)$. There are $\frac{1}{2}\cdot \binom{M}{\frac{M}{2}}^2$ such pairs and we assume that $G(i,j)$ has a uniform distribution, i.e. each pair occurs with probability $$p = 2 \cdot \binom{M}{\frac{M}{2}}^{-2}.$$ Each pair $(A,B)$ describes a distribution of edges between $G_i$ and $G_j$ such that for $x\in G_i, y \in G_j$ we have $$\begin{aligned} x \rightarrow^G y &\Leftrightarrow (x\in A, y \in G_j \setminus B) \vee (x \in G_i \setminus A, y \in B)\\ y \rightarrow^G x &\Leftrightarrow (x\in A, y \in B) \vee (x \in G_i \setminus A, y \in G_j \setminus B)\\ x \in G_i, y \in G_j &\Rightarrow (R^G(x,y) \Leftrightarrow y \in B)\\ x \in G_i, y \in G_j &\Rightarrow (R^G(y,x) \Leftrightarrow x \in A)\end{aligned}$$ In this way we obtain a random structure $\mathbf G = (G, \rightarrow^G, R^G) \in \mathcal{S}_R$. For the McDiarmid Inequality we take $\rho\expand{\HH} = \frac{1}{k! \cdot (M!)^k}$. Notice $\#(\mathbf G) = n!(M!)^n$ and $\#(\HH) = k!(M!)^k$. In particular we have: $$I(n,M,k) := \EE[\Nemb(\HH,\mathbf G)] = (n)_k \cdot (M!)^k \cdot p^{\binom{k}{2}}$$ Then for $$f(\mathbf G) := \frac{\Nemb(\HH,\mathbf G)}{I(n,M,k)}$$ we have $\EE[f(\mathbf G)] = 1$. Changing a single value of a single $G(i,j)$ changes $\Nemb(\HH,\mathbf G)$ by not more than: $$(k)_2 \cdot (n-2)_{k-2} \cdot (M!)^k,$$ and changes $f(\mathbf G)$ by at most $$\frac{(k)_2 \cdot (n-2)_{k-2} \cdot (M!)^k}{(n)_k \cdot (M!)^k \cdot p^{\binom{k}{2}}} = \frac{1}{n(n-1)}\cdot \frac{(k)_2}{p^{\binom{k}{2}}} \leq \frac{\epsilon_1}{n^2}$$ for a large enough $n$, a positive constant $\epsilon_1$ and a fixed $k$, since $p$ depends only on $M$. Now let $\expand{\HH},\expand{\mathbf G}\in \mathcal{S}_R$ where $H^* = (H, \leq^H)$ and $\mathbf G^* = (G, \leq^G)$. Without loss of generality we may assume that $H_1 <^H \ldots <^H H_k$ and $G_1 <^G \ldots <^G G_n$. For $$f^*(\mathbf G) = \frac{\Nexp(\HH^*,\mathbf G^*)}{I(n,M,k)}$$ a change in a single $G(i,j)$ will change $f^*(\mathbb G)$ by not more than $$\frac{(k)_2 \cdot (n-2)_{k-2}}{(n)_k \cdot (M!)^k \cdot p^{\binom{k}{2}}} = \frac{1}{n(n-1)}\cdot \frac{(k)_2}{ (M!)^k \cdot p^{\binom{k}{2}}} \leq \frac{\epsilon_2}{n^2}$$ for a large enough $n$, a positive constant $\epsilon_2$, and a fixed $k$. Thus we are finished by Proposition \[lem:QOP\_strategy\]. Comments about $\SS$. --------------------- The procedure outlined above fails for $\SS$. The major obstacle is that the number of expansions of a structure in $\mathcal{S}$ grows on the order of $O(2^{n^2})$, where $n$ is the number of columns. This invalidates the probabalistic argument presented in section \[sec:QOP\_strategy\], namely that the probability $p$ of finding a witness $\GG$ does not necessarily limit to $0$. The hypergraph method {#sec:hypergraph} ===================== In this section we discuss a method for proving the $\QOP$ by using hypergraphs. This method was introduced in [@AKL12], and is different from the one presented in Section \[sec:QOP\_strategy\]. The idea is to construct a large random object $\GG$ subject to some constraints. We first construct a hypergraph of large girth with many hyperedges. Then each hyperedge is replaced by a random copy of $\HH$. When checking the $\QOP$ for this structure we only examine embeddings that map $\HH$ entirely within a single hyperedge. We shall directly compute the $\QOP$ estimate and will only need a single application of McDiarmid’s Inequality. There is some subtlety in constructing $\GG$ from the hypergraph which is why we include proofs of unique ergodicity of $\Aut(\GG_n)$ and $\Aut(\FF(\T))$, even though similar statement appear in Section 5 of [@AKL12]. Our proof of Theorem \[thm:UE\_hyper\] should be compared to the proof of Theorem 5.2 in [@AKL12]. Unique Ergodicity. ------------------ \[thm:UE\_hyper\] Let $n \geq 2$ be a natural number and let $\T$ be a collection of finite tournaments, with $\vert T \vert \geq 3, \forall T \in \T$. Then $\Aut(\GG_n)$ and $\Aut(\FF(\T))$ are uniquely ergodic. Since it is more natural to have $n := \vert \GG \vert$ vary, we shall fix $m \geq 2$ and let $n$ vary. Since $(\F(\T), \F^*(\T))$ and $(\G_m, \G_m^*)$ are both excellent pairs, we may use Proposition \[threeone\] to establish amenability and unique ergodicity. Let $\K = \G_m$ or $\F(\T)$, and let $\K^* = \G_m^*$ or $\F(\T)^*$, as appropriate. Let $\AA$ be a structure in $\K$, and let $\expand{\AA}\in \K^*$. Consider $$\mu_\AA (\{\AA^*\}) := \frac{1}{\vert \AA \vert !}.$$ By [@AKL12 Proposition 9.3], $(\mu_\AA)$ is a consistent random expansion since the expansions $\K^*$ of $\K$ are just the linear orders. By Proposition \[threeone\] this ensures amenability of $\Aut(\Flim(\K))$. We check unique ergodicity by verifying the $\QOP$. Let $\HH \in \K, \expand{\HH} \in \K^*, \vert\HH\vert=k, \rho\expand{\HH} = \frac{1}{k!}$ and $\epsilon >0$. We will find a $\GG \in \K$ and a collection $\E$ of embeddings from $\HH$ into $\GG$ such that for all $\expand{\HH} \in \K^*$ and $\expand{\GG} \in \K^*$ we have: $$\left\vert \frac{\Nexp(\E,\HH^*,\GG^*)}{\vert \E \vert} - \frac{1}{k!} \right\vert < \epsilon.$$ There is a constant $C$, which depends only on $k$, such that for all $n \geq k$ there is a $k$-uniform hypergraph on $n$ vertices with at least $Cn^{\frac{4}{3}}$ hyperedges and with girth at least $4$, see Lemma 4.1 in [@AKL12]. Let $n$ be large enough and let $G$ be the underlying set of one such hypergraph, and let $E_1, \ldots, E_s$ be its hyperedges. Since the girth of the hypergraph is at least $4$, for all $i \neq j$ we have in particular that $\vert E_i \cap E_j \vert \leq 1$. \[$\K = \G_m$\] Let $x \neq y \in G$. Then: - If there is an $E_i$ such that $\{x,y\} \sse E_i$, there this is exactly one by girth, so then we choose uniformly at random an injective map $e_i: \HH \rightarrow E_i$ and take: $$x \rightarrow^G y \Leftrightarrow \left(e_i^{-1} (x) \rightarrow^H e_i^{-1} (y) \right).$$ - Otherwise, fix a directed edge between $x$ and $y$ arbitrarily. In this way we obtain a random directed graph $\mathbf G = (G, \rightarrow^G)$. The construction, and girth at least $4$, ensure that $\II_{m+1}$ can be embedded only in a subgraph induced by $E_i$. However, this is also impossible, since $\HH \in \G_m$, thus $\mathbf G \in \G_m$. In particular, the large girth ensures that $\mathbf G$ does not contain a copy of the three cycle $C_3$.  \[$\K = \F(\T)$\] Let $x \neq y \in G$. Then: - If there is an $E_i$ such that $\{x,y\} \sse E_i$, there this is exactly one by girth, so then we choose uniformly at random an injective map $e_i: \HH \rightarrow E_i$ and take: $$x \perp^G y \Leftrightarrow \left(e_i^{-1} (x) \rightarrow^H e_i^{-1} (y) \right).$$ - Otherwise, $x \perp^G y$. In this way we obtain a random directed graph $\mathbf = (G, \rightarrow^G)$. The construction, and girth at least $4$, ensure that an induced tournament can be embedded only in a subgraph induced by an $E_i$. However, this is also impossible, since $\HH \in \F(\T)$, thus $\mathbf G \in \F(\T)$. In particular, the large girth ensures that $\mathbf G$ does not contain a copy of the three cycle $C_3$. Now let us check the $\QOP$ estimate. Let $\E$ denote the collection of embeddings of $\HH$ into $\mathbf G$ whose image is completely in one of the $E_i$’s. Note that $\Nemb(\E, \HH,\mathbf G) = s \cdot L$, and $\Nexp (\E, \HH,\mathbf G)$ has a binomial distribution with parameters $(s, \frac{L}{k!})$ where $L = \vert \Aut(\HH) \vert$ and $s \geq Cn^{\frac{4}{3}}$ is the number of hyperedges of $\mathbf G$. Fix structures $\expand{\HH}, \expand{\mathbf G} \in \K^*$. For $$f(\mathbf G) := \frac{\Nexp (\E, \HH,\mathbf G)}{\Nemb(\E, \HH,\mathbf G)}$$ we have $\EE[f(\mathbf G)] = \frac{1}{k!}$. Changing a single value of a single $e_i$ changes $f(\mathbf G)$ by not more than: $$\frac{1}{s \cdot L}.$$ Thus by the McDiarmid inequality we have: $$\begin{aligned} P\left[\left\vert f(\mathbf G) - \frac{1}{k!} \right\vert \geq D \right] &\leq 2 \exp\left( \frac{-2D^2}{s \cdot \left(\frac{1}{s \cdot L}\right)^2}\right) \\ &\leq 2 \exp\left(-2 \cdot D^2 \cdot L^2 \cdot C\cdot n^{\frac{4}{3}}\right) \\ &= 2 \exp(-\delta \cdot n^{\frac{4}{3}})\end{aligned}$$ where $D, L, k$ and $C$ (hence $\delta$) do not depend on $n$. The same estimate holds for all expansions $\expand{\HH}$ and $\expand{\mathbf G}$ in $\K^*$. Therefore, since $\# (\HH) = k!$ and $\# (\mathbf G) = n!$, except on a set of measure $$k! \cdot n! \cdot 2 \exp(-\delta \cdot n^{\frac{4}{3}}),$$ which is less than $1$ for large $n$, we have $$\left\vert f(\mathbf G) - \frac{1}{k!} \right\vert \leq D.$$ In particular, choosing $D = \epsilon$ and $n$ large enough, we have our desired digraph $\GG$, which witnesses $\QOP$. Conclusion and open questions {#sec:conclusion} ============================= The most glaring open question is the following: Is $\Aut(\SS)$, the automorphism group of the semi-generic multipartite digraph, uniquely ergodic? Theorem \[thm:semi\_amen\] establishes that it is amenable, and Theorem \[thm:S\_exp\_UE\] gives us that the variant $\Aut(\SS_R)$ is uniquely ergodic. It seems as though there are just too many precompact expansions of $\SS$ for the probabilistic methods presented here to work. One approach would be to directly analyze the universal minimal flow of $\Aut(\SS)$. One could try a maximal chain construction that was successful for Uspenskij in a related context, see [@U00]. See Section 4 of [@K12], and Chapter 6 of [@pes06] for good surveys of the results and history relating to the universal minimal flow of the automorphism group of a Fraïssé structure. Give a concrete description of the universal minimal flow of $\Aut(\SS)$. In a separate direction, there are still many open questions relating to Hrushovski classes. The most pressing question here is the following, which was implicitly asked by Hrushovski in [@Hru92]. Is the class of all finite tournaments a Hrushovski class? This question seems to be quite challenging, given the relatively complicated nature of tournaments. The seemingly easier question about the class $\D_n$ of complete $n$-partite digraphs is also open and still interesting. In general, the known Hrushovski classes seem to all be relational classes with symmetric relations, so any example of a non-symmetric relation Hrushovksi class would be interesting. Appendix {#sec:appendix} ======== The appendix contains the proof of the expansion property for $\hat{\T}$. Expansion property for $\hat{\T}$. ---------------------------------- As promised in Section \[sec:def\_blowups\] we will check that $\hat{\T}^*$ satisfies **RP** and **EP**. \[lem:hatT\_Delta\] There is a map $\Delta: \hat{\T}^* \longrightarrow \T^*$ which is an injective assignment up to isomorphism, between structures in $\hat{\T}^*$ whose $\perp$-equivalence classes have exactly two elements, and the class $\T^*$ of finite ordered tournaments. Let $(A, \rightarrow^A) \in \hat{\T}$, with $\perp^A$-equivalence classes $A = A_1 \sqcup \ldots \sqcup A_k$ where each class has two elements. Consider a related structure $$\Delta(\AA) := (\{1,2, \ldots, k\}, \rightarrow^A, \leq^A)$$ such that $\leq^A$ is the natural ordering on the set $\{1,2, \ldots, k\}$ and for $1 \leq i < j \leq k$ we have: $$i \rightarrow^A j \Leftrightarrow \left(x \in A_i \wedge y \in A_j \wedge I_1^A(y) \wedge x \rightarrow^A y \right)$$ Clearly $\Delta(\AA) \in \T^*$. For $\BB = (B, \rightarrow^B, \leq^B) \in \T^*$ we may consider $$\Delta^{-1}(\BB) := (B \times \{C,P\}, \rightarrow^B, I_0^B, I_1^B, \leq^B) \in \hat{\T}^*$$ such that for $(x,i), (y,j) \in B \times \{C,P\}$ we have: - $I_1^B ((x,i)) \Leftrightarrow i = C$; - $I_0^B ((x,i)) \Leftrightarrow i = P$; - $(x,i) <^B (y,j) \Leftrightarrow \left( (x=y, i<j) \vee (x <^B y) \right)$; - $(x,i) \rightarrow^B (y,j) \Leftrightarrow \left( (y\rightarrow^B x, i=j) \vee (x \rightarrow^B y, i \neq j) \right)$. Let $\AA^\prime \leq \AA$ be such that $\AA^\prime$ and $\AA$ have the same number of $\perp$-equivalence classes. Then $\left\vert\binom{\AA}{\AA^\prime} \right\vert = 1$. Therefore, in order to verify the **RP** for $\hat{\T}^*$ it is enough to consider only structures in $\hat{\T}^*$ whose $\perp$-equivalence classes each have exactly two elements. \[thm:hatT\_Ramsey\] $\hat{\T}^*$ is a Ramsey Class. Let $n$ be a natural number and let $\AA,\BB \in \hat{\T}^*$ be such that $\binom{\BB}{\AA} \neq \emptyset$. Without loss of generality we may assume that all $\perp$-equivalence classes in $\AA$ and $\BB$ both have two elements each. Since $\T^*$ is a Ramsey class, see [@NR77; @NR83; @NR89] and [@AH78], there is a (large) $\CC \in \T^*$ such that: $$\CC \longrightarrow \left(\Delta(\BB)\right)^{\Delta(\AA)}_2.$$ Then we have: $$\Delta^{-1}(\CC) \longrightarrow \left(\BB\right)^{\AA}_2,$$ and so the verification of the Ramsey Property is complete. \[prop:hatT\_EP\] $\hat{\T}^*$ satisfies **EP** with respect to $\hat{\T}$. We will verify that for each $\AA = (A, \rightarrow^A, \leq^A, I_0^A, I_1^A) \in \hat{\T}^*$ there is an $\HH \in \hat{\T}$ such that for every $\expand{\HH}\in\hat{\T}$ we have $\AA \into \expand{\HH}$. Since $\hat{\T}^*$ satisfies the **JEP**, it is enough to obtain **EP**. Without loss of generality we may assume that each $\perp$-equivalence class in $\AA$ contains exactly two elements.  We make use of the structures $\XX = (X, \rightarrow^X, \leq^X, I_0^X, I_1^X) \in \hat{\T}^*$ and $\YY = (Y, \rightarrow^Y, \leq^Y, I_0^Y, I_1^Y) \in \hat{\T}^*$ such that: - $X = \{1,2\} \times \{P,C\}, Y = \{1\}$, - $I_1^X((1,P)), I_1^X((2,P)), I_1^Y(1)$, - $(1,P) <^X (1,C) <^X (2,P) <^X (2,C)$, - $(1,P) \rightarrow^X (2,P), (1,C) \rightarrow^X (2,C), (2,C) \rightarrow^X (1,P), (2,P) \rightarrow^X (1,C)$. Let $A_1, \ldots, A_k$ be $\perp^A$-equivalence classes which are linearly ordered such that $A_1 <^A A_2 <^A \ldots <^A A_k$. Then there is a $\BB \in \hat{\T}^*$ such that $\AA \leq \BB$ and for every $1 \leq i < k$ there is a $\perp$-equivalence class $B_i^\prime$ in $\BB$ such that: 1. $A_i <^B B_i^\prime <^B A_{i+1}$; and 2. $\BB \restrict (A_i \cup B_i^\prime) \cong \BB \restrict (B_i^\prime \cup A_{i+1}) \cong \XX$. Then there is a $\BB^\prime \in \hat{\T}^*$ such that $\BB^\prime$ and $\BB$ have the same underlying set and the same relations $I_0, I_1, \rightarrow$ but the linear ordering induced on the $\perp$-equivalence classes in $\BB^\prime$ are opposite to the linear ordering induced on the $\perp$-equivalence classes in $\BB$. Since $\hat{\T}^*$ satisfies the **JEP** there is a $\CC \in \hat{\T}^*$ such that $\BB \into \CC$ and $\BB^\prime \into \CC$. Without loss of generality we may assume that each $\perp$-equivalence class in $\CC$ contains exactly two elements. There is a $\CC^\prime \in \hat{\T}^*$ which has the same underlying set as $\CC$, the same linear ordering of equivalence classes, the same $\rightarrow$ relation, but $I_0$ and $I_1$ are inverted. Since $\hat{\T}^*$ satisfies the **JEP**. there is an $\EE \in \hat{\T}^*$ such that $\CC \into \EE$ and $\CC^\prime \into \EE$.  Since $\hat{\T}^*$ is a Ramsey class there are $\FF, \GG \in \hat{\T}^*$ such that: $$\FF \longrightarrow (\EE)^\XX_2 \textnormal{ and } \GG \longrightarrow (\FF)^\YY_2.$$ Let $\GG = (G, \rightarrow^G, \leq^G, I_0^G, I_1^G)$. We claim that $\HH = (G, \rightarrow^G)$ verifies the **EP** for $\AA$. Let $\HH^* := (G, \rightarrow^G, \leq^G, I_0^G, I_1^G)$ be such that $\expand{\HH} \in \hat{\T}^*$. Then consider the colouring: $$\xi_Y : \binom{\GG}{\YY} \longrightarrow \{0,1\}$$ such that $$\xi_Y (\YY^\prime) = 1 \Leftrightarrow I_1^H \restrict Y^\prime = I_1^G \restrict Y^\prime$$ Consider also the colouring: $$\xi_X : \binom{\GG}{\XX} \longrightarrow \{0,1\}$$ such that $$\begin{aligned} & \xi_X (\XX^\prime) = 1 \Leftrightarrow \\ & \leq^H \textnormal{ and } \leq^G \textnormal{ induce the same linear ordering on } \perp\textnormal{-equivalence classes in } \XX^\prime.\end{aligned}$$ From the construction there are $\FF^\prime \in \binom{\GG}{\FF}$ and $\EE^\prime \in \binom{\FF^\prime}{\EE}$ such that $\xi_Y$ is constantly $c_Y$ on $\binom{\FF^\prime}{\YY}$ and $\xi_X$ is constantly $c_X$ on $\binom{\EE^\prime}{\XX}$. In particular we have $\xi_Y$ is constant on $\binom{\FF^\prime}{\YY}$. Consider the following options for $(c_X, c_Y)$: - Here $I_0^H, I_1^H, \leq^H$ agree with $I_0^G, I_1^G, \leq^G$ on $\EE^\prime$ and we have that $\AA \into \EE^\prime$, so $\AA \into \expand{\HH}$. - Here $I_1^H$ and $I_0^H$ agree with $I_1^G$ and $I_0^G$ on $\BB^\prime$ respectively, but $\leq^H$ and $\leq^G$ induce opposite linear orderings on $\perp$-equivalence classes. Since $\BB^\prime \into \EE^\prime$, this embedding produces $\AA \into \expand{\HH}$. - Here $I_1^H$ and $I_0^H$ are opposite of $I_1^G$ and $I_0^G$ on $\EE^\prime$ respectively, while $\leq^H$ and $\leq^G$ agree on $\EE^\prime$. Since $\DD^\prime \into \EE^\prime$, this embedding shows that $\AA \into \expand{\HH}$. - Here $I_1^H, I_0^H, \leq^H$ are opposite of $I_1^G, I_0^G, \leq^G$ on $\EE^\prime$ respectively. Since we have that $\DD^\prime \into \EE^\prime$ and $\BB^\prime \into \EE^\prime$, there is an embedding of $\AA \into \expand{\HH}$. [^1]: **C**astor and **P**ollux are the Gemini twins. Using the set $\{C,P\}$ should help the reader, in the proofs that follow, distinguish what notation represents variables and what represents fixed objects

--- abstract: 'The classical Method of Successive Approximations (MSA) is an iterative method for solving stochastic control problems and is derived from Pontryagin’s optimality principle. It is known that the MSA may fail to converge. Using careful estimates for the backward stochastic differential equation (BSDE) this paper suggests a modification to the MSA algorithm. This modified MSA is shown to converge for general stochastic control problems with control in both the drift and diffusion coefficients. Under some additional assumptions linear rate of convergence is proved. The results are valid without restrictions on the time horizon of the control problem, in contrast to iterative methods based on the theory of forward-backward stochastic differential equations.' address: - '$^1$[Maxwell Institute Graduate School in Analysis and Applications](http://www.maxwell.ac.uk/migsaa)' - '$^2$[School of Mathematics, University of Edinburgh](https://www.maths.ed.ac.uk)' - '$^3$[Vega Protocol](https://vegaprotocol.io)' - '$^4$[Alan Turing Institute](https://www.turing.ac.uk)' author: - 'B. Kerimkulov$^{1,2}$' - 'D. Šiška$^{2,3}$' - '[Ł]{}. Szpruch$^{2,4}$' date: ', ' title: A modified MSA for stochastic control problems --- [10]{} R. Bellman, Functional equations in the theory of dynamic programming. V. Positivity and quasi-linearity, [*Proc. Nat. Acad. Sci. U.S.A.*]{}, 41(10):743–746, 1955. R. Bellman, [*Dynamic Programming*]{}, Princeton University Press, Princeton, NJ, USA, 1957. R. A. Howard, [*Dynamic Programming and Markov Processes*]{}. MIT Press, Cambridge, MA, 1960. V. G. Boltyanskii, R. V. Gamkrelidze, and L. S. Pontryagin, Theory of optimal processes. I. The maximum principle, [*Izv. Akad. Nauk SSSR Ser. Mat.*]{}, 24(1): 3–42, 1960. I. A. Krylov and F. L. Chernousko, On the method of successive approximations for solution of optimal control problems (in Russian), [*J. Comp. Mathem. and Mathematical Physics*]{}, 2:6, 1371–1382, 1962. V. D. Demyanov and A. M. Rubinov, [*Approximate Methods in Extremal Problems*]{} (in Russian), Leningrad, 1968. L. S. Pontryagin, [*Mathematical Theory of Optimal Processes*]{}, CRC Press, 1987. N. V. Krylov, [*Controlled diffusion processes*]{}, Springer, 1980. F. Antonelli, Backward-forward stochastic differential equations, [*Ann. Appl. Probab.*]{}, 3:777–793, 1993. J. Ma, P. Protter, and J. M. Yong, Solving forward-backward stochastic differential equations explicitly – a four step scheme, [*Probab. Theory Related Fields*]{}, 98:339–359, 1994. Y. Hu and S. Peng, Solution of forward-backward stochastic differential equations, [*Probab. Theory Related Fields*]{}, 103:273–283, 1995. J. Douglas, J. Ma, and P. Protter, Numerical methods for forward-backward stochastic differential equations, [*Ann. Appl. Probab.*]{}, 6(3):940–968, 1996. J. Yong, Finding adapted solutions of forward-backward stochastic differential equations: Method of continuation, [*Probab. Theory Related Fields*]{}, 107:537–572, 1997. E. Pardoux and S. Tang, Forward-backward stochastic differential equations and quasilinear parabolic PDEs, [*Probab. Theory Related Fields*]{}, 114(2):123–150, 1999. S. Peng and Z. Wu, Fully coupled forward-backward stochastic differential equations and applications to optimal control, [*SIAM J. Control Optim.*]{} 37:825–843, 1999. F. Delarue, On the existence and uniqueness of solutions to FBSDEs in a non-degenerate case, [*Stochastic Processes and their Applications*]{}, 99:209–289, 2002. F. Delarue and S. Menozzi, A forward-backward stochastic algorithm for quasi-linear PDEs, [*Ann. Appl. Probab.*]{}, 16(1):140–184, 2006. G. Milstein and M. Tretyakov, Discretization of forward-backward stochastic differential equations and related quasi-linear parabolic equations, [*IMA J. Numer. Anal.*]{}, 27(1):24–44, 2007. H. Dong and N. V. Krylov, The rate of convergence of finite-difference approximations for parabolic Bellman equations with Lipschitz coefficients in cylindrical domains, [*Appl. Math. Optim.*]{}, 56(1):37–66, 2007. J. Ma, J. Shen, and Y. Zhao, On numerical approximations of forward-backward stochastic differential equations, [*SIAM J. Numer. Anal.*]{}, 46(5):2636–2661, 2008. I. Gyöngy and D. Šiska, On Finite-Difference Approximations for Normalized Bellman Equations, [*Appl. Math. Optim.*]{}, 60:297–339, 2009. H. Pham, [*Continuous-time stochastic control and optimization with financial applications*]{}, Springer, 2009. W. Guo, J. Zhang, and J. Zhuo, A monotone scheme for high-dimensional fully nonlinear PDEs, [*Ann. Appl. Probab.*]{}, 25(3):1540–1580, 2015. R. Carmona, [*Lectures on BSDEs, Stochastic Control, and Stochastic Differential Games with Financial Applications*]{}, Society for Industrial and Applied Mathematics, 2016. J. Zhang, [*Backward Stochastic Differential Equations: From Linear to Fully Nonlinear Theory*]{}, Springer-Verlag, New York, 2017. S. D. Jacka and A. Mijatović, On the policy improvement algorithm in continuous time, [*Stochastics*]{}, 89(1):348–359, 2017. S. D. Jacka, A. Mijatović, and D. Siraj, Coupling and a generalised Policy Iteration Algorithm in continuous time, [*arXiv*]{}:1707.07834, 2017. J. Maeda and S. D. Jacka, Evaluation of the Rate of Convergence in the PIA, [*arXiv*]{}:1709.06466, 2017. M. Sabate Vidales, D. Šiška, and Ł. Szpruch, Unbiased deep solvers for parametric PDEs, [*arXiv*]{}:1810.05094v2, 2018. J. Han, A. Jentzen, and W. E, Solving high-dimensional partial differential equations using deep learning, [*Proceedings of the National Academy of Sciences*]{}, 115(34):8505–8510, 2018. Q. Li, L. Chen, C.Tai, and W. E, Maximum principle based algorithms for deep learning, [*The Journal of Machine Learning Research*]{}, 18(165), 1–29, 2018. S. Ji, S. Peng, Y. Peng, and X. Zhang, Three algorithms for solving high-dimensional fully-coupled FBSDEs through deep learning, [*IEEE Intelligent Systems*]{}, 35(3):71–84, 2020. B. Kerimkulov, D. Šiška, and Ł. Szpruch. Exponential convergence and stability of Howard’s policy improvement algorithm for controlled diffusions, [*SIAM J. Control Optim.*]{}, 58(3), 1314–1340, 2020. D. Šiška and Ł. Szpruch, Gradient Flows for Regularized Stochastic Control Problems, [*arXiv*]{}:2006.05956, 2020.

--- abstract: 'Let $S$ and $R$ be rings and $_SC_R$ a (faithfully) semidualizing bimodule. We introduce and study $C$-weak flat and $C$-weak injective modules as a generalization of $C$-flat and $C$-injective modules (J. Math. Kyoto Univ. [**47**]{}(2007), 781–808) respectively, and use them to provide additional information concerning the important Foxby equivalence between the subclasses of the Auslander class $\mathcal{A}_C(R)$ and that of the Bass class $\mathcal{B}_C(S)$. Then we study the stability of Auslander and Bass classes, which enables us to give some alternative characterizations of the modules in $\mathcal{A}_C(R)$ and $\mathcal{B}_C(S)$. Finally we consider an open question which is closely relative to the main results (Proc. Edinb. Math. Soc. [**48**]{}(2005), 75–90), and discuss the relationship between the Bass class $\mathcal{B}_C(S)$ and the class of Gorenstein injective modules.' author: - | Zenghui Gao[^1],  Tiwei Zhao[^2]\ [*1. College of Applied Mathematics, Chengdu University of Information Technology, Chengdu 610225, P.R. China*]{}\ [*2. Department of Mathematics, Nanjing University, Nanjing 210093, P.R. China*]{} date: title: ' **Foxby equivalence relative to $C$-weak injective and $C$-weak flat modules [^3] [^4]**' --- Introduction {#introduction .unnumbered} ============ Over a commutative Noetherian ring $R$, a finitely generated $R$-module $C$ is *semidualizing* if the natural homothety morphism $R\ra \Hom_R(C,C)$ is an isomorphism and $\Ext^i_R(C,C)=0$ for all $i\geq 1$. Semidualizing modules (under different names) were independently studied by Foxby, Golod and Vasconcelos (see [@Fo73; @Go84; @Va74]). In [@Ch01], Christensen extended this notion to semidualizing complexes. Araya, Takahashi and Yoshino [@ATY05] extended the notion of semidualizing modules to a pair of non-commutative, but Noetherian rings. Furthermore, Holm and White in [@HW07] generalized the notion of a semidualizing module to general associative rings, and defined and studied Auslander and Bass classes with respect to a semidualizing bimodule $C$. They obtained some beautiful characterizations of the modules in the Auslander and Bass classes in terms of $C$-injective, $C$-projective and $C$-flat modules and showed Foxby equivalence between the subclasses of the Auslander class and that of the Bass class. In particular, it was proven in [@HW07 Lemma 4.1] that the Auslander class $\mathcal{A}_C(R)$ contains all flat left $R$-modules and the Bass class $\mathcal{B}_C(S)$ contains all injective left $S$-modules. Recently, Bennis et al. continued a study of homological notions relative to an extension of a semidualizing module (see [@BGO16a; @BGO16b; @BGO16c]). More recently, Gao and his coauthors introduced and studied in [@GH15a; @GW15] a generalization of injective and flat modules, named weak injective and weak flat modules respectively, and generalized many homological results from coherent rings to arbitrary rings. In this process finitely presented modules are replaced by super finitely presented modules. In [@BGH14], Bravo, Gillespie and Hovey described how Gorenstein homological algebra should work for general rings, and the weak injective and weak flat modules were also called $FP_\infty$-injective (or absolutely clean) and level modules respectively. Following the above philosophy, the following question naturally arises in this situation: [**Question 1.**]{} Is it true that the Auslander class $\mathcal{A}_C(R)$ contains all weak flat left $R$-modules and the Bass class $\mathcal{B}_C(S)$ contains all weak injective left $S$-modules? In [@EJL05], Enochs, Jenda and López-Ramos proved that if $R$ and $S$ are right and left Noetherian rings respectively admitting a dualizing bimodule (see [@EJL05 Definition 3.1]), then all Gorenstein projective left $R$-modules are in $\mathcal{A}(R)$ ([@EJL05 Proposition 3.9]) and all Gorenstein injective left $S$-modules are in $\mathcal{B}(S)$ ([@EJL05 Proposition 3.8]). Moreover, if every flat left $R$-module has finite projective dimension, then a left $S$-module $N\in \mathcal{B}(S)$ if and only if $N$ has finite Gorenstein injective dimension by [@EJL05 Lemma 3.15 and Proposition 3.13]; dually, we can deduce that a left $R$-module $M\in \mathcal{A}(R)$ if and only if $M$ has finite Gorenstein projective dimension. In view of the relationship between the Auslander class (resp. Bass class) and the class of Gorenstein projective (resp. Gorenstein injective) modules in [@EJL05], it is natural to ask the following question: [**Question 2.**]{} Is there an appropriate semidualizing bimodule $C$ such that the Auslander class $\mathcal{A}_C(R)$ contains all Gorenstein projective $R$-modules and the Bass class $\mathcal{B}_C(S)$ contains all Gorenstein injective $S$-modules? The aim of this paper is to study these two questions, and we will define and investigate $C$-weak injective and $C$-weak flat modules with respect to a semidualizing bimodule $C$. Suppose that $C$ is a faithfully semidualizing bimodule. We provide additional information concerning the important Foxby equivalence between the subclasses of Auslander class $\mathcal{A}_C(R)$ and that of the Bass class $\mathcal{B}_C(S)$. In addition, we study the stability of the Auslander and Bass classes, and some new characterizations of the modules in the Auslander and Bass classes are given. We will answer Question 1 in Theorem \[2.2\], and give a partial answer to Question 2 at the end of the paper, that is, it is shown that if ${}_SC_R$ is a faithfully semidualizing bimodule with finite $S$-projective dimension, then every Gorenstein injective left $S$-module is in $\mathcal{B}_C(S)$. This paper is organized as follows. In Section 1, we give some terminology and some preliminary results. In Section 2, we introduce the notions of $C$-weak injective and $C$-weak flat modules with respect to a semidualizing bimodule $C$, and prove that the Auslander class $\mathcal{A}_C(R)$ contains all weak flat left $R$-modules and the Bass class $\mathcal{B}_C(S)$ contains all weak injective left $S$-modules. We show that $C$-weak injective and $C$-weak flat modules possess many nice properties analogous to that of $C$-injective and $C$-flat (or $C$-projective) modules as in [@HW07]. For example, we prove that the classes $\mathcal{WF}_C(S)$ and $\mathcal{WI}_C(R)$, consisting of all $C$-weak flat left $S$-modules and all $C$-weak injective left $R$-modules respectively, are closed under direct summands, direct products, direct sums and direct limits. Also, both of them are closed under pure submodules and pure quotients. As a consequence, we obtain that the classes $\mathcal{WF}_C(S)$ and $\mathcal{WI}_C(R)$ are covering and preenveloping. In Section 3, we investigate Foxby equivalence relative to $C$-weak injective and $C$-weak flat modules. The following is Theorem \[3.4\]. Here $\mathcal{WF}(R)_{\leq n}$ and $\mathcal{WI}(S)_{\leq n}$ stand for the class of left $R$-modules of weak flat dimension at most $n$ and the class of left $S$-modules of weak injective dimension at most $n$, respectively; and $\mathcal{WF}_C(S)_{\leq n}$ and $\mathcal{WI}_C(R)_{\leq n}$ denote the class of left $S$-modules of $C$-weak flat dimension at most $n$ and the class of left $R$-modules of $C$-weak injective dimension at most $n$, respectively. [**Theorem A.**]{} [(Foxby Equivalence)]{} [*There are equivalences of categories $$\xymatrix@R=20pt@C=60pt{ \mathcal{WF}(R) \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[d] & \mathcal{WF}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[d]\\ \mathcal{WF}(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[d] & \mathcal{WF}_C(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[d]\\ \mathcal{A}_C(R) \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{B}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\\ \mathcal{WI}_C(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[u] & \mathcal{WI}(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[u]\\ \mathcal{WI}_C(R) \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[u] & \mathcal{WI}(S). \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[u]\\}$$*]{} In Section 4, we characterize the stability of Auslander class $\mathcal{A}_C(R)$ and the Bass class $\mathcal{B}_C(S)$, and then give some applications of them. Motivated by [@SSW08 Theorem A], we show that an iteration of the procedure used to describe the Auslander class yields exactly the Auslander class, which generalizes [@HW07 Theorem 2]. That is, we set $[\mathcal{A}_C(R)]^1=\mathcal{A}_C(R)$, and inductively set $[\mathcal{A}_C(R)]^{n+1}=\{M\in \Mod R\mid$ there exists a $C\otimes_R-$ exact exact sequence $\cdots\ra W_1\ra W_0\ra W^0\ra W^1\ra\cdots$ in $\Mod R$ with all $W_i$ and $W^i$ in $[\mathcal{A}_C(R)]^{n}$ such that $M\cong\coker(W_1\ra W_0)\}$ for any $n\geq 1$. Similarly, we inductively set $[\mathcal{B}_C(S)]^{n}$. The following are Theorems \[4.7\] and \[4.8\]. [**Theorem B.**]{} [*$[\mathcal{A}_C(R)]^{n}=\mathcal{A}_C(R)$ and $[\mathcal{B}_C(S)]^{n}=\mathcal{B}_C(S)$ for any $n\geq 1$.* ]{} Preliminaries ============= In this section, we give some terminology and some preliminary results needed in the sequel. For more details the reader can consult [@EJ00; @GW15; @GT12; @HJ06; @HW07; @TW10]. [**1.1**]{} Throughout this paper, $R$ and $S$ are fixed associative rings with unites, and all modules are unitary. We use $\Mod R$ or $\Mod S$ to stand for the class of left $R$- or $S$-modules. Right $R$- or $S$-modules are identified with left modules over the opposite rings $R^{op}$ or $S^{op}$. The notation ${}_SM_R$ is used to indicate that $M$ is an $(S,R)$-bimodule, and the structures are compatible in the sense that $s(xr)=(sx)r$ for all $s\in S, r\in R, x\in M$. For a left or right $R$-module $M$, $M^+=\Hom_\mathbb{Z}(M,\mathbb{Q}/\mathbb{Z})$. [**1.2.**]{} A *degreewise finite projective resolution* of a left $R$-module $M$ is a projective resolution of $M$:  $\cdots\ra P_n\ra \cdots\ra P_1\ra P_0\ra M\ra 0$ in $\Mod R$ with each $P_i$ finitely generated projective. Note that a left $R$-module admitting a degreewise finite projective resolution is also called $FP_{\infty}$ in [@BGH14; @Br82; @HM09], *infinitely presented* in [@BM07], *strongly finitely presented* in [@GT12], and *super finitely presented* in [@GW15]. Also, it is shown that this class of modules plays a crucial role in the process of generalizing many homological results from coherent rings to arbitrary rings (see [@BGH14; @GH15a; @GW15]). [**1.3.**]{} An $(S,R)$-bimodule $C={}_SC_R$ is *semidualizing* if (a1) ${}_SC$ admits a degreewise finite projective resolution in $\Mod S$. (a2) $C_R$ admits a degreewise finite projective resolution in $\Mod R^{op}$. (b1) The homothety map ${}_SS_S\buildrel{{}_S\gamma}\over\lra\Hom_{R^{op}}(C,C)$ is an isomorphism. (b2) The homothety map ${}_RR_R\buildrel{\gamma_R}\over\lra\Hom_{S}(C,C)$ is an isomorphism. (c1) $\Ext_S^i(C,C)=0$ for all $i\geq 1$. (c2) $\Ext_{R^{op}}^i(C,C)=0$ for all $i\geq 1$. A semidualizing bimodule ${}_SC_R$ is *faithfully semidualizing* if it satisfies the following conditions for all modules ${}_SN$ and $M_R$: \(1) If $\Hom_S(C,N)=0$, then $N=0$. \(2) If $\Hom_{R^{op}}(C,M)=0$, then $M=0$. By definition, it follows that every semidualizing module is super finitely presented as a left $S$-module or a right $R$-module. It was shown in [@HW07 Proposition 3.1] that if $R=S$ is commutative, then every semidualizing $R$-module is faithfully semidualizing. Also in [@HW07] many examples of faithfully semidualizing bimodules were provided over a wide class of non-commutative rings. [**1.4.**]{} The *Auslander class* $\mathcal{A}_C(R)$ with respect to $C$ consists of all modules $M$ in $\Mod R$ satisfying: (A1) $\Tor_i^R(C,M)=0$ for all $i\geq 1$. (A2) $\Ext_S^i(C,C\otimes_RM)=0$ for all $i\geq 1$. (A3) The natural evaluation homomorphism $\mu_{_M}: M\longrightarrow\Hom_S(C,C\otimes_RM)$ is an isomorphism (of left $R$-modules). The *Bass class* $\mathcal{B}_C(S)$ with respect to $C$ consists of all modules $N\in\Mod S$ satisfying: (B1) $\Ext_S^i(C,N)=0$ for all $i\geq 1$. (B2) $\Tor^R_i(C,\Hom_S(C,N))=0$ for all $i\geq 1$. (B3) The natural evaluation homomorphism $\nu_{_N}: C\otimes_R\Hom_S(C,N)\longrightarrow N$ is an isomorphism (of left $S$-modules). It is an important property of Auslander and Bass classes that they are equivalent under the pair of functors ([@HW07 Proposition 4.1]): $$\xymatrix@C=80pt{ \mathcal{A}_C(R) \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{B}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}. }$$ [**1.5.**]{} Let $\mathcal{F}$ be a subcategory of $\Mod R$. The homomorphism $f:F\ra M$ in $\Mod R$ with $F\in \mathcal{F}$ is an *$\mathcal{F}$-precover* of $M$ if for any homomorphism $g:F_0\ra M$ in $\Mod R$ with $F_0\in \mathcal{F}$, there exists a homomorphism $h:F_0\ra F$ such that the following diagram commutes: $$\xymatrix{ & F_0 \ar[d]^{g} \ar@{-->}[ld]_{h}\\ F \ar[r]^{f} & M.}$$ The homomorphism $f:F\ra M$ is *right minimal* if an endomorphism $h:F\ra F$ is an automorphism whenever $f=fh$. An $\mathcal{F}$-precover $f:F\ra M$ is an *$\mathcal{F}$-cover* if $f$ is right minimal. We say that $\mathcal{F}$ is [*(pre)covering*]{} if every module in $\Mod R$ admits an $\mathcal{F}$-(pre)cover. Dually, the notions of an [*$\mathcal{F}$-preenvelope*]{}, a [*left minimal homomorphism*]{}, an [*$\mathcal{F}$-envelope*]{} and a [*(pre)enveloping subcategory*]{} are defined. [**1.6.**]{} We say that a sequence $\mathbf{X}=\ \ \cdots\ra X_1\ra X_0\ra X_{-1}\ra \cdots$ in $\Mod R$ (resp. in $\Mod S^{op}$) is $C\otimes_R-$ (resp. $-\otimes_S C$) exact if the complex $C\otimes_R\mathbf{X}$ (resp. $\mathbf{X}\otimes_S C$) is exact; and a sequence $\mathbf{X}$ in $\Mod S$ is $\Hom_S(C,-)$ (resp. $\Hom_S(-,C)$) exact if the complex $\Hom_S(C,\mathbf{X})$ (resp. $\Hom_S(\mathbf{X},C)$) is exact. We denote by $\mathcal{X}$ a fixed class of left $R$-modules. An *$\mathcal{X}$-resolution* of a left $R$-module $M$ is an exact sequence $\mathbf{X}=\ \ \cdots\ra X_1\ra X_0\ra M\ra 0$ in $\Mod R$ with $X_i\in \mathcal{X}$ for all $i\geq 0$. An *$\mathcal{X}$-coresolution* of a left $R$-module $M$ is an exact sequence $\mathbf{X}=\ \ 0 \ra M\ra X^0\ra X^1\ra\cdots$ in $\Mod R$ with $X^i\in \mathcal{X}$ for all $i\geq 0$. If the class $\mathcal{X}$ is precovering, then for any left $R$-module $M$, there exists an *augmented proper $\mathcal{X}$-resolution* of $M$, that is, a complex $$\mathbf{X}=\ \ \cdots\overset{\partial_{2}^\mathbf{X}}\lra X_1\overset{\partial_{1}^\mathbf{X}}\lra X_0\lra M\lra 0$$ in $\Mod R$ such that it is $\Hom_R(X,-)$ exact for each $X\in \mathcal{X}$. The truncated complex $$\mathbf{X}_M=\ \ \cdots\overset{\partial_{3}^\mathbf{X}}\lra X_2\overset{\partial_{2}^\mathbf{X}}\lra X_1\overset{\partial_{1}^\mathbf{X}}\lra X_0\lra 0$$ is a *proper $\mathcal{X}$-resolution* of $M$. In general, an augmented proper $\mathcal{X}$-resolution $\mathbf{X}$ need not be exact. However, the complex $\mathbf{X}$ is exact if $\mathcal{X}$ contains all projective left $R$-modules. Dually, the augmented coproper $\mathcal{X}$-coresolutions are defined, and they must be exact if the class $\mathcal{X}$ contains all injective left $R$-modules. [**1.7.**]{} A module in $\Mod S$ is *$C$-flat* (resp. *$C$-projective*) if it has the form $C\otimes_RF$ for some flat (resp. projective) module $F\in\Mod R$. A module in $\Mod R$ is *$C$-injective* if it has the form $\Hom_S(C,I)$ for some injective module $I\in\Mod S$. We set $$\begin{array}{ll}\vspace{0.1cm} &\mathcal{F}_C(S)=\{C\otimes_RF\mid F \mbox{ is a flat left $R$-module}\},\\\vspace{0.1cm} &\mathcal{P}_C(S)=\{C\otimes_RF\mid P \mbox{ is a projective left $R$-module}\},\\\vspace{0.1cm} &\mathcal{I}_C(R)=\{\Hom_S(C,I)\mid I \mbox{ is an injective left $S$-module}\}.\\ \end{array}$$ Over a commutative ring $R$, the notions of $C$-projective and $C$-injective dimensions of an $R$-module were introduced in [@TW10]. That is, the *$\mathcal{P}_C$-projective dimension* of an $R$-module $M$ is $$\mathcal{P}_C\mbox{-}\pd(M)=\mbox{inf}\{\mbox{sup}\{n\mid X_n\neq 0\}\mid \mbox{$X$ is a proper $\mathcal{P}_C$-projective resolution of $M$}\}.$$ The *$\mathcal{I}_C$-injective dimension*, denoted by $\mathcal{I}_C$-$\id(-)$, can be defined dually. It was also proven in [@TW10 Corollary 2.10] that \(a) $\mathcal{P}_C\mbox{-}\pd(M)\leq n$ if and only if there is an exact sequence $$0\ra C\otimes_RP_n\ra\cdots\ra C\otimes_RP_1\ra C\otimes_RP_0\ra M\ra 0$$ with each $P_i$ a projective $R$-module. \(b) $\mathcal{I}_C\mbox{-}\id(M)\leq n$ if and only if there is an exact sequence $$0\ra M\ra \Hom_R(C,I^0)\ra\Hom_R(C,I^1)\ra\cdots\ra\Hom_R(C,I^n)\ra 0$$ with each $I^i$ an injective $R$-module. [**1.8.**]{} A module $M$ in $\Mod R$ (resp. $N$ in $\Mod R^{op}$) is *weak injective* (resp. *weak flat*) if $\Ext_R^1(F,M)=0$ (resp. $\Tor^R_1(N,F)=0$) for any super finitely presented left $R$-module $F$. We use $\mathcal{WI}(R)$ (resp. $\mathcal{WF}(R^{op})$) to denote the full subcategory of $\Mod R$ (resp. $\Mod R^{op}$) consisting of weak injective modules (resp. weak flat modules). The weak injective dimension of a module $M$ in $\Mod R$, denoted by $\wid_R(M)$, is defined as $$\begin{aligned} \wid_R(M)= \inf\left\{ \right.&n\mid\Ext_R^{n+1}(F,M)=0\mbox{ for any super} \\ & \left.\mbox{finitely presented left $R$-module $F$}\right\}.\end{aligned}$$ $ $ If no such $n$ exists, we set $\wid_R(M)=\infty$. Dually, the weak flat dimension $\wid_R(-)$ of a module is defined. Given a short exact sequence $0\ra A\ra B\ra C\ra 0$ in $\Mod R$ where $A$ is a submodule of $B$ and $C$ is the corresponding quotient module. The sequence is said to be *pure exact* if $\Hom_R(P,B)\ra \Hom_R(P,C)\ra 0$ is exact for any finitely presented module $P$ in $\Mod R$, or equivalently, if $0\ra M\otimes_R A\ra M\otimes_R B$ is exact for any module $M$ in $\Mod R^{op}$. In this case, $A$ and $C$ are called a *pure submodule* and a *pure quotient* of $B$ respectively. $C$-weak injective and $C$-weak flat modules ============================================ In this section, we give a treatment of $C$-weak injective and $C$-weak flat modules with respect to a (faithful) semidualizing bimodule $C$. \[2.1\] A module in $\Mod S$ is called *$C$-weak flat* if it has the form $C\otimes_RF$ for some weak flat module $F\in \Mod R$. A module in $\Mod R$ is called *$C$-weak injective* if it has the form $\Hom_S(C,I)$ for some weak injecitve module $I\in \Mod S$. We set $$\begin{array}{ll}\vspace{0.1cm} &\mathcal{WF}_C(S)=\{C\otimes_RF\mid F \mbox{ is a weak flat left $R$-module}\},\\ &\mathcal{WI}_C(R)=\{\Hom_S(C,I)\mid I \mbox{ is a weak injective left $S$-module}\}.\\ \end{array}$$ The *$C$-weak flat dimension* of a module $M\in \Mod S$ is defined that $C\mbox{-}\wfd_S(M)\leq n$ if and only if there is an exact sequence $$0\ra C\otimes_RF_n\ra\cdots\ra C\otimes_RF_1\ra C\otimes_RF_0\ra M\ra 0$$ in $\Mod S$ with each $F_i$ in $\mathcal{WF}(R)$. If no such $n$ exists, set $C\mbox{-}\wfd_S(M)=\infty$. The *$C$-weak injective dimension* of a module $M\in \Mod R$ is defined that $C\mbox{-}\wid_R(M)\leq n$ if and only if there is an exact sequence $$0\ra M\ra \Hom_S(C,I^0)\ra\Hom_S(C,I^1)\ra\cdots\ra\Hom_S(C,I^n)\ra 0$$ in $\Mod R$ with each $I^i$ in $\mathcal{WI}(S)$. If no such $n$ exists, set $C\mbox{-}\wid_R(M)=\infty$. The following theorem gives an affirmative answer to Question 1. \[2.2\] [The Auslander class $\mathcal{A}_C(R)$ contains all weak flat modules in $\Mod R$, and the Bass class $\mathcal{B}_C(S)$ contains all weak injective modules in $\Mod S$.]{} We only show that $\mathcal{A}_C(R)$ contains all weak flat modules in $\Mod R$, the other proof is dual. Let $M$ be a weak flat module in $\Mod R$. Then $\Tor_i^R(C,M)=0$ for all $i\geq 1$ by [@GW15 Proposition 3.1]. Since ${}_SC_R$ is semidualizing, it follows that $\Ext_S^i(C,C)=0$ for all $i\geq 1$, and there exists a $\Hom_S(-,C)$-exact exact sequence $$\cdots\lra P_n\lra\cdots\lra P_1\lra P_0\lra C\lra 0\eqno{(2.1)}$$ in $\Mod S$ with all $P_i$ finite generated projective. This gives rise to the exactness of $$0\ra\Hom_S(C,C)\ra \Hom_S(P_0,C)\ra \cdots\ra \Hom_S(P_n,C)\ra\cdots.$$ One easily checks that each $\Hom_S(P_i,C)$ is super finitely presented by applying $\Hom_S(P_i,-)$ to the sequence (2.1). Now that $\Hom_S(C,C)\cong R$ since ${}_SC_R$ is semidualizing, it follows that all $\im(\Hom_S(P_i,C)\ra \Hom_S(P_{i+1},C))$ are super finitely presented by [@HM09 Lemma 2.3]. Because $M$ is a weak flat left $R$-module, one gets the following exact sequence $$0\ra\Hom_S(C,C)\otimes_RM\ra \Hom_S(P_0,C)\otimes_RM\ra\Hom_S(P_1,C)\otimes_RM\ra\cdots.$$ By the tensor evaluation morphism (cf.[@HW07 1.10]), we have the following isomorphisms: $$\omega_{P_iCM}:\Hom_S(P_i,C)\otimes_RM\buildrel{\cong}\over\lra\Hom_S(P_i,C\otimes_RM) \ \ \mbox{for all } i\geq 0$$ since $P_i$ is finitely generated projective. Thus we obtain the following commutative diagram with exact rows: $$\xymatrix@C=10pt@R=15pt{0 \ar[r] & \Hom_S(C,C)\otimes_RM \ar[d]_{\omega_{CCM}} \ar[r] & \Hom_S(P_0,C)\otimes_RM \ar[d]_{\omega_{P_0CM}}^{\cong}\ar[r] & \Hom_S(P_1,C)\otimes_RM \ar[d]_{\omega_{P_1CM}}^{\cong} \\ 0 \ar[r] & \Hom_S(C,C\otimes_RM) \ar[r] & \Hom_S(P_0,C\otimes_RM) \ar[r] &\Hom_S(P_1,C\otimes_RM).}$$ Hence $\omega_{CCM}:\Hom_S(C,C)\otimes_RM\ra\Hom_S(C,C\otimes_RM)$ is an isomorphism by the five lemma, which implies that $\mu_M:M\ra\Hom_S(C,C\otimes_RM)$ is an isomorphism. Now consider the following commutative diagram with the upper row exact: $$\xymatrix@C=10pt@R=15pt{0 \ar[r] & \Hom_S(C,C)\otimes_RM \ar[d]^{\cong} \ar[r] & \Hom_S(P_0,C)\otimes_RM \ar[d]^{\cong}\ar[r] & \Hom_S(P_1,C)\otimes_RM \ar[d]^{\cong} \ar[r] & \cdots\\ 0 \ar[r] & \Hom_S(C,C\otimes_RM) \ar[r] & \Hom_S(P_0,C\otimes_RM) \ar[r] &\Hom_S(P_1,C\otimes_RM)\ar[r] & \cdots.}$$ Then we obtain the exactness of $$0\ra\Hom_S(C,C\otimes_RM)\ra \Hom_S(P_0,C\otimes_RM)\ra\Hom_S(P_1,C\otimes_RM)\ra\cdots.$$ It follows that $\Ext_S^i(C,C\otimes_RM)=0$ for all $i\geq 1$, and hence $M\in \mathcal{A}_C(R)$. In what follows, $C={}_SC_R$ always stands for a faithfully semidualizing bimodule. By Theorem \[2.2\] and [@HW07 Theorem 6.3], we immediately get the following result. \[2.3\] [The Auslander class $\mathcal{A}_C(R)$ contains all modules in $\Mod R$ of finite weak flat dimension, the Bass class $\mathcal{B}_C(S)$ contains all modules in $\Mod S$ of finite weak injective dimension. ]{} The following result plays a fundamental role in this paper. \[2.4\] The following statements hold for modules ${}_SV$ and ${}_RU$: $\mathrm{(1)}$ $V\in\mathcal{WF}_C(S)$ if and only if $V\in\mathcal{B}_C(S)$ and $\Hom_S(C,V)$ is weak flat over $R$. $\mathrm{(2)}$ $U\in\mathcal{WI}_C(R)$ if and only if $U\in\mathcal{A}_C(R)$ and $C\otimes_RU$ is weak injective over $S$. We only prove (1), and (2) is dual. “Only if" part. Let $V\in\mathcal{WF}_C(S)$. Then $V=C\otimes_RF$ for some weak flat left $R$-module $F$. Note that $F\in \mathcal{A}_C(R)$ by Theorem \[2.2\], it follows that $V\in \mathcal{B}_C(S)$ by [@HW07 Proposition 4.1]. It is clear that $\Hom_S(C,C\otimes_RF)\cong F$, and so $\Hom_S(C,V)$ is a weak flat left $R$-module. “If" part. Suppose that $V\in\mathcal{B}_C(S)$ and $\Hom_S(C,V)$ is a weak flat left $R$-module. Then it is clear that $V\cong C\otimes_R\Hom_S(C,V)$, and thus $V$ is a $C$-weak flat left $S$-module. \[2.5\] The following statements hold. $\mathrm{(1)}$ The class $\mathcal{WF}_C(S)$ is closed under extensions and kernels of epimorphisms. $\mathrm{(2)}$ The class $\mathcal{WI}_C(R)$ is closed under extensions and cokernels of monomorphisms. We only prove (1), and (2) is the dual of (1). Let $$0\ra M'\ra M\ra M''\ra 0 \eqno{(2.2)}$$ be a short exact sequence in $\Mod S$. If $M', M''\in \mathcal{WF}_C(S)$, then $M', M''\in \mathcal{B}_C(S)$ and $\Hom_S(C,M')$ and $\Hom_S(C,M'')$ are weak flat left $R$-modules by Proposition \[2.4\]. It follows that $M\in\mathcal{B}_C(S)$ by [@HW07 Theorem 6.2]. Also, we have $\Ext^1_S(C,M')=0$. Using $\Hom_S(C,-)$ to the sequence (2.2), we obtain the following exactness of $$0\ra \Hom_S(C,M')\ra \Hom_S(C,M)\ra \Hom_S(C,M'')\ra 0$$ in $\Mod R.$ Then $\Hom_S(C,M)$ is weak flat by [@GH15a Proposition 2.6(2)]. Hence $M\in\mathcal{WF}_C(S)$ by Proposition \[2.4\](1). Assume that $M, M''\in \mathcal{WF}_C(S)$ in the sequence (2.2), then $M, M''\in \mathcal{B}_C(S)$, and $\Hom_S(C,M)$ and $\Hom_S(C,M'')$ are weak flat left $R$-modules. By [@HW07 Theorem 6.3], we have $M'\in\mathcal{B}_C(S)$, and so $\Ext^1_S(C,M')=0$. Applying $\Hom_S(C,-)$ to the sequence (2.2), one gets the following exact sequence $$0\ra \Hom_S(C,M')\ra \Hom_S(C,M)\ra \Hom_S(C,M'')\ra 0$$ in $\Mod R$. Then, by [@GH15a Proposition 2.6(2)], we have $\Hom_S(C,M')$ is weak flat. It follows from Proposition \[2.4\](1) that $M'\in\mathcal{WF}_C(S)$, as desired. The following result generalizes [@GW15 Theorem 2.10, and Remark 2.2(2)]. \[2.6\] The following statements hold for modules ${}_SV$ and ${}_RU$. $\mathrm{(1)}$ $V\in \mathcal{WF}_C(S)$ if and only if $V^+\in \mathcal{WI}_C(S^{op})$. $\mathrm{(2)}$ $U\in \mathcal{WI}_C(R)$ if and only if $U^+\in \mathcal{WF}_C(R^{op})$. \(1) “Only if" part. Let $V\in \mathcal{WF}_C(S)$. Then by definition $V=C\otimes_RF$ for some weak flat left $R$-module $F$. Since $F^+$ is weak injective in $\Mod R^{op}$ by [@GW15 Remark 2.2(2)], it follows that $V^+\cong\Hom_{R^{op}}(C,F^+)\in \mathcal{WI}_C(S^{op})$. “If" part. Let $V^+\in \mathcal{WI}_C(S^{op})$. Then we have $V^+\in\mathcal{A}_C(S^{op})$ and $V^+\otimes_SC$ is weak injective over $R^{op}$ by Proposition \[2.4\](2). It is not hard to check that $V\in\mathcal{B}_C(S)$ by a non-commutative version of [@HW07 Proposition 7.2(b) and Remark 4]. So we have the isomorphism: $V^+\otimes_SC\cong\Hom_S(C,V)^+$ by [@GT12 Lemma 2.16] since $C$ is finitely presented. It follows from [@GW15 Remark 2.2(2)] that $\Hom_S(C,V)$ is weak flat. Thus $V\in \mathcal{WF}_C(S)$ by Proposition 2.4(1). \(2) The proof is similar to that of (1). \[2.7\] The following statements hold. $\mathrm{(1)}$ $V\in \mathcal{WF}_C(S)$ if and only if $V^{++}\in \mathcal{WF}_C(S)$. $\mathrm{(2)}$ $U\in \mathcal{WI}_C(R)$ if and only if $U^{++}\in \mathcal{WI}_C(R)$. The assertions follows immediately from Proposition \[2.6\]. \[2.8\] [The classes $\mathcal{WF}_C(S)$ and $\mathcal{WI}_C(R)$ are closed under direct summands, direct products, direct sums and direct limits.]{} To prove $\mathcal{WF}_C(S)$ is closed under direct summands, we assume that $$0\ra X_1\ra X_2\ra X_3\ra 0$$ is a split exact sequence in $\Mod S$ with $X_2\in \mathcal{WF}_C(S)$. Then $X_2\in \mathcal{B}_C(S)$ and $\Hom_S(C,X_2)$ is weak flat by Proposition \[2.4\](1). It follows from [@HW07 Proposition 4.2] that $X_1,X_3\in \mathcal{B}_C(S)$. One easily checks that the sequence $$0\lra \Hom_S(C,X_1)\lra \Hom_S(C,X_2)\lra \Hom_S(C,X_3)\lra 0$$ in $\Mod R$ is split exact. Then $\Hom_S(C,X_1)$ and $\Hom_S(C,X_3)$ are weak flat since the weak flat modules are closed under direct summands by [@GW15 Proposition 2.3]. It follows that $X_1,X_3\in \mathcal{WF}_C(S)$ by Proposition \[2.4\](1) again. Let $\{F_\lambda\}_{\lambda\in \Lambda}$ be a family of $C$-weak flat modules in $\Mod S$. Then $F_\lambda\in \mathcal{B}_C(S)$ and $\Hom_S(C,F_\lambda)$ is weak flat in $\Mod R$ for any $\lambda\in \Lambda$ by Proposition \[2.4\](1). By [@Ro79 Theorem 2.6], we have the isomorphism $$\prod_{\lambda\in \Lambda}\Hom_R(C,F_\lambda)\cong\Hom_R(C,\prod_{\lambda\in \Lambda}F_\lambda).$$ Note that $\prod_{\lambda\in \Lambda}\Hom_R(C,F_\lambda)$ is weak flat by [@GW15 Theorem 2.13], and $\prod_{\lambda\in \Lambda}F_\lambda\in\mathcal{B}_C(S)$ by [@HW07 Proposition 4.2], it follows that $\prod_{\lambda\in \Lambda}F_\lambda\in \mathcal{WF}_C(S)$ by Proposition \[2.4\](1). Therefore $\mathcal{WF}_C(S)$ is closed under direct products. Since $C$ is finitely presented and the weak flat modules are closed under direct sums, similar to the arguments above, one can deduce that the class $\mathcal{WF}_C(S)$ is closed under direct sums. Let $\{F_i\}_{i\in I}$ be a direct system of $C$-weak flat modules in $\Mod S$. Then $F_i\in \mathcal{B}_C(S)$ for any $i\in I$ and $\{\Hom_S(C,F_i)\}_{i\in I}$ is a direct system of weak flat modules in $\Mod R$ by Proposition 2.4(1). By [@HW07 Proposition 4.2], we get that $\lim F_i\in \mathcal{B}_C(S)$. Note that the weak flat modules are closed under direct limits since $\Tor$ commutes with direct limits, it follows that $\lim \Hom_S(C,F_i)$ is weak flat. By [@GT12 Lemma 2.7], we have the isomorphism $$\lim\Hom_S(C,F_i)\cong\Hom_S(C,\lim F_i).$$ Then $\Hom_S(C,\lim F_i)$ is a weak flat left $R$-module. Therefore, we have $\lim F_i\in \mathcal{WF}_C(S)$ by Proposition \[2.4\](1). Thus the class $\mathcal{WF}_C(S)$ is closed under direct limits. As a similar argument to the above, we can deduce that the class $\mathcal{WI}_C(R)$ is closed under direct summands, direct products, direct sums and direct limits. The following lemma is stated in [@TW10] for a commutative ring, but the proof is valid in the present context. \[2.9\] ([@TW10 Theorem 2.8]) The following statements hold. $\mathrm{(1)}$ If $M\in \Mod S$, then $M\in \mathcal{B}_C(S)$ if and only if $\Hom_S(C,M)\in\mathcal{A}_C(R)$. $\mathrm{(2)}$ If $M\in \Mod R$, then $M\in \mathcal{A}_C(R)$ if and only if $C\otimes_RM\in\mathcal{B}_C(S)$. \[2.10\] The following statements hold. $\mathrm{(1)}$ $\Hom_S(C,I)\in\mathcal{WI}_C(R)$ if and only if $I\in \mathcal{WI}(S)$. $\mathrm{(2)}$ $C\otimes_RF\in \mathcal{WF}_C(S)$ if and only if $F\in \mathcal{WF}(R)$. \(1) “If" part is by definition. “Only if" part. Since $\Hom_S(C,I)\in\mathcal{WI}_C(R)$, then we have $\Hom_S(C,I)\in \mathcal{A}_C(R)$ by Proposition \[2.4\](2), and so $I\in\mathcal{B}_C(S)$ by Lemma \[2.9\](1). On the other hand, there exists a weak injective left $S$-module $I'$ such that $\Hom_S(C,I)=\Hom_S(C,I')$. Also, we have $I'\in\mathcal{B}_C(S)$ by Theorem \[2.2\]. Thus we have the following isomorphism: $$I\cong C\otimes_R\Hom_S(C,I)\cong C\otimes_R\Hom_S(C,I')\cong I'.$$ Hence $I$ is a weak injective left $S$-module, as desired. \(2) Similar to the proof of (1). It was shown in [@BGH14 Propositions 2.7(2) and 2.10(2)] that the classes of weak flat and weak injective modules are both closed under pure submodules and pure quotients. Here we have \[2.11\] The following statements hold. $\mathrm{(1)}$ The class $\mathcal{WF}_C(S)$ is closed under pure submodules and pure quotients. $\mathrm{(2)}$ The class $\mathcal{WI}_C(R)$ is closed under pure submodules and pure quotients. \(1) Let $$\mathbb{Y}=\ \ 0\ra Y_1\ra Y_2\ra Y_3\ra 0$$ be a pure exact sequence in $\Mod S$ with $Y_2\in \mathcal{WF}_C(S)$. We will show that $Y_1,Y_3\in\mathcal{WF}_C(S)$. Since $C$ is a finitely presented left $S$-module, the sequence $$\Hom_S(C,\mathbb{Y})=\ \ 0\lra \Hom_S(C,Y_1)\lra \Hom_S(C,Y_2)\lra \Hom_S(C,Y_3)\lra 0$$ is exact in $\Mod R$. We claim that $\Hom_S(C,\mathbb{Y})$ is pure exact. Let $Q$ be a finitely presented left $R$-module. It is easy to check that $C\otimes_RQ$ is a finitely presented left $S$-module. By the natural isomorphism $$\Hom_R(Q,\Hom_S(C,\mathbb{Y}))\cong\Hom_S(C\otimes_RQ,\mathbb{Y}).$$ Then we have $\Hom_S(C\otimes_RQ,\mathbb{Y})$ is an exact sequence since $\mathbb{Y}$ is pure exact. It follows that $\Hom_R(Q,\Hom_S(C,\mathbb{Y}))$ is exact, and hence $\Hom_S(C,\mathbb{Y})$ is a pure exact sequence. In the pure exact sequence $\Hom_S(C,\mathbb{Y})$, the module $\Hom_S(C,Y_2)$ is a weak flat left $R$-module by Proposition \[2.4\](1) since $Y_2$ is $C$-weak flat. Note that the class of weak flat modules is closed under pure submodules and pure quotients, it follows that $\Hom_S(C,Y_1)$ and $\Hom_S(C,Y_3)$ are weak flat left $R$-modules. Hence $\Hom_S(C,Y_1)$ and $\Hom_S(C,Y_3)$ belong to $\mathcal{A}_C(R)$ by Theorem \[2.2\]. By Lemma \[2.9\](1), we have $Y_1,Y_3$ belong to $\mathcal{B}_C(S)$. Consequently, $Y_1,Y_3\in \mathcal{WF}_C(S)$ by Proposition \[2.4\](1). \(2) By analogy with the proof of (1), one can deduce that the class $\mathcal{WI}_C(R)$ is closed under pure submodules and pure quotients. \[2.12\] The following statements hold. $\mathrm{(1)}$ The class $\mathcal{WF}_C(S)$ is covering and preenveloping. $\mathrm{(2)}$ The class $\mathcal{WI}_C(R)$ is covering and preenveloping. \(1) Since the class $\mathcal{WF}_C(S)$ is closed under pure quotients by Proposition \[2.11\] and is closed under direct sums by Proposition \[2.8\], one gets directly that $\mathcal{WF}_C(S)$ is covering by [@HJ08 Theorem 2.5]. On the other hand, it follows from Proposition \[2.11\] and [@HJ08 Proposition 3.2] that $\mathcal{WF}_C(S)$ is a Kaplansky class. Also, the class $\mathcal{WF}_C(S)$ is closed under direct limits by Proposition \[2.8\]. Therefore $\mathcal{WF}_C(S)$ is preenveloping by [@EL02 Theorem 2.5]. \(2) The proof is similar to that of (1). Foxby equivalence ================= In this section we investigate Foxby equivalence relative to $C$-weak injective and $C$-weak flat modules. Some known results in [@HW07] are generalized. \[3.1\] [There are equivalences of categories $$\xymatrix@C=80pt{\mathcal{WF}(R) \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{WF}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\\ \mathcal{WI}_C(R) \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{WI}(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}. }$$]{} It suffices to prove the first assertion. Dually, we get the second one. We have that the functor $C\otimes_R-$ maps $\mathcal{WF}(R)$ to $\mathcal{WF}_C(S)$ by definition, and the functor $\Hom_S(C,-)$ maps $\mathcal{WF}_C(S)$ to $\mathcal{WF}(R)$ by Proposition \[2.4\](1). On the other hand, if $M\in\mathcal{WF}(R)$ and $N\in\mathcal{WF}_C(S)$, then $M\in \mathcal{A}_C(R)$ by Theorem \[2.2\] and $N\in\mathcal{B}_C(S)$ by Proposition \[2.4\](1). Then there exist natural isomorphisms: $M\cong \Hom_S(C,C\otimes_RM)$ and $N\cong C\otimes_R\Hom_S(C,M)$. Thus the assertion follows. Let $n$ be a non-negative integer. For convenience, we set $$\begin{array}{ll}\vspace{0.1cm} \mathcal{WF}(R)_{\leq n}=\mbox{the class of left $R$-modules of weak flat dimension at most $n$,}\\\vspace{0.1cm} \mathcal{WI}(S)_{\leq n}=\mbox{the class of left $S$-modules of weak injective dimension at most $n$,}\\\vspace{0.1cm} \mathcal{WF}_C(S)_{\leq n}=\mbox{the class of left $S$-modules of $C$-weak flat dimension at most $n$,}\\\vspace{0.1cm} \mathcal{WI}_C(R)_{\leq n}=\mbox{the class of left $R$-modules of $C$-weak injective dimension at most $n$.}\\ \end{array}$$ \[3.2\] [There are equivalences of categories $$\xymatrix@C=80pt{\mathcal{WF}(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{WF}_C(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\\ \mathcal{WI}_C(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{WI}(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}. }$$]{} We only prove the first assertion, and the second one is dual. The case $n=0$ holds by Proposition \[3.1\]. Now suppose that $n\geq 1$ and $M\in\mathcal{WF}(R)_{\leq n}$. Then there exists an exact sequence $$0\lra F_n\overset{f_n}\lra \cdots\overset{f_2}\lra F_1\overset{f_1}\lra F_0\lra M\lra 0\eqno{(3.1)}$$ in $\Mod R$ with each $F_i$ weak flat. By Corollary \[2.3\], we have $\coker(f_i)\in\mathcal{A}_C(R)$ for $1\leq i\leq n$. Applying the functor $C\otimes_R-$ to (3.1), we get an exact sequence $$0\lra C\otimes_RF_n\lra \cdots\lra C\otimes_RF_1\lra C\otimes_RF_0\lra C\otimes_RM\lra 0$$ in $\Mod S$. Note that each $C\otimes_RF_i$ is $C$-weak flat, it follows that $C$-$\wfd_S(C\otimes_RM)\leq n$, and thus $C\otimes_RM\in\mathcal{WF}_C(S)_{\leq n}$. Conversely, assume that $M\in\mathcal{WF}_C(S)_{\leq n}$. Then there is an exact sequence $$0\lra C\otimes_RQ_n\overset{1_C\otimes_Rf_n}\lra \cdots\overset{1_C\otimes_Rf_2}\lra C\otimes_RQ_1\overset{1_C\otimes_Rf_1}\lra C\otimes_RQ_0\lra M\lra 0\eqno{(3.2)}$$ in $\Mod S$ with each $Q_i$ a weak flat left $R$-module. Because $Q_i\in\mathcal{A}_C(R)$ by Theorem \[2.2\], we have each $C\otimes_RQ_i\in \mathcal{B}_C(S)$ by [@HW07 Proposition 4.1]. It follows from [@HW07 Corollary 6.3] that all $\coker(1_C\otimes_Rf_i)$ in (3.2) are in $\mathcal{B}_C(S)$. So we get the following exact sequence $$0\ra \Hom_S(C,C\otimes_RQ_n)\ra \cdots\ra \Hom_S(C,C\otimes_RQ_0)\ra \Hom_S(C,M)\ra 0.$$ Note that $\mu_{Q_i}:Q_i\lra \Hom_S(C,C\otimes_RQ_i)$ is an isomorphism for any $1\leq i\leq n$, which gives rise to the exactness of $$0\lra Q_n\lra \cdots\lra Q_1\lra Q_0\lra \Hom_S(C,M)\lra 0.$$ Thus $\Hom_S(C,M)\in\mathcal{WF}(R)_{\leq n}$, as desired. \[3.3\] For any integer $n\geq 0$, the following statements hold. $\mathrm{(1)}$ If $C$-$\wfd_S(M)\leq n$, then $M\in \mathcal{B}_C(S)$. $\mathrm{(2)}$ If $C$-$\wid_R(M)\leq n$, then $M\in \mathcal{A}_C(R)$. It suffices to prove the first assertion. Dually, we get the second one. If $n=0$, then the assertion follows by Proposition \[2.4\](1). Now suppose $n\geq 1$, then there exists an exact sequence $$0\ra C\otimes_RF_n\ra\cdots\ra C\otimes_RF_1\ra C\otimes_RF_0\ra M\ra 0 \eqno{(3.3)}$$ in $\Mod S$ with each $F_i$ a weak flat left $R$-module. Since $C\otimes_RF_i\in\mathcal{B}_C(S)$ for any $0\leq i\leq n$, we have every cokernel in (3.3) is in $\mathcal{B}_C(S)$ by [@HW07 Theorem 6.2]. Thus $M\in\mathcal{B}_C(S)$. The following theorem is one of main results in this paper. \[3.4\] [(Foxby Equivalence)]{} [There are equivalences of categories $$\xymatrix@R=20pt@C=60pt{ \mathcal{WF}(R) \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[d] & \mathcal{WF}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[d]\\ \mathcal{WF}(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[d] & \mathcal{WF}_C(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[d]\\ \mathcal{A}_C(R) \ar@<+4pt>[r]^{C\otimes_R-} & \mathcal{B}_C(S) \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\\ \mathcal{WI}_C(R)_{\leq n} \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[u] & \mathcal{WI}(S)_{\leq n} \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[u]\\ \mathcal{WI}_C(R) \ar@<+4pt>[r]^{C\otimes_R-}\ar@{^{(}->}[u] & \mathcal{WI}(S). \ar@<+5pt>[l]^{\Hom_S(C,-)}_{\sim}\ar@{^{(}->}[u]\\}$$]{} This follows directly from Propositions \[3.1\], \[3.2\] and \[3.3\]. \[3.5\] The following equalities hold. $\mathrm{(1)}$ $\wfd_R(M)=C$-$\wfd_S(C\otimes_RM)$ for any left $R$-module $M$. $\mathrm{(2)}$ $\wid_S(M)=C$-$\wid_R(\Hom_S(C,M))$ for any left $S$-module $M$. $\mathrm{(3)}$ $C$-$\wid_R(M)=\wid_S(C\otimes_RM)$ for any left $R$-module $M$. $\mathrm{(4)}$ $C$-$\wfd_S(M)=\wfd_R(\Hom_S(C,M))$ for any left $S$-module $M$. We only prove (1), and (2)–(4) can be proved similarly. Let $\wfd_R(M)=t<\infty$. Then $M\in\mathcal{A}_C(R)$ by Corollary \[2.3\], and so $\Tor_i^R(C,M)=0$ for all $i\geq 1$. On the other hand, there is a weak flat resolution of $M$: $$0\ra F_t\ra\cdots\ra F_1\ra F_0\ra M\ra 0,$$ which gives rise to the exactness of $$0\ra C\otimes_RF_t\ra\cdots\ra C\otimes_RF_1\ra C\otimes_RF_0\ra C\otimes_RM\ra 0,$$ where each $C\otimes_RF_i$ is $C$-weak flat. This implies that $C$-$\wfd_S(C\otimes_RM)\leq t=\wfd_R(M)$. Conversely, assume that $C$-$\wfd_S(C\otimes_RM)=s<\infty$. Then, by Proposition \[3.3\](1), we have $C\otimes_RM\in \mathcal{B}_C(S)$. It follows that $M\in\mathcal{A}_C(R)$ by Proposition \[2.4\](2), which implies that $\mu_{_{M}}:M\ra \Hom_S(C,C\otimes_RM)$ is an isomorphism. On the other hand, there exists an exact sequence $$\mathbb{X}=\ \ 0\ra C\otimes_RF_s\ra C\otimes_RF_{s-1}\ra\cdots\ra C\otimes_RF_0\ra C\otimes_RM\ra 0$$ in $\Mod S$ with each $F_i$ a weak flat left $R$-module. Since $C\otimes_RF_i\in\mathcal{B}_C(S)$ for any $0\leq i\leq s$, we have that $\Hom_S(C,\mathbb{X})$ is exact by [@HW07 Corollary 6.3]. Consider the following commutative diagram with the lower row exact: $${\small\xymatrix@R=20pt@C=10pt{0 \ar[r] & F_s \ar[d]^{\cong} \ar[r] & \cdots\ar[r] & F_0 \ar[d]^{\cong} \ar[r] & M\ar[d]^{\cong} \ar[r] & 0\\ 0 \ar[r] & \Hom_S(C,C\otimes_RF_s) \ar[r] & \cdots\ar[r] &\Hom_S(C,C\otimes_RF_0) \ar[r] &\Hom_S(C,C\otimes_RM)\ar[r] & 0.}}$$ Then the upper row is exact in $\Mod R$ with each $F_i$ weak flat, and hence $\wfd_R(M)\leq s$. We finish this section with some applications of Theorem \[3.4\], which is of independent interest. For convenience, we assume that $R$ is a commutative ring. In view of Theorem \[3.4\], it is natural to ask whether there is a relationship between the subcategory $\mathcal{WI}_C(R)$ and the subcategory of $\mathcal{WF}_C(S)$. The answer is positive, that is, a suitable functor $\Hom_R(-,E)$ is discovered for any injective $R$-module $E$. Here we have \[3.6\] Let $R$ be a commutative ring and $E$ an injective $R$-module. Then $\mathrm{(1)}$ $W\in \mathcal{WI}_C(R)$ implies $\Hom_R(W,E)\in \mathcal{WF}_C(R)$. $\mathrm{(2)}$ $W\in \mathcal{WF}_C(R)$ implies $\Hom_R(W,E)\in \mathcal{WI}_C(R)$. \(1) Let $W\in \mathcal{WI}_C(R)$. Then there exists a weak injective $R$-module $I$ such that $W=\operatorname{Hom}_R(C,I)$. Let $E$ be an injective $R$-module. One can easily check that $\operatorname{Hom}_R(I,E)$ is weak flat by the isomorphism ([@HW07 Lemma 1.2]): $$\mbox{Tor}^R_1(N,\mbox{Hom}_R(I,E))\cong\mbox{Hom}_R(\mbox{Ext}^1_R(N,I),E)$$ for any super finitely presented $R$-module $N$. Thus we have $$\operatorname{Hom}_R(W,E)=\operatorname{Hom}_R(\operatorname{Hom}_R(C,I),E)\cong C\otimes_R\operatorname{Hom}_R(I,E).$$ It follows that $\operatorname{Hom}_R(W,E)$ is $C$-weak flat by definition. \(2) Let $W\in \mathcal{WF}_C(R)$. Then $W=C\otimes_RF$ for some weak flat $R$-module $F$. For any injective $R$-module $E$, one easily gets that $\operatorname{Hom}_R(F,E)$ is weak injective by the isomorphism ([@GT12 Lemma 2.16]): $$\mbox{Ext}^1_R(N,\operatorname{Hom}_R(F,E))\cong\mbox{Hom}_R(\mbox{Tor}^R_1(N,F),E).$$ Thus we have $$\operatorname{Hom}_R(W,E)=\operatorname{Hom}_R(C\otimes_RF,E)\cong \operatorname{Hom}_R(C,\operatorname{Hom}_R(F,E)).$$ Therefore $\operatorname{Hom}_R(W,E)$ is $C$-weak injective, as desired. For convenience, we will denote by ${_R(-,-)}:=\Hom_R(-,-)$ in the following commutative diagrams. By Theorem \[3.4\] and Proposition \[3.6\], we have the following two commutative diagrams for any injective $R$-module $E$: $${\small \xymatrix@C=1.5cm{ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}_C(R)\ar[d]^{{_R(C,-)}}\\ \mathcal{WI}(R) \ar[r]^{{_R(-,E)}} & \mathcal{WF}(R), } \xymatrix@C=1.5cm{ \mathcal{WF}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WI}_C(R)\ar[d]^{C\otimes_R-}\\ \mathcal{WF}(R) \ar[r]^{{_R(-,E)}} & \mathcal{WI}(R) }}$$ which give rise to the commutative diagram for all injective $R$-modules $E$ and $E'$: $${\small \xymatrix@C=1.5cm{ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}_C(R)\ar[r]^{{_R(-,E')}}&\mathcal{WI}_C(R)\\ \mathcal{WI}(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}(R) \ar[r]^{{_R(-,E')}}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}& \mathcal{WF}_C(R)\ar[r]^{{_R(-,E')}}&\mathcal{WI}_C(R).\ar[u]_{C\otimes_R-} }}$$ \[3.7\] Let $R$ be a commutative ring and $F$ a flat $R$-module. Then $\mathrm{(1)}$ $W\in \mathcal{WI}_C(R)$ implies $W\otimes_RF\in \mathcal{WI}_C(R)$. $\mathrm{(2)}$ $W\in \mathcal{WF}_C(R)$ implies $W\otimes_RF\in \mathcal{WF}_C(R)$. \(1) Let $W\in \mathcal{WI}_C(R)$. Then there exists a weak injective $R$-module $I$ such that $W=\operatorname{Hom}_R(C,I)$. For any super finitely presented $R$-module $N$, we have that $I\otimes_RF$ is weak injective by the isomorphism ([@HW07 Lemma 1.1]): $$\mbox{Ext}_R^1(N,I\otimes_RF)\cong\mbox{Ext}^1_R(N,I)\otimes_RF.$$ On the other hand, by the tensor evaluation morphism ([@HW07 1.10]), one easily gets that $$W\otimes_RF=\operatorname{Hom}_R(C,I)\otimes_RF\cong \operatorname{Hom}_R(C,I\otimes_RF)$$since $C$ is a semidualizing module. Hence $W\otimes_RF$ is $C$-weak injective by definition. \(2) Let $W\in \mathcal{WF}_C(R)$. Then we have $W=C\otimes_RQ$ for some weak flat $R$-module $Q$. For any super finitely presented $R$-module $N$, one gets that $Q\otimes_RF$ is weak flat by the isomorphism ([@Ro79 Theorem 9.48]): $$\Tor^R_1(N,Q\otimes_RF)\cong\Tor_1^R(N,Q)\otimes_RF.$$ On the other hand, we have that $$W\otimes_RF=(C\otimes_RQ)\otimes_RF\cong C\otimes_R(Q\otimes_RF).$$ Therefore $W\otimes_RF$ is $C$-weak flat, as desired. By Theorem \[3.4\] and Proposition \[3.7\], we have the following commutative diagrams for any flat $R$-module $F$: $${\small \xymatrix@C=1.5cm{ \mathcal{WI}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[d]^{C\otimes_R-}\\ \mathcal{WI}(R) \ar[r]^{-\otimes_RF} & \mathcal{WI}(R), } \xymatrix@C=1.5cm{ \mathcal{WF}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[d]^{{_R(C,-)}}\\ \mathcal{WF}(R) \ar[r]^{-\otimes_RF} & \mathcal{WF}(R). }}$$ Moreover, for any flat $R$-module $F$ and any injective $R$-module $E$, we obtain the following commutative diagrams: $${\scriptsize \xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WF}_C(R)\\ \mathcal{WI}(R) \ar[r]^{-\otimes_RF}\ar[d]_{{_R(C,-)}} & \mathcal{WI}(R) \ar[r]^{{_R(-,E)}}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WI}_C(R) \ar[r]^{-\otimes_RF}& \mathcal{WI}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WF}_C(R),\ar[u]_{{_R(C,-)}} } \!\!\!\xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WI}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WI}_C(R)\\ \mathcal{WF}(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WI}(R) \ar[r]^{-\otimes_RF}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WF}_C(R) \ar[r]^{{_R(-,E)}}& \mathcal{WI}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WI}_C(R),\ar[u]_{C\otimes_R-} }}$$ $${\scriptsize \xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WI}_C(R)\\ \mathcal{WF}(R) \ar[r]^{-\otimes_RF}\ar[d]_{C\otimes_R-} & \mathcal{WF}(R) \ar[r]^{{_R(-,E)}}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WF}_C(R) \ar[r]^{-\otimes_RF}& \mathcal{WF}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WI}_C(R),\ar[u]_{C\otimes_R-} } \!\!\!\xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WF}_C(R)\\ \mathcal{WI}(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}(R) \ar[r]^{-\otimes_RF}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}& \mathcal{WF}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WF}_C(R).\ar[u]_{{_R(C,-)}} }}$$ \[3.8\] Let $R$ be a commutative ring and $P$ a projective $R$-module. Then $\mathrm{(1)}$ $W\in \mathcal{WF}_C(R)$ implies $\Hom_R(P,W)\in \mathcal{WF}_C(R)$. $\mathrm{(2)}$ $W\in \mathcal{WI}_C(R)$ implies $\Hom_R(P,W)\in \mathcal{WI}_C(R)$. \(1) Let $W\in \mathcal{WF}_C(R)$. Then there exists a weak flat $R$-module $F$ such that $W=C\otimes_RF$. For any super finitely presented $R$-module $N$, we have an exact sequence $$0\rightarrow N'\rightarrow P_0\rightarrow N\rightarrow 0$$ with $P_0$ finitely generated projective and $N'$ super finitely presented. Now consider the following commutative diagram $$\xymatrix@C=15pt{0\ar[r] & \operatorname{Hom}_R(P,\operatorname{Tor}_1^R(N,F))\ar[r]\ar[d] & \operatorname{Hom}_R(P,N'\otimes_RF) \ar[r]\ar[d]^\cong & \operatorname{Hom}_R(P,P_0\otimes_RF)\ar[d]^\cong\\ 0\ar[r] & \operatorname{Tor}^R_1(\operatorname{Hom}_R(P,F),N)\ar[r] & \operatorname{Hom}_R(P,F)\otimes_RN' \ar[r] & \operatorname{Hom}_R(P,F)\otimes_RP_0. }$$ Since $N'$ and $P_0$ are finitely generated and $P$ is projective, one gets that the right two morphisms are isomorphic by [@SSW Appendix A, Lemma 1.4]. Hence we have $$\mbox{Tor}^R_1(\mbox{Hom}_R(P,F),N)\cong \mbox{Hom}_R(P,\operatorname{Tor}^R_1(N,F))=0$$ since $F$ is weak flat. It follows that $\operatorname{Hom}_R(P,F)$ is weak flat. On the other hand, we have $$\operatorname{Hom}_R(P,W)=\operatorname{Hom}_R(P,C\otimes_RF)\cong \operatorname{Hom}_R(P,F)\otimes_RC\cong C\otimes_R\operatorname{Hom}_R(P,F).$$ Thus $\operatorname{Hom}_R(P,W)$ is $C$-weak flat. \(2) Let $W\in \mathcal{WI}_C(R)$. Then $W=\operatorname{Hom}_R(C,I)$ for some weak injective $R$-module $I$. For any super finitely presented $R$-module $N$, we have an exact sequence $$0\rightarrow N'\rightarrow P_0\rightarrow N\rightarrow 0$$ with $P_0$ finitely generated projective. Consider the following commutative diagram $${\small\xymatrix@C=10pt{ \operatorname{Hom}_R(P,\operatorname{Hom}_R(P_0,I))\ar[r]\ar[d]^\cong & \operatorname{Hom}_R(P,\operatorname{Hom}_R(N',I)) \ar[r]\ar[d]^\cong & \operatorname{Hom}_R(P,\operatorname{Ext}^1_R(N,I))\ar[d]\ar[r] &0\\ \operatorname{Hom}_R(P_0,\operatorname{Hom}_R(P,I))\ar[r] & \operatorname{Hom}_R(N',\operatorname{Hom}_R(P,I)) \ar[r] & \operatorname{Ext}^1_R(N,\operatorname{Hom}_R(P,I))\ar[r]&0. }}$$ Then the left two morphisms in the above diagram are isomorphic viewed as the swap maps. It follows that $\operatorname{Ext}^1_R(N,\operatorname{Hom}_R(P,I))\cong\operatorname{Hom}_R(P,\operatorname{Ext}^1_R(N,I))$. Note that $I$ is weak injective, we get that $\operatorname{Hom}_R(P,I)$ is weak injective. By Theorem \[3.4\] and Proposition \[3.8\], we have the following commutative diagrams for any projective $R$-module $P$: $${\small \xymatrix@C=1.5cm{ \mathcal{WF}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[d]^{{_R(C,-)}}\\ \mathcal{WF}(R) \ar[r]^{{_R(P,-)}} & \mathcal{WF}(R), } \xymatrix@C=1.5cm{ \mathcal{WI}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[d]^{C\otimes_R-}\\ \mathcal{WI}(R) \ar[r]^{{_R(P,-)}} & \mathcal{WI}(R). }}$$ Consequently, we obtain the following commutative diagrams for any injective $R$-module $E$, any flat $R$-module $F$ and any projective $R$-module $P$: $${\scriptsize \xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WI}_C(R)\\ \mathcal{WF}(R) \ar[r]^{{_R(P,-)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}(R) \ar[r]^{{_R(-,E)}}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WF}_C(R) \ar[r]^{{_R(P,-)}}& \mathcal{WF}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WI}_C(R),\ar[u]_{C\otimes_R-} } \!\!\! \xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WF}_C(R)\\ \mathcal{WI}(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}(R) \ar[r]^{{_R(P,-)}}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WI}_C(R) \ar[r]^{{_R(-,E)}}& \mathcal{WF}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WF}_C(R),\ar[u]_{{_R(C,-)}} }}$$ $${\scriptsize \xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WF}_C(R)\\ \mathcal{WF}(R) \ar[r]^{{_R(P,-)}}\ar[d]_{C\otimes_R-} & \mathcal{WF}(R) \ar[r]^{-\otimes_RF}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WF}_C(R) \ar[r]^{{_R(P,-)}}& \mathcal{WF}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WF}_C(R),\ar[u]_{{_R(C,-)}} } \!\!\! \xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{{_R(C,-)}} & \mathcal{WF}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WF}_C(R)\\ \mathcal{WF}(R) \ar[r]^{-\otimes_RF}\ar[d]_{C\otimes_R-} & \mathcal{WF}(R) \ar[r]^{{_R(P,-)}}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WF}_C(R) \ar[r]^{-\otimes_RF}& \mathcal{WF}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WF}_C(R),\ar[u]_{{_R(C,-)}} }}$$ $${\scriptsize \xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WF}_C(R)\\ \mathcal{WI}(R) \ar[r]^{{_R(P,-)}}\ar[d]_{{_R(C,-)}} & \mathcal{WI}(R) \ar[r]^{{_R(-,E)}}&\mathcal{WF}(R)\ar[u]_{C\otimes_R-}\\ \mathcal{WI}_C(R) \ar[r]^{{_R(P,-)}}& \mathcal{WI}_C(R)\ar[r]^{{_R(-,E)}}&\mathcal{WF}_C(R),\ar[u]_{{_R(C,-)}} } \!\!\! \xymatrix@C=15pt{ \mathcal{WF}_C(R) \ar[r]^{{_R(-,E)}}\ar[d]_{{_R(C,-)}} & \mathcal{WI}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WI}_C(R)\\ \mathcal{WF}(R) \ar[r]^{{_R(-,E)}}\ar[d]_{C\otimes_R-} & \mathcal{WI}(R) \ar[r]^{{_R(P,-)}}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WF}_C(R) \ar[r]^{{_R(-,E)}}& \mathcal{WI}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WI}_C(R),\ar[u]_{C\otimes_R-} }}$$ $${\scriptsize \xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{{_R(P,-)}}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WI}_C(R)\\ \mathcal{WI}(R) \ar[r]^{{_R(P,-)}}\ar[d]_{{_R(C,-)}} & \mathcal{WI}(R) \ar[r]^{-\otimes_RF}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WI}_C(R) \ar[r]^{{_R(P,-)}}& \mathcal{WI}_C(R)\ar[r]^{-\otimes_RF}&\mathcal{WI}_C(R),\ar[u]_{C\otimes_R-} } \!\!\! \xymatrix@C=15pt{ \mathcal{WI}_C(R) \ar[r]^{-\otimes_RF}\ar[d]_{C\otimes_R-} & \mathcal{WI}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WI}_C(R)\\ \mathcal{WI}(R) \ar[r]^{-\otimes_RF}\ar[d]_{{_R(C,-)}} & \mathcal{WI}(R) \ar[r]^{{_R(P,-)}}&\mathcal{WI}(R)\ar[u]_{{_R(C,-)}}\\ \mathcal{WI}_C(R) \ar[r]^{-\otimes_RF}& \mathcal{WI}_C(R)\ar[r]^{{_R(P,-)}}&\mathcal{WI}_C(R).\ar[u]_{C\otimes_R-} }}$$ Stability of the Auslander and Bass classes ============================================ In this section, we show that an iteration of the procedure used to describe the Auslander class (resp. Bass class) yields exactly the Auslander class (resp. Bass class), which generalize [@HW07 Theorems 2 and 6.1] and [@EH09 Propositions 3.6 and 3.7]. This enables us to provide more beautiful characterizations of the modules in the Auslander and Bass classes. Finally, special attention is paid to giving a partial answer to Question 2. \[4.1\] ([@HW07 Theorems 2 and 6.1]) A left $R$-module $M\in \mathcal{A}_C(R)$ if and only if there exists an exact sequence $$\mathbb{X}=\ \ \cdots\lra P_1\lra P_0\lra U^0\lra U^1\lra\cdots$$ in $\Mod R$ with each $P_i$ projective (or flat) and $U^i$ $C$-injective such that $M\cong\coker(P_1\ra P_0)$ and the complex $C\otimes_R\mathbb{X}$ is exact. A left $S$-module $N\in \mathcal{B}_C(S)$ if and only if there exists an exact sequence $$\mathbb{Y}=\ \ \cdots\lra W_1\lra W_0\lra I^0\lra I^1\lra\cdots$$ in $\Mod S$ with each $I^i$ injective and $W_i$ $C$-projective such that $N\cong\ker(I^0\ra I^1)$ and the complex $\Hom_S(C,\mathbb{Y})$ is exact. \[4.2\] Since ${}_SC_R$ is a faithfully semidualizing bimodule, it is straightforward to get that all kernels and cokernels of $\mathbb{X}$ in Lemma \[4.1\] are in $\mathcal{A}_C(R)$ by [@HW07 Lemmas 4.1, 5.1 and Corollary 6.3]. Similarly, one can easily obtain that all kernels and cokernels of $\mathbb{Y}$ in Lemma \[4.1\] are in $\mathcal{B}_C(S)$. In the following, $\mathscr{A}$ denotes an abelian category, all subcategories are full subcategories of $\mathscr{A}$ closed under isomorphisms. We fix subcategories $\mathscr{X}$ and $\mathscr{Y}$ of $\mathscr{A}$. Recall from [@SSW08] that a subcategory $\mathscr{X}$ of $\mathscr{Y}$ is called a *generator* (resp. *cogenerator*) for $\mathscr{Y}$ if for any object $Y$ in $\mathscr{Y}$, there exists an exact sequence $0\ra Y'\ra X\ra Y\ra 0$ (resp. $0\ra Y\ra X\ra Y'\ra 0$) in $\mathscr{Y}$ with $X$ an object in $\mathscr{X}$. We use gen $\mathscr{Y}$ (resp. cogen $\mathscr{Y}$) to denote a generator (resp. cogenerator) for a subcategory $\mathscr{Y}$. The following two results play a crucial role in the section, which give a generalization of [@Hu13 Theorem 5.3]. \[4.3\] Let $\mathscr{X}$ be closed under extensions and $$\cdots\ra G_n\ra\cdots \ra G_1\ra G_0\ra M\ra 0\eqno{(4.1)}$$ be an exact sequence in $\mathscr{A}$ with all $G_i$ objects in $\mathscr{X}$. Then we have the following $\mathrm{(1)}$ There exists an exact sequence $$\cdots\ra P_n\ra \cdots\ra P_1\ra P_0\ra M\ra 0 \eqno{(4.2)}$$ in $\mathscr{A}$ with all $P_i$ objects in gen $\mathscr{X}$. $\mathrm{(2)}$ Let $\mathscr{A}=\Mod R$, and $D$ be an object in $\mathscr{A}$ such that any short exact sequence in $\mathscr{X}$ is $D\otimes_R-$ exact. If [(4.1)]{} is $D\otimes_R-$ exact, then so is [(4.2)]{}. $\mathrm{(3)}$ Let $D$ be an object in $\mathscr{A}$ such that any short exact sequence in $\mathscr{X}$ is $\Hom_\mathscr{A}(D,-)$ exact [(]{}resp. $\Hom_\mathscr{A}(-,D)$ exact[)]{}. If [(4.1)]{} is $\Hom_\mathscr{A}(D,-)$ exact [(]{}resp. $\Hom_\mathscr{A}(-,D)$ exact[)]{}, then so is [(4.2)]{}. (1) Let $$\cdots\ra G_n\ra\cdots\ra G_1\ra G_0\ra M\ra 0$$ be an exact sequence in $\mathscr{A}$ with all $G_i\in\mathscr{X}$. Put $K_1=\im(G_1\ra G_0)$. Then we get the following two exact sequences $$\cdots\ra G_n\ra\cdots\ra G_1\ra K_1\ra 0\ \ \mbox{ and }\ \ 0\ra K_1\ra G_0\ra M\ra 0.$$ On the other hand, there exists a short exact sequence $$0\ra G \ra P_0\ra G_0\ra 0$$ in $\mathscr{A}$ with $P_0$ an object in gen $\mathscr{X}$ and $G\in \mathscr{X}$. Consider the following pull-back diagram: $$\xymatrix@R=16pt@C=16pt{& 0 \ar[d] & 0 \ar[d] & &\\ & G \ar@{=}[r] \ar[d]& G \ar[d] & & \\ 0 \ar[r] & N \ar[r] \ar[d] & P_0 \ar[r] \ar[d] & M \ar[r] \ar@{=}[d] & 0\\ 0 \ar[r] & K_1 \ar[r] \ar[d] & G_0 \ar[r] \ar[d] & M \ar[r] & 0\\ & 0 & ~0. & &}$$ Let $K_2=\im(G_2\ra G_1)$. Then we have the following exact sequences $$\cdots\ra G_n\ra\cdots\ra G_2\ra K_2\ra 0\ \ \mbox{ and }\ \ 0\ra K_2\ra G_1\ra K_1\ra 0.$$ Now consider the following pull-back diagram: $$\xymatrix@R=16pt@C=16pt{& & 0 \ar[d] & 0 \ar[d]& &\\ & & G \ar@{=}[r] \ar[d] & G \ar[d]& &\\ 0 \ar[r] & K_2 \ar@{=}[d] \ar[r] & G'_1 \ar[d] \ar[r] &N \ar[d] \ar[r] & 0\\ 0 \ar[r] & K_2 \ar[r] & G_1 \ar[r] \ar[d] & K_1 \ar[d] \ar[r] & 0 &\\ & & 0 & ~0. & & }$$ In the sequence $0\ra G\ra G'_1\ra G_1\ra 0$, both $G$ and $G_1$ belong to $\mathscr{X}$, then so is $G'_1$. Thus we obtain the following exact sequences $$0\ra N\ra P_0\ra M\ra 0$$ and $$\cdots\ra G_n\ra\cdots\ra G_2\ra G'_1\ra N\ra 0,\eqno{(4.3)}$$ where $P_0$ is an object in gen $\mathscr{X}$, and $G'_1$ and $G_i\ (i=2,3,\cdots)$ belong to $\mathscr{X}$. By repeating the above step to (4.3) and so on, one gets the desired exact sequence (4.2). \(2) Let $D$ be an object in $\mathscr{A}$ such that any short exact sequence in $\mathscr{X}$ is $D\otimes_R-$ exact. Then the middle columns in the above two diagrams are $D\otimes_R-$ exact. If $(4.1)$ is $D\otimes_R-$ exact, then both third rows in the above two diagrams are $D\otimes_R-$ exact. Thus both the middle rows in these two diagrams are also $D\otimes_R-$ exact. Hence the sequence (4.3) is $D\otimes_R-$ exact. Continuing this process, one can easily deduce that (4.2) is $D\otimes_R-$ exact. \(3) The proof is similar to that of (2). Dually, we have the following \[4.4\] Let $\mathscr{X}$ be closed under extensions and let $n\geq 1$ and $$0\ra M\ra G^0\ra G^1\ra \cdots \ra G^n\ra\cdots\eqno{(4.4)}$$ be an exact sequence in $\mathscr{A}$ with all $G^i$ objects in $\mathscr{X}$. Then we have the following $\mathrm{(1)}$ There exists an exact sequence $$0\ra M\ra I^0\ra I^1\ra \cdots \ra I^n\ra\cdots\eqno{(4.5)}$$ in $\mathscr{A}$ with all $I^i$ objects in cogen $\mathscr{X}$. $\mathrm{(2)}$ Let $\mathscr{A}=\Mod R$, and $D$ be an object in $\mathscr{A}$ such that any short exact sequence in $\mathscr{X}$ is $D\otimes_R-$ exact. If [(4.4)]{} is $D\otimes_R-$ exact, then so is [(4.5)]{}. $\mathrm{(3)}$ Let $D$ be an object in $\mathscr{A}$ such that any short exact sequence in $\mathscr{X}$ is $\Hom_\mathscr{A}(D,-)$ exact [(]{}resp. $\Hom_\mathscr{A}(-,D)$ exact[)]{}. If [(4.4)]{} is $\Hom_\mathscr{A}(D,-)$ exact [(]{}resp. $\Hom_\mathscr{A}(-,D)$ exact[)]{}, then so is [(4.5)]{}. By Propositions \[4.3\] and \[4.4\], we get immediately the following two results. \[4.5\] $\mathrm{(1)}$ A left $R$-module $M$ has a projective resolution which is $C\otimes_R-$ exact if and only if there exists a $C\otimes_R-$ exact exact sequence $\cdots\ra G_1\ra G_0\ra M\ra 0$ with each $G_i\in\mathcal{A}_C(R)$. $\mathrm{(2)}$ A left $S$-module $N$ has a $C$-projective resolution which is $\Hom_R(C,-)$ exact if and only if there exists a $\Hom_R(C,-)$ exact exact sequence $\cdots\ra V_1\ra V_0\ra N\ra 0$ with each $V_i\in\mathcal{B}_C(S)$. \[4.6\] $\mathrm{(1)}$ A left $R$-module $M$ has a $C$-injective coresolution which is $C\otimes_R-$ exact if and only if there exists a $C\otimes_R-$ exact exact sequence $0\ra M\ra G^0\ra G^1\ra \cdots$ with each $G^i\in\mathcal{A}_C(R)$. $\mathrm{(2)}$ A left $S$-module $N$ has an injective coresolution which is $\Hom_R(C,-)$ exact if and only if there exists a $\Hom_R(C,-)$ exact exact sequence $0\ra N\ra V^0\ra V^1\ra \cdots$ with each $V^i\in\mathcal{B}_C(S)$. Set $[\mathcal{A}_C(R)]^{1}=\mathcal{A}_C(R)$, and inductively set $[\mathcal{A}_C(R)]^{n+1}=\{M\in \Mod R\mid$ there exists a $C\otimes_R-$ exact exact sequence $\cdots\ra G_1\ra G_0\ra G^0\ra G^1\ra\cdots$ in $\Mod R$ with all $G_i$ and $G^i$ in $[\mathcal{A}_C(R)]^n$ such that $M\cong\coker(G_1\ra G_0)\}$ for any $n\geq 1$. \[4.7\] [$[\mathcal{A}_C(R)]^n=\mathcal{A}_C(R)$ for any $n\geq 1$. ]{} It is easy to see that $\mathcal{A}_C(R)\subseteq[\mathcal{A}_C(R)]^2\subseteq[\mathcal{A}_C(R)]^3\subseteq\cdots$ is an ascending chain of subcategories of $\Mod R$. Let $A\in[\mathcal{A}_C(R)]^2$. Then there exists a $C\otimes_R-$ exact exact sequence $$\cdots\ra G_1\ra G_0\ra G^0\ra G^1\ra\cdots$$ in $\Mod R$ with all $G_i$ and $G^i$ in $\mathcal{A}_C(R)$ such that $A\cong\coker(G_1\ra G_0)$. It follows from Corollary \[4.5\] that $A$ has a projective resolution which is $C\otimes_R-$ exact. On the other hand, $A$ has a $C$-injective coresolution which is $C\otimes_R-$ exact by Corollary \[4.6\]. Hence $A\in\mathcal{A}_C(R)$ by Lemma \[4.1\], and so $[\mathcal{A}_C(R)]^2\subseteq\mathcal{A}_C(R)$. Thus we have that $[\mathcal{A}_C(R)]^2=\mathcal{A}_C(R)$. By using induction on $n$ we can easily get the assertion. Set $[\mathcal{B}_C(S)]^{1}=\mathcal{B}_C(S)$, and inductively set $[\mathcal{B}_C(S)]^{n+1}=\{M\in \Mod S\mid$ there exists a $\Hom_S(C,-)$ exact exact sequence $\cdots\ra V_1\ra V_0\ra V^0\ra V^1\ra\cdots$ in $\Mod S$ with all $V_i$ and $V^i$ in $[\mathcal{B}_C(S)]^n$ such that $M\cong\ker(V^0\ra V^1)\}$ for any $n\geq 1$. Dual to Theorem \[4.7\], one easily gets the following result. \[4.8\] [$[\mathcal{B}_C(S)]^n=\mathcal{B}_C(S)$ for any $n\geq 1$. ]{} As applications of Theorems \[4.7\] and \[4.8\], we can immediately obtain the following two results, which give some equivalent characterizations of the modules in the Auslander and Bass classes in terms of weak flat, weak injective, $C$-weak injective and $C$-weak flat modules. For a non-negative integer $n$, we use $\mathcal{P}(R)_{\leq n}$ (resp., $\mathcal{F}(R)_{\leq n}$, $\mathcal{I}(S)_{\leq n}$ and $\mathcal{FI}(S)_{\leq n}$) to denote the class of left $R$- (or left $S$-) modules of projective (resp., flat, injective and FP-injective) dimension at most $n$. \[4.9\] The following statements are equivalent for a module $M\in\Mod R$. $\mathrm{(1)}$ $M\in \mathcal{A}_C(R)$. $\mathrm{(2)}$ There exists a $C\otimes_R-$ exact exact sequence $$\cdots\ra G_1\ra G_0\ra G_{-1}\ra \cdots$$ in $\Mod R$ with all $G_i\in \mathcal{P}(R)_{\leq n}$ or $\mathcal{I}_C(R)_{\leq n}$ such that $M\cong\coker(G_1\ra G_0)$. $\mathrm{(3)}$ There exists a $C\otimes_R-$ exact exact sequence $$\cdots\ra G_1\ra G_0\ra G_{-1}\ra\cdots$$ in $\Mod R$ with all $G_i\in\mathcal{F}(R)_{\leq n}$ or $\mathcal{I}_C(R)_{\leq n}$ such that $M\cong\coker(G_1\ra G_0)$. $\mathrm{(4)}$ There exists a $C\otimes_R-$ exact exact sequence $$\cdots\ra G_1\ra G_0\ra G_{-1}\ra\cdots$$ in $\Mod R$ with all $G_i\in\mathcal{F}(R)_{\leq n}$ or $\mathcal{FI}_C(R)_{\leq n}$ such that $M\cong\coker(G_1\ra G_0)$. $\mathrm{(5)}$ There exists a $C\otimes_R-$ exact exact sequence $$\cdots\ra G_1\ra G_0\ra G_{-1}\ra\cdots$$ in $\Mod R$ with all $G_i\in\mathcal{WF}(R)_{\leq n}$ or $\mathcal{WI}_C(R)_{\leq n}$ such that $M\cong\coker(G_1\ra G_0)$. \[4.10\] The following statements are equivalent for a module $N\in\Mod S$. $\mathrm{(1)}$ $N\in \mathcal{B}_C(S)$. $\mathrm{(2)}$ There exists a $\Hom_S(C,-)$ exact exact sequence $$\cdots\ra V_1\ra V_0\ra V_{-1}\ra \cdots$$ in $\Mod S$ with all $V_i\in \mathcal{I}(S)_{\leq n}$ or $\mathcal{P}_C(S)_{\leq n}$ such that $N\cong\coker(V_1\ra V_0)$. $\mathrm{(3)}$ There exists a $\Hom_S(C,-)$ exact exact sequence $$\cdots\ra V_1\ra V_0\ra V_{-1}\ra\cdots$$ in $\Mod R$ with all $V_i\in\mathcal{I}(S)_{\leq n}$ or $\mathcal{F}_C(S)_{\leq n}$ such that $N\cong\coker(V_1\ra V_0)$. $\mathrm{(4)}$ There exists a $\Hom_S(C,-)$ exact exact sequence $$\cdots\ra V_1\ra V_0\ra V_{-1}\ra\cdots$$ in $\Mod R$ with all $V_i\in\mathcal{FI}(S)_{\leq n}$ or $\mathcal{F}_C(S)_{\leq n}$ such that $N\cong\coker(V_1\ra V_0)$. $\mathrm{(5)}$ There exists a $\Hom_S(C,-)$ exact exact sequence $$\cdots\ra V_1\ra V_0\ra V_{-1}\ra\cdots$$ in $\Mod S$ with all $V_i\in\mathcal{WI}(S)_{\leq n}$ or $\mathcal{WF}_C(S)_{\leq n}$ such that $N\cong\coker(V_1\ra V_0)$. We round off the paper by giving a partial answer to Question 2, which may be viewed as an illustration of the usefulness of the stability of the Auslander and Bass classes. Before that, recall from [@GW12] that a left $R$-module $M$ is called *Gorenstein FP-injective* if there exists an exact sequence of FP-injective left $R$-modules $\cdots\ra E_1\ra E_0\ra E^0\ra E^1\ra \cdots$ with $M=\ker(E^0\ra E^1)$ such that $\Hom_R(P,-)$ leaves this sequence exact whenever $P$ is a finitely presented left $R$-module of finite projective dimension. \[4.11\] [If ${}_SC_R$ is a faithfully semidualizing bimodule with finite $S$-projective dimension, then every Gorenstein injective modules in $\Mod S$ is in $\mathcal{B}_C(S)$. ]{} Let ${}_SC_R$ be faithfully semidualizing with finite $S$-projective dimension. Then it is (super) finitely presented, and so every Gorenstein FP-injective module in the sense of [@GW12] belongs to $\mathcal{B}_C(S)$ by Corollary \[4.10\]. It follows that all Gorenstein injective modules in $\Mod S$ are in $\mathcal{B}_C(S)$ by [@GW12 Proposition 2.5]. [99]{} T. Araya, R. Takahashi, and Y. Yoshino, Homological invariants associated to semi-dualizing bimodules, J. Math. Kyoto Univ. [**45**]{}(2005), 287–306. D. Bennis, J. R. García Rozas, and L. Oyonarte, Relative Gorenstein dimensions, Mediterr. J. Math. [**13**]{}(2016), 65–91. D. Bennis, J. R. García Rozas, and L. Oyonarte, Relative projective and injective dimensions, Comm. Algebra [**44**]{}(2016), 3383–3396. D. Bennis, J. R. García Rozas, and L. Oyonarte, When do Foxby classes coincide with modules of finite Gorenstein dimension?, Kyoto J. Math. [**56**]{}(2016), 785–802. D. Bennis and N. Mahdou, Strongly Gorenstein projective, injective, and flat modules, J. Pure Appl. Algebra, [**210**]{}(2007), 437–445. D. Bravo, J. Gillespie, and M. Hovey, The stable modules category of a general ring, arXiv:1405.5768v1. K. S. Brown, Cohomology of Groups, Springer-Verlag, New York, 1982. L. W. Christensen, Semi-dualizing complexes and their Auslander categories, Trans. Amer. Math. Soc. [**353**]{}(2001), 1839–1883. E. E. Enochs and H. Holm, Cotorsion pairs asociated with Auslander categories, Israel J. Math. [**174**]{} (2009), 253–268. E. E. Enochs and O. M. G. Jenda, Relative Homological Algebra, de Gruyter Expositions in Math. [**30**]{}, Walter de Gruyter GmbH & Co. KG, Berlin, 2000. E. E. Enochs, O. M. G. Jenda, and J. A. López-Ramos, Dualizing modules and $n$-perfect rings, Proc. Edinb. Math. Soc. [**48**]{}(2005), 75–90. E. E. Enochs and J. A. López-Ramos, Kaplansky classes, Rend. Sem. Mat. Univ. Padova [**107**]{}(2002), 67–79. H.-B. Foxby, Gorenstein modules and related modules, Math. Scand. [**31**]{}(1973), 267–284. Z. H. Gao and Z. Y. Huang, Weak injective covers and dimension of modules, Acta Math. Hungar. [**147**]{}(2015), 135–157. Z. H. Gao and F. G. Wang, Coherent rings and Gorenstein FP-injective modules, Comm. Algebra [**40**]{}(2012), 1669–1679. Z. H. Gao and F. G. Wang, Weak injective and weak flat modules, Comm. Algebra [**43**]{}(2015), 3857–3868. E. S. Golod, G-dimension and generalized perfect ideals, Trudy Mat. Inst. Steklov [**165**]{}(1984), 62–66. R. Göbel and J. Trlifaj, Approximations and Endomorphism Algebras of Modules, de Gruyter Expositions in Math. [**41**]{}, 2nd revised and extended edition, Walter de Gruyter GmbH & Co. KG, Berlin-Boston 2012. H. Holm and P. J[ø]{}rgensen, Semidualizing modules and related Gorenstein homological dimensions. J. Pure Appl. Algebra [**205**]{}(2006), 423–445. H. Holm and P. J[ø]{}rgensen, Covers, precovers and purity, Illinois J. Math. [**52**]{}(2008), 691–703. H. Holm and D. White, Foxby equivalence over associative rings, J. Math. Kyoto Univ. [**47**]{}(2007), 781–808. Z. Y. Huang, Proper resolutions and Gorenstein categories, J. Algebra [**393**]{}(2013), 142–169. L. Hummel and T. Marley, The Auslander-Bridger formula and the Gorenstein property for coherent rings. J. Commut. Algebra [**1**]{}(2009), 283-314. J. J. Rotman, An Introduction to Homological Algebra, Academic Press, New York, 1979. S. Sather-Wagstaff, [Semidualizing Modules,]{} <https://www.ndsu.edu/pubweb/~ssatherw/DOCS/survey.pdf>. S. Sather-Wagstaff, T. Sharif, and D. White, Stability of Gorenstein categories, J. London Math. Soc. [**77**]{}(2008), 481–502. R. Takahashi and D. White, Homological aspects of semidualizing modules, Math. Scand. [**106**]{}(2010), 5–22. W. V. Vasconcelos, Divisor Theory in Module Categories, North-Holland Mathematics Studies, Vol. 14, North-Holland Publ. Co., Amsterdam, 1974. [^1]: E-mail address: gaozenghui@cuit.edu.cn [^2]: E-mail address: tiweizhao@hotmail.com [^3]: 2010 Mathematics Subject Classification: 18G05,16E30,18G20 [^4]: Keywords: (faithfully) semidualizing bimodule, Auslander class, Bass class, $C$-weak injective module, $C$-weak flat module, Foxby equivalence, cover, preenvelope.

--- abstract: 'We study the relation among some security parameters for vectorial Boolean functions which prevent attacks on the related block cipher. We focus our study on a recently-introduced security criterion, called weak differential uniformity, which prevents the existence of an undetectable trapdoor based on imprimitive group action. We present some properties of functions with low weak differential uniformity, especially for the case of power functions and 4-bit S-Boxes.' author: - Riccardo Aragona - Marco Calderini - Daniele Maccauro - Massimiliano Sala date: 'Received: date / Accepted: date' title: On weak differential uniformity of vectorial Boolean functions as a cryptographic criterion --- Introduction {#intro} ============ Differential and linear attacks are major cryptanalytic tools which apply to most cryptographic algorithms. Therefore, functions which guarantee a high resistance to these attacks have been extensively studied. In particular, those with low differential uniformity and high non-linearity, e.g. Almost Perfect Nonlinear (APN) functions or Almost Bent (AB) functions, have received a lot of attention. Since in the design of a block cipher an invertible S-Box of even dimension is usually needed, there is strong interest in non-linear permutations. However, we know examples of APN permutations in even dimension only for dimension 6, for more details see [@apn]. For the highly interesting cases of dimension $4$ and $8$, the cipher designer will certainly use $4$-differentially uniform S-Boxes, but she will also look at other security criteria, if applicable, although it is not obvious which. Besides, even a $4$-differentially uniform S-Box can hide a trapdoor in the related cipher, if not carefully chosen. Algebraic trapdoors can be very dangerous, especially when they are undetectable [@preenel]. We are investigating the security criterion introduced recently in [@CVS], called weak differential uniformity. As shown in [@CVS], any cipher (with a prescribed structure) possessing a weakly-APN vectorial Boolean function as S-Box cannot be successfully attacked using a trapdoor based on imprimitive group action. Indeed, ciphers suffering from such a trapdoor have been built in [@CGC-cry-art-paterson1] and might be used as standards without anyone realizing the trapdoor existence. So, a designer would have advantage in choosing an invertible S-Box which is simultaneously weakly APN and $4$-differentially uniform, which exists for dimension $4$ and $8$ (and actually for any dimension). Results in [@CVS] are generalized on any field in [@ACVS], where again the notion of weakly APN plays an important security role. In Section 2 we recall the attack [@CGC-cry-art-paterson1] that can be mounted on an AES-like cipher when an imprimitive group action is present. We recall also how a weakly APN S-Box would make this attack ineffective [@CVS], motivating thus this security criterion. In Section 3 we present some first results on weak differential uniformity. In Section 4 we specialize to the case of monomial functions, where we see an interesting connection with the property of having the image of a function derivative as an affine space, which is an unexpected weakness within the underlying algorithms (see for instance [@crooked; @onan]). In Section 5 we relate the weak differential uniformity with other algebraic properties of vectorial Boolean functions, in particular with the degrees of both the function components and the function derivative components. We can thus improve some results given in [@onwAPN] and give a formal proof of Fact 4 in [@onwAPN]. In Section 6, we give some results about the partially bent (quadratic), components of a weakly APN permutation and we note that in even dimension weakly APN functions cannot be partially bent (quadratic), behaving thus as APN functions [@Ny; @SZZ]. In Section 7 we give some other properties of vectorial Boolean functions whose derivatives have no constant components, allowing also a deeper understanding of $4$-bit S-Boxes. Cryptographic motivations for studying weak differential uniformity {#sec:0} =================================================================== Most block ciphers used for real-life applications are *iterated block ciphers*, i.e. obtained by a composition of several key-dependent permutations of the message space called “round functions”. Let $\mathcal{C}$ be a block cipher, i.e. a set of permutations $\{\varphi_k\}_{k\in\mathcal{K}}$ of the message space $V$, where $\mathcal{K}$ is the key space. An interesting problem is determining the properties of the permutation group $\Gamma_{\infty}(\mathcal{C})=\Gamma_{\infty}$ generated by the round functions of $\mathcal{C}$ that imply weaknesses of the cipher. A property of $\Gamma_{\infty}$ considered undesirable is the imprimitivity. Paterson [@CGC-cry-art-paterson1] showed that if this group is imprimitive, then it is possible to embed a trapdoor in the cipher. On the other hand, if the group is primitive no such trapdoor can be inserted. We give the idea of the basic chosen-plaintext attack by Paterson. First we recall what [it ]{}is an imprimitive group. Let $G$ be a finite group acting transitively on a set $V$. We will write the action of $g\in G$ on an element $v\in V$ as $v g$. A *partition* $\mathcal{B}$ of $V$ is said to be $G$-*invariant* if $B g\in\mathcal{B}$, for every $B\in\mathcal{B}$ and $g\in G$. A partition $\mathcal{B}$ is *trivial* if $\mathcal{B}=\{V\}$ or $\mathcal{B}=\{\{v\}\, |\, v\in V\}$. A non-trivial $G$-invariant partition $\mathcal{B}$ of $V$ is called a *block system* for the action of $G$ on $V$. Each $B\in\mathcal{B}$ is called a *block of imprimitivity*. $G$ is called *imprimitive* in its action on $V$ if it admits a block system, otherwise it is called *primitive*. Now we suppose that $\Gamma_\infty$ is imprimitive. Let us fix any $k\in\mathcal{K}$ and let $\varphi_k\in\Gamma_\infty$ be the related encryption function. Let $B_1,\dots,B_r$ be a non-trivial block system for the group $\Gamma_\infty$. This attack works only if we know an efficient algorithm (*block sieving*) with input any vector $v\in V$ and output the (unique) block $B_l$ containing $v$. The classical case is when the block system is formed by all the cosets of a known vector subspace of $V$. Paterson gives this trapdoor for a DES-like cipher (for more details on DES see [@specDES]), but it can be extended to the case of AES-like ciphers. We now describe the attack.\  \  \ `Preprocessing performed ones per key`\ We choose one plaintext $m_i$ in each set $B_i$, obtaining the corresponding ciphertext $c_i$. Then the effect of $\varphi_k$ on each block $B_i$ is determined, $$c_i=m_i\varphi_k \in B_j \Rightarrow B_i \varphi_k=B_j.$$ `Real-time processing`\ Given any ciphertext $c$, we can compute $l$ such that $c\in B_l$ via the block sieving. Then, we can find the plaintext $m$ of $c$ by examining the block $B_l\varphi_k^{-1}$.\ `Attack cost`\ The preprocessing costs $r$ encryptions. For any intercepted ciphertext, the search for the corresponding plaintext is limited to a block, whose size is $\frac{|V|}{r}$, requiring at most $\frac{|V|}{r}$ encryptions.  \ Moreover, a cipher $\mathcal{C}$ may have another weakness if $\Gamma_{\infty}(\mathcal{C})$ is of small size, since not every possible permutation of the message space can be realized by the cipher [@CG; @EG83]. Attacks on ciphers whose encryptions generate small groups were given in [@KJRS88]. In [@CVS] the authors define a class of iterated block ciphers, called translation based ciphers ([@CVS], Definition 3.1), large enough to include many common ciphers (as AES [@specAES], SERPENT [@specSERPENT] and PRESENT [@specPRESENT]), and provide some conditions on the S-Boxes of these ciphers that guarantee the primitivity of $\Gamma_{\infty}$. In particular, in Theorem 4.4 of [@CVS], it is proved that if $\mathcal{C}$ is a translation based cipher such that any S-Box satisfies, for some integer $r$, - the weak $2^r$-differential uniformity (see Definition \[def:weakly\] in the next section), and - the strongly $r$-anti-invariance (it means that no S-Box of $\mathcal{C}$ sends a proper subspace of codimension $l$ of the plaintext space to another proper subspace of codimension $l$), then $\Gamma_{\infty}(\mathcal{C})$ is primitive. In Theorem 2 of [@onan] , under the same hypotheses plus an additional cryptographic assumption, i.e. none of the images of the derivatives of any S-Box is a coset of a linear subspace of the message space, it is proved that $\Gamma_{\infty}$ is the alternating group, and so, in other words, $\Gamma_\infty$ is huge. Starting from these cryptographic motivations, in [@onwAPN] the authors provide a deep analysis of 4-bit vectorial Boolean functions focusing on the weak differential uniformity. They determine several conditions, computational and theoretical, which are either sufficient or necessary for a 4-bit vectorial Boolean function to be weakly $2$-differential uniform (weakly APN). Moreover they consider two non-linearity measures, $\hat{n}(f)$ and $n_i(f)$ where $f$ is a vectorial Boolean function (see Section \[sec:degree\]), and they give some relations between such measures and the weakly APNness. If the image of a derivative of an S-Box of a cipher $\mathcal{C}$ is an affine space then this can be another weakness of $\mathcal{C}$. For example, in the yet unpublished PhD Thesis [@phd] the author shows how this condition could induce a weakness based on the action of an alternative operation, called *hidden sum*, for which the vector space structure of the message space is preserved. In [@calsala] some differential properties for permutations in the affine group of the message space with respect to a hidden sum are investigated. In [@ACVS] the authors present conditions on the S-box able to prevent a type of trapdoors based on this alternative operation. One of these conditions is that the derivatives of the S-box do not map the space to an affine subspace. Moreover, also for the hash function case, in [@crooked] the authors show an attack on a SHA-3 candidate (Maraca) [@specMARACA], which is especially effective when the associated vectorial Boolean function has this feature. In Section \[sec:powfun\] we will show a sufficient condition for monomial weakly APN vectorial Boolean functions in order to have that none of their derivatives sends the message space to a proper affine subspace (Corollary \[cor:monomial\]). Weak differential uniformity {#sec:1} ============================ Let $\FF=\f2$. Let $m\ge 1$, any vectorial Boolean function (vBf) $f$ from $\FF^m$ to $\FF^m$ can be expressed uniquely as a univariate polynomial in $\fm[x]$. Any time we write that $f$ is a vBf, we will implicit mean $f:\FF^m\to\FF^m$. When $f$ is also invertible we call it a vBf permutation. We denote the *derivative* of $f$ by $\hat{f}_a(x)=f(x+a)+f(x)$, where $a\in\FF^m\setminus\{0\}$, and the *image* of $f$ by $\Im(f) = \{f(x) \mid x \in\FF^m\}$.\ A notion of non-linearity for S-Boxes that has attracted a lot of research. Let $m\ge 1$. Let $f$ be a vBf, for any $a,b\in \FF^m$ we define $${\delta}_f(a,b)=|\{x\in\FF^m \mid \hat{f}_a(x)=b\}|.$$ The *differential uniformity* of $f$ is $${\delta}(f)=\max_{a,b\in \FF^m\\ a\neq 0}{\delta}_f(a,b).$$ $f$ is said ${\delta}$-*differentially uniform* if ${\delta}={\delta}(f)$.\ Those functions with ${\delta}(f)=2$ are said *Almost Perfect Nonlinear (APN)*. There is a generalization of differential uniformity presented recently in [@CVS], which we recall in the following definition. \[def:weakly\] Let $f$ be a vBf. $f$ is *weakly* ${\delta}$-*differential uniform* if $$|\Im(\hat{f}_a)|>\frac{2^{m-1}}{{\delta}}, \quad \forall\, a \in\FF^m\setminus\{0\}.$$ If $f$ is weakly $2$-differential uniform, it is said *weakly APN*. As shown in [@CVS], a ${\delta}$-differentially uniform map is weakly ${\delta}$-differential uniform. Moreover the following result holds The weak ${\delta}$-differential uniformity is affine-invariant. If $f$ is weakly ${\delta}$-differential uniform and $g(x)=D(f(Cx+c))+d$, for some $m\times m$ matrices $C$ and $D$ with coefficients in $\FF^m$ and for some $c,d\in\FF^m$, then we have $$\begin{aligned} \hat{g}_a(x)&=D (f(C(x+a)+c))+d+D (f(Cx+c))+d\\ &=D (f(Cx+Ca+c))+D (f(Cx+c))\\ &=D (\hat{f}_{Ca}(Cx+c)), \end{aligned}$$ for any $a\in\FF^m$, and so $\Im(\hat{g}_a)=D(\Im(\hat{f}_{Ca}))$.\ Since $C$ and $D$ are permutations, we have that weak ${\delta}$-differential uniformity is affine-invariant. In [@bcc] another non-linearity notion, *the locally almost perfect nonlinearity (locally APN)*, is introduced. Note that, in general, the local-APN property is not equivalent to the weak-APN property. For example, the monomial function $x^{11}$ defined over $\FF^6$ is weakly APN but it is not locally APN. However, for any dimension there exist Boolean functions that are both locally APN and weakly APN, e.g. the patched inversion. \[rem:2.7\] Suppose that $f$ is not a monomial function. If $f$ is weakly ${\delta}$-differential uniform then $f^{-1}$ is not necessarily weakly ${\delta}$-differential uniform. We provide the following example. Let $f:\FF^{4} \to \FF^{4}$ be [$$\begin{aligned} f(x)&=x^{14}+ e^{10}x^{13}+ex^{12}+e^2x^{11}+e^{9}x^{10}+e^8x^9+e^3x^8+e^5x^7\\ &+e^5x^6+e^{11}x^5+e^8x^3+e^{10}x^2+ex+e^{12}, \end{aligned}$$ ]{} where $e$ is a primitive element of $\mathbb{F}_{16}$ such that $e^4=e+1$. The inverse of $f$ is [$$\begin{aligned} f^{-1}(x)&=x^{14}+ e^{10}x^{13}+e^{14}x^{12}+e^8x^{11}+e^{7}x^{10}+e^{10}x^9+x^8+e^5x^7+e^{14}x^6\\ &+e^{2}x^5+e^7x^4+e^5x^3+e^{14}x^2+e^{11}x+e^{14}. \end{aligned}$$ ]{} We have that $f$ is weakly APN while $f^{-1}$ is only weakly 4-differential uniform. We recall that two vBf’s $f$ and $g$ are called CCZ-equivalent (Carlet-Charpin-Zinoviev equivalent) if their graphs $G_f=\{(x,f(x)) \mid x\in\FF^m\}$ and $G_{g}=\{(x,g(x)) \mid x\in\FF^m\}$ are affine equivalent, while they are called EA-equivalent (Extended Affine equivalent) if there exist three affine functions $\lambda$, $\lambda'$ and $\lambda''$ such that $g=\lambda'\circ f\circ \lambda+\lambda''$. Remark \[rem:2.7\] and the fact that a vBf $f$ is CCZ-equivalent to $f^{-1}$ imply the following result. The weak differential uniformity is not CCZ invariant. On the other hand, weak differential uniformity behaves well with respect to EA invariance, as shown below. The weak differential uniformity is EA invariant. Let $f$ and $g$ be EA equivalent and let $f$ be weakly ${\delta}$-differential uniform.\ Then, $g=\lambda'\circ f\circ \lambda+\lambda''=g'+\lambda''$, with $g'$ affine equivalent to $f$ (and ${\lambda}''$ is an affinity over $\FF^m$).\ Since weak differential uniformity is affine invariant, we have $|\Im(\hat{g'}_a)|>2^{m-1}/ {\delta}$ for all $a \in \FF^m\setminus \{0\}$.\ Note that $\Im(\hat{g}_a)=\{b+{\lambda}''(a) \mid b\in \Im(\hat{g'}_a)\}= \Im(\hat{g'}_a)+{\lambda}''(a)$ and so $|\Im(\hat{g}_a)|=|\Im(\hat{g'}_a)|>2^{m-1}/ {\delta}$ for any $a \in \FF^m\setminus \{0\}$. As seen in Section \[sec:0\], if the image of a derivative of an S-Box is an affine space, then there may be a weakness in the cipher. In this direction the following theorem can be useful. Moreover, in Section \[sec:powfun\] we prove a stronger result regarding the monomial functions (Corollary \[cor:monomial\]). \[weaknotAPNcoset\] Let $f$ be a vBf on $\FF^m$ that is weakly APN but not APN. Then, there exists $a\in \FF^m\setminus\{0\}$ such that $\Im(\hat{f}_{a})$ is not a coset of a subspace $W\subseteq \FF^m$. By contradiction suppose that for any $a\neq 0$ we have $ \Im(\hat{f}_{a})=w+ W$ for some $w \in \FF^m$ and $W$ vector subspace of $\FF^m$. Since $f$ is weakly APN, $|\Im(\hat{f}_{a})|>2^{m-2}$, thus $\dim_{\FF}(W)=m-1$. Therefore, we have that $\hat{f}_a$ is a $2$-to-$1$ function for all $a\neq 0$, which means $f$ is APN, contradicting our assumption. Power functions {#sec:powfun} =============== In this section we focus on monomial functions, also called *power functions*. In particular we prove that the weak differential uniformity of a function $f$ is equal to that of $f^{-1}$, and we show some properties of $\Im(\hat{f}_a)$. In this section when we write $f=x^d$ we mean that $f$ is a power function on $\fm$ for any $0\leq d\leq 2^{m-1}$. We will also identify $\fm$ and $\FF^m$ without any further comments. The following result is well-known (see for instance [@power]). \[prop:1\] Let $f(x)=x^d$. For any $a,a'\in \FF^m$, with $a,a'\neq 0$, and any $0\le i\le 2^{m}$, we have $$|\{b\in\FF^m \mid {\delta}_f(a',b)=i\}|=|\{b\in\FF^m \mid {\delta}_f(a,b)=i\}|.$$ In other words, the differential characteristics of a monomial function depend only on $b$. Let $f(x)=x^d$ and $0\le i\le 2^m$. We denote by ${\omega}_i$ the number of output differences of $b$ that occur $i$ times, that is $${\omega}_i(f)=|\{b \in \FF^m \mid {\delta}_f(1,b)=i\}|.$$ The *differential spectrum* of $f$ is the sequence of ${\omega}_i(f)$’s, denoted by $\S(f)$. Note that if a monomial function $f$ has $2^s$-to-$1$ derivatives then it is weakly $2^s$-differential uniform, since $|\Im(\hat{f}_a)|=2^{m-s}$ for any $a\in\FF^m\setminus\{0\}$. The following lemma is well-known, for instance see Lemma 1 in [@power]. \[lm:1\] Let $f(x)=x^d$ and $g(x)=x^e$. If - $\gcd(2^m-1,d)=1$ and $ed\equiv {1} \mod (2^m-1)$, - or - $e\equiv 2^kd \mod (2^m-1)$, for some $k$, $0\leq k\leq m$, then $\quad \S(f)=\S(g)$. From Lemma \[lm:1\] we obtain our first result. Let $f(x)=x^d$ with $\gcd(2^m-1,d)=1$. Then $f$ is weakly ${\delta}$-differential uniform if and only if $f^{-1}$ is weakly ${\delta}$-differential uniform. For a power function we have $$|\Im(\hat{f}_a)|=|\Im(\hat{f}_1)| = 2^m-{\omega}_0,\quad\forall a \neq 0.$$ From Lemma \[lm:1\] we have ${\omega}_0(f)={\omega}_0(f^{-1})$, and this concludes the proof. Consider the following lemma for a power function (not necessarily a permutation). \[prop:2\] Let $f(x)= x^d$. If there exists $a \in \FF^m$, $a\neq 0$, such that $\Im(\hat{f}_a)$ is a coset of a subspace of $\FF^m$, then $\Im(\hat{f}_{a'})$ is a coset of subspace of $\FF^m$ for all $a'\neq 0$. We have $\Im(\hat{f}_a)=w+W$, where $W$ is a $\FF$-vector subspace of $\FF^m$ and $w\in \FF^m$. If we now consider $a'\in \FF^m\setminus\{0\}$ we have $$\hat{f}_{a'}(x)=(x+a')^d+x^d=\left(\frac{a'}{a}\right)^d\left[\left(x\frac{a}{a'}+a\right)^d+\left(x\frac{a}{a'}\right)^d\right]=\left(\frac{a'}{a}\right)^d\hat{f}_a\left(x\frac{a}{a'}\right).$$ Therefore, $\Im(\hat{f}_{a'})=\left( \frac{a'}{a} \right)^d \Im(\hat{f}_{a})=\left(\frac{a'}{a}\right)^d w+\left(\frac{a'}{a}\right)^dW=w'+W'$. Since $W'=(a'/a)^d W$, also $W'$ is an $\FF$-vector subspace of $\FF^m$ and our claim is proved. Here we give a sufficient condition for a power function to thwart the aforementioned weakness.\ The following result is an obvious consequence of Theorem \[weaknotAPNcoset\] and Lemma \[prop:2\]. \[cor:monomial\] Let $f$ be a vBf permutation on $\FF^m$ that is weakly APN but not APN. If $f(x)=x^d$, then for all $a\in \FF^m\setminus\{0\}$, $\Im(\hat{f}_{a})$ is not a coset of a subspace $W\subseteq \FF^m$. Weakly APN functions and degrees of derivatives {#sec:degree} =============================================== Without loss of generality, in the sequel we consider only vBf’s such that $f(0) = 0$. Let $v\in\FF^m\setminus\{0\}$, we denote by $\langle f,v\rangle$ the component $\sum_{i=1}^m v_if_i$ of $f$, where $f_1,\dots,f_m$ are the coordinate functions of $f$. The *algebraic degree* of a vectorial Boolean function $f$ is the maximal algebraic degree of the coordinate functions of $f$ and is denoted by $\deg(f)$. We recall the following non-linearity measures, as introduced in [@onwAPN]: $$n_i(f):=|\{v\,\in\FF^m \setminus\{0\}\mid\deg(\langle f,v\rangle)=i\}|,$$ and $$\hat n (f):=\max_{a\in\FF^m\setminus\{0\}}|\{v\,\in\FF^m \setminus\{0\}\mid\deg(\langle\hat{f}_a,v\rangle)=0\}|.$$ In other words, $n_i(f)$ corresponds to the number of component functions of $f$ which are of degree $i$ and $\hat n (f)$ corresponds to the number of components of the derivative functions of $f$ which are constant.\ We state two lemmas useful to extend some results of [@onwAPN]. First, we recall that the algebraic degree of a permutation $f(x)=x^d$ is the Hamming weight of the binary representation of $d$, denoted by $\mathrm{w}(d)$. \[Kyu07\] Let $f(x) = x^d$ be a permutation. Then $\hat{f}_1$ has at least one constant component if and only if $\mathrm{deg}(f) = 2$. \[Her05\] Let $f(x) = x^d$, with $d = 2^{2k} - 2^k + 1$ (Kasami exponent), $\gcd(k, n) = s$ and $\frac{n}{s}$ odd. Then $\hat{f}_1$ is a $2^s$-to-$1$ function. \[th:wapn\] Let $f$ be a vBf permutation such that $\hat{n}(f)=0$, i.e. no component of $f$ has linear structure. Then \(i) if $m=3$ then $f$ is weakly APN;\ (ii) if $m=4$ then $f$ is weakly APN;\ (iii) if $m = 2n$, with $n$ odd, then $f$ is not necessarily weakly APN. \(i) Let $\FF^3=\{x_1,\dots,x_8\}$ and let $M_a$ be the matrix of dimension $3\times8$, whose columns are $m_j=\hat{f}_a(x_j)$ for $1\le j\le 8$. We claim that $\hat n (f)=0$ implies $\mathrm{rank}(M_a)=3$ for any $a$. Otherwise, we could obtain $(0,\dots,0) \in \FF^8$ from a combination of the rows of $M_a$, and the corresponding component of $\hat{f}_a$ would be identically $0$.\ If $f$ is not weakly APN, we have $|\mathrm{Im}(\hat{f}_a)| \le 2$ for some $a\in\FF^3\setminus{\{0\}}$. So we have at most $2$ distinct columns, which implies $\mathrm{rank}(M_a)\le 2$ and contradicts $\mathrm{rank}(M_a)=3$. \(ii) See [@onwAPN] Proposition $2$ . \(iii) Let $t>0$ be such that $\gcd(2^{2^{t+1}} - 2^{2^t} + 1, 2^m - 1) = 1$, and consider the power function $f(x) = x^d$, with $d = 2^{2^{t+1}} - 2^{2^t} + 1$. By hypothesis $\gcd(2^t,m) = 2$, thus, by Lemma \[Her05\], $f$ is $4$-differentially uniform and thus weakly $4$-differential uniform. Since in our case $d=2^{2^t}(2^{2^t}-1)+1$, then $\mathrm{w}(d)=\mathrm{w}(2^{2^t}-1)+1$ which is strictly bigger than $2$ for $t>0$. So $f$ is not quadratic and then, by Lemma \[Kyu07\], $\hat{n}(f) = 0$. In [@onwAPN] it was shown that a weakly APN function $f$ over $\mathbb{F}^4$ has $n_3(f)\in\{12,14,15\}$, moreover by a computer check on the class representatives the authors exclude the case $n_3(f)=12$ (Fact $4$ in [@onwAPN]).\ We are now able to provide a formal proof. Let $f: \mathbb{F}^4 \to \mathbb{F}^4$ be a weakly APN permutation. Then $n_3(f)\in\{14,15\}$. Let $f = (f_1, f_2, f_3, f_4)$ with $f_i : \FF^4\to\FF$, and assume by contradiction that $\deg(S) \le 2$ for three distinct linear combinations $S=\sum_{i}v_if_i$, that we call $S_1, S_2, S_3$. From the theory of quadratic Boolean functions (see for instance [@quadratic]) $\hat{S}_a$ is constant for every $a \in V(S)$ where $V(S)\subseteq \FF^4 $ is a vector subspace, called the *set of linear structures* of $S$. It is well-known that $V(S)$ has dimension $0$ if and only if $S$ is bent, it has dimension $4$ if and only if $S$ is linear (affine), and it has dimension $2$ otherwise. Since $V(S)$ is a vector space, $S_3=S_1+S_2$. If there exists $a\in V(S_i)\cap V(S_j)$ different from $0$ for some $i\ne j$, then $\hat{n}(f)\ge 2$. But $f$ weakly APN implies $\hat{n}(f)\le 1$ (see [@onwAPN] Theorem $1$). Therefore, $V(S_i)\cap V(S_j)=\{0\}$ and $\dim(V(S_i))\le 2$ for any $i$.\ For any $i$, since $f$ is a permutation, then $S_i$ is balanced, so $S_i$ is not bent, and then $$\label{S3} \dim(V(S_i))=2, \quad i=1,2,3 \,.$$ Summarizing, $\deg(S_i)=2$ for any $i$ and $V(S_i)\cap V(S_j)=\{0\}$ for any $i\ne j$.\ Up to an affine transformtion, since $V(S_1)\oplus V(S_2)=\FF^4$, we can assume $V(S_1)=\mathrm{Span}( (1,0,0,0),(0,1,0,0))$ and $V(S_2)=\mathrm{Span}((0,0,1,0),(0,0,0,1))$. Let $S_1(x)=\sum_{i<j}c_{i,j}x_ix_j+\sum_i c_i x_i$. Since $S_1(x+(1,0,0,0))+S_1(x)$ is constant we have that $c_{i,j}=0$ if $i$ or $j$ equals $1$. Similarly, since $S_1(x+(0,1,0,0))+S_1(x)$ is constant we have $c_{i,j}=0$ if $i$ or $j$ equals $2$. Then $S_1(x)=x_3x_4+\sum_i c_i x_i$ and analogously we have $S_2(x)=x_1x_2+\sum_i c_i' x_i$, for some $c_i'$’s.\ So, $S_3(x)=x_1x_2+x_3x_4+\sum_i b_i x_i$, $b_i=c_i+c_i'$, and we can compute the derivative of $S_3$ with respect to $a=(a_1,a_2,a_3,a_4)\in\FF^4\setminus\{0\}$ as $$\hat{(S_3)}_a(x)=a_2x_1+a_1x_2+a_4x_3+a_3x_4+c, \mbox{ where $c$ is constant.}$$ Hence $\hat{(S_3)}_a(x)$ is constant if and only if $a=0$, so $S_3$ is bent and $\dim(V(S_3))=0$, contradicting . Quadratic and partially bent functions {#sec:part} ====================================== \[quadraticweakly\] A quadratic function is APN if and only it is weakly APN. Let $f$ be weakly APN and $a\ne 0$ arbitrary. Then by definition, $|\Im(\hat{f}_{a})|>2^{m-2}$. Since $f$ is quadratic, $\hat{f}_{a}$ is affine. Then $\Im(\hat{f}_{a})$ is an affine subspace. Hence its size is a power of $2$, the only possibility of being equal $2^{m-1}$. So $|\Im(\hat{f}_{a})|=2^{m-1}$ for any non-zero $a$, which means that $\hat{f}_{a}$ is $2$-to-$1$ for all non-zero $a$. As was shown in [@SZZ] there is no APN quadratic permutation over $\FF^m$ for $m$ even, and so, by previous theorem, there is no weakly APN quadratic permutation over $\FF^m$ for $m$ even. This result was extended by Nyberg [@Ny] to the case of permutations with all components partially bent (for $m$ even), in other words there is no APN partially bent permutation. We are able to extend these results by relaxing the condition *APN permutations* with the condition *weakly APN permutation*. A Boolean function $f$ is partially bent if there exists a linear subspace $V(f)$ of $\FF^m$ such that the restriction of $f$ to $V(f)$ is affine and the restriction of $f$ to any complementary subspace $U$ of $V(f)$, $V(f)\oplus U =\FF^m$, is bent. In that case, $f$ can be represented as a direct sum of the restricted functions, i.e., $f(y + z) = f(y) + f(z)$, for all $z\in V(f)$ and $y\in U$. \[rk:U\] The space $V(f)$ is formed by the linear structures of $f$, in fact $$f(x+a)+f(x)=f(y+z+a)+f(y+z)=f(y)+f(z)+f(a)+f(y)+f(z)=f(a)$$ where $z,a\in V(f)$ and $y\in U$. Moreover, since bent function exist only in even dimension, $m-\dim(V(f))$ is even. That means if $m$ is even, the dimension of $V(f)$ is even. \[th:pb\] For $m$ even, a weakly APN permutation has at most $\frac{2^m-1}{3}$ partially bent components. In particular $f$ cannot have all partially bent components. Let $f$ be a weakly APN permutation. Assume by contradiction that $f$ has more than $\frac{2^m-1}{3}$ partially bent components, and denote those with $f_1,\dots,f_s$. $f$ is a permutation, then $\dim(V(f_i))\neq 0$ for all $1\le i\le s$, otherwise $f_i$ is bent and it is not balanced. From Remark \[rk:U\] we have that there exist at least three nonzero vectors in each $V(f_i)$. So $$\sum_{i=1}^s |V(f_i)|\ge 3\,s>2^m-1.$$ Thus, there exist $i$ and $j$ such that $a\in V(f_i)\cap V(f_j)$ with $a\neq 0$. This implies $\hat{n}(f)\ge 2$, which contradicts that $f$ is weakly APN, since in that case $\hat n(f)\le 1$ ([@onwAPN] Theorem $1$). From the fact that a quadratic Boolean function is partially bent (see for instance [@Ny]), we have immediately the following result. Let $m$ even. Let $f$ be a weakly APN permutation. Then $f$ has at most $2^{m-2}-1$ quadratic components. That depends on the fact that the set of components of degree less or equal to $2$ is a vector space. Properties linked to $\hat{n}(f)$ ================================= In this last part of the paper we give some properties linked to the value of $\hat{n}(f)$ of a vBf. For all $a\in \FF^m\setminus\{0\}$, let $V_a$ be the vector space $\{v\,\in\FF^m\,:\,\deg(\langle\hat{f}_a,v\rangle)=0\}$. By definition, if $t=\max_{a\in\FF^m\setminus\{0\}}\dim(V_a)$, then $\hat{n}(f)=2^t-1$. Let $f$ be a vBf and $a\in \FF^m\setminus\{0\}$. $f(a)+V_a^{\perp}$ is the smallest affine subspace of $\FF^m$ containing $\Im(\hat{f}_a)$. In particular, $\hat{n}(f)=0$ if and only if there does not exist a proper affine subspace of $\FF^m$ containing $\Im(\hat{f}_a)$, for all $a\in \FF^m\setminus\{0\}$. Let $a\in\FF^m\setminus\{0\}$. Note that $V_a=\{v\,\in\FF^m\,:\,\langle\hat{f}_a,v\rangle \mbox{ is constant} \}$. Let $x\in \FF^m$, then $\hat f_a(x)=f(a)+w$, for some $w\in \FF^m$, and $\langle\hat{f}_a(x),v\rangle=c\in\FF$ for all $v \in V_a$. In particular $c=\langle\hat{f}_a(0),v\rangle\,=\,\langle f(a),v\rangle$ and so $\langle w,v\rangle=0$, that is, $w\in V_a^\perp$. Then we have $\Im(\hat f_a)\subseteq f(a)+V_a^\perp$. Now, let $A$ be an affine subspace containing $\Im(\hat{f}_a)$, then $A=f(a)+V$, for some vector subspace $V$ in $\FF^m$. For all $v\in V^{\perp}$, we have $\langle\hat{f}_a,v\rangle=\langle f(a),v\rangle=c\in\FF$ and so, by definition, $V^{\perp}\subseteq V_a$. Then $A$ contains $f(a)+V_a^{\perp}$. Finally, $\hat{n}(f)=0$ if and only if $V_a=\{0\}$ for all $a\in\FF^m\setminus\{0\}$, and so our claim follows. The proposition above gives a sufficient condition, *i.e.* $\hat{n}(f)=0$, such that the derivates do not map the message space to an affine subspace; and so a type of trapdoors can be avoided, as noted in Section \[sec:1\]. The following proposition is well-known, but we propose a proof in our context. \[prop:15\] Let $f: \mathbb{F}^m \to \mathbb{F}^m$ be a Boolean permutation such that $\hat{n}(f)=0$. Then $f$ has no partially bent (quadratic) components. $\hat{n}(f)=0$ implies that the linear structures set of any component contains only $0$. So if there exists a partially bent (quadratic) component, then it is bent. But $f$ is a permutation, then this is not possible. For the particular case of $4$-bit S-Boxes we obtain two more results. Let $f: \mathbb{F}^4 \to \mathbb{F}^4$ be a vBf permutation.\ (i) If $\hat{n}(f)=0$. Then $f$ is weakly APN and $n_3(f)=15$.\ (ii) If $f$ is weakly APN and $n_3(f)=14$. Then $\hat{n}(f)=1$. Let $f$ be weakly APN, so $\hat{n}(f)\leq 1$ (see [@onwAPN]). From Proposition \[prop:15\], the claim follows. So for weakly APN function $f:\FF^4\to \FF^4$ we have all the three cases. Below we provide an example for each case reporting the algebraic normal form of the components of $f$: - $\hat{n}(f)=0$ and $n_3(f)=15$: $$\begin{aligned}[l] &f_1=x_1x_2x_3+ x_2x_3x_4 + x_1x_3 + x_2x_3+ x_1 + x_2 + x_3 + x_4\\ &f_2=x_1x_2x_4 + x_1x_2 + x_1x_3 + x_2x_3 + x_2x_4 + x_4\\ &f_3=x_1x_3x_4 +x_1x_2 + x_1x_3 + x_1x_4 + x_3 + x_4\\ &f_4=x_2x_3x_4 +x_1x_4 + x_2x_4 + x_2 + x_3x_4 + x_3 + x_4 \end{aligned}$$ - $\hat{n}(f)=1$ and $n_3(f)=15$: $$\begin{aligned}[l] &f_1=x_1x_3x_4 + x_2x_3x_4 + x_2x_3 + x_2x_4 + x_3x_4+ x_1\\ &f_2=x_1x_2x_4 + x_1x_3 + x_1x_4 + x_2x_3 + x_2\\ &f_3=x_1x_2x_3 + x_1x_2x_4+ x_1x_3x_4 + x_2x_3x_4 + x_1x_2 +x_3x_4 + x_3\\ &f_4=x_2x_3x_4 +x_1x_2 + x_1x_4 + x_2x_3 + x_4 \end{aligned}$$ - $\hat{n}(f)=1$ and $n_3(f)=14$: $$\begin{aligned}[l] &f_1=x_1x_2x_3 + x_1x_2x_4 + x_1x_3 + x_1 + x_2x_3x_4 + x_2x_3 + x_3x_4\\ &f_2=x_1x_2x_4 + x_1x_2 + x_1x_3x_4 + x_1x_3 + x_1x_4 + x_2\\ &f_3=x_1x_2x_4 + x_1x_2 + x_1x_3x_4 + x_1x_3 + x_2x_4 + x_3\\ &f_4=x_1x_3 + x_1x_4 + x_2x_3x_4 + x_2x_4 + x_4 \end{aligned}$$ Conclusions =========== As reported in Section \[sec:0\] and \[sec:1\], weak differential uniformity and the cryptographic condition that the image of the derivatives of an S-Box is never a coset of a subspace of the message space are useful to prevent hiding certain type of trapdoors in the related cipher. First we study the algebraic structure of the image of the derivatives of a vectorial Boolean function. In particular we prove that for any vBf that is weakly APN but not APN, there is at least one derivative whose image is not an affine subspace (Theorem \[weaknotAPNcoset\]). In the case of power functions, to be weakly APN but not APN guarantees that none of the image of the derivatives is an affine subspace (Corollary \[cor:monomial\]). An interesting open problem is to find families of vBf that are not monomial having this property. Then we show that the sufficient condition $\hat{n}(f)=0$, that ensures weakly APNness for the 4-bit vBf’s, does not guarantee this property for $m$-bit vBf’s with $m>4$ (Theorem \[th:wapn\]). It would be interesting to find sufficient conditions that imply weakly APNness for any $m>4$. In Section \[sec:part\] we extend some results known for the (quadratic) partially bent components of an APN permutation to the case of weakly APN permutations. In the last section we report some other results linked to the value of $\hat{n}(f)$, in particular we prove that this value is zero if and only if the derivates of $f$ do not map the message space to an affine subspace. Acknowledgements {#acknowledgements .unnumbered} ================ We are grateful to the unknown referees for their suggestions, which were decisive in order to improve and clarify the exposition. In particular we would like to thank one of the referees for the Theorem \[quadraticweakly\] and its proof. [100]{} R. Aragona, A. Caranti, F. Dalla Volta, M. Sala, On the group generated by the round functions of translation based ciphers over arbitrary finite fields, Finite Fields and Appl. 25, 293–305 (2014). E. Biham, R. Anderson, L. Knudsen. Serpent: A New Block Cipher Proposal, in: Fast Software Encryption, LNCS 1372, pp 222–238 (1998). C. Blondeau, A. Canteaut, and P. Charpin. Differential properties of power functions, Int. J. Inf. Coding Theory 1 (2), 149–170 (2010). C. Blondeau, A. Canteaut, and P. Charpin, Differential Properties of $x\mapsto x^{2^{t}-1}$, IEEE Trans. Inform. Theory, 57(12), 8127–8137 (2011). A. Bogdanov, L. R. Knudsen, G. Leander, C. Paar, A. Poschmann, M. J. B. Robshaw, Y. Seurin, and C. Vikkelsoe. PRESENT: An Ultra-Lightweight Block Cipher, CHES’07, LNCS 4727, pp 450–466 (2007). K. Browning, J. Dillon, M. McQuistan, A. Wolfe, An APN permutation in dimension six, in Finite Fields: Theory and Applications - FQ9. Providence, RI: AMS, vol. 518, Contemporary Mathematics, pp. 33–42 (2010). M. Calderini, On Boolean functions, symmetric cryptography and algebraic coding theory, PhD Thesis, University of Trento (2015). M. Calderini and M. Sala, On differential uniformity of maps that may hide an algebraic trapdoor, in Algebraic Informatics, LNCS 9270, pp. 70–78 (2015). A. Canteaut, P. Charpin and G. M. Kyureghyan, A new class of monomial bent functions, Finite Fields and Appl. 14 (1), 221–241 (2008). A. Canteaut and M. Naya-Plasencia, Structural weakness of permutations with a low differential uniformity and generalized crooked functions, Finite Fields: Theory and Applications-Selected Papers from the 9th International Conference Finite Fields and Applications, Contemporary Mathematics, 518 (2009). A. Caranti, F. Dalla Volta, and M. Sala, On some block ciphers and imprimitive groups, Appl. Algebra Engrg. Comm. Comput. 20 (5-6), 339–350 (2009). A. Caranti, F. Dalla Volta, and M. Sala, An application of the O’Nan-Scott theorem to the group generated by the round functions of an AES-like cipher, Des. Codes Cryptogr. 52(3), 293–301 (2009). C. Carlet, Partially-bent functions, Des. Codes Cryptogr. 3(2), 135–145 (1993). D.Coppersmith, E.Grossman, Generators for certain alternating groups with applications to cryptography, SIAM J. Appl. Math. 29(4), 624–627 (1975). J. Daemen and V. Rijmen. The design of Rijndael: AES-the advanced encryption standard. Springer Science & Business Media, 2013. S. Even, and O. Goldreich, Des-like functions can generate the alternating group., IEEE Trans. Inform. Theory 29(6), 863–865 (1983). FIPS PUB 46-3. Data Encryption Standard (DES), NIST, 1999. <http://csrc.nist.gov/publications/fips/fips46-3/fips46-3.pdf> C. Fontanari, V. Pulice, A. Rimoldi, M. Sala, On weakly APN function and 4-bit S-boxes, Finite Fields and Appl. 18, 522–528 (2012). D. Hertel, A note on the Kasami power function, Cryptology ePrint Archive (2005), <https://eprint.iacr.org/2005/436.pdf>. R. J. Jenkins Jr. Maraca - algorithm specification, Submission to NIST, 2008. <http://burtleburtle.net/bob/crypto/maraca/nist/Supporting_Documentation/specification.pdf> B. S. Kaliski Jr, R. L. Rivest, and A. T. Sherman, Is the Data Encryption Standard a group? (Results of cycling experiments on DES), J. Cryptology 1(1), 3–36 (1988). G. M. Kyureghyan, Crooked maps in $\F_{2^n}$, Finite Fields and Appl. 13 (3), 713–726 (2007). K. Nyberg, S-boxes and Round Functions with Controllable Linearity and Differential Uniformity, in: Fast Software Encryption, LNCS 1008, pp. 111–130 (1995). K.G.Paterson, Imprimitive permutation groups and trapdoors in iterated block ciphers, in: Fast Software Encryption, LNCS 1636, pp. 201–214 (1999). V. Rijmen, B. Preneel, A family of trapdoor ciphers, in: Fast Software Encryption LNCS 1267, pp. 139–148 (1997). J. Seberry, X. Zhang, and Y. Zheng. Pitfalls in designing substitution boxes, in: Advances in Cryptology - CRYPTO ’94, LNCS 839, pp. 383–396 (1994).

--- abstract: | The purpose of this note is to report on the discovery of the primes of the form $p=1+n!\sum n$, for some natural numbers $n>0$. The number of digits in the prime p are approximately equal to $\lfloor log_{10}(1+n!\sum n)\rceil+1$.\ **Key Words:** Primes, factorials, sum of the first $n$ natural numbers\ **AMS Subject Classification:** 11A41,05A10,97F40 author: - | [By]{}\ Maheswara Rao Valluri[ ]{} title: 'Primes of the form [$p=1+n!\sum n,$ for some $n\in\mathbb{N}^{+}$]{}' --- Introduction {#introduction .unnumbered} ============ A natural number is a prime if it has only factors of $1$ and itself. There are, by Euclidean theorem (about 350BC) infinitely many primes. There are many patterns of primes amongst which the classical known primes are the Mersenne primes of the form $2^{p}-1$, where $p$ is a prime [\[]{}Da2011[\]]{}, and the Fermat primes of the form $2^{2^{n}}+1$ for a natural number $n\geq0$ [\[]{}Da2011[\]]{}. We omit the other classes of primes except the factorial primes. There was a known fact that early 12th century Indian scholars knew about factorials. In 1677, a British mathematician, Fabian Stedman described factorials for music. In 1808, a French mathematician, Christian Kramp introduced notation $!$ for factorials. The factorial of $n$ can be described as product of all positive integers less than or equal to $n$. In the Christian Kramp’s notation, $n!=n(n-1)(n-2)......3.2.1.$ Factorials of $0$ and $1$ can be written as $0!=1$ and $1!=1$, respectively. There are dozens of prime factorials. However, we recall only few of them, particularly, factorial primes of the form $(p!\pm1)$ [\[]{}Bor1972, BCP1982[\]]{}, double factorial primes of the form $n!!\pm1$ for some natural number $n$ [\[]{}Mes1948[\]]{}, Wilson primes: $p$ for which $p^{2}$ divides $(p-1)!+1$ [\[]{}Bee1920[\]]{}, and primorial primes of the form $(p\#\pm1)$ which means as the product of all primes up to and including the prime [\[]{}Dub1987, Dub1989[\]]{}. Further, a class of the Smarandache prime is of the form $n!\times S_{n}(n)+1$ , where $S_{n}(n)$ is smarandache consecutive sequence [\[]{}Ear2005[\]]{}. The purpose of this note is to report on the discovery of the primes of the form $p=1+n!\sum n$ for some natural numbers $n>0$. Primes of the form $p=1+n!\sum n$ for some $n\in\mathbb{N}^{+}$ {#primes-of-the-form-p1nsum-n-for-some-ninmathbbn .unnumbered} =============================================================== We list in the table 1, the primes of the form $p=1+n!\sum n$ for some natural numbers $n>0$. They are verified up-to $n=950$ and primes are found when $n=1,2,3,4,5,6,$ $7,8,9,10,12,13,14,19,24$,251,374. The above primes can also be expressed as $p=1+\frac{(n+1)!n}{2},$ for some natural numbers $n>0$. The author has used python software to search and verify the above form primes and could verify up-to $n=950$. The author conjectures that there are infinitely many primes of the form $p=1+n!\sum n$ for some $n\in\mathbb{N}^{+}$ .   ------- ----- ---------------------------------------------- --------------- S.No. n Prime, $p$ Number of digits in $p$ 1 1 2 1 2 2 7 1 3 3 37 2 4 4 241 3 5 5 1801 4 6 6 15121 5 7 7 141151 6 8 8 1451521 7 9 9 16329601 8 10 10 199584001 9 11 12 37362124801 11 12 13 566658892801 12 13 14 9153720576001 13 14 19 23112569077678080001 20 15 24 186134520519971831808000001 27 16 251 25662820338985371726..Omitted..000000000001 500 17 374 22873802587990440054..Omitted..0000000000001 807 ------- ----- ---------------------------------------------- --------------- : Primes List Size of the prime of the form $p=1+n!\sum n$ for some $n\in\mathbb{N}^{+}$ {#size-of-the-prime-of-the-form-p1nsum-n-for-some-ninmathbbn .unnumbered} -------------------------------------------------------------------------- To compute the size of primes of above form, we use the Stirling ’s formula [\[]{}Sec 2.2, CG2001[\]]{}: $log\,n!=(n+\frac{1}{2})log\,n-n+\frac{1}{2}log2\pi+O(\frac{1}{n})$. Simply, we can also write $log\,n!\sim n(logn-1)$, if necessary. The size of the prime, $p=1+n!\sum n$ is approximately equal to $\lfloor log_{10}(1+n!\sum n)\rceil+1$. Conclusion {#conclusion .unnumbered} =========== In this note, the author conjectures that there are infinitely many primes of the form $p=1+n!\sum n$ for some natural numbers $n>0$. The author has found the primes for $n=1,2,3,4,5,5,6,7,8,9,10,12,13,14,19,24,251,374$, when they are searched and verified up-to $n=950$. The number of digits in the primes of the form $(1+n!\sum n)$ are approximately equal to $\lfloor log_{10}(1+n!\sum n)\rceil+1$. Furthermore, an investigation will be required for finding such primes. [BCP1982]{} J. P. Buhler, R. E. Crandall and M. A. Penk, Primes of the form $n!\pm1$ and $2\cdot3\cdot5...p\pm1$, Math. Comp., Vol.38(158), 1982, Pages 639-643. A. Borning, Some results for $k!\pm1$ and $2\cdot3\cdot5...p\pm1$, Math. Comp., Vol. 26, 1972, Pages 567-570. H. Dubner, Factorial and primorial primes, J. Rec. Math., Vol.19(3), 1987, Pages 197-203. D.M.Burton, “Elementary Number Theory”, McGraw-Hill 7th Edition, 2011. B.E.Meserve. Double factorials, Amer. Math. Monthly, Vol.55, 1948, Pages 425-426. H. Dubner, A new primorial prime, J. Rec.Math., Vol.21(4), 1989, Page 276. N. G. W. H. Beeger, “On the congruence $(p-1)!\equiv-1(mod\,p^{2})$”, Messenger of Mathematics, Vol.49, 1920, Pages 177-178. J.Earls, “On a new class of Smarandache prime numbers”, Scientia Magna, Vol.01(01), 2005. C.K.Caldwell and Y.Gallot, “On the primality of $n!\pm1$ and $2\times3\times5\times\cdot\cdot\cdot\times p\pm1$”, Math.Comp., Vol.71(237), 2001, Pages 441-448. ----------------------------------------------- School of Mathematical and Computing Sciences Fiji National University P.O.Box:7222, Derrick Campus, Suva, Fiji E-mail: maheswara.valluri@fnu.ac.fj -----------------------------------------------

--- abstract: 'In completely local settings, we establish that a dynamically evolving black hole horizon can be assigned a Hawking temperature. Moreover, we calculate the Hawking flux and show that the radius of the horizon shrinks.' author: - Ayan Chatterjee - Bhramar Chatterjee - Amit Ghosh title: Hawking radiation from dynamical horizons --- The laws of black hole mechanics in general relativity are remarkably analogous to the laws of thermodynamics [@Bardeen:1973gs]. This analogy is exact when quantum effects are taken into account. Indeed, Hawking’s semiclassical analysis establishes that quantum mechanically, a stationary black hole with surface gravity $\kappa$ radiates particles to infinity with a perfect black body spectrum at temperature $\kappa/2\pi$ [@Hawking:1974sw]. Consequently, asymptotic observers perceive a thermal state and assign a physical temperature to the black hole. The precise match to thermodynamics is complete when the thermodynamic entropy of the black hole is identified with a quarter of its area [@Bekenstein:1973ur]. The original calculation of Hawking is independent of the gravitational field equations. It relies only on the behavior of quantum fields in a specific spacetime geometry describing a stationary black hole formed due to a gravitational collapse. Over the years, several other techniques have been developed to study spontaneous particle emission and the Hawking temperature for more general spacetimes. For example, the Hartle-Hawking proposal [@Hartle:1976tp] and the Euclidean approach [@Gibbons:1976pt] have been extensively used to associate thermal states to spacetimes with bifurcate Killing horizons. In fact, it has been established that in any globally hyperbolic spacetime with bifurcate Killing horizon, there can exist a vacuum thermal state at temperature $\kappa/2\pi$ which remains invariant under the isometries generating the horizon [@Kay:1988mu]. Although these constructions are elegant, they are quite restrictive, inapplicable even for spacetimes with superradiance [@Kay:1988mu]. These formulations also do not indicate how such a thermal state may arise as a result of some version of physical process. In addition, their existence requires knowledge of global structure of spacetime. As a result, they do not appear very useful to study thermal properties of local horizons. On the other hand, the laws of black hole mechanics apply equally well to black hole horizons which can been proved using only local geometrical properties of null surfaces, without any assumptions on the global development of the spacetime in which the horizon is embedded [@Hayward:1993wb; @Hayward:1997jp; @Ashtekar:2000sz; @Booth:2003ji]. It has also been established that such horizons can be assigned an entropy proportional to the area of the local horizon [@Ashtekar:1997yu; @Ghosh:2011fc]. Thus, it seems to be a reasonable physical expectation that even with a local definition of black hole horizon one should be able to establish the analogy to thermodynamics. More precisely, such horizons should have a temperature of $\kappa/2\pi$. Incidentally, this question has been investigated in a semiclassical approach which treats Hawking radiation as a quantum tunneling phenomenon [@Parikh:2000ht]-[@Mitra:2008hj]. The method involves calculating the imaginary part of the action for the (classically forbidden) process of s-wave emission, from inside and through the horizon (see [@Vanzo:2011tm] for more details). Using the WKB-approximation the tunnelling probability for such a classically forbidden trajectory is calculated to be, $\Gamma = e^{-2{\rm Im} S}$ where, $S$ is the classical action of the trajectory to leading order in $\hbar$. This is equated to the Boltzmann factor $ e^{-{\beta}E}$ to extract the inverse Hawking temperature $\beta$. The main advantage of this formalism is that the calculations involve only the local geometry and hence can be applied to any local horizon. Indeed, tunnelling method has been applied to local dynamical black hole horizons and the temperature is found to be $\kappa/2\pi$ where $\kappa$ is the *dynamical* surface gravity [@Criscienzo:2007db; @Hayward:2008jq]. Still there are some problems with the method itself and some issues which have not been addressed in this treatment of dynamical horizons. First, the approach depends heavily on the semiclassical approximation and though it is argued that this remains valid near the horizon, it would be better to devise a more general formalism which does not rely on WKB-like approximations. Secondly, in calculating the imaginary part of the semiclassical action $S$ from the Hamilton-Jacobi equation, a singular integral appears with a pole at the horizon. While for the static case the result is standard, for the dynamical horizon it is not clear how the integration is to be performed since the position of the horizon changes in a dynamical process. Lastly, in all these treatments of radiation from dynamical horizons the evolution of the horizon itself is never addressed. In other words, it is not clear how the horizon loses area due to emission of a flux of radiation. The local formalisms of black hole horizon should be able to address these issues. In this paper, a formalism is developed to establish two basic issues. First, that one can associate a temperature to local dynamical horizons without the need for any WKB-like approximation schemes. Second, that there exists a precise relation between the radiation emitted by the horizon and area loss, i.e., flux of outgoing radiation through the horizon in between two partial Cauchy slices exactly equals the difference of radii of the sphere that foliates the horizon at those two instances. We elucidate our arguments as follows. To calculate temperature for local dynamical horizons, we begin by considering the Kodama vector field [@Kodama:1979vn]. For dynamical spacetimes, this vector field provides a preferred timelike direction and is parallel to the Killing vector at spatial infinity which we assume to be flat. We can construct well-behaved positive frequency field modes on both sides of the horizon by considering the Kodama vector field but the outgoing modes exhibit logarithmic singularities on the horizon under some approximation. However, if considered as distribution valued, these modes can be interpreted as horizon crossing and the probability current for these modes remain well defined. The Hawking temperature is determined if one equates the conditional probability, that modes incident on one side is emitted to the other side, to the Boltzmann factor [@Damour:1976jd; @Chatterjee:2012jh], P\_[(emission|incident)]{} = = e\^[-E]{}. Since this method does not depend on the entire evolution of the field modes in the spacetime, it is ideally suited for our purpose. To evaluate the Hawking flux, we recall that there are two well known (and related) definitions of local black hole horizon, the future outward trapping horizon (FOTH) [@Hayward:1993wb; @Hayward:1997jp] and the Dynamical Horizon [@Ashtekar:2002ag; @Ashtekar:2004cn] (or its equilibrium version called the isolated horizon). In these local settings, black hole horizons are a stack of apparent horizons which, under suitable energy conditions, are either null or spacelike. As such, energy flux can only remain on the surface or flow into the horizon. In order that matter fields flow out off such a surface requires that the surface must be timelike in some affine interval. However, to achieve a timelike evolution of the horizon, some energy conditions need to be violated. This is only natural since Hawking radiation necessarily associates, with the thermal emission of particles, a positive flux of energy flowing to infinity (we shall assume that the spacetime is asymptotically flat) and a corresponding flux of negative energy flowing into the black hole (this negative energy flux can also be motivated by the fact that the expectation values of stress energy tensor of quantum fields generically violate energy conditions). In this process the horizon looses area and energy. The plan of the paper is as follows: First, we will discuss the geometrical setup which is based on future outer trapping horizon (FOTH). Next, we show that how the Hawking temperature is proportional to the dynamical surface gravity associated with the Kodama vector. Finally, we will calculate the flux of energy radiated in a dynamical process. We begin with definitions. We follow the conventions of [@Hayward:1997jp]. Consider a four dimensional spacetime ${\mathfs {M}}$ with signature $(-, +,+,+)$. A three-dimensional submanifold $\Delta$ in ${\mathfs {M}}$ is said to be a [*future outer trapping horizon*]{} (FOTH) if [1)]{} It is foliated by a preferred family of topological two-spheres such that, on each leaf $S$, the expansion $\theta_+$ of a null normal $l^{a}_+$ vanishes and the expansion $\theta_-$ of the other null normal $l^{a}_-$ is negative definite, [2)]{} The directional derivative of $\theta_+$ along the null normal $l^{a}_-$ (i.e., $\lie_{l_-}\theta_+$) is negative definite. Thus, $\Delta$ is foliated by marginally trapped two-spheres. According to a theorem due to Hawking, the topology of $S$ is necessarily spherical in order that matter or gravitational flux across $\Delta$ is non-zero. If these fluxes are identically zero then $\Delta$ becomes a Killing or isolated horizon. Even though our arguments will remain local, for definiteness, we choose a spherically symmetric background metric ds\^2=-2e\^[-f]{}dx\^+dx\^-+r\^2(d\^2+\^2d\^2)\[spt\]where both $f$ and $r$ are smooth functions of $x^\pm$. The expansions of the two null normals are $\theta_\pm=(2/r)\,\del_\pm r$ respectively where $\del_\pm=\partial/\del x^{\pm}$. In this coordinate system, the second requirement for FOTH translates to $\del_-\theta_+<0$ on $\Delta$. Let the vector field $t^{a}=l^{a}_+ + h\,l^{a}_-$ be tangential to the FOTH for some smooth function $h$. Then the Raychaudhuri equation for $l^{a}_+$ and the Einstein equation implies \_+\_+=-h\_-\_+=-8 T\_[++]{}.\[rceq\]where $T_{++}=T_{ab}\,l^{a}_+ l^{b}_+$ and $T_{ab}$ is the energy momentum tensor. Several consequences follow from this equation. First, the FOTH is degenerate (or null) if and only if $T_{++}=0$ on $\Delta$. In that case, the FOTH is generated by $l^{a}_+$. Degenerate FOTH is not interesting for Hawking radiation because this implies $\del_+r=0$. As a consequence, the area, $A=4\pi r^2$ of $S$, and the Misner-Sharp energy for this spacetime, given by $E=\half r$, also remains unchanged. Secondly, since $t^2=-2h\,e^{-f}$, a FOTH becomes spacelike if and only if $T_{++}>0$ and is timelike if and only if $T_{++}<0$. For a timelike FOTH, several consequences follow. Here, $\lie_tr <0$, and hence, $\Delta$ is timelike if and only if the area $A$ and the Misner-Sharp energy $E$ decreases along the horizon. This is also expected on general grounds since the horizon receives an incoming flux of negative energy, $T_{++}<0$. As we have emphasized before, in the dynamical spacetime (\[spt\]) the Kodama vector field plays the analog role of the Killing vector. For this spacetime, it is given by K\^[a]{}=e\^f(\_-r) \^a\_+-e\^f(\_+r)\^a\_-.\[kodama\]The surface gravity is defined through $K^a\nabla_{[b} K_{a]}=\kappa\, K_b$ and is $k=-e^f\,\del_-\del_+r$. The FOTH condition $\del_-\theta_+<0$ implies $k>0$. Let us now determine the positive frequency modes of the Kodama vector. It is easy to see that any smooth function of $r$ is a zero-mode of the Kodama vector. Once, a zero-mode is obtained, other positive frequency eigenmodes are evaluated using iKZ\_=Z\_\[modeeq\] Here, $Z_\omega$ are the eigenfunctions corresponding to the positive frequency $\omega$. For simplification, let us introduce new coordinates, $y=x^-$ and $r$ and two new functions, $\bar Z_\omega(y,r)=Z_\omega(x^+,x^-)$ and $G(y,r)=e^f\,(\del_+r)$. As a result, the eigenvalue equation reduces to G\_y|Z\_=i|Z\_.Integrating and transforming back to old coordinates, the above equation gives Z\_=F(r)(i\_r)\[modeeqnew\]where $F(r)$ is an arbitrary smooth function of $r$ and the subscript $r$ under the integral sign denotes that while doing the integration $r$ is kept fixed. To evaluate the integral in , we multiply the numerator and the denominator by $(\del_-\theta_+)$ and use the fact that for any fixed $r$ surface, $e^f\,(\del_-\theta_+)=-2k/r$, (although the strict interpretation of $k$ as the surface gravity holds only for surfaces with $\theta_+ =0$, it exists as a function in any neighbourhood of the horizon). Thus, in some neighbourhood of the horizon we get \_r=-\_rWe now assume (this is the only assumption we make in this calculation) that during the dynamical evolution $k$ is a slowly varying function in some small neighbourhood of the horizon (the zeroth law takes care of it on the horizon, but we also assume it to hold in a small neighbourhood of the horizon). This gives Z\_=F(r) \_+\^[-]{} & \_+&gt;0\ (-|\_+|)\^[-]{} & [for]{} \_+&lt;0. \[umode\]where the spheres are not trapped ‘outside the trapping horizon’ ($\theta_+>0$) and fully trapped ‘inside’ ($\theta_+<0$). These are precisely the modes which are defined outside and inside the dynamical horizon respectively but not on the horizon. Now we have to keep in mind the modes (\[umode\]) are not ordinary functions, but are distribution-valued. Comparing with the spherically symmetric static case [@Chatterjee:2012jh], we find for $\epsilon\to 0^+$ (\_++i)\^= \_+\^& \_+&gt;0\ |\_+|\^e\^[i]{} & \_+&lt;0 \[plusdist\]for the choice $\lambda=-i\omega/k$. The distribution (\[plusdist\]) is well-defined for all values of $\theta_+$ and $\lambda$, and it is differentiable to all orders. The modes $Z^*_\omega$ are given by the complex conjugate distribution. We wish to calculate the probability density in a single particle Hilbert space for positive frequency solutions across the dynamical horizon ()= -=Z\_\^\*Z\_.\[numberd\]A straightforward calculation gives, apart from a positive function of $r$, $$\begin{aligned} \varrho(\omega)&=\omega(\theta_++i\epsilon)^{-\frac{i\omega}{k}} (\theta_+-i\epsilon)^ {\frac{i\omega}{k}}.\nn &=\begin{cases} \omega & \text{for}\;\theta_+>0\\ \omega e^{\frac{2\pi\omega}{k}} & \text{for}\;\theta_+<0.\end{cases}\end{aligned}$$ The conditional probability that a particle emits when it is incident on the horizon from inside is, P\_[(emission|incident)]{} = e\^[-]{} This gives the correct Boltzmann weight with the temperature $k/2\pi$, which is the desired value. We now show that as the horizon evolves, the radius of the 2-sphere foliating the horizon shrinks in precise accordance with the amount of flux radiated by the horizon. To study the flux equation, consider new coordinates, ($x^+,x^-)\mapsto (\theta_+,\tilde x^-)$ where $\tilde x^-=x^-$. On FOTH, $(\del_-\theta_+)/(\del_+\theta_+)$ is equal to $-(\del_-\del_+r)/(4\pi r\,T_{++})$ and negative definite. As a result, the derivatives are related to each other by \_-=\_-+()\_+.\[tilded\]It is not difficult to show that $\tilde\del_-$ is proportional to the tangent vector $t^a$ to the FOTH. Observe that the normal one-form to $\Delta$ must be proportional to $(dr-2\,dE)$, which on the horizon is equal to the one-form (8e\^fr\^2T\_[++]{}-2re\^f\_-\_+r)\_-r dx\^-.In arriving at the above identity we have made use of two Einstein’s equations [@Hayward:1997jp] $$\begin{aligned} \label{Einsteineqn} r\,\del_-\del_+r+\del_+r\,\del_-r+\half e^{-f}&=&4\pi r^2\,T_{-+},\\ \nonumber \del_+^2r+\del_+f\,\del_+r&=&-4\pi r\,T_{++},\end{aligned}$$ and energy equations \_E=2e\^f r\^3(T\_[-+]{}\_-T\_\_).As a result, the normal vector $n^a$ is proportional to \_+-()\_-=\_+-h\_-,so that the tangent vector $t^a=\del^a_++h\del^a_-$, which is clearly proportional to (\[tilded\]). So $\tilde x^-,\theta,\phi$ are natural coordinates on FOTH. The line-element (\[spt\]) induces a line-element on $\Delta$ ds\^2=-2e\^[-f]{}h\^[-1]{}(dx\^-)\^2+r\^2(d\^2+\^2d\^2).Consequently, the volume element on $\Delta$ is given by $d\mu=\sqrt{2e^{-f}h^{-1}}r^2\sin\theta\,d\tilde x^-d\theta d\phi$. We can now calculate the flux of matter energy that crosses the dynamical horizon—it is an integral on a slice of horizon bounded by two spherical sections $S_1$ and $S_2$ F=dT\_[ab]{}n\^a K\^bwhere $\hat n^a$ is the unit normal vector n\^a=(\_+\^a-h\_-\^a)and $K^a$ is the Kodama vector. Using spherical symmetry, eqn. (\[tilded\]) and eqn. , we get $$\begin{aligned} {\mathfs {F}}&=\int d\tilde x^-\;4\pi r^2(\frac{1}{h}T_{++}-T_{+-})e^f\del_-r\nn &=\int d\tilde x^-\;4\pi r^2(\frac{1}{4\pi r}\del_+\del_-r-T_{+-})e^f\del_-r. \end{aligned}$$ \[flux1\] Making use of the Einstein equation (\[Einsteineqn\]) on the horizon and (\[tilded\]), we get $$\begin{aligned} {\mathfs {F}}&=-\int d\tilde x^-\;\frac{1}{2}\del_-r=-\int d\tilde x^-\;\frac{1}{2}\tilde\del_-r\nn &=-\frac{1}{2}(r_2-r_1)\label{flux} \end{aligned}$$ where $r_1,r_2$ are respectively the two radii of $S_1,S_2$. Since the area is decreasing along the horizon, $r_2<r_1$ where $S_2$ lies in the future of $S_1$. As a result, the outgoing flux of matter energy radiated by the dynamical horizon is positive definite (and the ingoing flux of matter energy is negative definite). The flux formula (\[flux\]) differs from that given in [@Ashtekar:2002ag]. Since the Kodama vector field provides a timelike direction and is null on the horizon, it seems more appropriate to use $K^a$ for the dynamical horizon. The derivation of Hawking temperature and the flux law depends on two assumptions. First, that the Kodama vector exists in the spacetime. For spherically symmetric spacetimes, the Kodama vector field exists unambiguously and the Misner-Sharp energy is well defined. For more general spacetimes, a Kodama-like vector field is not known, however, one can still define some mass for such cases that reduces to the Misner-Sharp energy in the spherical limit [@Mukohyama:1999sp]. The second assumption, the existence of a slowly varying $k$ can also be motivated for large black holes. In such cases, the horizon evolves slowly enough so that the surface gravity function should vary slowly in some small neighbourhood of the horizon. Alternatively, we can conclude that the Hawking temperature for a dynamically evolving large black hole is $k/2\pi$ if the dynamical surface gravity is slowly varying in the vicinity of the horizon. The set-up described in this paper can be further developed to model dynamically evaporating black hole horizons through Hawking radiation, analytically as well as numerically. Over the years, several models have been constructed which study radiating black holes, formed in a gravitational collapse, based on the imploding Vaidya metric with a negative energy-momentum tensor, show that a timelike apparent horizon forms due to violation of energy conditions [@Hiscock:1980ze]. However, such models are based on global considerations of event horizons, while local structures like that used in [@Hayward:2005gi] might be useful for a better understanding of Hawking radiation and computations of quantum field theoretic effects (see also [@Pranzetti:2012pd]). It is also interesting to speculate on the extension of the present method for other diffeomorphism invariant theories of gravity. While the zeroth and the first law hold for any arbitrary such theory, the second law has only been proved for a class of such theories [@Chatterjee:2011wj]. If the present formalism can be extended to other theories of gravity, it will lend a support to the existence of the area increase theorem for such theories. While more interesting and deeper issues can only be understood in a full quantum theory of gravity, the present framework can elucidate the suggestions of [@Ashtekar:2005cj] and provide a better understanding of the Hawking radiation process. [100]{} J. M. Bardeen, B. Carter and S. W. Hawking, Commun. Math. Phys.  [**31**]{}, 161 (1973). S. W. Hawking, Commun. Math. Phys.  [**43**]{}, 199 (1975) J. D. Bekenstein, Phys. Rev.  D [**7**]{}, 2333 (1973). R. M. Wald, [*Chicago, USA: Univ. Pr. (1994) 205 p*]{} J. B. Hartle and S. W. Hawking, Phys. Rev.  D [**13**]{}, 2188 (1976). G. W. Gibbons and M. J. Perry, Proc. Roy. Soc. Lond.  A [**358**]{}, 467 (1978). B. S. Kay and R. M. Wald, Phys. Rept.  [**207**]{}, 49 (1991). S. A. Hayward, Phys. Rev.  D [**49**]{}, 6467 (1994). S. A. Hayward, Class. Quant. Grav.  [**15**]{}, 3147 (1998) \[arXiv:gr-qc/9710089\]. A. Ashtekar, C. Beetle, O. Dreyer, S. Fairhurst, B. Krishnan, J. Lewandowski and J. Wisniewski, Phys. Rev. Lett.  [**85**]{}, 3564 (2000) \[arXiv:gr-qc/0006006\]. I. Booth and S. Fairhurst, Phys. Rev. Lett.  [**92**]{}, 011102 (2004) \[arXiv:gr-qc/0307087\]. A. Ashtekar, J. Baez, A. Corichi and K. Krasnov, Phys. Rev. Lett.  [**80**]{}, 904 (1998) \[arXiv:gr-qc/9710007\]. A. Ghosh and A. Perez, Phys. Rev. Lett.  [**107**]{}, 241301 (2011) \[arXiv:1107.1320 \[gr-qc\]\]. M. K. Parikh and F. Wilczek, Phys. Rev. Lett. [**85**]{}, 5042-5 (2000). S. Shankaranarayanan, T. Padmanabhan and K. Srinivasan, Class. Quantum  Grav. [**19**]{}, 2671-88 (2002). M. Angheben, M. Nadalini, L. Vanzo and S. Zerbini, J. High Energy Phys. [**0505**]{} 014 (2005). R. Kerner and R. B. Mann, Phys. Rev. D [**73**]{} 104010 (2006). B. Chatterjee, A. Ghosh and P. Mitra, Phys. Lett. B [**661**]{} 307-11 (2008). L. Vanzo, G. Acquaviva and R. Di. Criscienzo Class. Quantum  Grav. [**28**]{} 183001 (2011). R. Di. Criscienzo, M. Nadalini, L. Vanzo, S. Zerbini and G. Zoccatelli, Phys. Lett. B [**657**]{} 107-11 (2007). S. A. Hayward, R. Di Criscienzo, L. Vanzo, M. Nadalini and S. Zerbini, Class. Quant. Grav.  [**26**]{}, 062001 (2009) \[arXiv:0806.0014 \[gr-qc\]\]. H. Kodama, Prog. Theor. Phys.  [**63**]{}, 1217 (1980). T. Damour and R. Ruffini, Phys. Rev.  D [**14**]{}, 332 (1976). B. Chatterjee and A. Ghosh, arXiv:1201.4017 \[gr-qc\]. A. Ashtekar and B. Krishnan, Phys. Rev. Lett.  [**89**]{}, 261101 (2002) \[arXiv:gr-qc/0207080\]; Phys. Rev.  D [**68**]{}, 104030 (2003) \[arXiv:gr-qc/0308033\]. A. Ashtekar and B. Krishnan, Living Rev. Rel.  [**7**]{}, 10 (2004) \[arXiv:gr-qc/0407042\]. S. Mukohyama and S. A. Hayward, Class. Quant. Grav.  [**17**]{}, 2153 (2000) \[arXiv:gr-qc/9905085\]. W. A. Hiscock, Phys. Rev.  D [**23**]{}, 2813 (1981); R. Balbinot, R. Bergamini and B. Giorgini, Nuovo Cim.  B [**70**]{} (1982) 201. S. A. Hayward, Phys. Rev. Lett.  [**96**]{}, 031103 (2006). D. Pranzetti, arXiv:1204.0702 \[gr-qc\]. A. Chatterjee and S. Sarkar, Phys. Rev. Lett.  [**108**]{}, 091301 (2012) \[arXiv:1111.3021 \[gr-qc\]\]; S. Kolekar, T. Padmanabhan and S. Sarkar, arXiv:1201.2947 \[gr-qc\]. A. Ashtekar and M. Bojowald, Class. Quant. Grav.  [**22**]{}, 3349 (2005) \[arXiv:gr-qc/0504029\]; A. Ashtekar, V. Taveras and M. Varadarajan, Phys. Rev. Lett.  [**100**]{}, 211302 (2008) \[arXiv:0801.1811 \[gr-qc\]\].

--- abstract: 'We prove sufficient conditions for Hausdorff convergence of the spectra of sequences of closed operators defined on varying Hilbert spaces and converging in norm-resolvent sense, i.e. $\|J_\eps(1+A_\eps)^{-1} - (1+A)^{-1}J_\eps\|\to 0$ as $\eps\to 0$, where $J_\eps$ is a suitable identification operator between the domains of the operators. This is an extension of results of [@MNP], who proved absence of spectral pollution for sectorial operators.' author: - 'F. Rösler[^1]' title: A Note on Spectral Convergence in Varying Hilbert Spaces --- [***Keywords:*** Operator Theory; Spectral Approximation; Norm-Resolvent convergence]{} Introduction ============ Convergence of spectra of sequences of operators has long been a subject of intense interest in asymptotic analysis. Many techniques, such as homogenisation, or dimensional reduction rely on convergence theorems for the spectra of sequences of operators. Classical theorems from perturbation theory (cf [@Kato]) give partial answers to the question of spectral convergence under different assumptions on the sequence. One central classical result is that under norm-resolvent convergence, so-called *spectral pollution* is not possible, i.e. if $A_n$ converges to $A$ in norm resolvent sense, then the “limit” of the spectra of $A_n$ cannot be larger than the spectrum of $A$. More precisely, one has \[th:classical\] If $A_n,A$ are closed operators on a Hilbert space $\h$ and there exists $\mu\in\rho(A)$ such that for $n$ large enough one has $\mu\in\rho(A_n)$ and $\|(\mu-A_n)^{-1}-(\mu-A)^{-1}\|\to 0$, as $n\to\infty$, then for any $\lambda\in\rho(A)$ there exists $n_0\in\N$ such that $\lambda\in\rho(A_n)$ for all $n>n_0$. However, the converse of Theorem \[th:classical\] is *not* true, meaning that there exist sequences of operators for which the spectrum of the spectrum of the limit operator $A$ is much larger than the spectrum of any $A_n$ for finite $n$. Indeed, let us demonstrate this with an example (cf. [@Kato Ch. IV.3.1]). #### Example. Let $\h=l^2(\mathbb{Z})$ and let $\{e_n\}$ be its canonical basis. For $n\in\N$ define $T_n\in L(\h)$ by $$\begin{aligned} T_ne_0 &:= n^{-1} e_{-1}\\ T_ne_i &:= e_{i-1},\qquad i\neq 0.\end{aligned}$$ A straightforward calculation shows that $T_n-\lambda$ is boundedly invertible for every $\lambda\in\C$ with $|\lambda|<1$. Since also clearly $\|T_n\|\leq 1$ for all $n$, we conclude that $\sigma(T_n)\subset S^1$, the unit circle in $\C$. Now consider the limit of $(T_n)$. By definition, $T_n$ is a rank-one perturbation of the operator $T$ defined by $$\begin{aligned} T_ne_0 &:= 0\\ T_ne_i &:= e_{i-1},\qquad i\neq 0.\end{aligned}$$ It follows that $\|T_n-T\|_{L(\h)} = \f1n$ and hence $T_n$ converges to $T$ in operator norm. However, the spectrum of $T$ is considerably larger than the unit circle $S^1$. Indeed, another straightforward calculation shows that for every $\lambda\in\C$ with $|\lambda|<1$ the vector $x:=\sum_{n=0}^\infty \lambda^ne_n$ is an eigenvector of $T$ and thus $\sigma(T)$ contains the closed unit disk. The above example shows that even under operator norm convergence, *spectral inclusion* can fail, i.e. there may exist points $\lambda\in\sigma(A)$ such that there exists no sequence $\lambda_n\in\sigma(A_n)$ with $\lambda_n\to\lambda$. There exist other examples demonstrating this lower semi-discontinuity of the spectrum (cf. [@Rickart p. 282], [@aupetit Ch.1, §5] for an example in which the spectrum collapses from a disk to a point). Therefore, further assumptions are necessary in order to obtain spectral inclusion. In the next section, we will set the scene and present our main results. The following sections contain the proofs of these results. Main Results ============ In order to aim for generality, we consider families of operators $A_\eps$ which are not necessarily defined on the same Hilbert space. More precisely, for $\eps>0$ let $\h,\h_\eps$ be Hilbert spaces and let $A_\eps:\h_\eps\supset\dom(A_\eps)\to\h_\eps$ and $A:\h\supset\dom(A)\to\h$ be closed operators. Let us denote $\V_\eps:=\big(\dom(A_\eps),\|\cdot\|_{A_\eps}\big)$ and $\V:=\big(\dom(A),\|\cdot\|_{A}\big)$, where $\|\cdot\|_A$ denotes the graph norm of $A$, that is, $\|u\|_{\V}^2 := \|u\|_A := \|u\|_{\h}+\|Au\|_{\h}$ (analogously for $\|\cdot\|_{\V_\eps}$). \[def:extended\_norm-resolvent\] Assume that there exists $z_0\in\bigcap_{\eps>0}\rho(A_\eps)\cap\rho(A)$ and operators $J_\eps: \h_\eps\to \h$ and $ I_\eps:\h\to\h_\eps$ such that $\|I_\eps J_\eps - \id_{\h_\eps}\|_{L(\V_\eps,\h_\eps)}\to 0$ as $\eps\to 0$,\[(i)old\] $J_\eps I_\eps \to \id_{\h}$ strongly as $\eps\to 0$,\[(ii)new\] $\|I_\eps\|_{ L(\h,\h_\eps)},\|J_\eps\|_{ L(\h_\eps,\h)}\leq M$ for some $M>0$ uniformly in $\eps$,\[(ii)old\] $\bigl\|J_\eps(z_0\,\id_{\h_\eps}-A_\eps)^{-1} -(z_0\,\id_{\h}-A)^{-1}J_\eps\bigr\|_{ L(\h_\eps,\h)}\to 0$ as $\eps\to 0$.\[(iii)old\] Then we say that the sequence $(A_\eps)$ converges to $A$ in the *extended norm resolvent sense*. Note the asymmetry between the assumptions (i) and (ii) above, where we require convergence with respect to the operator norm $\|\cdot \|_{L(\V_\eps,\h_\eps)}$ and only *strong* convergence for $J_\eps I_\eps$, allowing a great deal of freedom for the construction of $I_\eps,J_\eps$ in applications. Moreover, note that if $\h_\eps\equiv \h$ for all $\eps>0$ and $I_\eps=J_\eps=\id_{\h}$ for all $\eps>0$, this definition reduces to the classical definition of norm resolvent convergence. The assumptions on the identification operators $I_\eps,J_\eps$ cover many cases relevant in applications. Examples include *Projection onto subspaces.* Let $\h_n$ be an increasing sequence of closed subspaces such that the orthogonal projection $P_n:\h\to\h_n$ converges strongly to the identity. Then defining $I_n:=P_n$ and $J_n:\h_n\hookrightarrow\h,\, J_n(x)=x$ satisfy assumptions (i)-(iii) of Definition \[def:extended\_norm-resolvent\]. Indeed, it is easy to check that $I_n J_n=\id_{\h_n}$, while the strong convergence of $J_\eps I_\eps$ follows from the strong convergence $P_n\to\id_\h$. *Perforated domains.* Let $\Om\subset\R^d$ be open and $T_\eps\subset\Om$ be a collection of closed subsets such that $|T_\eps|\to 0$ as $\eps\to 0$, where $|\cdot|$ denotes the Lebesgue measure. Let $\h:=L^2(\Om), \h_\eps:=L^2(\Om\setminus T_\eps)$ and $\V:=H^1(\Om)$. Define $I_\eps,J_\eps$ by $$\begin{aligned} J_\eps &: \h_\eps \to \h, &J_\eps f(x) &= \begin{cases} f(x), & x\in\Om\setminus T_\eps, \\[0.1em] 0, & x\in T_\eps \end{cases} \\[1mm] I_\eps &: \h \to \h_\eps, &I_\eps g(x) &= g|_{\Om\setminus T_\eps} \end{aligned}$$ In this case we have again that $I_\eps J_\eps = \mathrm{id}_{\h_\eps}$ and $ \|J_\eps I_\eps - \mathrm{id}_{\h}\|_{ L(\V,\h)} \to 0$. Indeed, the first equality is trivial, while the second follows by the following calculation. Let $f\in \V$. Then we have $\|f-J_\eps I_\eps f\|_{L^2(\Om)} = \|f\|_{L^2(T_\eps)}$. To show that this quantity converges to 0 uniformly in $f$, note that $$\begin{aligned} \|f\|_{L^2(T_\eps)}^2 &\leq \left\|1\right\|_{L^{2p}( T_\eps)}^2 \left\|f\right\|_{L^{2q}(T_\eps)}^2 \end{aligned}$$ for $p,q>1$ with $p^{-1}+q^{-1}=1$, by Hölder’s inequality. Since $f\in H^1(\Om)$, we can use the Gagliardo-Sobolev-Nierenberg inequality to conclude (for suitable $q$) that $$\begin{aligned} \|f\|_{L^2(T_\eps)}^2 &\leq \left\|1\right\|_{L^{2p}(T_\eps)}^2 C \left\|f\right\|_{H^1(\Om)}^2\\ &= C |T_\eps|^{\nicefrac 1 p} \left\|f\right\|_{H^1(\Om)}^2 \end{aligned}$$ with some suitable $p>0$. Since $|T_\eps|\to 0$ as $\eps\to 0$, the desired convergence follows. By density of $\V$ in $\h$ we readily conclude the strong convergence $J_\eps I_\eps\to\id_\h$. Indeed, the main result of this paper complements the proof of spectral convergence in perforated domains in [@CDR], which was only sketched there. *Dimensional reduction.* Consider a domain $\Om_\eps\subset\R^{d+1}$ of the form $\Om_\eps = (\eps U)\times(0,1)$, where $U\subset\R^d$ is open and bounded. Define $\h_\eps:=L^2(\Om_\eps),\;\h:=L^2((0,1))$ and $\V_\eps:=H^1(\Om_\eps)$. For $f\in\h$, define $(I_\eps f)(x,t):=f(t)$ and for $u\in\h_\eps$ define $(J_\eps u)(t):=|\eps U|^{-d}\int_{\eps U}u(x,t)\,dx$, which is well-defined for almost every $t\in(0,1)$ by Fubini’s theorem. This time, it is easily checked that $J_\eps I_\eps=\id_\h$. Moreover, one has $$\begin{aligned} (u-I_\eps J_\eps u)(x,t) &= u(x,t) - |\eps U|^{-d}\int_{\eps U}u(y,t)\,dy\\ &= |\eps U|^{-d}\int_{\eps U}(u(x,t) - u(y,t))\,dy\\ \end{aligned}$$ and hence $$\begin{aligned} \|(1-I_\eps J_\eps)u\|_{L^2(\Om_\eps)}^2 &=\int_{\eps U}\int_0^1\left||\eps U|^{-d}\int_{\eps U}(u(x,t) - u(y,t))\,dy\right|^2 dx\,dt\\ &\leq \int_{\eps U}\int_0^1|\eps U|^{-d}\int_{\eps U}|u(x,t) - u(y,t)|^2\,dy\,dx\,dt\\ &= |\eps\operatorname{diam}(U)|^2\int_{\eps U}\int_0^1|\eps U|^{-d} \left\|\nabla u(\cdot,t)\right\|^2_{L^2(\eps U)}\,dx\,dt\\ &= \eps^2\operatorname{diam}(U)^2 \left\|\nabla u\right\|^2_{L^2(\Om_\eps)}, \end{aligned}$$ where we have used Jensen’s inequality. The above inequality shows that we have $\|\id_{\h_\eps}-I_\eps J_\eps\|_{L(\V_\eps,\h_\eps)}\leq C\eps$. The main result of this article is the following. \[th:mainth\] Let $A_\eps,A$ be closed operators on $\h_\eps,\h$ respectively, and assume that $A_\eps$ converges to $A$ in extended norm resolvent sense. Then one has If $\rho(A)$ is connected, then for every compact $K\subset\rho(A)$ there exists $\eps_0>0$ such that $K\subset\rho(A_\eps)$ for all $\eps\in(0,\eps_0)$. Assume that $K\subset\C$ is compact, connected such that $K\subset\rho(A_{\eps})$ and $\|(z-A_\eps)^{-1}\|_{L(\h_\eps)}\leq C$ for all $\eps>0,\,z\in K$. Assume further that $K$ can be connected to $\{z_0\}$ by a curve $\gamma$ lying in $\rho(A_{\eps})$ for all $\eps>0$. Then one has $K\subset\rho(A)$. For every isolated point $\lambda\in\sigma(A)$ such that $B_\delta(\lambda)\setminus\{\lambda\}$ is in the same connected component of $\rho(A)$ as $z_0$ there exists a sequence $\lambda_\eps\in\sigma(A_\eps)$ such that $\lambda_\eps\to\lambda$. The first part of the above theorem shows the absence of spectral pollution, while the second part shows spectral inclusion under the additional assumption that $\|(z-A_\eps)^{-1}\|_{L(\h_\eps)}$ be uniformly bounded. We remark that a similar statement to part (i) in the above theorem has already been proven in [@MNP; @Post2006]. Our result is an extension of theirs in three ways. First, they considered only *sectorial* operators, which can be defined via a sesquilinear form, whereas we treat general closed operators. Second, our assumptions on the identification operators $I_\eps,\,J_\eps$ are less restrictive. Third, the spectral inclusion results (ii) and (iii) are not at all considered in [@MNP]. Furthermore, convergence of spectra and pseudospectra of operators on varying spaces have been studied in [@B16; @B18; @BoegliSiegl; @Hansen11] in the situation where all spaces $\h,\h_\eps$ are subspaces of a common “large” Hilbert space $\h_0$ and $I_\eps$ plays the role of a projection operator. In this situation, an analogue of Theorem \[th:mainth\] has been shown in [@B16]. Note that we do not assume any connection between the spaces $\h$ and $\h_\eps$ besides that introduced by Definition \[def:extended\_norm-resolvent\]. From Theorem \[th:mainth\] we immediately recover two classical results about spectral convergence. \[cor:classical1\] If $A_\eps,A$ are selfadjoint and bounded below for almost all $\eps>0$ and $A_\eps\to A$ in extended norm resolvent sense, then for all bounded open $V\subset\mathbb C,$ one has $\sigma(A_\eps )\cap V\rightarrow\sigma(A )\cap V$ in Hausdorff sense. By selfadjointness and boundedness from below of the operators involved, we have that $\rho(A)$ is connected. For given $r>0$, let $\delta>0$ and define the compact set $K:=\overline{B_r(0)}\setminus U_\delta(\sigma(A ))$, where $U_\delta(\cdot)$ denotes the open $\delta$-neighbourhood. By (i) we have that $K\subset \rho(A_\eps )$ for $\eps$ small enough. This shows that $\overline{B_r(0)\cap\sigma(A_\eps )}\subset \overline{B_r(0)\cap U_\delta(\sigma(A ))}$. To see the converse inclusion $B_r(0)\cap \sigma(A )\subset B_r(0)\cap U_\delta(\sigma(A _\eps))$, let us argue by contradiction and suppose that there exists $\delta_0>0$ such that $$\overline{B_r(0)\cap\sigma(A )}\nsubseteq \overline{B_r(0)\cap U_{\delta_0}(\sigma(A _\eps))}\qquad\forall\eps>0.$$ By this assumption, there exists a sequence $(\lambda_\eps)$ in $\overline{B_r(0)\cap\sigma(A)}$ such that $\dist(\lambda_\eps,\sigma(A_\eps))\ge\delta_0$ for all $\eps>0$. Since $(\lambda_\eps)$ is bounded, we can extract a subsequence $\lambda_{\eps'}\to\lambda_0\in \overline{B_r(0)}\cap\sigma(A)$. It follows that $$\overline{B_{\nicefrac{\delta_0}{2}}(\lambda_0)}\subset\rho(A _{\eps'})\qquad \text{for almost all }\eps'>0.$$ Since for all $\eps>0$ we have $\sigma(A_\eps)\subset[0,\infty)$, we can connect $B_{\nicefrac{\delta_0}{2}} $ to $\{z_0\}$ by a curve lying in $B_r(0) \cap\rho(A_\eps)$ and use Theorem \[th:mainth\] (ii) to conclude that $\overline{B_{\nicefrac{\delta_0}{2}}(\lambda_0)}\subset\rho(A )$, which contradicts the fact that $\lambda_0\in\sigma(A )$. Since $\delta>0$ was arbitrary, the desired Hausdorff convergence follows. \[cor:classical2\] If $A_\eps\to A$ in extended norm resolvent sense and $(z_0-A)^{-1},\,(z_0-A_\eps)^{-1}$ are compact for all $\eps>0$ then for all bounded open $V\subset\mathbb C,$ one has $\sigma(A_\eps )\cap V\rightarrow\sigma(A )\cap V$ in Hausdorff sense. Compactness of the resolvent implies that for all $\eps>0$, $\rho(A_\eps),\,\rho(A)$ are connected and the spectra of $A_\eps,\,A$ consist of isolated points only. Hence the assertion follows from (i), (iii) of Theorem \[th:mainth\]. Classical proofs of the statements in Corollaries \[cor:classical1\] \[cor:classical2\], in the situation where $\h_\eps\equiv\h$ for all $\eps>0$ can be found in [@RS1; @Kato]. Proof of Theorem \[th:mainth\] ============================== In this section we will prove Theorem \[th:mainth\]. Although the main ideas in the proof of the first part (i) are the same as in [@MNP], we repeat the reasoning here to account for our differing notation and our more general hypotheses. #### Proof of (i). By assumption we have $z_0\in\rho(A_\eps)$ for all $\eps>0$ and $z_0\in\rho(A)$ and the operator norms $\bigl\|(z_0-A_\eps)^{-1}\bigr\|_{L(\h_\eps,\V_\eps)}$ are finite. Indeed, we have even more: \[lem:HVeps&lt;HHeps\] For $z\in\rho(A_\eps)$ one has $$\begin{aligned} \label{eq:(3.10)_Post} \bigl\|(z-A_\eps)^{-1}\bigr\|_{L(\h_\eps,\V_\eps)} \leq 1 + (1+|z|) \bigl\|(z-A_\eps)^{-1}\bigr\|_{L(\h_\eps)}. \end{aligned}$$ Let $z\in\rho(A_\eps)$. Then $$\begin{aligned} \bigl\|(z-A_\eps)^{-1}u\bigr\|_{\V_\eps} &= \|(z-A_\eps)^{-1}u\|_{\h_\eps}+\|A_\eps(z-A_\eps)^{-1}u\|_{\h_\eps}\\ &= \|(z-A_\eps)^{-1}u\|_{\h_\eps}+\|u-z(z-A_\eps)^{-1}u\|_{\h_\eps}\\ &\leq \|u\|_{\h_\eps} + (1+|z|)\|(z-A_\eps)^{-1}u\|_{\h_\eps}\end{aligned}$$ The next lemma is technical, but central to the argument. It shows that if $\eps$ is small and $\l\|\res{z,A}\r\|_{ L(\h)}$ is uniformly bounded, then $\l\|\res{z,A_\eps}\r\|_{ L(\h_\eps)}$ is uniformly bounded. \[lemma:R&lt;l=&gt;R\_eps&lt;L\] For every $l,r>0$ there exist $\delta>0$ and $L>0$ such that if $\bigl\|J_\eps(z_0-A_\eps)^{-1} -(z_0-A)^{-1}J_\eps\bigr\|_{ L(\h_\eps,\h)} < \delta, $ $\l\|\res{z,A}\r\|_{ L(\h)}\leq l$, $\|\id_{\h_\eps}-I_\eps J_\eps\|_{L(\V_\eps,\h_\eps)}<\f{1}{2(|z_0|+r)}$, $z \in \rho(A_\eps)\cap\rho(A)\cap B_r(0)$ then $\l\|\res{z,A_\eps}\r\|_{ L(\h_\eps)}\leq L$. The useful point in this lemma is that $L$ does not depend on $z$ as long as $z\in\rho(A_\eps)\cap\rho(A)\cap B_r(0)$ and $\l\|\res{z,A}\r\|_{ L(\h)}\leq l$. We use the shorthand notation $R_\eps(z):=(z-A_\eps)^{-1}$ and $R(z):=(z-A)^{-1}$. For $z\in\rho(A_\eps)\cap\rho(A)\cap B_r(0)$ define $$\begin{aligned} V(z) := J_\eps R_\eps(z) - R(z)J_\eps. \end{aligned}$$ The resolvent identity can be used to show that $$\begin{aligned} \bigl( R(z_0) - R(z) \bigr)J_\eps R_\eps(z)R_\eps(z_0) = R(z)R(z_0)J_\eps \bigl( R_\eps(z_0) - R_\eps(z) \bigr) \end{aligned}$$ which implies $$\begin{aligned} R(z_0)V(z)R_\eps(z_0) = R(z)V(z_0)R_\eps(z) \end{aligned}$$ or $$\begin{aligned} \label{eq:V(z)} V(z) &= (z_0-A)R(z)V(z_0)R_\eps(z)(z_0-A_\eps) \\ &= \bigl(\id_{\h} - (z-z_0)R(z)\bigr)V(z_0)\bigl(\id_{\h_\eps} - (z-z_0)R_\eps(z)\bigr) \end{aligned}$$ on $\dom(A)$ and thus on $\h_\eps$ by density. Using our assumptions we deduce that $$\begin{aligned} \label{eq:||V(z)||} \l\|V(z)\r\|_{ L(\h_\eps,\h)} &\leq \delta\bigl( 1+|z-z_0|\|R_\eps(z)\|_{ L(\h_\eps)} \bigr)\bigl( 1+|z-z_0|l\bigr). \end{aligned}$$ Now, decompose $R_\eps(z)$ as $$\begin{aligned} \label{eq:R_eps_representation} R_\eps(z) &= I_\eps(J_\eps R_\eps(z) - R(z)J_\eps) + I_\eps R(z)J_\eps + (\id_{\h_\eps}-I_\eps J_\eps)R_\eps(z) \end{aligned}$$ This representation, together with shows that $$\begin{aligned} \|R_\eps(z)\|_{ L(\h_\eps)} &\leq \|I_\eps\|_{ L(\h,\h_\eps)} \l\|V(z)\r\|_{ L(\h_\eps,\h)} + \|I_\eps\|_{ L(\h,\h_\eps)}\|J_\eps\|_{ L(\h_\eps,\h)} \|R(z)\|_{ L(\h)}\\ &\qquad + \|\id_{\h_\eps}-I_\eps J_\eps\|_{L(\V_\eps,\h_\eps)}\|R_\eps(z)\|_{L(\h_\eps,\V_\eps)}\\ &\leq \delta M \bigl( 1+|z-z_0|\|R_\eps(z)\|_{ L(\h_\eps)} \bigr)\bigl( 1+|z-z_0|l\bigr) + M^2 l \\ &\qquad + \f{1}{2(|z_0|+r)}\|R_\eps(z)\|_{L(\h_\eps,\V_\eps)} \end{aligned}$$ To estimate the last term on the right hand side we apply Lemma \[lem:HVeps&lt;HHeps\] to obtain $$\begin{aligned} \|R_\eps(z)\|_{ L(\h_\eps)} &\leq \delta M( 1+|z-z_0|l)|z-z_0|\|R_\eps(z)\|_{ L(\h_\eps)} + \delta M( 1+|z-z_0|l) + M^2l \\ &\qquad + \f{1}{2(|z_0|+r)} \big( 1+(1+|z|)\|R_\eps(z)\|_{L(\h_\eps)} \big)\\ &\leq \|R_\eps(z)\|_{L(\h_\eps)}\left[ \delta M( 1+(|z_0|+r)l)(|z_0|+r)+ \f{1}{2(|z_0|+r)}(|z_0|+r) \right] \\ &\qquad + \delta M( 1+(|z_0|+r)l) + M^2l + \f{1}{2(|z_0|+r)} \end{aligned}$$ Thus, if we choose $\delta<\f{1}{4M( 1+(|z_0|+r)l)(|z_0|+r)}$, we obtain the estimate $$\begin{aligned} \|R_\eps(z)\|_{ L(\h_\eps)} &\leq \|R_\eps(z)\|_{L(\h_\eps)}\left[ \f14 + \f12 \right] + \delta M( 1+(|z_0|+r)l) + M^2l + \f{1}{2(|z_0|+r)} \end{aligned}$$ and hence $$\begin{aligned} \|R_\eps(z)\|_{ L(\h_\eps)} &\leq 4\left(\delta M( 1+(|z_0|+r)l) + M^2l + \f{1}{2(|z_0|+r)}\right) \\ &= 4M^2l + \f{3}{2(|z_0|+r)} \\ &=:L \end{aligned}$$ uniformly for $z\in\rho(A_\eps)\cap\rho(A)\cap B_r(0)$. \[pr:Generalised\_spectral\_convergence1\] Let $A_\eps:\h_\eps\supset\dom(A_\eps)\to\h_\eps$ converge to $A:\h\supset\dom(A)\to\h$ in extended norm resolvent sense. Then for every compact, connected $K\subset\rho(A)$ such that $K\cap\rho(A_\eps)\neq\emptyset$ for $\eps$ small enough there exists $\eps_0>0$ such that $K\subset\rho(A_\eps)$ for all $\eps\in(0,\eps_0)$. We use the notation from the previous proof. Let $K\subset\rho(A)$ be compact and choose $r>0$ such that $K\subset B_r(0)$. Denote $$\begin{aligned} l:=\sup_{z\in K}\|R(z)\|_{L(\h_\eps)}<\infty \end{aligned}$$ and choose $\delta>0$ as in Lemma \[lemma:R&lt;l=&gt;R\_eps&lt;L\] and $\eps_0>0$ such that $\bigl\|J_\eps(z_0-A_\eps)^{-1} - (z_0-A)^{-1}J_\eps\bigr\|_{L(\h_\eps,\h)}<\delta$ and $\|\id_{\h_\eps}-I_\eps J_\eps\|_{L(\V_\eps,\h_\eps)}<\f{1}{2(|z_0|+r)}$ for all $\eps\in(0,\eps_0)$, which is possible by norm resolvent convergence. Let $K_\eps:=\rho(A_\eps)\cap K$, which is non-empty by assumption and by definition relatively open in $K$. We will show that $K_\eps$ is also relatively closed in $K$ which by connectedness of $K$ implies $K_\eps=K$. To this end, let $(z_n)$ be a sequence in $K_\eps$ converging to $z\in K$. By Lemma \[lemma:R&lt;l=&gt;R\_eps&lt;L\], the sequence $\big(\|R_\eps(z_n)\|_{L(\h_\eps)}\big)_{n\in\N}$ is bounded and hence $z\in\rho(A_\eps)$. Hence, $K_\eps$ is closed in $K$ and the proof is completed. Proposition \[pr:Generalised\_spectral\_convergence1\] is almost what we want. It only remains to remove the assumptions that $K$ be connected and that $K\cap\rho(A_\eps)\neq\emptyset$. This will be done in the following. #### *Conclusion of Part (i).* Let $K\subset\rho(A)$ be compact. We decompose $K$ into its connected components $K=\bigcup_{i\in I}K_i$, where $I$ is some appropriate index set. Next, choose for each $i\in I$ a connected, open, bounded set $U_i$ such that $\overline{K_i}\subset U_i\subset\rho(A)$. Then for each $i\in I$, the set $\overline {U_i}$ is connected, compact and contained in $\rho(A)$. Next, choose a curve $\gamma$ in $\rho(A)$ that connects $\overline {U_i}$ with $\{z_0\}$. Then the set $K':=\overline{U_i}\cup\gamma$ is compact, connected and contained in $\rho(A)$. Moreover, since $z_0\in K'$, one has $\rho(A_\eps)\cap K'\neq\emptyset$ for all $\eps>0$ and applying Proposition \[pr:Generalised\_spectral\_convergence1\] we conclude that there exists $\eps_i>0$ such that $\overline{U_i}\cup\gamma\subset \rho(A_{\eps})$ for all $\eps\in(0,\eps_i)$. Since $i\in I$ was arbitrary, such an estimate exists for every $i$. But since $K$ is compact and the $U_i$ form an open covering of $K$, there exists a finite subset $F\subset I$ such that $K\subset \bigcup_{i\in F}U_i$. It follows immediately that $K\subset \rho(A_{\eps})$ for all $\eps$ smaller than $\min\{\eps_i\,|\,i\in F\}$. #### Proof of (ii). The proof of part (ii) of Theorem \[th:mainth\] is similar to the previous one, but has some crucial differences that we will highlight in due course. We start with an analogue of Lemma \[lemma:R&lt;l=&gt;R\_eps&lt;L\]. \[lemma:R\_eps&lt;l=&gt;R&lt;L\] For every $l,r>0$, there exists $L>0$ such that if $\bigl\|J_\eps(z_0-A_\eps)^{-1} -(z_0-A)^{-1}J_\eps\bigr\|_{ L(\h_\eps,\h)} \to 0$ as $\eps\to 0$, $K\subset\C$ compact with $K\subset \rho(A_\eps)\cap\rho(A)\cap B_r(0)$ for almost all $\eps>0$, $\l\|\res{z,A_\eps}\r\|_{ L(\h_\eps)}\leq l$ for all $z \in K$, then one has $\l\|\res{z,A}\r\|_{ L(\h)}\leq L$ for all $z \in K$. Defining $V(z)$ as in the proof of Lemma \[lemma:R&lt;l=&gt;R\_eps&lt;L\], the same reasoning leads to the estimate $$\begin{aligned} \l\|V(z)\r\|_{L(\h_\eps,\h)} &\leq \bigl\|J_\eps R_\eps(z_0) - R(z_0)J_\eps\bigr\|_{ L(\h_\eps,\h)} \bigl( 1+|z-z_0|\|R(z)\|_{L(\h)} \bigr)\bigl( 1+|z-z_0|l\bigr). \end{aligned}$$ for $z\in K$. Now we decompose $R(z)$ as $$\begin{aligned} R(z) = (R(z)J_\eps - J_\eps R_\eps(z))I_\eps + J_\eps R_\eps(z) I_\eps + R(z)(\id_{\h}-J_\eps I_\eps) \end{aligned}$$ For $u\in\h$ with $\|u\|_{\h}=1$ we immediately obtain $$\begin{aligned} \|R(z)u\|_{L(\h)} &\leq \|I_\eps\|_{L(\h,\h_\eps)}\|V(z)\|_{L(\h_\eps,\h)} + \|R(z)(u-J_\eps I_\eps u)\|_\h \\ &\qquad + \|I_\eps\|_{L(\h,\h_\eps)}\|J_\eps\|_{L(\h_\eps,\h)}\|R_\eps(z)\|_{L(\h_\eps)}\\ &\leq M \bigl\|J_\eps R_\eps(z_0) - R(z_0)J_\eps\bigr\|_{ L(\h_\eps,\h)} \bigl( 1+|z-z_0|\|R(z)\|_{ L(\h)} \bigr)\bigl( 1+|z-z_0|l\bigr)\\ &\qquad + M^2l + \|R(z)\|_{L(\h)} \|(u-J_\eps I_\eps u)\|_\h \end{aligned}$$ Next, choose $\eps=\eps(u)$ so small that $\|u-J_{\eps(u)} I_{\eps(u)} u\|_\h<\f12$. Without loss of generality we can assume that $\bigl\|J_{\eps(u)} R_{\eps(u)}(z_0) - R(z_0)J_{\eps(u)}\bigr\|_{ L(\h_{\eps(u)},\h)}$ is smaller than $\delta:=\f{1}{4M( 1+(|z_0|+r)l)(|z_0|+r)}$. We obtain $$\begin{aligned} \f12 \|R(z)u\|_{L(\h)} &\leq M \delta \bigl( 1+|z-z_0|\|R(z)\|_{ L(\h)} \bigr)\bigl( 1+|z-z_0|l\bigr) + M^2l \\ &\leq M \delta \bigl( 1+(|z_0|+r)\|R(z)\|_{ L(\h)} \bigr)\bigl( 1+(|z_0|+r)l\bigr) + M^2l\\ &= \delta M \bigl( 1+(|z_0|+r)l\bigr)(|z_0|+r)\|R(z)\|_{ L(\h)} + M\delta\bigl( 1+(|z_0|+r)l\bigr) + M^2l\\ &= \f14\|R(z)\|_{ L(\h)} + \f{1}{4(|z_0|+r)} + M^2l, \end{aligned}$$ where the right hand side does not depend on $u$. Taking the supremum over all $u$ with $\|u\|_\h=1$ we conclude that $$\begin{aligned} \|R(z)\|_{L(\h)} &\leq \f{1}{4(|z_0|+r)} + M^2l\\ &=:L \end{aligned}$$ uniformly for $z\in K$. \[pr:Generalised\_spectral\_convergence2\] Let $A_\eps\to A$ in extended norm resolvent sense, and assume that $K\subset\C$ is connected, compact such that $K\subset\rho(A_\eps)$, $\|(z-A_\eps)^{-1}\|_{L(\h_\eps)}$ uniformly bounded in $z\in K$ and $\eps>0$ and $K\cap\rho(A)\neq\emptyset$ and there exists $r>0$ such that $K\subset B_r(0)$ for almost all $\eps>0$. Then one has $K\subset\rho(A)$. Let $K\subset\C$ be compact, connected such that $K\subset\rho(A_\eps)$ and $K\cap\rho(A)\neq\emptyset$ for almost all $\eps>0$. Choose $l>0$ such that $\|(z-A_\eps)^{-1}\|_{L(\h_\eps)}\leq l$ for all $z\in K,\,\eps>0$. Choose $r>0$ such that $K\subset B_r(0)$ for almost all $\eps>0$. By assumption, we have $K':=K\cap\rho(A)\neq\emptyset$ and $K'$ is relatively open in $K$. We will show that $K'$ is also relatively closed in $K$. Let $(z_n)\subset K'$ be a sequence such that $z_n\to z\in K$. Then by Lemma \[lemma:R\_eps&lt;l=&gt;R&lt;L\] the sequence $\|(z_n-A)^{-1}\|_{L(\h)}$ is uniformly bounded and hence $z\in\rho(A)$. Indeed, since $K'\subset\rho(A)\cap\rho(A_\eps)\cap B_r(0)$, it satisfies the assumptions of Lemma \[lemma:R\_eps&lt;l=&gt;R&lt;L\]. Since $K'\neq\emptyset$ and $K$ is connected, we conclude that $K'=K$. #### *Conclusion of Part (ii).* Let $K\subset\C$ be compact and connected such that $K\subset\rho(A_\eps)$ for all $\eps>0$. By assumption we may choose a curve $\gamma$ such that $z_0\in K\cup\gamma$. But now $K\cup\gamma$ satisfies the assumptions of Proposition \[pr:Generalised\_spectral\_convergence2\] and hence $K\subset K\cup\gamma\subset\rho(A)$. #### Proof of (iii). Let $\lambda\in\sigma(A)$ be an isolated point and let $\delta>0$ small enough that $B_\delta(\lambda)\setminus\{\lambda\}\subset\rho(A)$. Define $K:=\overline{B_\delta(\lambda)}\setminus B_{\frac{\delta}{2}}(\lambda)$. Then by (i) of Theorem \[th:mainth\] we know that $K\subset\rho(A_\eps)$ for $\eps$ small enough and there exists $L>0$ such that $\|(z-A_\eps)^{-1}\|_{L(\h_\eps)}\leq L$ for all $z\in K$ and $\eps>0$ (cf. Lemma \[lemma:R&lt;l=&gt;R\_eps&lt;L\]). Next, define $K':=\overline{B_\delta(\lambda)}$. Then - either there exists $\eps_0>0$ such that $K'\cap\sigma(A_\eps)\neq\emptyset$ for all $\eps\in(0,\eps_0)$, or - there exists a sequence $\eps_n\searrow 0$ such that $K'\subset\rho(A_{\eps_n})$ for all $n$. In the first case, we conclude that there is a spectral point of $A_\eps$ in $\overline{B_\delta(\lambda)}$. In the second case, we argue as follows. By construction, $\|R_\eps(z)\|_{L(\h_\eps)}\leq L$ on $K'\setminus B_{\f{\delta}{2}}(\lambda)$. We know that $\|R_\eps(z)\|_{L(\h_\eps)}$ cannot be uniformly bounded on all of $K'$, since Theorem \[th:mainth\] (ii) would imply $K'\subset\rho(A)$, which is false, since $\lambda\in K'$. Hence we must have that $\|R_\eps\|_{L(\h_\eps)}$ is unbounded on $B_{\f{\delta}{2}}(\lambda)$, i.e. there exists a sequence $(z_n)\subset B_{\f{\delta}{2}}(\lambda)$ such that $\|R_{\eps_n}(z_n)\|_{L(\h_\eps)}\to\infty$ as $n\to\infty$. If for infinitely many $n$ there is no spectral point of $A_{\eps_n}$ in $B_{\f{\delta}{2}}(\lambda)$, then we conclude by the maximum principle for subharmonic functions that there exists another sequence of points $\tilde z_n$ on the boundary of $B_{\f{\delta}{2}}(\lambda)$ such that $\|R_{\eps_n}(\tilde z_n)\|_{L(\h_{\eps_n})}\to\infty$. But the boundary of $B_{\f{\delta}{2}}(\lambda)$ is included in $K$ on which we have $\|R_\eps\|_{L(\h_\eps)}\leq L$ for all $\eps>0$ - a contradiction. Hence there must be a spectral point of $A_\eps$ in $B_{\f{\delta}{2}}(\lambda)$ for $\eps$ small enough. We have shown that in either case, we necessarily have $B_{\delta}(\lambda)\cap\sigma(A_\eps)\neq\emptyset$ for $\eps$ small enough. Since $\delta>0$ was arbitrary, the claim follows. Concluding Remarks ================== We conclude with a few remarks on the hypotheses in Definition \[def:extended\_norm-resolvent\]. It has been shown in [@B16] that spectral inclusion in fact holds under the milder assumption of *strong* resolvent convergence. We will now show that an analogous statement is also true in the present situation. \[prop:ConcProp1\] Assume that there exists $z_0\in\rho(A)$ such that $z_0\in\rho(A_\eps)$ for almost all $\eps>0$ and for all $u\in\h$ $$\begin{aligned} \left\| \left(I_\eps(z_0-A)^{-1} - (z_0-A_\eps)^{-1}I_\eps\right)u \right\|_{L(\h,\h_\eps)}\to 0. \end{aligned}$$ Then conclusion (ii) of Theorem \[th:mainth\] holds. The proof of Proposition \[prop:ConcProp1\] merely requires a version of Lemma \[lemma:R\_eps&lt;l=&gt;R&lt;L\]: \[lemma:R\_eps&lt;l=&gt;R&lt;LSTRONG\] For every $l,r>0$, there exists $L>0$ such that if For all $u\in\h$ one has $\left\| \left(I_\eps(z_0-A)^{-1} - (z_0-A_\eps)^{-1}I_\eps\right)u \right\|_{L(\h,\h_\eps)}\to 0$ as $\eps\to 0$, $K\subset\C$ compact with $K\subset \rho(A_\eps)\cap\rho(A)\cap B_r(0)$ for almost all $\eps>0$, $\l\|\res{z,A_\eps}\r\|_{ L(\h_\eps)}\leq l$ for all $z \in K$, then one has $\l\|\res{z,A}\r\|_{ L(\h)}\leq L$ for all $z \in K$. For $z\in\rho(A_\eps)\cap\rho(A)\cap B_r(0)$, define $$\begin{aligned} V_\eps(z):=I_\eps(z_0-A)^{-1} - (z_0-A_\eps)^{-1}I_\eps. \end{aligned}$$ An analogous calculation to the one above eq. leads to the identity $$\begin{aligned} V_\eps(z) = (\id_{\h_\eps} - (z-z_0)R_\eps(z))V_\eps(z_0)(\id_\h - (z-z_0)R(z)). \end{aligned}$$ This implies that for any $u\in\h$ one has the inequality $$\begin{aligned} \|V_\eps(z)u\|_{L(\h,\h_\eps)}\leq (1+|z-z_0|l)\left\| V_\eps(z_0)(u-(z-z_0)R(z)u) \right\|_{L(\h,\h_\eps)}. \end{aligned}$$ Decomposing $R(z)$ as $$\begin{aligned} R(z) = J_\eps(I_\eps R(z)-R_\eps(z)I_\eps) + J_\eps R_\eps I_\eps + (\id_\h - J_\eps I_\eps)R(z) \end{aligned}$$ we find that for all $u\in\h$ with $\|u\|_\h=1$ and all $\eps>0$ $$\begin{aligned} \|R(z)u\|_{L(\h)} &\leq M\|V_\eps(z)u\|_{\h} + M^2\|R_\eps(z)\|_{L(\h_\eps)} + \|(\id_\h - J_\eps I_\eps)R(z)u\|_{L(\h)}\\ &\leq M(1+|z-z_0|l)\left\| V_\eps(z_0)(u-(z-z_0)R(z)u) \right\|_{L(\h,\h_\eps)} + M^2l \\ &\qquad\qquad\quad +\|(\id_\h - J_\eps I_\eps)R(z)u\|_{L(\h)}. \end{aligned}$$ We immediately conclude that $$\begin{aligned} \|R(z)u\|_{L(\h)} &\leq \limsup_{\eps\to 0} \Big(M(1+|z-z_0|l)\left\| V_\eps(z_0)(u-(z-z_0)R(z)u) \right\|_{L(\h,\h_\eps)} \\ &\qquad\qquad\quad + M^2l + \|(\id_\h - J_\eps I_\eps)R(z)u\|_{L(\h)}\Big)\\ &\leq M^2l, \end{aligned}$$ by the strong convergences $V_\eps(z_0)\to 0$ and $\id_\h - J_\eps I_\eps\to 0$. Hence, $\|R(z)u\|_{L(\h)}$ is uniformly bounded for $z\in\rho(A_\eps)\cap\rho(A)\cap B_r(0)$ and $u\in\h$ with $\|u\|_\h=1$, which implies the assertion. The rest of the proof of Proposition \[prop:ConcProp1\] follows that of Theorem \[th:mainth\] (ii) verbatim. [RWW09]{} B. Aupetit. . Lecture Notes in Mathematics. Springer Berlin Heidelberg, 2006. S. Boegli. Convergence of sequences of linear operators and their spectra. , 88(4):559–599, 2017. S. Boegli. Local convergence of spectra and pseudospectra. , 8(3):1051–1098, 2018. S. Bögli and P. Siegl. Remarks on the convergence of pseudospectra. , 80(3):303–321, 2014. K. Cherednichenko, P. Dondl, and F. Rösler. Norm-resolvent convergence in perforated domains. , 110(3-4):163–184, 2018. A. C. Hansen. On the solvability complexity index, the [$n$]{}-pseudospectrum and approximations of spectra of operators. , 24(1):81–124, 2011. T. Kato. . Springer, reprint of the corr. 2nd edition, 1995. D. Mugnolo, R. Nittka, and O. Post. Norm convergence of sectorial operators on varying [H]{}ilbert spaces. , 7(4):955–995, 2013. O. Post. Spectral convergence of quasi-one-dimensional spaces. , 7(5):933–973, 2006. C. Rickart. . University series in higher mathematics. Van Nostrand, 1960. M. [Reed]{} and B. [Simon]{}. . cademic [P]{}ress, [I]{}nc., 1980. R. Rockafellar, M. Wets, and R. Wets. . Grundlehren der mathematischen Wissenschaften. Springer Berlin Heidelberg, 2009. [^1]: School of Mathematics, Cardiff University Email: [RoslerF@Cardiff.ac.uk](mailto:roslerf@cardiff.ac.uk)

--- abstract: 'The explosive growth of Web 2.0, which was characterized by the creation of online social networks, has reignited the study of factors that could help us understand the growth and dynamism of these networks. Various generative network models have been proposed, including the Barabási-Albert and Watts-Strogatz models. In this study, we revisit the problem from a perspective that seeks to compare results obtained from these generative models with those from real networks. To this end, we consider the dating network Skout Inc. An analysis is performed on the topological characteristics of the network that could explain the creation of new network links. Afterwards, the results are contrasted with those obtained from the Barabási-Albert and Watts-Strogatz generative models. We conclude that a key factor that could explain the creation of links originates in its cluster structure, where link recommendations are more precise in Watts-Strogatz segmented networks than in Barabási-Albert hierarchical networks. This result reinforces the need to establish more and better network segmentation algorithms that are capable of clustering large networks precisely and efficiently.' address: - 'Universidad Técnica Federico Santa María, Santiago, Chile' - 'Skout Inc., Santiago, Chile' author: - Marcelo Mendoza - Matías Estrada title: 'Revisiting Link Prediction: Evolving Models and Real Data Findings' --- Link Prediction ,Network Evolving Models ,Discrete Dynamics in Networks Introduction {#intro} ============ The explosive growth of Web 2.0, which is characterized by its consolidation of online social networks, has brought millions of users into participating in these networks, sharing their life experiences and establishing new relationships through these platforms. The relationships created in these networks can be modeled as large graphs. Their evolution and dynamism represent an interesting challenge for the complex network community; this community is dedicated to proposing and discussing the significance of generative models that could precisely explain the growth of these networks. The most widely discussed generative models are those developed by Barabási-Albert [@Barabasi99] and Watts-Strogatz [@watts98]; each of these models has their own strengths and weaknesses in their ability to explain topological characteristics of real networks. Specifically, the Watts-Strogatz model is able to model large but small-diameter networks, which are properties that can be effectively observed in real networks, whereas the Barabási-Albert networks follow an evolving model of preferential attachment with a tendency to generate networks with few vertices that contain a high number of links, where the majority of vertices have low connectivity. This approach has been used as a very effective model for biological networks [@ravasz2002]. A growing interest in establishing better methods for link predictions is using a microscale approximation of the network evolution problem. This approximation has driven investigations by the complex network community in their efforts to not only find models that effectively fit the dynamic laws of social networks but can also provide the possibility of link recommendation, which is also known as link prediction and is particular useful in online social networks. In this study, we perform analyses on link prediction methods based on the topological characteristics of real data from the dating network Skout. Subsequently, we contrast the results obtained with those observed from the Barabási-Albert and Watts-Strogatz generative models. A number of experimental configurations were evaluated, including intra-cluster recommendations in segmented networks. The contribution of this study is the evaluation of the effectiveness of these models in light of the results observed from data of a real network. The problem is analyzed from a purely topological standpoint, which contrasts the classical standpoint of FOAF (friend-of-a-friend) recommendations with those based on segmented networks. This article is organized as follows. In Section \[sec:relwork\] we discuss related work. Background is summarized in Section \[back\]. The case study is presented in Section \[case-study\]. Results from synthetic data are presented in Section \[synth\] and discussed in Section \[disc\]. Finally, we conclude in Section \[conc\] with a brief discussion about some open questions and future work. Related Work {#sec:relwork} ============ Link prediction is mainly addressed using transitivity properties of the graph also known as friend-of-a-friend approaches [@silva2010]. The rationale of this kind of approaches is the following: If a user A is friend of a user B and a user B is friend of a user C, then A and C probably will be friends [@kossinets2006]. However, the success of link prediction approaches is limited because real online social networks tend to be very sparse [@delgenio2011], meaning that the total amount of potential links to be created is much greater than the links that are actually created. Sparsity turns the link prediction problem into a very dificult problem returning success improvements over a random predictor from 3% to 54% [@liben2003]. A number of different types of features are available for this problem [@lin1998]. By exploiting locality we do not require the full graph from being stored. One of the simplest locality features is the Common Neighboors index [@newman2001]. This index considers the amount of common neighboors that two nodes have. When the number of common neighboors is high, then is more likely that those nodes will create a link in the future [@kossinets2006]. However, this feature is biased by the number of neighboors that a given node registers. A more robust feature is the Jaccard index [@jaccard1912] which compares the common neighboors cardinality with the cardinality of the union of both neighboors, getting a value that represents the proportion between the cardinality of these two sets. Other features used are the Salton index [@salton1983] and the Sorensen index [@sorensen1948] that are mainly used in ecological networks [@linyuan2010]. Finally, another well known locality measure is the Adamic & Adar index [@ada2003]. Graph-based features considers the whole network to be calculated. SimRank index [@jeh2002] is calculated using a random walk process which is propagated through the graph with a decay factor. It is an expensive feature because it needs several passes through the whole network. Then, unexpensive and conventional graph-based features as degree or transitivity are more recommended for link prediction [@linyuan2010]. Finally, the combination of global/local scope is also explored for this task, using approaches as HITS (Hypertext Induced Topic Search) which comes from the information retrieval community, proposed by Kleinberg [@kleinberg] for web page ranking. An undirected version of HITS can be explored for link prediction purposes in undirected graphs, giving us an authority feature for each node of the graph. There are more information sources that are useful for link prediction tasks. For instance, by recovering text from user messages, it is possible to create more sophisticated user descriptors [@roth2010]. It is also possible to characterize neighboors by recovering text and then using this information for user description [@zhou2009]. Transactional registers has been also explored for link prediction [@xlang10]. These methods tend to involve high computational costs due to the necessity of maintain content-based indexes or log-repositories. These methods are out of the scope of this study. Regarding network segmentation methods, there has been recent interest in the exploration of spectral clustering methods for link prediction. [@symeonidis] explored the use of segmented networks for link prediction, proving in synthetic networks that this approach is feasible. The study is close in aim to our article but it does not consider an evaluation regarding evolving graph models. In addition the comparison is constrained only to the top-1 recommended node. Link prediction has been explored in several domains, showing that is still a relevant task. For instance, different algorithms has been developed to predict new links in protein networks [@lu2009], energy grid networks [@murata2008], medical co-author networks [@plos14], disease networks [@disease14], among others. Further details about this topic can be read in the survey of Lú [@lu11]. Background {#back} ========== Link prediction is the problem of inferring whether potential edges between pairs of vertices in a graph will be present or absent in the near future. In a link prediction task, we first assume the existence of an original observed graph, with a completely known set of vertices and a partially known set of edges. Accordingly, the link prediction task is to infer the rest of the edges. Let $G = (V,E)$ be an undirected graph. At a given point in time $t_0$, we assume that all vertices $v \in V$ are known but only a subset of $E$ is known. Let $E^{obs}$ be the subset of edges $e \in E$ that is known at $t_0$, and let $E^{miss}$ be the subset of unobserved edges at $t_0$, such that $E^{miss} = E \setminus E^{obs}$. Link prediction is to predict the edges in $E^{miss}$ from $G = (V,E^{obs})$. A natural way to address link prediction is to use a ranking-based strategy. A ranking-based strategy considers two steps to rank links. For a given vertex $u$, a vertex retrieval step is conducted, using a locality constraint to recover a list of proximal vertices to $u$. Then, the list is sorted according to a given scoring measure. A number of graph-based measures can be used for scoring. We explore the perfomance of three of them: Normalized degree, authority, and transitivity. These features are query independent, that is to say, these measures define a collection of pointwise estimations at vertex level. The value of a measure of this collection is the same for the whole graph. The definitions of these measures are given below. #### Degree coefficient Let $\Gamma (u)$ be the neighborhood of a vertex $u$ and let $\mid \Gamma (u) \mid$ be the cardinality of this set, also known as the degree of $u$. We define a degree coefficient by normalizing $\mid \Gamma (u) \mid$ with the maximum degree of $G$. Formally: $$\tt{Degree}(u) = \frac{\mid \Gamma (u) \mid}{\tt{Max}_{v \in G} \mid \Gamma (v) \mid}.$$ #### Transitivity coefficient Let $\Gamma (u)$ be the neighborhood of a vertex $u$. The transitivity coefficient $\tt{Transitivity}(u)$ (a.k.a. clustering coefficient) is the ratio between the number of links in $\Gamma (u)$ and the maximum number of such links. If $\Gamma (u)$ has $e_u$ links, we have: $$\tt{Transitivity}(u) = \frac{e_u}{\frac{\mid \Gamma (u) \mid \cdot (\mid \Gamma (u) \mid - 1)}{2}} .$$ #### Authority HITS (Hypertext Induced Topic Search) coefficients come from the information retrieval community and were proposed by [@kleinberg] for web page ranking. The idea is that pages that have many links pointing to them are called authorities and pages that have many outgoing links are called hubs. Good hubs point to good authority pages, and vice-versa. Lets $\tt{hub}(u)$ and $\tt{auth}(u)$ be hub and authority coefficients for a vertex $u$. The following equations can be solved through an iterative algorithm that addresses the fixed point problem defined by: $$\tt{hub(u) = \sum_{v \in G \mid u \rightarrow v} \tt{auth}(v)} ,$$ $$\tt{auth(u) = \sum_{v \in G \mid v \rightarrow u} \tt{hub}(v)} .$$ In undirected graphs, both coefficients have the same value. In this study, we denote this feature as authority, given that a dating network is an undirected graph. Link prediction in a real network: the Skout Inc. case {#case-study} ====================================================== We explore properties of the problem of link prediction in a real network provided directly by Skout Inc.. After recording each link created until January 1st 2014, we generated a network with 3,855,389 links among 1,920,015 active users. We define an active user as one that, after creating their account, has added at least one friend during the period. Afterwards, we recorded links created during the first 25 days of January 2014 between active users. A total of 582,119 new links were recorded solely from users who were active during 2013. A total of 76,848 users added at least one new friend during the observation period. The distribution of new friends per user is shown in Figure \[fig-2\]. ![Number of links created per active user during the observation period.[]{data-label="fig-2"}](users_links.eps){width="0.7\columnwidth"} As shown in Figure \[fig-2\], the distributions of new friends per user follow the law of “the rich get richer”: only a few users added a high number of new friends; the majority of users added only a few new friends, confirming that link prediction is a very difficult task due to the imbalance between potential friends and actual friends. The collection of links that could have potentially been created during this period corresponds to the group of links that was not created until January 1st 2014 among active users. Then, we analyzed the properties of separability that exist between the group of potential links created versus the group of not-created links; the logic of the methodology used in this study is explained in the scheme shown in Figure \[fig-1\]. ![Case study methodology. Blue links and nodes depict the graph at $t_0$. Red links depict links created during the observation period. Links depicted with dashed lines indicates potential unobserved links. In our strategy, dashed links are labeled as false link instances and red links as real link instances.[]{data-label="fig-1"}](figure2.eps){width="0.7\columnwidth"} Each link in the group of potential links is classified as either real (created) or not (not created), where the authority, degree, and transitivity scores for both the target user and the candidate user are calculated. In this experiment, the target user is the user who is looking for a new friend, and the candidate user is the potential new friend that could be recommended to the target user. Table \[tab:inf\] shows the results of information gain from separating the two classes. Feature Information Gain ---------------- ------------------ Authority 2 0.9155 Authority 1 0.5367 Transitivity 2 0.1192 Degree 2 0.1054 Transitivity 1 0.0392 Degree 1 0.0227 : Information Gain values for each feature considered in our data set. Target users features are depicted with a subindex equals to one. Candidate users are depicted with a subindex equals to two.[]{data-label="tab:inf"} As shown in Table \[tab:inf\], the most relevant characteristic for this problem is the authority of the candidate user. Because the first vertex corresponds to the target user (the one who accepts or declines the new relationship), the authority of the candidate user is a measure of the visibility (and therefore popularity) of the user for the rest of the network. However, the characteristics of the target user are not extremely relevant, which indicates that the present connectivity of the user is not necessarily an indicator of how connected the user will be in the future. We should note that an algorithm that predicts links should be capable of making recommendations to newly target users in the network, i.e., users with few neighbors and a short user history; this challenge is also referred to as a cold start. Because of this, the characteristics of the target user are not as relevant as those of the candidate user. In this strategy, we always consider that it is the user with fewer friends who gets the most recommendations to equalize the size of the network neighborhoods. This focus is paralleled with information retrieval, where the query (the target user) is considered as a document with few descriptors, and therefore, the descriptors of the candidate documents are primarily exploited on being recommended in the list of answers. Results from synthetic data {#synth} =========================== Networks with 10,000 vertices were generated based on three generative link models: random-based graphs models, also known as Erdós-Rényi graphs [@erdos60], Watts-Strogatz [@watts98], with a rewired probability equal to 0.1, and Barabási-Albert [@Barabasi99]. For each of these models, we performed a pruning process according to a factor taken as an experimental parameter that could control the ratio of links eliminated to all the links in the network. Later, link prediction was performed on the eliminated links, where a recommended and eliminated link was considered a success, and a recommended but kept (not-eliminated) link is considered a failure. In this way, the pruning factor controls the ratio of real solutions to the total candidates. A low pruning factor, which corresponds to a smaller ratio of links to be predicted over the total candidates, makes the problem more difficult. Each of the networks was segmented with a directed k-clustering algorithm (k-means for list of arcs), with k values ranging between 3 and 7. Then, using one of the three candidate ranking functions performed on the Skout network (authority, degree, and transitivity), an intra-cluster ranking process was performed. We were thus able to evaluate the impact of network segmentation on the quality of the recommendations. Tables 2 and 3 show the average precision and recall results for the first ten recommendations, which are micro-averages, i.e., the results correspond to averages for all the nodes in a network. Table 2 shows the results for a pruning factor of 0.1, and Table 3 shows the results for a pruning factor of 0.25. More configurations were examined in these experiments, which incorporated a greater number of small networks and different pruning factors. The P@N and R@N curves, with N values ranging between 1 and 10 for all configurations, can be seen in greater detail for some small networks (up to 1,000 nodes) at the following url \[<http://104.236.107.4:17264>\] [^1]. \[tab:I\] [90]{} -- --- ------------- ----------------- ---------------- ------------- ----------------- ---------------- k Erdós-Rényi Barabási-Albert Watts-Strogatz Erdós-Rényi Barabási-Albert Watts-Strogatz P@10 P@10 P@10 R@10 R@10 R@10 3 0.0126 0.0604 0.0516 0.0001 0.0022 0.0081 4 0.0121 0.0462 0.0798 0.0005 0.0015 0.0169 5 0.0121 0.0467 0.1114 0.0005 0.0010 0.0174 6 0.0110 0.0328 0.1310 0.0005 0.0007 0.0232 7 0.0104 0.0424 0.1552 0.0008 0.0010 0.0298 3 0.0110 0.0612 0.0312 0.0001 0.0026 0.0072 4 0.0110 0.0464 0.0513 0.0004 0.0012 0.0143 5 0.0110 0.0465 0.0400 0.0005 0.0010 0.0143 6 0.0112 0.0326 0.0746 0.0005 0.0007 0.0212 7 0.0112 0.0412 0.1066 0.0008 0.0010 0.0268 3 0.005­0 0.0406 0.0236 0.0001 0.0026 0.0074 4 0.0120 0.0412 0.0311 0.0005 0.0014 0.0143 5 0.0102 0.0405 0.0400 0.0005 0.0010 0.0122 6 0.0126 0.0333 0.0441 0.0005 0.0007 0.0204 7 0.0114 0.0384 0.0472 0.0008 0.0010 0.0212 -- --- ------------- ----------------- ---------------- ------------- ----------------- ---------------- **Table 2**. Precision and recall results with pruning rate = 0.10. \[tab:II\] [90]{} -- --- ------------- ----------------- ---------------- ------------- ----------------- ---------------- k Erdós-Rényi Barabási-Albert Watts-Strogatz Erdós-Rényi Barabási-Albert Watts-Strogatz P@10 P@10 P@10 R@10 R@10 R@10 3 0.0272 0.0712 0.0988 0.0018 0.0033 0.0144 4 0.0241 0.0688 0.1548 0.0021 0.0025 0.0254 5 0.0280 0.0612 0.2028 0.0022 0.0019 0.0255 6 0.0282 0.0575 0.2616 0.0014 0.0018 0.0339 7 0.0276 0.0510 0.2901 0.0017 0.0013 0.0399 3 0.0264 0.0791 0.0744 0.0018 0.0033 0.0122 4 0.0224 0.0689 0.1135 0.0020 0.0024 0.0178 5 0.0264 0.0616 0.1486 0.0022 0.0018 0.0231 6 0.0270 0.0579 0.2235 0.0014 0.0018 0.0336 7 0.0272 0.0520 0.2735 0.0017 0.0013 0.0377 3 0.0244 0.0512 0.0485 0.0016 0.0033 0.0102 4 0.0212 0.0522 0.0603 0.0022 0.0026 0.0124 5 0.0270 0.0614 0.0850 0.0022 0.0019 0.0146 6 0.0292 0.0507 0.1015 0.0014 0.0018 0.0147 7 0.0264 0.0472 0.1177 0.0017 0.0013 0.0141 -- --- ------------- ----------------- ---------------- ------------- ----------------- ---------------- **Table 3**. Precision and recall results with pruning rate = 0.25. Discussion {#disc} ========== In this study, we addressed the problem of link prediction, which segments networks to later analyze each resulting partition in terms of degree, authority, and transitivity. Several of the results from the experiments using model data demonstrated certain relevant characteristics. One characteristic is related to the link pruning parameter because it is also related to the network evolution time. That is, a low pruning factor emulates a short period of observation and thus, a more difficult problem because worse results would be obtained compared with that of higher pruning factors. This result is due to predictability improving as the limits of observation increase for prediction problems. Specifically for our experiments, increasing the pruning ratio raises the number of possible links that are candidates for selection; each of the candidate links added corresponds to occurrences of links already existing in the original network, and thus, it is expected that the results would improve when they are selected. The results demonstrate that the process of segmenting the networks is key to link prediction because dissimilar results are obtained from different segmentation routines. Such is the case with the Watts-Strogatz generative model, which achieved a precision of approximately 0.1 with three clusters and approximately 0.3 with seven clusters. The optimum number of clusters depends on the nature of the network, the number of vertices, and the density of the network, among other factors. Most of the methods of link prediction use locality from a FOAF (friend-of-a-friend) standpoint. However, the results obtained in this research based on transitivity exhibited worse results than those based on authority and degree. It is possible that when using locality, transitivity restricts the predictions to the graph of nearest neighbors, whereas clustering uses global network characteristics, which are only partially approximated by a FOAF standpoint. Barabási-Albert-type networks give worse results from segmentation, where prediction results worsen as the number of clusters increases. Barabási-Albert networks are essentially hierarchical, and a flat clustering segmentation method is therefore not extremely suitable for this type of structure. The measures used to characterize and select candidates that behave best depend on the network type. In Barabási-Albert-type networks, predictions based on degree are slightly better than those based on authority. The opposite is observed in Watts-Strogatz-type networks; predictions based on authority are somewhat better than those based on degree. In both network types, the worst precision is achieved using transitivity. When comparing these results to those obtained from Skout Inc., we should remember that in the latter case, the best properties of prediction were shown in the authority, degree, and transitivity scores (in that order). These results are similar to those obtained from a Watts-Strogatz-type generative model, which suggests that this model could approximate the structure of the Skout network, provide a means to perform link prediction on a segmented network, and complement FOAF strategies with ranking strategies based on intra-cluster authority. Conclusions {#conc} =========== This study revisited the problem of link prediction from a perspective that aims to unify results obtained from real networks and those from generative models. We conclude that the recommendations provided from FOAF are overvalued because nearest-neighbor structures are unable to explain the growth and dynamism of a network on their own. Instead, the use of network segmentation methods can explain the creation of links between users that might not be among the nearest-neighbors, which highlights how the dynamism and evolution of a network is explained via a combination of local (nearest neighbor) and global (clusters) factors. These results emphasize the need for more and better network segmentation algorithms capable of working on large data collections. There is also a need to consolidate link prediction models capable of effectively combining local and global factors. Acknowledgment {#acknowledgment .unnumbered} ============== Marcelo Mendoza was supported by project FONDECYT 11121435 and by FB0821 Centro Científico Tecnológico de Valparaíso. [00]{} Adamic, L. A., Adar, E. (2003) *Friends and neighbors on the Web*, Social Networks, 25(3), 211-230. Barabási, A. L., Albert, R. (1999) *Emergence of scaling in random networks*. Science, 286, 509-512. Erdós, P., Rényi, A. (1960) *On the evolution of random graphs*, Publications of the Mathematical Institute of the Hungarian Academy of Sciences, 5, 17-61. Genio, C. I. Del, Gross, T., Bassler, K. E. (2011) *All scale-free networks are sparse*, Physical Review Letters, 107, 178701, 1-4. Jaccard, P. (1912) *The distribution of the flora in the alphine zone*, The New Phytologist, XI, 37-50. Jeh, G., Widom, J. (2002) *SimRank: a measure of structural-context similarity*, In Proceedings of the ACM International Conference on Information and Knowledge Management, KDD, 1-11. Kaya, B., Poyraz, M. (2014) *Supervised link prediction in symptom networks with evolving case*, Measurement, 56, 231-238. Kleinberg, J. (1998) *Authoritative Sources in a Hyperlinked Environment*, in Proceedings of the Symposium on Discrete Algorithms, SODA, 668-677. Kossinets, G. (2006) *Effects of missing data in social networks*, Social Networks, 28(3), 247-268. Liben-Nowell, D., Kleinberg, J. (2003) *The link prediction problem for social networks*, In Proceedings of the ACM Conference on Information Knowledge and Management, CIKM, 556-559. Lin, D. (1998) *An Information-Theoretic Definition of Similarity*, in Proceedings of the International Conference on Machine Learning, ICML, 296-304. Linyuan, L. (2011) *Link Prediction in Complex Networks : A Survey*. Physica A: Statistical Mechanics and its Applications, 390(6), 1150-1170. Lú, L., Jin, C.-H., Zhou, T. (2009) *Similarity index based on local paths for link prediction of complex networks*, Physical Review. E, Statistical, Nonlinear, and Soft Matter Physics, 80, 046122. Lú, L., Zhou, T. (2011) *Link prediction in complex networks: A survey*, Physica A: Statistical Mechanics and its Applications, 390(6), 1150-1170. Murata, T., Moriyasu, S. (2008) *Link Prediction based on Structural Properties of Online Social Networks*, New Generation Computing, 26(3), 245-257. Newman, M. E. (2001) *Clustering and preferential attachment in growing networks*, Physical Review. E, Statistical, Nonlinear, and Soft Matter Physics, 64, 025102. Ravasz, E., Somera, A. L., Mongru, D. A., Oltvai, Z. N., Barabási, A. L. (2002) *Hierarchical organization of modularity in metabolic networks*. Science, 297, 1551-1555. Roth, M., Flysher, G., Matias, Y., Leichtberg, A., Merom, R. (2010) Suggesting Friends Using the Implicit Social Graph Categories and Subject Descriptors. In Proceedings of the ACM International Conference on Knowledge Discovery and Data Mining, KDD, 233-242. Salton, G., McGill, M. J. (1983) *Introduction to Modern Information Retrieval*, New York (Vol. 22, p. xv, 448 p.). Silva, N. B., Tsang, I.-R., Cavalcanti, G. D. C., Tsang, I.-J. (2010) *A graph-based friend recommendation system using Genetic Algorithm*, In Proceedings of the IEEE Congress on Evolutionary Computation, 1-7. Sorensen, T. (1948) *A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on Danish commons*, Kongelige Danske Videnskabernes Selskab, 5(4), 1-34. Symeonidis, P., Iakovidou, N., Mantas, N., Manolopoulos, Y. (2013) *From biological to social networks: Link prediction based on multi-way spectral clustering*, Data & Knowledge Engineering, 87, 226-242. Watts, D., Strogatz, S. (1998) *Collective dynamics of ’small-world’ networks*, Nature, 393, 6684, 440-442. Xiang, R., Neville, J., Rogati, M. (2010) *Modeling relationship strength in online social networks*, In Proceedings of the ACM World Wide Web Conference, WWW, 981-990. Yu Q., Long C., Lv Y., Shao H., He P. (2014) *Predicting Co-Author Relationship in Medical Co-Authorship Networks*, PLoS ONE 9(7): e101214. Zhou, T., Lü, L., Zhang, Y.-C. (2009) *Predicting missing links via local information*. The European Physical Journal B, 71(4), 623-630. [^1]: We will replace the ip number by a .org url in the camera ready version

--- abstract: 'We present tunneling spectroscopy and transport measurements on disordered indium oxide films that reveal the existence of a superconducting gap in an insulating state. Two films on both sides of the disorder induced superconductor to insulator transition (SIT) show the same energy gap scale at low temperatures. This energy gap persists up to relatively high magnetic fields and is observed across the magnetoresistance peak typical of disordered superconductors. The results provide useful information for understanding the nature of the insulating state in the disorder induced SIT.' author: - 'D. Sherman$^{1}$' - 'G. Kopnov$^{2}$, D. Shahar$^{2}$ and A. Frydman$^{1}$' date: - - title: '**A Superconducting Gap in an Insulator**' --- Increasing the disorder of a metallic system causes the localization of its electronic wave function. If the ground state of the system is a superconductor, increasing the disorder leads to a transition from a superconductor to an insulator (SIT). This transition has gained increasing attention lately due to the experimental observations of a number of dramatic features near the SIT such as simple activated temperature dependence of the resistance on the insulating side [@shahar_ovadyahu], a large peak in the magneto-resistance [@hebard; @gant; @Sambandamurthy; @Steiner; @BaturinaC], peculiar I-V characteristics [@iv1; @iv2] and traces of superconductivity at temperatures above $T_C$ [@sacepe0; @sacepe2; @armitage; @pratap]. Another reason for renewed interest in this field is that the SIT may be a basic realization of a quantum phase transition which occurs at T = 0 as a function of disorder or magnetic field and is driven by quantum rather than thermal fluctuations. Despite this growing interest, the mechanism of the SIT and, in particular, the nature of the insulating state are not understood. Recently a few indications for the presence of electronic pairs in the insulator have been reported [@sacepe2; @hollen; @shachaf], inspiring further theoretical effort. Two general ideas have been put forward to try to explain some of the observations. One relies on the role of the disorder in generating inhomogeneity in structurally homogeneous samples [@kowal; @efrat; @dubi; @imry; @sacepe1]. Within this framework, the crossover from insulating to superconducting behavior occurs when the Josephson-coupling between superconducting islands succeeds in forming a percolation path throughout the system. A second scenario invokes the existence of uncorrelated pre-formed electron pairs which do not constitute a condensate but are characterized by an energy gap that is associated with the pair binding energy [@feigel]. Other models adapt concepts from both pictures and suggest that the insulating film is composed of small superconducting islands that are uncorrelated and are too small to sustain bulk superconductivity [@nandini; @nandini2]. Clearly, additional experimental results are required to help shed light on this “supercoducting insulator”. In this letter we present an experimental study of the DOS and corresponding transport characteristics of two disordered films on both sides of the transition. These measurements show that a similar energy gap exists in both the superconducting and the insulating states. We present the dependence of this gap on disorder and magnetic field and discuss the possible implications of these experimental results towards the understanding of superconductivity in highly disordered films. For measuring the DOS of the films we fabricated tunnel junctions in the following way: A 30 nm Al stripe was thermally evaporated on a Si/SiO substrate and was allowed to oxidize for a few hours in ambient conditions. Subsequently, a 31 nm thick indium oxide (InO) stripe was e-gun evaporated perpendicular to the Al stripe, thus forming a planar tunnel junction with barrier dimensions of 1mm\*1mm. In order to produce InO films with different disorder, dry oxygen was injected into the evaporation chamber at different partial oxygen pressure [@zvi_86]. This resulted in amorphous yet structurally homogeneous films with different degrees of disorder characterized by their sheet resistance $R_{\Box}$. The results presented in this letter were obtained on two amorphous InO films; one exhibiting superconducting transport (sample S) and the other exhibiting insulating behavior (sample I). Fig. 1 shows the resistance versus temperature of both films. It is seen that while sample S shows a clear superconducting transition with $T_C \sim 3K$, the resistance of sample I increases rapidly with lowering T down to our base temperature of $25mK$, thus showing insulating properties. A fit to Arrhenius law: $$\begin{aligned} R=R_{0}exp(T_{0}/T).\end{aligned}$$ is shown in Fig 1. From this fit one yields an activation temperature $T_{0} = 0.4K$ [@rem0], and $R_{0} = 11.5k\Omega$ which is larger than the quantum resistance for pairs, $1/G_0=h/2e^{2} = 6.45k\Omega$. \[rt1\] {width="\columnwidth"} Despite this clear difference in the transport properties, the tunneling spectra of the two films appears to be surprisingly similar. Fig. 2 depicts dI/dV versus V curves for both samples. All tunneling measurements presented here were performed by standard lock-in techniques while making sure that the junction resistance was at least an order of magnitude larger than the InO sheet resistance so that the film could be regarded as an equi-potential electrode. Since these are disordered films one has to take into account that electronic interactions cause a suppression of the density of states (DOS) at low energies either due to the Altsuler Aronov [@AA] zero bias anomaly (ZBA) mechanism for weak disorder or due to the coulomb gap [@ES] for high disordered systems. In order to isolate the superconducting contribution to the DOS we normalized the curves by the tunneling spectra of the films taken at a magnetic field, H, of 11T. The justification for this procedure relies on the assumptions that superconductivity is fully suppressed at $H=11T$, and that the magnetic field hardly affects the normal state DOS. Indeed, the dI/dV-V curves of both samples exhibit very small magnetic field dependence at energies larger than the superconductive gap, and for $H \geq 7.5T$ (where superconductivity is suppressed) the curves are practically indistinguishable. The dI/dV-V curves shown in Fig. 2 demonstrate that a superconducting gap exists in both samples, despite the fact that sample I clearly shows insulating transport behavior. In order to extract the value of the superconducting gap, $\Delta$, we fit these curves to the BSC expression modified by a broadening parameter $\Gamma$ that accounts for the finite scattering time of the superconducting quasi-particles [@dynes]: $$\label{eq_BCS} N_S(E)/N_N(E)=\Re\left\{(E-i\Gamma)/ [(E-i\Gamma)^{2}-\Delta^{2}]^{1/2} \right\}$$ The fits depicted in Fig. 2 show that the experimental curves of both films deviate from the BSC theoretical predictions. In particular, the so called “coherence peaks” at the gap edges are suppressed in our disordered films. Nevertheless, the best fits yield $\Delta = 0.7mV$ *for both films* [@rem1]. The only observed differences between the samples are a smooth suppression of the coherence peaks and an increase of the sub-gap conductance. {width="50.00000%"} \[dos\] . The fact that the superconductive gap persists into the insulating side of the SIT was experimentally implied in the past. Scanning tunneling microscopy measurements of the local DOS were performed on disordered TiN [@sacepe1; @sacepe0] and InO [@sacepe2] superconductors with different degrees of disorder. In both cases it was found that $\Delta$ does not decrease with disorder as would be expected from the decrease of $T_C$ extracted from the transport measurements. In the TiN films $\Delta$ decreased with increasing disorder, however $\frac{T_C}{\Delta}$ was found to decrease as the films approached the SIT. In the InO samples the average $\Delta$ showed no clear dependence on disorder. In both cases an extrapolation would predict a finite gap in the insulator. Our measurements confirm this trend since we observe the same gap magnitude on the two sides of the transition. This observation is in accordance with a recent theoretical work [@nandini2] which predicts that $\Delta$ is expected to remain unchanged through the SIT and may even grow in magnitude deep in the insulator. {width="50.00000%"} {width="50.00000%"} \[dos\_h\] As may be expected, the magnetic field, H, has a significant effect on $\Delta$. Fig 3 shows the dependence of dI/dV versus V curves on H for samples S and I. It is seen that for both samples $\Delta$ decreases monotonically with H and is wiped out altogether at $H \sim 7.5T$. It is interesting to relate these results to the magnetoresistance of the samples. Fig 3 shows that both samples exhibit a magnetoresistance peak similar to that reported in the past [@hebard; @gant; @Sambandamurthy; @Steiner; @BaturinaC]. The magnetic field at which the peak occurs, $H_P$, is lower in sample I (1.7T at 1K compared to 6.5T for sample S). This is consistent with the trend seen in superconducting samples where $H_P$ was found to decrease with increasing disorder [@baturina]. Comparing the tunneling and the magnetoresistance results reveals the fact that while in sample S the gap is wiped out at fields lower than $H_P$, in sample I $\Delta$ persists up to fields that are larger than $H_P$. It appears, therefore, that the DOS is unaffected by the magnetoresistance details, and particularly by the resistance peak, and only depends on the value of H. This demonstrates, again, that while the resistance of the two samples may behave differently, their DOS is very similar. It has been suggested [@dubi] that the MR peak is due to the formation of superconducting islands with sizes that depends on the magnetic field. In this picture, $H_P$ corresponds to the magnetic field at which it is preferable for the current to flow through normal regions rather than through the superconducting islands. Our results are consistent with this picture since it implies that the resistance would be sensitive to $H_P$ but the DOS would not be affected by it. The results presented above strongly suggest that the disorder driven superconductor to insulator transition is of a geometrical percolation nature. The fact that the DOS changes smoothly across the transition and that the order parameter is hardly affected by it is consistent with a model of superconducting islands of which the effective sizes gradually decreases with disorder. The observation of coherence peaks in the insulator suggests that superconductive grains that sustain a coherent condensate are present in this phase. We envision that the average size of these islands decreases with increasing disorder so that a larger percentage of islands falls below the Anderson limit for superconductivity. This causes the coherence peaks to shrink with disorder. Further increasing of the disorder is expected to cause the coherence peaks to vanish altogether, however, it is very difficult to perform tunneling junction experiments in this regime of high film resistance. {width="50.00000%"} \[normal\] This “granular” model is further supported by the results of the DOS in the normal state. As noted, for $H > 7.5T$ the superconductive gap is fully suppressed. At these fields dI/dV is found to be proportional to ln(V) (see Fig. 4). This behavior is consistent with the zero bias anomaly model for weak disorder [@AA]. Due to the nearby Al electrode the long range Coulomb interactions are screened. Hence for low voltage and temperature, such that the thickness of the sample is smaller than the thermal length, the DOS is expected to follow the screened 2D expression: $$\begin{aligned} \nu(\epsilon, T) -\nu (\infty)=-\frac{\nu (\infty)}{4 \pi^{2} \hbar g}\ln\frac{2\kappa b}{k_{2}}\ln\frac{\tilde{\varepsilon}\tau}{\hbar} \end{aligned}$$ where $\tilde{\varepsilon}=max(V,T)$, $\nu (\infty)$ is the DOS at high energies, $\tau$ is the inelastic relaxation time, $k$ is the dielectric function, $\kappa$ is the inverse screening length, $b$ is the barrier thickness and g is the dimensional conductance $G/G_{0}$. Since the slopes of the dI/dV versus ln(V) curves are proportional to 1/g, our results enable us to extract the conductance ratio between the two films. This yields $\frac{g_S}{g_I}=1.6$. We stress that this ratio does not depend on the exact model for ZBA as long as the 1/g dependence is valid. On the other hand, the conductance ratio extracted from the transport at H=11T gives $\frac{g_S}{g_I}=11$. This dramatic difference can be interpreted as an indication for granularity in the film. While the transport g is sensitive to the current carrying network, tunneling may take place into the metallic regions which have a relatively high conductance even in the insulating phase. In this picture the difference between the DOS in the metal and in the insulator may be very small. In conclusion, We have provided direct evidence for the existence of a superconductive gap in an insulator state. The gap amplitude is hardly affected by the dramatic change in the global transport properties. The results strongly suggest that homogeneously disordered superconducting films near the SIT contain superconducting islands with bulk properties that are similar on both sides of the transition. The DOS is not influenced by the SIT nor by the magnetoresistance peak indicating that the insulator contains regions of finite order parameter amplitude that are similar to those in the superconductor. We are grateful for the useful discussions with D.B. Gutman, B. Sacépé, E. Shimshoni and N. Trivedi. This research was supported by the US Israel binational fund (grant No. 2008299) D. Shahar, Z. Ovadyahu Phys. Rev. B [**46**]{}, 10917 (1992). M. A. Paalanen, A. F. Hebard, and R. R. Ruel, Phys. Rev. Lett. **69**, 1604 (1992). V. F. Gantmakher, M. V. Golubkov, V. T. Dolgopolov, G. E. Tsydynzhapov, and A. A. Shashkin, JETP Lett. **68**, 363 (1998). G. Sambandamurthy, L. W. Engel, A. Johansson and D. Shahar, Phys. Rev. Lett. **92**, 107005 (2004); G. Sambandamurthy, L. W. Engel, A. Johansson, E. Peled and D. Shahar, Phys. Rev. Lett. **94**, 017003 (2005). M. A. Steiner, G. Boebinger, and A. Kapitulnik, Phys. Rev. Lett. [**94**]{}, 107008 (2005). T. I. Baturina, J. Bentner, C. Strunk, M. R. Baklanov, and A. Satta, Physica B **359**, 500 (2005); T. I. Baturina, C. Strunk, M. R. Baklanov, and A. Satta, Phys. Rev. Lett. **98**, 127003 (2007). V.M. Vinokur, T.I. Baturina, M.V. Fistul, A.Y. Mironov, M.R. Baklanov, and C. Strunk, Nature **452**, 613 (2008). O. Cohen, M. Ovadia and D. Shahar, Phys. Rev. B. **84**, 100507 (2011). B. Sacépé, C. Chapelier, T.I Baturina, V.M. Vinokur, M.R. Baklanov and M. Sanquer. Nature Commun. **1**, 140 (2010). B. Sacépé, T. Dubouchet, C. Chapelier, M. Sanquer, M. Ovadia, D. Shahar, M.V. Feigel’man and L. Ioffe, Nature Physics [**7**]{}, 239 (2011). R. Crane, N. P. Armitage, A. Johansson, G. Sambandamurthy, D. Shahar, and G. Gruner, Phys. Rev. B **75**, 184530 (2007) M. Mondal, A. Kamlapure, M. Chand, G. Saraswat, S. Kumar, J. Jesudasan, L. Benfatto, V. Tripathi, and P. Raychaudhuri, Phys. Rev. Lett. **106**, 047001 (2011). S.M. Hollen, H.Q. Nguyen, E. Rudisaile, M.D. Stewart, J. Shainline, J.M. Xu, and J.M. Valles, Phys. Rev. B **84**, 064528 (2011). S. Poran, E. Shimshoni, and A. Frydman, Phys. Rev. B [**84**]{}, 014529 (2011). D. Kowal and Z. Ovadyahu, Solid State Comm., **90**, 783 (1994); D. Kowal and Z. Ovadyahu, Physica C **468** 322 (2008). E. Shimshoni, A. Auerbach, and A. Kapitulnik, Phys. Rev. Lett. **80**, 3352 (1998). Y. Dubi, Y. Meir and Y. Avishai, Phys. Rev. B. **73** 054509 (2006); Nature **449**, 876 (2007). Y. Imry, M. Strongin and C.C. Homes, Physica C **468** 288 (2008). B. Sacépé, C. Chapelier, T.I. Baturina, V.M. Vinokur, M.R. Baklanov and M. Sanquer, Phys. Res. Lett. [**101**]{}, 157006 (2008). M.V. Feigel’man, L.B. Ioffe, V.E. Kravtsov, and E.A. Yuzbashyan, Phys. Rev. Lett, **98**, 027001 (2007). M.V. Feigel’man, L.B. Ioffe, V.E. Kravtsov, Cuevas, Annals of Physics **325**, 1390 (2010). N. Trivedi, R. T. Scalettar, and M. Randeria, Phys. Rev. **B54**, R3756 (1996); A. Ghosal, M. Randeria, and N. Trivedi, ibid. **65**, 014501 (2001). K. Bouadim, Y.L. Loh, M. Randeria and N. Trivedi, Nature [**7**]{}, 884 (2011). Z. Ovdyahu J., Phys. C. Sol. State Phys., **19** 5187 (1986). Noth that $T_0$ is much smaller than the energy gap of this film as shown below. B. L. Altshuler, A.G. Aronov, M.E. Gershenson, and Yu.V. Sharvin, Sov. Sci. Rev. A [**9**]{}, 223 (1987). A. L. Efros and B. I. Shklovskii, J. Phys. C **8**, L49 (1975); B. I. Shklovskii and A. L. Efros, Electronic Properties of Doped Semiconductors (Springer, New York, 1984). R. C. Dynes, V. Narayanamurti, and J. P. Garno, Phys. Rev. Lett. [**41**]{}, 1509 (1978). Practically, the dI/dV curves represent a convolution of the DOS of the indium oxide and the Al electrodes. However, the Al gap energy is less than a fifth of the indium oxide gap. In addition, the Al was prepared such that its $T_C$ was $0.4K$ in Sample I, and $0.9K$ in sample S. The presented measurements are at higher temperatures, except for the curves in Fig. 3(b), in which the effect of the Al gap can be seen in a log representation in the 0T curve. T.I. Baturina, A.Y. Mironov, V.M. Vinokur, M.R. Baklanov, and C. Strunk, Phys. Rev. Lett. [**99**]{}, 257003 (2007).

--- abstract: 'We compute the quark-mass dependence of the baryon octet and decuplet masses within the $\chi$-MS scheme at the one-loop level. The results are confronted with recent lattice QCD simulations of the MILC collaboration. A fair reproduction of the quark-mass dependence as suggested by the lattice simulations is obtained up to pion masses of about $700$ MeV. Based on the chiral one-loop results we predict that the dependence of the baryon octet and decuplet masses on the quark-masses is discontinuous. Typically the pion-mass dependence is smooth up to about 300 MeV only. This is a consequence of self consistency imposed on the partial summation, i.e. the masses used in the loop functions are identical to those obtained from the baryon self energies. The ’mysterious’ quark-mass dependence of the $\Xi$ mass predicted by the MILC collaboration is recovered in terms of a discontinuous chiral extrapolation.' address: | Gesellschaft für Schwerionenforschung (GSI)\ Planck Str. 1, 64291 Darmstadt, Germany author: - 'A. Semke' - 'and M.F.M. Lutz' title: | Discontinuous quark-mass dependence\ of baryon octet and decuplet masses --- Introduction ============ The present-day interpretation of QCD lattice simulations requires a profound understanding of the dependence of observable quantities on the light quark masses. A powerful tool to derive such dependencies is the chiral Lagrangian, an effective field theory based on the chiral properties of QCD. The application of strict chiral perturbation theory to the SU(3) flavor sector of QCD is plagued by poor convergence properties for processes involving baryons [@Jenkins:1992; @Bernard:Kaiser:Meissner:1993; @Borasoy:Meissner:1997; @Ellis:Torikoshi:1999; @Lehnhart:Gegelia:Scherer:2005]. Thus it is important to establish partial summation schemes that enjoy improved convergence properties and that are better suited for chiral extrapolations of lattice simulations. It was suggested by Donoghue and Holstein [@Donoghue:Holstein:1998; @Donoghue:Holstein:Borasoy:1999; @Borasoy:Holstein:Lewis:Ouimet:2002] to improve the convergence properties by introducing a finite cutoff into the heavy-baryon effective field theory of Jenkins and Manohar [@Jenkins:Manohar:1991]. Another scheme was suggested in [@Lutz:2000; @Lutz:Kolomeitsev:2002; @Semke:Lutz:2005], the construction of which was guided by keeping manifestly covariance, analyticity and causality. The computation of the baryon octet and decuplet masses in this scheme was worked out recently indicating convincing convergence properties of the chiral loop expansion [@Semke:Lutz:2005]. For studies of chiral extrapolations of the nucleon mass within the chiral SU(2) Lagrangian we refer to the work by Procura, Hemmert and Weise [@Muenchen:2004]. The latter study applied the partial summation scheme of Becher and Leutwyler [@Becher:Leutwyler:1999]. It was based on simulations of the CP-PACS collaboration that used dynamical u- and d-quarks only [@CP-PACS:2003]. The application of the scheme by Becher and Leutwyler [@Becher:Leutwyler:1999] to the chiral SU(3) Lagrangian is questioned by poor convergence properties as was demonstrated by Ellis and Torikoshi at hand of the baryon octet masses [@Ellis:Torikoshi:1999]. The recent work of Pascalutsa and Vanderhagen [@Pascalutsa:Vanderhagen:2005] suggested to take serious the partial summation implied by the extended on-mass shell scheme (EOMS) introduced by Gegelia and Japaridze [@Gegelia:Japaridze:1999]. Their computation addressed so far the chiral extrapolation of the nucleon and isobar mass only. It is the purpose of the present work to confront the results of [@Semke:Lutz:2005] to recent lattice QCD simulation of the MILC collaboration [@MILC:2001; @MILC:2004], that use dynamical u-,d- and s-quarks in the staggered approximation. We do not aim at a quantitative extrapolation of the lattice simulation, which would require a continuum limit extrapolation and a quantitative control of systematic errors. Rather, we want to use the available partial lattice results to make predictions for future QCD lattice simulations and possibly obtain rough constraints on some chiral parameters. An analysis of the recent MILC results, similar in spirit, was undertaken by Frink and Meißner [@Frink:Meissner:2005] based on the cutoff scheme of Donoghue and Holstein. Adjusting the values of the $Q^2$ counter terms to the physical masses we predict a discontinuous dependence of the baryon masses on the pion mass. This is a consequence of self consistency imposed on the partial summation approach, i.e. the masses used in the loop function are identical to those obtained from the baryon self energy. The latter is a crucial requirement since the loop functions depend sensitively on the precise values of the baryon masses. Our results may explain the mysterious quark-mass dependence observed by the MILC collaboration for the $\Xi$ mass [@MILC:2001; @MILC:2004]. Chiral interaction terms ======================== We collect the terms of the chiral Lagrangian that determine the leading orders of baryon octet and decuplet self energies [@Krause:1990; @Bernard:Hemmert:Meissner:2003]. Up to chiral order $Q^2$ the baryon propagators follow from $$\begin{aligned} {\mathcal L} &=& \mathrm{tr} \,\Big( \bar{B}\,\big[i\, \slash{\partial}\,- \stackrel{\circ}{M}_{[8]}\big]\,B \Big) \nonumber \\ &-& \mathrm{tr}\,\Big(\bar{\Delta}_\mu \cdot \Big(\big[i\,\slash{\partial}\, -\stackrel{\circ}{M}_{[10]}\big]\,g^{\mu\nu} -i\,(\gamma^\mu \partial^\nu + \gamma^\nu \partial^\mu) + \gamma^\mu\,\big[i\,\slash{\partial} + \stackrel{\circ}{M}_{[10]}\big]\,\gamma^\nu \Big)\,\Delta_\nu\Big) \nonumber \\ & -& 2\,d_0\, \mathrm{tr}\Big(\bar{\Delta}_\mu \cdot \Delta^\mu \Big)\, \mathrm{tr}\Big(\chi_0\Big) -2\,d_D\, \mathrm{tr}\Big( (\bar{\Delta}_\mu \cdot \Delta^\mu)\, \chi_0\Big) \nonumber \\ &+&2\,b_0 \,\mathrm{tr}\Big(\bar{B}\,B\Big)\, \mathrm{tr}\Big(\chi_0\Big) + 2\,b_F\,\mathrm{tr}\Big(\bar{B}\,[\chi_0,B]\Big) + 2\,b_D\,\mathrm{tr}\Big(\bar{B}\,\{\chi_0,B\}\Big) \,, \nonumber\\ \nonumber\\ && \chi_0 = \left( \begin{array}{ccc} m_\pi^2 & 0 & 0 \\ 0 & m_\pi^2 & 0 \\ 0 & 0 & 2\,m_K^2-m_\pi^2 \end{array}\right)\,. \label{chiral-L}\end{aligned}$$ We assume perfect isospin symmetry through out this work. The fields are decomposed into isospin multiplets $$\begin{aligned} \Phi &=& \tau \cdot \pi + \alpha^\dagger \!\cdot \! K + K^\dagger \cdot \alpha + \eta \,\lambda_8 \;, \nonumber\\ \sqrt{2}\,B &=& \alpha^\dagger \!\cdot \! N+\lambda_8 \,\Lambda+ \tau \cdot \Sigma +\Xi^t\,i\,\sigma_2 \!\cdot \!\alpha \, ,\end{aligned}$$ with the Gell-Mann matrices, $\lambda_i$, and the isospin doublet fields $K =(K^+,K^0)^t $ and $\Xi = (\Xi^0,\Xi^-)^t$. The isospin Pauli matrices $\sigma=(\sigma_1,\sigma_2,\sigma_3)$ act exclusively in the space of isospin doublet fields $(K,N,\Xi)$ and the matrix valued isospin doublet $\alpha$, $$\begin{aligned} && \alpha^\dagger = {\textstyle{1\over\sqrt{2}}}\left( \lambda_4+i\,\lambda_5 , \lambda_6+i\,\lambda_7 \right) \;,\;\;\;\tau = (\lambda_1,\lambda_2,\lambda_3)\;. \label{def-alpha}\end{aligned}$$ The evaluation of the baryon self energies to order $Q^3$ probes the meson-baryon vertices $$\begin{aligned} {\mathcal L} &=& \frac{F}{2f}\, \mathrm{tr} \,\Big( \bar{B}\, \gamma_5 \gamma^\mu \,[\partial_\mu \Phi,\,B] \Big) + \frac{D}{2f}\, \mathrm{Tr}\,\Big( \bar{B}\, \gamma_5 \gamma^\mu \{\partial_\mu \Phi,B\} \Big) \nonumber\\ &-& \frac{C}{2f}\, \mathrm{tr}\,\Big(\bar{\Delta}_\mu \cdot (\partial_\nu \Phi)\, \big[g^{\mu\nu}-\frac{1}{2}\,Z\,\gamma^\mu \,\gamma^\nu\big]\, B + \mathrm{h.c.} \Big) \nonumber\\ &-& \frac{H}{2f} \mathrm{tr} \,\Big(\big[ \bar{\Delta}^\mu \cdot \gamma_5\,\gamma_\nu\, \Delta_\mu \big]\,(\partial^\nu \Phi)\,\Big)\,, \label{chiral-FD}\end{aligned}$$ where we apply the notations of [@Lutz:Kolomeitsev:2002]. We use $f = 92.4$ MeV in this work. The values of the coupling constants $F,D,C$ and $H$ may be correlated by a large-$N_c$ operator analysis [@Dashen; @Jenkins; @Jenkins:Manohar]. At leading order the coupling constants can be expressed in terms of $F$ and $D$ only. We employ the values for $F$ and $D$ as suggested in [@Okun; @Lutz:Kolomeitsev:2002]. All together we use $$\begin{aligned} && F = 0.45 \,, \qquad D= 0.80 \,, \qquad H= 9\,F-3\,D \,,\qquad C=2\,D \,, \label{large-Nc}\end{aligned}$$ in this work. We take the parameter $Z=0.72$ from a detailed coupled-channel study of meson-baryon scattering that was based on the chiral SU(3) Lagrangian [@Lutz:Kolomeitsev:2002]. The latter parameter is an observable quantity within the chiral SU(3) approach: it contributes at order $Q^2$ to the meson-baryon scattering amplitudes but cannot be absorbed into the available $Q^2$ counter terms [@Lutz:Kolomeitsev:2002]. Chiral loop expansion at the one-loop order =========================================== We briefly recall the results of [@Semke:Lutz:2005]. A summation approach was defined by performing a chiral loop expansion rather than a strict chiral expansion: for a given truncation of the relativistic chiral Lagrangian we take the loop expansion that is defined in terms of the approximated Lagrangian seriously. The number of loops we would consider is correlated with the chiral order to which the Lagrangian is constructed. A renormalization based on the Passarino-Veltman reduction was devised that installs the correct minimal chiral power of a given loop function. The residual dependence on the renormalization scales is used to monitor the convergence properties of the expansion and therewith estimate the error encountered at a given truncation. The loop contribution to the baryon octet and decuplet masses read: $$\begin{aligned} &&\Delta M^{\rm loop}_{B \in [8]} = \sum_{Q\in [8], R\in [8]} \left(\frac{G_{QR}^{(B)}}{2\,f} \right)^2 \Bigg\{ \frac{M_R^2-M_B^2}{2\,M_B}\, \bar I_Q \nonumber\\ && \quad - \frac{(M_B+M_R)^2}{E_R+M_R}\, p^2_{QR}\, \Big(\bar I_{QR} + \frac{\bar I_Q}{M_R^2-m_Q^2}\,\Big)\Bigg\} \label{octet}\\ && \qquad \quad \;\,\,+\sum_{Q\in [8], R\in [10]} \left(\frac{G_{QR}^{(B)}}{2\,f} \right)^2 \, \Bigg\{ \Bigg( \frac{(M_R-M_B)\,(M_R+M_B)^3+m_Q^4}{12\,M_B\,M^2_R}\, \nonumber\\ && \quad - \frac{(Z\,(Z+2)-5)\,M_B^2+2\,(2\,Z\,(Z-1)-3)\,M_R\,M_B+2\,M_R^2}{12\,M_B\,M_R^2}\,m_Q^2\Bigg)\, \bar I_Q \nonumber\\ && \quad - \frac{2}{3}\,\frac{M_B^2}{M_R^2}\,\big(E_R+M_R\big)\,p_{QR}^{\,2}\, \Big(\bar I_{QR} + \frac{\bar I_Q}{M_R^2-m_Q^2}\Big) \Bigg\} \,, \nonumber\end{aligned}$$ and $$\begin{aligned} &&\Delta M^{\rm loop}_{B\in [10]} = \sum_{Q\in [8], R\in [8]} \left(\frac{G_{QR}^{(B)}}{2\,f} \right)^2 \Bigg\{ \Bigg( \frac{(M_R-M_B)\,(M_R+M_B)^3+m_Q^4}{24\,M^3_B}\, \nonumber\\ && \quad - \frac{3\,M_B^2+2\,M_R\,M_B+2\,M_R^2}{24\,M^3_B}\,m_Q^2\Bigg)\, \bar I_Q \nonumber\\ && \quad -\frac{1}{3}\,\big( E_R +M_R\big)\,p_{QR}^{\,2}\, \Big(\bar I_{QR}+ \frac{\bar I_Q}{M_R^2-m_Q^2}\Big) \Bigg\} \label{decuplet} \\ && \qquad \quad \;\,\,+\sum_{Q\in [8], R\in [10]} \left(\frac{G_{QR}^{(B)}}{2\,f} \right)^2 \, \Bigg\{ \Bigg( \frac{(M_B+M_R)^2\,m_Q^4}{36\,M_B^3\,M_R^2} \nonumber\\ &&\quad +\frac{3\,M_B^4-2\,M^3_B\,M_R+3\,M_B^2\,M_R^2-2\,M_R^4}{36\,M_B^3\,M_R^2} \,m_Q^2 \nonumber\\ &&\quad +\frac{M_R^4+M_B^4+12\,M_R^2\,M_B^2-2\,M_R\,M_B\,(M_B^2+M_R^2)}{36\,M^3_B\,M^2_R}\, (M^2_R-M^2_B) \Bigg)\,\bar I_Q \nonumber\\ && \quad -\frac{(M_B+M_R)^2}{9\,M_R^2}\,\frac{2\,E_R\,(E_R-M_R)+5\,M_R^2}{E_R+M_R}\, p_{QR}^{\,2}\,\Big(\bar I_{QR}(M_B^2)+ \frac{\bar I_Q}{M_R^2-m_Q^2}\Big) \Bigg\}\,, \nonumber\end{aligned}$$ where $$\begin{aligned} && \bar I_Q =\frac{m_Q^2}{(4\,\pi)^2}\, \ln \left( \frac{m_Q^2}{\mu_{\,UV}^2}\right)\,, \nonumber\\ && \bar I_{Q R}=\frac{1}{16\,\pi^2} \left\{ \frac{2\,\pi\,\mu_{IR}}{M_R}+\left(\frac{1}{2}\,\frac{m_Q^2+M_R^2}{m_Q^2-M_R^2} -\frac{m_Q^2-M_R^2}{2\,M_B^2} \right) \,\ln \left( \frac{m_Q^2}{M_R^2}\right) \right. \nonumber\\ && \;\quad \;\,+\left. \frac{p_{Q R}}{M_B}\, \left( \ln \left(1-\frac{M_B^2-2\,p_{Q R}\,M_B}{m_Q^2+M_R^2} \right) -\ln \left(1-\frac{M_B^2+2\,p_{Q R}\,M_B}{m_Q^2+M_R^2} \right)\right) \right\}\;, \nonumber\\ && p_{Q R}^2 = \frac{M_B^2}{4}-\frac{M_R^2+m_Q^2}{2}+\frac{(M_R^2-m_Q^2)^2}{4\,M_B^2} \,,\qquad E_R^2=M_R^2+p_{QR}^2 \,.\end{aligned}$$ The sums in (\[octet\], \[decuplet\]) extend over the intermediate Goldstone bosons ($Q\in[8]$) baryon octet ($R\in [8]$) and decuplet states ($R\in[10]$). The various coupling constants $G_{QR}^{(B)}$ are determined by the parameters $F,D,C,H$. They are listed in [@Semke:Lutz:2005]. We emphasize that (\[octet\], \[decuplet\]) depend on the physical meson and baryon masses $m_Q$ and $M_{R}$. This defines a self consistent summation since the masses of the intermediate baryon states in (\[octet\], \[decuplet\]) should match the total masses. At order $Q^3 $ the latter are the sum of the tree-level contributions linear in the parameter $b_{0},b_D, b_F$ or $d_0, d_D$ and the loop contribution (\[octet\], \[decuplet\]). The mesonic tadpole $\bar I_Q$ enjoys a logarithmic dependence on the ultraviolet renormalization scale $\mu_{UV}$ and the one-loop master function $\bar I_{QR}$ a linear dependence on the infrared renormalization scale $\mu_{IR} $. By construction the results (\[octet\], \[decuplet\]) are necessarily consistent with all chiral Ward identities as discussed in [@Semke:Lutz:2005]. As emphasized by the authors the parameters $b_0, b_D$, $b_F$ and $d_0, d_D$ are strongly dependent on the infrared scale $\mu_{IR}$. Applying a further chiral expansion to (\[octet\], \[decuplet\]) it was demonstrated that the physical masses are scale independent as they should be. However, the convergence properties reflect the choice of the renormalization scales. Good convergence properties can only be expected for natural values thereof. For the infrared scale $\mu_{IR} \sim Q$ a natural window 350 MeV $< \mu_{IR} < $ 550 MeV was suggested in [@Semke:Lutz:2005]. Keeping the partial summation as defined by (\[octet\], \[decuplet\]) a residual dependence on the renormalization scales remains. This is analogous to the residual cutoff dependence of the scheme of Donoghue and Holstein [@Donoghue:Holstein:1998]. As long as such dependencies are small and decreasing as higher order terms are included they do not pose a problem, rather, they offer a convenient way to estimate the error encountered at a given truncation. Quark-mass dependence of the baryon masses ========================================== We discuss the implications of (\[octet\], \[decuplet\]) for chiral extrapolations of lattice simulations of the baryon masses [@MILC:2001; @MILC:2004]. Since we are not aiming at a chiral extrapolation of the lattice simulations down to physical quark masses we adjust part of the parameters to empirical data directly. The goal of the present study is a qualitative understanding of the quark-mass dependence of the baryon masses. For given values of the infrared and ultraviolet renormalization scales the parameters $b_D, b_F$ and $d_D$ are fitted to the mass differences of the octet states and decuplet states. The absolute mass scale of the octet and decuplet states can be reproduced by appropriate values of the bare baryon masses. This procedure leaves undetermined the two parameters $b_0$ and $d_0$. Good representations of the physical baryon masses can be obtained for a wide range of the latter parameters. The parameter $b_0$ may be used to dial a given pion-nucleon sigma term at physical pion masses. Similarly the unknown parameter $d_0$ may be determined as to reproduce a given pion-delta sigma term. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- $\mu_{IR}=350$ MeV $ \mu_{IR}=450$ MeV $ \mu_{IR}=550$ MeV ----------------------------- -------------------------------------------------- -------------------------------------------------- -------------------------------------------------- $b_0\; \mathrm{[GeV^{-1}]}$ $-0.89$ $-0.63$ $-0.38$ $b_D\; \mathrm{[GeV^{-1}]}$ $+0.29$ $+0.19$ $+0.10$ $b_F\; \mathrm{[GeV^{-1}]}$ $-0.34$ $-0.25$ $-0.15$ $d_0\; \mathrm{[GeV^{-1}]}$ $-0.22$ $-0.15$ $-0.08$ $d_D\; \mathrm{[GeV^{-1}]}$ $-0.35$ $-0.30$ $-0.24$ $M_N$ \[MeV\] $\phantom{1}750+\phantom{1}310-\phantom{1}121 $ $ $ \phantom{1}875+\phantom{1}153-\phantom{11}89 $ \phantom{1}813+\phantom{1}232-\phantom{1}105 $ $=\phantom{1}939 $ $ =\phantom{1}939 $ $ =\phantom{1}939 $ $M_\Lambda$ \[MeV\] $\phantom{1}750+\phantom{1}536-\phantom{1}150 $ $ $ \phantom{1}875+\phantom{1}260-\phantom{111}9 $ \phantom{1}813+\phantom{1}398-\phantom{111}79 $ $=1136 $ $ =1131 $ $ =1126 $ $M_\Sigma$ \[MeV\] $\phantom{1}750+\phantom{1}882-\phantom{1}425 $ $\phantom{1}813+\phantom{1}630-\phantom{1}239 $ $\phantom{1}875+\phantom{1}379-\phantom{11}54 $ $=1207 $ $=1203 $ $=1200 $ $M_\Xi$ \[MeV\] $\phantom{1}750+\phantom{1}934-\phantom{1}361 $ $\phantom{1}813+\phantom{1}680-\phantom{11}171 $ $\phantom{1}875+\phantom{1}427+\phantom{11}19 $ $=1323 $ $=1321 $ $=1320 $ $M_\Delta$ \[MeV\] $1082+\phantom{1}241-\phantom{11}91 $ $1108+\phantom{1}164-\phantom{11}39$ $1133 +\phantom{11}87+\phantom{11}12$ $=1232 $ $ =1232 $ $ =1232 $ $M_\Sigma$ \[MeV\] $1082+\phantom{1}347-\phantom{11}49 $ $ $1133+\phantom{1}160+\phantom{11}83 $ 1108+\phantom{1}253+\phantom{11}17 $ $=1380 $ $ =1378 $ $ =1376 $ $M_\Xi$ \[MeV\] $1082+\phantom{1}453-\phantom{111}4 $ $1108+\phantom{1}343+\phantom{11}78 $ $1133+\phantom{1}233+\phantom{1}160 $ $=1530 $ $=1528 $ $=1526 $ $M_\Omega$ \[MeV\] $1082+\phantom{1}558+\phantom{11}34 $ $1108+\phantom{1}432+\phantom{1}134$ $1133+\phantom{1}305+\phantom{1}235$ $=1674 $ $=1674 $ $=1674 $ $\sigma_{\pi N}$ \[MeV\] 52.4 53.9 55.7 $\sigma_{K^- p}$ \[MeV\] 384.0 380.1 386.3 $\sigma_{K^- n}$ \[MeV\] 359.4 354.8 361.5 -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- : The parameters are fitted so that (\[octet\], \[decuplet\]) reproduces the baryon masses at physical pion masses as well as the SU(3) limit values $M_{[8]} \simeq 1575$ MeV and $M_{[10]}\simeq 1710$ MeV at $m_{\pi}\simeq 690$ MeV. We use $\mu_{UV}=800$ MeV. The masses are decomposed into their chiral moments. []{data-label="tab:parameter"} In this work $b_0$ and $d_0$ are adjusted as to reproduce the baryon octet and decuplet masses in the SU(3) limit at $m_{\pi}\simeq 690$ MeV. The MILC simulations suggest the values $M_{[8]} \simeq 1575$ MeV and $M_{[10]}\simeq 1710$ MeV. This procedure is biased to the extent that it assumes that the chiral one-loop results will be applicable at such high quarks masses. On the other hand as long as there are no continuum limit results of the MILC collaboration available this is an economical way to minimize the influence of lattice size effects. The latter are expected to be smaller at large quark masses. The procedure may be justified in retrospect if it turns out that the extrapolation recovers the behavior predicted by the lattice simulation. The parameters used in this work are collected in Tab. \[tab:parameter\] together with the implied masses of the baryon octet and decuplet states. A fair representation of the physical baryon masses is obtained. For simplicity the values quoted in Tab. \[tab:parameter\] correspond to a computation where the intermediate baryon masses are put to their empirical values. Since the resulting masses are very close to the physical masses this is well justified. As discussed in detail in [@Semke:Lutz:2005] the parameters $b_{0,D,F}$ and $d_{0,D}$ show a strong dependence on the infrared scale $\mu_{IR}$. In contrast the physical baryon masses suffer from a weak dependence only. For natural values of the infrared scale the chiral expansion appears well converging as indicated by the decomposition of the baryon masses into their moments. We emphasize that the values of the pion-nucleon sigma term shown in Tab. \[tab:parameter\] are an immediate consequence of the MILC simulation of the SU(3) symmetric point defined by $m_\pi\simeq 690$ MeV as discussed above. No further results of the MILC simulation are used. Taking the residual scale dependence of the pion-nucleon sigma term as a naive error estimate we obtain $\sigma_{\pi N} = 54 \pm 2$ MeV. The latter value is somewhat larger than the canonical value of Gasser, Leutwyler and Sainio $\sigma_{\pi N} \simeq 45$ MeV [@Gasser:Leutwyler:Sainio:1991]. Our value is, however, quite compatible with a recent analyses of Frink and Meissner who suggest $\sigma_{\pi N} \simeq 52$ MeV [@Frink:Meissner:2005]. Also Pascalutsa and Vanderhagen [@Pascalutsa:Vanderhagen:2005] arrive at a somewhat larger value $\sigma_{\pi N} \simeq 57$ MeV based on a chiral SU(2) analysis of the MILC simulation points. ![The pion mass dependence of baryon octet and decuplet masses predicted by the chiral loop expansion taking the parameters of Tab. \[tab:parameter\]. The lines represent the masses for the infrared scale put at $\mu_{IR}=450$ MeV. The solid squares are the simulation points of the MILC collaboration. []{data-label="fig:1"}](fig-1.eps){width="15cm"} We turn to the pion-mass dependence of the baryon octet and decuplet masses. Our results are collected in Fig. \[fig:1\] based on the parameters of Tab. \[tab:parameter\]. We show results for the particular choice $\mu_{IR}= 450$ MeV. It should be emphasized that the baryon masses are a solution of a set of coupled and non-linear equations in the present scheme. This is a consequence of self consistency imposed on the partial summation approach. The latter is a crucial requirement since the loop functions depend sensitively on the precise values of the baryon masses. As a consequence of the non-linearity for a given parameter set there is no guarantee for a unique solution to exist, nor that solutions found are continuous in the quark masses. Indeed as illustrated by Fig. 1 the pion-mass dependence predicted by the chiral loop expansion is quite non-trivial exhibiting various discontinuities. Upon inspecting the energy dependence of the baryon self energies this is readily understood: the baryon self energies, as implied by (\[octet\], \[decuplet\]), allow in certain cases for multiple solutions of the mass equation, $$\begin{aligned} {\FMslash p}={\stackrel{\circ}{M}}+\Sigma_B({\FMslash p}) \,, \label{mass-equation}\end{aligned}$$ even though the self consistency requirement singles out a particular solution at ${\FMslash p}=M_B$. Such a scenario appears inconsistent at first: the meson baryon loop functions have to be evaluated with all states provided by (\[mass-equation\]). In contrast the computation of the meson-baryon loop functions (\[octet\], \[decuplet\]) takes the existence of uniquely defined baryon states for granted. In addition it is assumed that the mass poles of the baryon propagators have residuum one. Thus our computation considers implicitly additional counter terms of the form: $$\begin{aligned} \Sigma_{B}^{wave-function}({\FMslash p}) = \zeta_B\,({\FMslash p}-M_B )\,, \label{wave-function}\end{aligned}$$ where $M_B$ denotes the baryon masses that solve the self consistent system. If the term (\[wave-function\]) is added to the self energies it does not affect the solutions of the self consistent system. However, it does affect the energy dependence of the final baryon self energy. By adjusting the parameters $\zeta_B$, it is possible to arrive at uniquely defined baryon states. Within the scheme described above it is intuitive that for a given set of parameters multiple solutions of the self consistent system may exist. A priori it is not evident which of the possible solutions one should select. Our strategy is to continuously evolve the solution established at physical pion masses to larger pion masses. Similarly we evolve down to smaller pion masses the solution that reproduces the lattice simulations at the SU(3) symmetric point defined by $m_\pi\simeq $ 690 MeV. Both solutions reach end points beyond which we do not find any continuation. In the region of intermediate pion masses we, however, find another solution, suggesting a discontinuous dependence of the baryon masses on the quark masses of QCD. One may expect that the physical solutions are those of lowest mass always. The three branches discussed above are shown in Fig. 1 and confronted with the simulation points of the MILC collaboration [@MILC:2001; @MILC:2004]. It is striking to see that we reproduce the ’mysterious’ pion-mass dependence of the $\Xi$ mass, i.e. the quite flat behavior which does not seem to smoothly approach the physical mass. Given the present uncertainties from finite lattice spacing, the staggered approximation and the theoretical uncertainties implied by higher order contributions, we would argue that we arrive at a fair representation of the lattice simulation points for all baryons with some reservation concerning the nucleon. We note that the parameter set used before in [@Semke:Lutz:2005] leads to results qualitative similar to those shown in Fig. 1 for the baryon octet masses. While the description of the nucleon is improved showing no discontinuity at large pion masses, the sizeable jump of the $\Xi$ mass at small pion masses remains. However, the MILC simulation points of the isobar mass disfavor the latter choice. An overestimate of the lattice simulations of the isobar mass at larger quark masses would be the consequence. Incorporating the many $Q^4$ counter terms offered by the chiral Lagrangian it is reasonable to expect that the latter will further improve the picture. However, as long as there is no detailed analysis available that performs the continuum limit and estimates the uncertainty from the staggered fermion approximation there is not much point considering the $Q^4$ counter terms. Summary ======= We evaluated the baryon octet and decuplet self energies at the one-loop level applying the relativistic chiral Lagrangian. Adjusting the $Q^2$ counter terms an excellent representation of the physical baryon octet and decuplet masses is obtained. The size of the pion-nucleon sigma term was estimated by including in the fit the baryon octet and decuplet masses at the SU(3) symmetric point defined by $m_\pi\simeq 690$ MeV. For the latter the MILC collaboration suggests the values $M_{[8]} \simeq 1575$ MeV and $M_{[10]}\simeq 1710$ MeV. As a result we predict $\sigma_{\pi N} \simeq 54$ MeV. The kaon-nucleon sigma terms are $\sigma_{K^-p} \simeq 380$ MeV and $\sigma_{K^-n}\simeq 355 $ MeV. Given this scenario the pion-mass dependence of the baryon octet and decuplet masses were evaluated. The latter are a solution of a set of coupled and non-linear algebraic equations. This is a direct consequence of self consistency imposed on the partial summation, i.e. the masses used in the loop functions are identical to those obtained from the baryon self energies. As a striking consequence we predict a discontinuous dependence of the baryon masses on the pion mass. Typically the baryon masses jump at pion masses as low as 300 MeV. Most spectacular is the behavior of the $\Xi$ mass. At small pion masses it decreases with increasing pion masses. At a critical pion mass of about 300-400 MeV it jumps up to a value amazingly close to the prediction of the MILC collaboration. For all baryon masses for which the MILC collaboration published simulation points our results are reasonably close to the lattice estimates, given the present uncertainties from finite lattice spacing, the staggered approximation and the theoretical uncertainties implied by higher order contributions to the baryon self energies. Acknowledgments {#acknowledgments .unnumbered} =============== M.F.M.L. acknowledges fruitful discussions with B. Friman and Ch. Redlich. [9]{} E. Jenkins, Nucl. Phys. [**B 368**]{} (1992) 190. V. Bernard, N. Kaiser and U.-G. Meißner, Z. Phys. [**C 60**]{} (1993) 111. B. Borasoy, U.-G. Meißner, Ann. Phys. [**254**]{} (1997) 192. P.J. Ellis and K. Torikoshi, Phys. Rev. [**C 61**]{} (1999) 015205. B.C. Lehnhart, J. Gegelia and S. Scherer, J. Phys. [**G 31**]{}(2005) 89. J.F. Donoghue and B.R. Holstein, Phys. Lett. [**B 436**]{} (1998) 331. J.F. Donoghue, B.R. Holstein and B. Borasoy, Phys. Rev. [**D 59**]{} (1999) 036002. B. Borasoy et al., Phys. Rev. [**D 66**]{} (2002) 094020. E. Jenkins and A. Manohar, Phys. Lett. [**B 255**]{} (1991) 558. M. Lutz, Nucl. Phys. [**A 677**]{} (2000) 241. M.F.M. Lutz and E. E. Kolomeitsev, Nucl. Phys. [**A 700**]{} (2002) 193. A. Semke and M.F.M. Lutz, nucl-th/0511061. C. Bernard et al., Phys. Rev. [**D 64**]{} (2001) 054506. C. Aubin et al., Phys. Rev. [**D 70**]{} (2004) 094505. M. Frink and U.-G. Meißner, hep-lat/0501024. M. Procura, T.R. Hemmert and W. Weise, Phys. Rev. [**D 69**]{} (2004) 034505. T. Becher and H. Leutwyler, Eur. Phys. J. [**C 9**]{} (1999) 643. CP-PACS Collaboration, A. Ali Khan et al., Phys. Rev. [**D 65**]{} (2002) 054505; Erratum-ibid. [**D 67**]{} (2003) 059901. V. Pascalutsa and M. Vanderhaeghen, hep-ph/0511261. J. Gegelia and G. Japaridze, Phys. Rev. [**D 60**]{} (1999) 114038. A. Krause, Helv. Phys. Acta [**63**]{} (1990) 3. V. Bernard, T.R. Hemmert and U.-G. Meißner, Phys. Lett. [**B 565**]{} (2003) 137. R. F. Dashen, E. Jenkins and A.V. Manohar, Phys. Rev. [**D 51**]{} (1995) 3697. E. Jenkins, Phys. Rev. [**D 53**]{} (1996) 2625. E. Jenkins and A. Manohar, Phys. Lett. [**B 259**]{} (1991) 353. L.B. Okun, [*Leptons and Quarks*]{}, Amsterdam, North-Holland (1982). J. Gasser, H. Leutwyler, M.E. Sainio, Phys. Lett. [**B 253**]{} (1991) 252.

--- author: - 'Eduardo L. Martín, Antonio Magazzù[^1], Xavier Delfosse, Robert D. Mathieu' date: 'Received /Accepted' title: 'The pre-main sequence spectroscopic binary UZ Tau East: improved orbital parameters and accretion phase dependence [^2] ' --- Introduction ============ UZ Tau was one of the first 11 stars originally identified as members of the class of T Tauri variable stars, and it was also noticed to be a wide pair with angular separation 36 (Joy & van Biesbroek 1944). UZ Tau East and West are a pair of classical T Tauri stars (CTTSs), i.e. very young stars with an emission line spectrum that indicates active mass accretion from a circumstellar disk (Appenzeller & Mundt 1989; Bertout 1989). UZ Tau E is a CTTS with spectral type M1, strong H$_\alpha$ emission, strong excess continuum emission at infrared and submillimiter wavelengths (Jensen, Mathieu & Fuller 1996) and a Keplerian disk (Simon et al. 2000). UZ Tau W is itself a pair with an angular separation of 034 (47.6 AU at 140 pc; Simon et al. 1995). UZ Tau W does not have as strong an infrared excess as E, and it also has much less submillimiter emission (Jensen et al. 1996). Prior to our observations, UZ Tau E was thought to be a single star. In 1994 we started high-resolution spectroscopic observations of this star and we noticed the radial-velocity variability. Follow-up observations were obtained in several runs. An early report of the orbital parameters of UZ Tau E was given by Mathieu, Martín & Magazzù (1996). Prato et al. (2002) obtained radial-velocity measurements of both components of UZ Tau E using near-infrared spectra, and estimated the mass ratio of the system. After AK Sco (Andersen et al. 1989), GW Ori (Mathieu et al. 1991), DQ Tau (Mathieu et al. 1997), V4046 Sgr (Quast et al. 2000), and RX J0530.7-0434 (Covino et al. 2001), UZ Tau E is only the fifth CTTS known to be a spectroscopic binary. Another CTTS spectroscopic binary, namely KH 15D, has recently been reported by Johnson et al. (2004). In this paper we present our spectroscopic observations of UZ Tau E obtained from 1994 to 1996. In Sect. 2 we describe the observations. Section 3 contains the main results, including the orbital solution found using our data and those in the literature. Section 4 provides a discussion of the properties of this spectroscopic binary. Observations and Data Reduction. ================================ Most of the spectroscopic observations presented in this paper were carried out with the 2.5-meter Isaac Newton Telescope (INT) equipped with the Intermediate Dispersion Spectrograph (IDS). We used the cameras and gratings listed in Table 1 which provided the full width half maximum (FWHM) spectral resolutions given in the same table. We placed both UZ Tau E and W along the same IDS long slit. The initial identification of radial-velocity variations was made in the analysis of the IDS spectra. In order to confirm the results and to improve the orbital parameters, we supplemented the INT data with observations obtained with the Hamilton echelle at the Shane 3-meter telescope at Lick Observatory. We measured the radial velocity of UZ Tau E in a spectrum obtained in 1988, which has been used to derive veiling by Basri & Batalha (1990) and to measure the strength of the lithium resonance line by Basri, Martín & Bertout (1991). Additional observations were obtained at the ESO 3.6-meter telescope in January 1996, using the CASPEC echelle spectrograph, and again at Lick with the Hamilton spectrograph in March 1996. Table 1 gives the log of the spectroscopic observations. IDS data reduction was done with IRAF[^3] routines in the same manner as described in Martín et al. (1992). Hamilton data reduction was made with IDL routines as explained in Basri & Batalha (1990). CASPEC data reduction was carried out with the echelle package in IRAF. The data were unbiased and flatfielded before extracting the spectrum. Wavelength calibration was performed using a ThAr lamp spectrum obtained the same night. Results. ======== Using the IDS spectra we measured radial velocities for UZ Tau E by cross-correlation with the spectra of UZ Tau W observed simultaneously (the slit was aligned with the visual binary axis). After several tests, we found that the spectral region between 640 nm and 650 nm gave the most precise radial velocities. Figure 1 displays one example of the spectral region that we used for cross-correlating the spectra of UZ Tau E and W. Table 2 lists the radial velocities measured in our spectra. These velocities are heliocentric, and relative to an assumed heliocentric radial velocity for UZ Tau W of 19.3 km s$^{-1}$. The radial-velocity error bars were obtained from the cross-correlation fit given by the IRAF task fxcor. {width="\textwidth"} We also measured H$\alpha$ equivalent widths and optical veiling in the IDS spectra, and we provide the values in Table 2. We did not measure these quantities in the echelle spectra (from CASPEC and Hamilton) because their resolution is much higher than that of IDS. It is well know that equivalent width measurements are affected by the spectral resolution. Moreover, we only have a few echelle spectra, and hence they do not contribute significantly to the study of the variability of H$_\alpha$ emission and veiling as a function of orbital phase. Nevertheless, in Table 2 we do quote the veiling measured by Basri & Batalha (1989). Our H$\alpha$ equivalent widths are estimated to be accurate within 10% at a 3 $\sigma$ confidence level. The veiling at a given wavelength (denoted as v$_{\lambda}$) is defined as the ratio of the excess flux to the photospheric flux. As a consequence, we have measured veiling as the ratio of two equivalent widths, i.e. v$_{\lambda}$=EW(LiI)$_{\rm phot}$/EW(LiI)$_{\rm obs}$ - 1, where EW(LiI)$_{\rm obs}$ is the observed equivalent width of the lithium doublet at 6708 Å. We could not obtain a value of veiling for the spectrum at JD=49972.730 using this method, because the spectral range of it did not include the lithium feature. Thus, the value given in Table 2 was obtained from the same spectral region as was used for the radial velocity measurements. We artificially veiled the spectrum of UZ Tau E obtained at JD=50040.528 by a small amount until we found a good match. We used EW(LiI)$_{\rm phot}$=720 mÅ, obtained by Basri et al. (1991). This value is close to the photospheric lithium equivalent width typical of early M-type T Tauri stars (Martín 1997). While the absolute veiling depends on this assumption, we note that our main interest is to measure the variability of the relative veiling as a function of orbital phase. The error bars in the veiling values were obtained in the same manner as explained in Basri et al. (1991). We combined our radial-velocity data with those of Prato et al. (2002) to improve the orbital parameters of the binary system. We used the ORBIT program (Forveille et al. 1999) to find the best orbital solution to the data, which has a chi-square value of 192. The results of the best orbital fits to the data are given in Table 3 and are shown in Fig. 2. We note that the radial velocity point of Prato et al. (2002) obtained at JD=2,452,311.2 significantly increased the error bars in the orbital elements and decreased the quality of the orbital fit. When this point was removed from the dataset, the chi-square value of the best orbital fit was 87, and the estimated masses of the components were changed significantly. In particular, the mass ratio q changed from 0.26 to 0.31. More radial velocity points for the secondary of UZ Tau E are clearly needed to improve the confidence of the mass estimates. {width="\textwidth"} Discussion. =========== The emission line spectrum, optical veiling, spectral energy distribution and millimeter emission of UZ Tau E are different facets of an accretion disk. Valenti et al. (1993) derived a mass accretion rate of 3$\times$10$^{-7}$ M$_\odot$ yr$^{-1}$ using blue low-resolution spectroscopy. In order to maintain such an accretion rate, circumbinary material must be flowing across the binary orbit. There is no evidence in the spectral energy distribution for a gap in the accretion disk, but the gap may be masked by a small amount of inflow material (Jensen et al. 1996). These authors infer a circumbinary disk mass of 0.06 M$_\odot$ from the millimeter continuum emission. The CTTS properties of UZ Tau E can be explained with the model developed by Artymowicz & Lubow (1996) where accretion streams across the binary orbit allow large accretion rates and suppress orbital shrinkage by the flow of high-angular-momentum circumbinary material onto the stars. This model has also been invoked to explain the properties of the CTTS spectroscopic binary DQ Tau. A key prediction of this model is that accretion onto the stars in a binary system is modulated with the orbital phase. For high eccentricy binaries, most of the accretion takes place during periastron passage, but not for orbits with lower eccentricities. Observational support for this prediction was found in DQ Tau, where outbursts of emission lines and veiling were seen during periastron passages, although they were not present one third of the time (Basri et al. 1997; Mathieu et al. 1997). Moreover, in DQ Tau there is a permanent level of accretion going on at all orbital phases. {width="\textwidth"} {width="\textwidth"} It is interesting to compare UZ Tau E and DQ Tau. The orbital periods are similar (18.9 and 15.8 days, respectively), but the eccentricities (0.14 and 0.55, respectively) and mass ratios (0.26 and 0.97, respectively) are very different. A comparison between the UZ Tau E and DQ Tau phased H$_\alpha$ emission datasets (without veiling corrections) is presented in Fig. 3. UZ Tau E has a level of permanent mass accretion comparable to DQ Tau as judged from the H$_\alpha$ equivalent width, which is on average 51.7 Å, while in DQ Tau it is 79 Å. However, we have not observed in UZ Tau E any outburst as strong as those seen in DQ Tau. We have one observation at JD=2,449,641.67 close to periastron passage (phase = 0.88) where we find the strongest H$_\alpha$ emission and largest veiling in our dataset. On the other hand, we do not witness any significant emission or veiling enhancement in the periastron passage from JD=2,450,039.49 to JD=2,450,042.66. In Fig. 4, we present a comparison of the dependence of veiling on orbital phase between UZ Tau E and DQ Tau. The veiling values for DQ Tau have been taken from Basri et al. (1997). The enhanced accretion events sometimes seen near periastron in DQ Tau are not present in our data of UZ Tau E. We conclude that the orbital modulation of accretion in UZ Tau E, if present, is not as clear as in DQ Tau. This could be due to the lower eccentricity of UZ Tau E or to the lower mass ratio of the two components, or to a combination of both parameters. Figure 2 of Artimowicz & Lubow (1996) shows that their hydrodynamical models predict that accretion is much more enhanced at periastron passage in a binary with e=0.5 than in a binary with e=0.1. Hence, the difference in behaviour between these two CTTS binaries is qualitatively consistent with the models. Few other studies of the orbital dependence of accretion diagnostics in pre-main sequence spectroscopic binaries exist. Alencar et al. (2003) failed to find evidence for enhanced accretion near periastron in the pre-main sequence spectroscopic binary AK Sco which has an eccentricity of e=0.47 and an orbital period of 13.6 days. Stempels & Gahm (2004) also did not find periastron activity in V4046 Sgr, a 2.4 day binary with a circular orbit. Future studies of UZ Tau E are needed to improve the phase coverage of the accretion diagnostics, and to check the repeatability in its behaviour. More CTTS spectroscopic binaries must be studied to determine observationally the role of the binary parameters on the time dependence of the accretion from the circumbinary disk. Enhanced periastron activity has clearly been seen in DQ Tau only, amongst the 4 pre-main sequence spectroscopic binaries observed so far. DQ Tau has the highest eccentricity among them, and hence it seems that enhanced accretion at periastron may occur only for extremely high eccentricities (e$>$0.5). We thank Juan Manuel Alcalá for observing UZ Tau E with CASPEC and Gibor Basri for observing it at Lick and sharing the radial velocity measurements with us. We thank the observatory staff at La Palma observatory for assistance during the observations. Alencar, S.H.P., Melo, C.H.F., Dullemond, C.P. et al. 2003, A&A, 409, 1037 Andersen, J., Lindgren, H., Hazen, M.L., & Mayor, M. 1989, A&A, 219, 142 Appenzeller I., & Mundt R., 1989, A&A Rev. 1, 291 Artymowicz, P. & Lubow, S.H. 1996, ApJ, 467, L77 Basri, G. & Batalha, C. 1990, ApJ, 363, 654 Basri, G., Martín E.L. & Bertout, C. 1991, A&A, 252, 625 Basri, G., Johns-Krull, C.M.J. & Mathieu, R.D. 1997, AJ, 114, 781 Bertout, C. 1989, ARA&A, 27, 351 Covino, E., Melo, C., Alcalá, J.F. et al. 2001, A&A, 375, 130 Forveille, T., Beuzit, J.-L., Delfosse, X. et al. 1999, A&A, 351, 619 Jensen, E.L.N., Mathieu, R.D., & Fuller, G. A. 1996, AJ, 458, 312 Johnson, J.A., Marcy, G.W., Hamilton, C.M., Herbst, W., Johns-Krull, C.M. 2004, AJ, 128, 1265 Joy, A. H., van Biesbroeck, G. 1944, PASP, 56, 123 Martín E.L., Magazzù, A., & Rebolo, R. 1992, A&A, 257, 186 Martín E.L. 1997, A&A, 321, 492 Mathieu, R.D., Adams, F.C., & Latham, D.W. 1991, AJ, 101, 2184 Mathieu, R.D., Martín E.L., & Magazzù, A. 1996, BAAS, 188, 6005 Prato, L., Simon, M., Mazeh, T., Zucker, S., & McLean, I.S. 2002, ApJ, 579, L99 Quast, G.R., Torres, C.A.O., de La Reza, R., da Silva, L. 2000, in IAU Symp. 200 Poster Proceedings, Birth and Evolution of Binary Stars, ed. B. Reipurth & H. Zinnecker (Postdam: Astrophys. Inst.), 28 Simon, M., Dutrey, A., & Guilloteau, S. 2000, ApJ, 545, 1034 Simon, M., Ghez, A.M., Leinert, Ch. et al. 1995, ApJ, 443, 625 Stempels, H.C., & Gahm, C. 2004, A&A, 421, 1159 Valenti, J.A., Basri, G., & Johns, C.M. 1993, AJ, 106, 2024 UT date JD (+2,400,000) Instrument Exp. (s) Wav. range (nm) FWHM (Å) -------------- ----------------- ------------------ ---------- ----------------- ---------- 30 Nov 1988 47495.735 Hamilton echelle 2000 510-870 0.2 16 Oct 1994 49641.672 IDS/500mm/R600 420 593-674 2.4 6 Aug 1995 49936.723 IDS/235mm/R1200 1200 653-673 1.6 7 Aug 1995 49937.730 IDS/235mm/R1200 1200 653-673 1.6 8 Aug 1995 49938.727 IDS/235mm/R1200 1200 653-673 1.6 9 Aug 1995 49939.730 IDS/235mm/R1200 1200 653-673 1.6 12 Sep 1995 49972.730 IDS/500mm/R1200 600 625-665 0.8 19 Nov 1995 50039.495 IDS/500mm/R1200 1000 635-674 0.8 20 Nov 1995 50040.528 IDS/500mm/R1200 900 635-674 0.8 21 Nov 1995 50041.415 IDS/500mm/R1200 900 635-674 0.8 21 Nov 1995 50041.696 IDS/500mm/R1200 900 635-674 0.8 22 Nov 1995 50042.433 IDS/500mm/R1200 900 635-674 0.8 22 Nov 1995 50042.665 IDS/500mm/R1200 900 635-674 0.8 30 Nov 1995 50051.401 IDS/500mm/R1200 1200 635-674 0.8 12 Jan 1996 50095.449 IDS/235mm/H1800 1250 624-677 1.1 31 Jan 1996 50113.572 CASPEC 1800 553-773 0.2 3 March 1996 50145.653 Hamilton echelle 2000 510-870 0.2 JD(+2,400,000) RV (km s$^{-1}$) EW(H$_\alpha$) (Å) v$_{\lambda}$ Phase ---------------- ------------------ -------------------- --------------- ------- 47495.735 7.3$\pm$1.2 0.70$\pm$0.25 0.81 49641.672 4.9$\pm$4.8 112.1 0.89$\pm$0.20 0.88 49936.723 34.9$\pm$4.1 57.5 0.46$\pm$0.20 0.43 49937.730 36.8$\pm$4.1 53.8 0.54$\pm$0.20 0.48 49938.727 32.5$\pm$4.1 34.6 0.66$\pm$0.20 0.53 49939.730 28.6$\pm$4.0 37.4 0.20$\pm$0.20 0.58 49972.730 22.2$\pm$3.1 43.5 0.40$\pm$0.20 0.33 50039.495 15.1$\pm$3.3 36.6 0.46$\pm$0.15 0.84 50040.528 6.0$\pm$3.0 45.8 0.38$\pm$0.20 0.90 50041.415 7.7$\pm$3.1 53.0 0.41$\pm$0.15 0.95 50041.696 5.0$\pm$2.7 52.6 0.50$\pm$0.15 0.95 50042.433 3.8$\pm$2.7 47.9 0.50$\pm$0.20 0.00 50042.665 2.8$\pm$2.5 46.7 0.50$\pm$0.20 0.00 50051.401 36.7$\pm$3.0 54.6 0.76$\pm$0.20 0.47 50095.449 12.7$\pm$4.1 47.6 0.71$\pm$0.15 0.79 50113.572 15.6$\pm$2.6 0.74 50145.653 29.5$\pm$3.0 0.44 [^4] Element Value ---------------------- -------------------------- Period 18.979$\pm$0.007 days Gamma 14.9$\pm$1.2 km s$^{-1}$ K1 15.5$\pm$2.0 km s$^{-1}$ K2 58.2$\pm$5.7 km s$^{-1}$ ecc 0.14$\pm$0.05 omega 217$^o$.9$\pm$30.1 T$_{\rm periastron}$ 2,451,314.09$\pm$1.66 JD M1 sin$^3$(i) 0.60$\pm$0.16 M$_\odot$ M2 sin$^3$(i) 0.16$\pm$0.04 M$_\odot$ a1 sin(i) 0.02675 AU a2 sin(i) 0.10056 AU [^1]: On leave from INAF-Osservatorio Astrofisico di Catania [^2]: Based on observations made with the Isaac Newton Telescope, operated on the island of La Palma by the Isaac Newton Group in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astro[fí]{}sica de Canarias, the ESO 3.6-meter telescope at La Silla Observatory in Chile, and the Shane 3-meter telescope at Lick observatory in California. [^3]: The IRAF package is distributed by the National Optical Astronomical Observatories. [^4]: Veling value from Basri & Batalha (1990).

--- abstract: | In this paper, we study high order correctors in stochastic homogenization. We consider elliptic equations in divergence form on ${\mathbb{Z}}^d$, with the random coefficients constructed from i.i.d. random variables. We prove moment bounds on the high order correctors and their gradients under dimensional constraints. It implies the existence of stationary correctors and stationary gradients in high dimensions. As an application, we prove a two-scale expansion of the solutions to the random PDE, which identifies the first and higher order random fluctuations in a strong sense. <span style="font-variant:small-caps;">Keywords:</span> quantitative homogenization, high order corrector, two-scale expansion, random fluctuation. address: 'Department of Mathematics, Building 380, Stanford University, Stanford, CA, 94305, USA' author: - Yu Gu title: 'High order correctors and two-scale expansions in stochastic homogenization' --- Main result =========== Quantitative stochastic homogenization has witnessed important progress in recent years, and a major contribution of the groundbreaking work of Gloria-Otto is to prove the high order moment estimates on the corrector [@gloria2011optimal; @gloria2014quantitative]. The result in the discrete setting can be described as follows. Let ${\mathbb{B}}$ be the set of nearest neighbor edges in ${\mathbb{Z}}^d$, and $\{e_i,i=1,\ldots,d\}$ be the canonical basis of ${\mathbb{Z}}^d$. On a probability space $(\Omega,\mathcal{F},{\mathbb{P}})$, we have a sequence of i.i.d. non-degenerate random conductances, denoted by $\{\omega_e\}_{e\in{\mathbb{B}}}$ with $\omega_e\in (\delta,1)$ for some $\delta>0$. We define $a:{\mathbb{Z}}^d\to {\mathbb{R}}^{d\times d}$ as a random diagonal matrix field such that $$a(x)=\mathrm{diag}(a_1(x),\ldots,a_d(x))=\mathrm{diag}(\omega_{(x,x+e_1)},\ldots,\omega_{(x,x+e_d)}).$$ The *regularized* corrector equation in the direction $\xi\in{\mathbb{R}}^d$ says $$(\lambda+\nabla^* a(x)\nabla) \phi_\xi^\lambda(x)=-\nabla^* a(x)\xi, \ \ x\in {\mathbb{Z}}^d. \label{eq:1stcor}$$ Here the discrete gradient and divergence for $f:{\mathbb{Z}}^d\to {\mathbb{R}}$ and $F:{\mathbb{Z}}^d\to {\mathbb{R}}^d$ are defined as $$\nabla f=(\nabla_1f,\ldots,\nabla_d f), \ \ \nabla^*F=\sum_{i=1}^d \nabla_i^* F_i,$$ with $$\nabla_i f(x)=f(x+e_i)-f(x), \ \ \nabla_i^* F_i=F_i(x-e_i)-F_i(x).$$ It was shown in [@gloria2011optimal Proposition 2.1] that when $d{\geqslant}3$, for any $p{\geqslant}1$, there exists a constant $C=C(d,\delta,p,\xi)>0$ such that $${\langle}|\phi_\xi^\lambda|^p{\rangle}{\leqslant}C \label{eq:gobound}$$ uniformly in $\lambda>0$, where ${\langle}\cdot {\rangle}$ denotes the expectation on $(\Omega,\mathcal{F},{\mathbb{P}})$. In particular, it implies the existence of a stationary corrector when $d{\geqslant}3$: there exists a zero-mean stationary random field $\phi_\xi$ solving $$-\nabla^* a(x)\nabla \phi_\xi(x) =\nabla^* a(x)\xi. \label{eq:1stcoreq}$$ The first goal of this paper is to go beyond the first order correctors, and present a proof of for high order correctors. As an application of high order correctors, we identity the first and higher order fluctuations in stochastic homogenization in a strong sense that will be specified later. High order correctors --------------------- We will stay in the same setting but further assume $\{\omega_e\}_{e\in{\mathbb{B}}}$ satisfies the log-Soblev inequality $$\left{\langle}\zeta^2 \log \frac{\zeta^2}{{\langle}\zeta^2{\rangle}}\right{\rangle}{\leqslant}\frac{1}{\rho}{\left\langle}\sum_e |\partial_e \zeta|^2{\right\rangle}\label{eq:logsob}$$ for some $\rho>0$. Here $\partial_e$ is the weak derivative with respect to $\omega_e$ and $\zeta:\Omega\to {\mathbb{R}}$ is any function so that the r.h.s. of makes sense. Through a formal two-scale expansion in Section \[s:2scale\], we define the regularized $n-$th order corrector, and denote them by $\psi_n^\lambda$. The following is our first main result. Fix any $n{\geqslant}2$. When $d{\geqslant}2n-1$, for any $p{\geqslant}1$, there exists $C=C(d,\delta,\rho,p)>0$ such that $${\langle}|\nabla \psi_n^\lambda|^p {\rangle}{\leqslant}C $$ uniformly in $\lambda>0$. When $d{\geqslant}2n+1$, the same result holds for $\psi_n^\lambda$, i.e., $${\langle}|\psi_n^\lambda|^p {\rangle}{\leqslant}C$$ uniformly in $\lambda>0$. In particular, for the $n-$th order corrector, it has a stationary gradient when $d{\geqslant}2n-1$ and it is stationary when $d{\geqslant}2n+1$. \[t:mainTh\] The log-Soblev inequality holds for i.i.d. random variables with a continuous density, hence can not deal with distributions with atoms. A weaker version is presented in [@marahrens2013annealed Equation (7)], covering all possible i.i.d. ensembles. We believe that the approach in this paper can be applied with the weaker version (with extra technicalities), and Theorem \[t:mainTh\] holds for any i.i.d. ensemble. To keep the presentation simple, we choose to work with . The study of stochastic homogenization started from the early work of Kozlov [@kozlov1979averaging] and Papanicolaou-Varadhan [@papanicolaou1979boundary], and revived recently from various quantitative perspectives [@caffarelli2010rates; @gloria2011optimal; @gloria2012optimal; @gloria2013fluctuation; @gloria2013quantification; @marahrens2013annealed; @gloria2014optimal; @gloria2014quantitative; @gloria2014regularity; @gloria2014improved; @gloria2015optimal; @mourrat2011variance; @mourrat2012kantorovich; @armstrong1; @armstrong2; @armstrong3; @armstrong4]. While the first order correctors have been analyzed extensively due to their role in determining the effective coefficients and proving convergence in homogenization, the high order correctors have been receiving less attention. Our interest in the high order correctors comes from the comparison between a pointwise two-scale expansion and a large scale central limit theorem derived for the solutions to the random PDE $$-\nabla\cdot a(\frac{x}{{\varepsilon}})\nabla u_{\varepsilon}(x)=f(x). \label{eq:spde}$$ The results in [@gu-mourrat; @gu-mourratmms] showed that when $d{\geqslant}3$, the first order corrector represents the local fluctuation, which is measured by $u_{\varepsilon}(x)-{\langle}u_{\varepsilon}(x){\rangle}$ for fixed $x\in{\mathbb{R}}^d$, but does not suggest the global large scale fluctuation, which is measured weakly in space by $\int (u_{\varepsilon}-{\langle}u_{\varepsilon}{\rangle}) g$ with test function $g$. We expect the surprising phenomenon may be explained by high order correctors which only become visible in the weak sense due to strong correlations; see a discussion in Section \[s:compa\]. In Section \[s:2scale\], we will construct high order correctors directly from a formal two-scale expansion, and there is a slightly different way of characterizing the high order correctors as the “high order intrinsic polynomials” that come out of the Liouville theorem: as the first order corrector $\phi_\xi$ is defined so that $\xi\cdot x+\phi_\xi(x)$ is $a-$harmonic, the high order correctors correct the $a_{\hom}-$harmonic high order polynomials to be $a-$harmonic. It seems the two ways of construction are equivalent although we do not attempt to prove it here. For our purpose, it is more convenient to directly start from the formal expansion. For the recent breakthrough in the direction of regularity theory of random operators in the continuous setting, we refer to [@armstrong1; @gloria2014regularity; @armstrong2; @fo; @armstrong3; @gloria2015optimal; @armstrong4]. Our strategy of proving moment estimates in Theorem \[t:mainTh\] follows [@gloria2011optimal], i.e., by using a spectral gap inequality (see below) and estimating the sensitivity of the correctors to the individual conductance; see also the unpublished work of Naddaf and Spencer [@ns]. A key quantity to control is $\partial_e \psi_n^\lambda$, which describes the dependence of the $n-$th order corrector on the conductance $\omega_e$. It involves the first and second order derivatives of the Green’s function of $\nabla^*a(x)\nabla$, and we use the $p-$th moment estimates derived in [@marahrens2013annealed], which came from the log-Soblev inequality together with the result of Delmotte and Deuschel on the lower order moment [@dd]. One of the difficulties is to obtain some a priori estimate on the gradient $\nabla\psi_n^\lambda$. For the first order corrector, the bound on the second moment of $\nabla \phi_\xi^\lambda$ comes directly from thanks to the divergence form of its r.h.s. source term. For high order correctors, we prove that the source term can also be expressed in divergence form. Once we have the $p-$th moment estimates on $\nabla\psi_n^\lambda$, Theorem \[t:mainTh\] follows from a straightforward induction argument. Two-scale expansions -------------------- As an application of high order correctors, we prove an expansion of solutions to the random PDE . If homogenization is viewed as a law of large numbers result, here we are looking for the next order random fluctuations that may or may not lead to a central limit theorem. A classical two-scale expansion indicates that the solutions to take the form $$u_{\varepsilon}=u_0+{\varepsilon}u_1+{\varepsilon}^2 u_2+\ldots, \label{eq:forex}$$ where $u_0,u_1,u_2,\ldots$ are constructed by equating the power-like terms in ${\varepsilon}$ upon substituting into . Since is only a formal series, it is a priori unclear whether $u_1$ indeed represents the “correct” first order fluctuation (it was shown not true when $d=1$ [@Gu2016]), and if it does, we need to understand in which sense holds. The second goal of the paper is to give an answer to the above questions and justify the formal two-scale expansion under appropriate dimensional constraint. To avoid the effects from boundary layers, we work on the equation $$(\alpha+\nabla_{\varepsilon}^* a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon}) u_{\varepsilon}(x)=f(x), x\in {\varepsilon}{\mathbb{Z}}^d, \label{eq:heteeq}$$ where $\alpha>0$ is a fixed constant and $f(x)\in {\mathcal{C}}_c^\infty({\mathbb{R}}^d)$. The discrete gradient and divergence are defined for $g:{\varepsilon}{\mathbb{Z}}^d\to {\mathbb{R}}$ and $G:{\varepsilon}{\mathbb{Z}}^d\to{\mathbb{R}}^d$ as $$\nabla_{\varepsilon}g=(\nabla_{{\varepsilon},1}g,\ldots,\nabla_{{\varepsilon},d}g) \mbox{ with } \nabla_{{\varepsilon},i}g(x)=[g(x+{\varepsilon}e_i)-g(x)]/{\varepsilon},$$ and $$\nabla_{\varepsilon}^* G=\sum_{i=1}^d \nabla_{{\varepsilon},i}^* G_i \mbox{ with } \nabla_{{\varepsilon},i}^*G_i(x)=[G_i(x-{\varepsilon}e_i)-G_i(x)]/{\varepsilon}.$$ It is well-known that $u_{\varepsilon}$ converges in a certain sense to $u_{0}$ solving $$(\alpha+\nabla_{\varepsilon}^*a_{\hom}\nabla_{\varepsilon}) u_0(x)=f(x), x\in {\varepsilon}{\mathbb{Z}}^d, \label{eq:homoeq}$$ where the effective coefficient matrix is given by $$a_{\hom}={\langle}a(\mathrm{I}_d+\nabla\phi){\rangle}=\bar{a} \mathrm{I}_d \mbox{ with }\nabla\phi=[\nabla\phi_{e_1},\ldots,\nabla\phi_{e_d}],$$ and $\phi_{e_k}$ is the first order corrector in the direction $e_k$. Our goal is to obtain the first and higher order fluctuations in $u_{\varepsilon}\to u_0$. By the formal two-scale expansion, the first order correction takes the form $$u_1(x,\frac{x}{{\varepsilon}})=\sum_{j=1}^d \nabla_{{\varepsilon},j}u_0(x)\phi_{e_j}(\frac{x}{{\varepsilon}}), x\in{\varepsilon}{\mathbb{Z}}^d.$$ Typically, one compare $u_{\varepsilon}$ with $\bar{u}_0$ solving the equation in the continuous space: $$(\alpha-\nabla\cdot a_{\hom}\nabla)\bar{u}_0(x)=f(x), x\in{\mathbb{R}}^d.$$ In this paper, we do not analyze $u_0-\bar{u}_0$, which is only an error due to discretization. For functions $f:{\varepsilon}{\mathbb{Z}}^d\times \Omega\to {\mathbb{R}}^d$, we define the $L^2({\varepsilon}{\mathbb{Z}}^d\times \Omega)$ norm $$\|f\|_{2,{\varepsilon}}:=\bigg({\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|f(x,\omega)|^2{\rangle}\bigg)^{\frac12}.$$ The following is our main result for the first order fluctuations. When $d{\geqslant}3$, we have $$\|u_{\varepsilon}-u_0-{\varepsilon}u_1\|_{2,{\varepsilon}}=o({\varepsilon}) \label{eq:1stex}$$ as ${\varepsilon}\to 0$. \[t:2scale\] Since ${\langle}u_1{\rangle}=0$, Theorem \[t:2scale\] in particular implies that the deterministic bias ${\langle}u_{\varepsilon}{\rangle}-u_0$ vanishes in the order of ${\varepsilon}$. For the errors quantified in the strong sense, by the $\|\cdot\|_{2,{\varepsilon}}$ norm here, Theorem \[t:2scale\] shows that the first order correction is given by the first order corrector that comes from the formal expansion. This is consistent with the pointwise result in the continuous setting [@gu-mourrat], and both results suggest that we do not expect a central limit theorem for $(u_{\varepsilon}(x)-u_0(x))/{\varepsilon}$. It should be contrasted with the low dimensional case, e.g., when $d=1$, it was shown in [@BP-AA-99; @Gu2016] that $(u_{\varepsilon}(x)-u_0(x))/\sqrt{{\varepsilon}}$ converges in distribution to a Gaussian process. For the errors quantified in the weak sense, i.e., after taking a spatial average with a test function, central limit theorems are obtained for the first order corrector $u_1$ and the solution $u_{\varepsilon}$ [@mourrat2014correlation; @MN; @gu-mourratmms; @gloria2016random; @armstrong2016scaling]. Various approximations to the effective coefficient matrix $a_{\hom}$ also exhibit Gaussian fluctuations; see [@nolen2011normal; @biskup2014central; @rossignol2012noise; @nolen2014normal; @GN2014CLT]. In higher dimensions, a higher order expansion similar to can also be obtained. To make sense of the expansion, the number of terms we are permitted to include in the expansion depends on the dimension in light of Theorem \[t:mainTh\]. Fix any $n{\geqslant}2$. When $d{\geqslant}2n+1$, there exists $\{u_k\}_{k=1}^n$ and $\{v_k\}_{k=1}^n$ such that $u_k$ is random with ${\langle}u_k{\rangle}=0$, $v_k$ is deterministic, and $$\bigg\|u_{\varepsilon}-u_0-\sum_{k=1}^n {\varepsilon}^ku_k-\sum_{k=1}^n {\varepsilon}^kv_k\bigg\|_{2,{\varepsilon}}=o({\varepsilon}^n)$$ as ${\varepsilon}\to 0$. \[t:highex\] Theorem \[t:2scale\] can be viewed as a special case of Theorem \[t:highex\] when $n=1$, in which case we have $v_1\equiv 0$. It can be seen from the proof of Theorem \[t:highex\] that the $k-$th order random fluctuation $u_k$ takes the form $$u_k(x)= \sum_i g_i(x)\varphi_i(\frac{x}{{\varepsilon}}),$$ where $g_i$ is some derivative of $u_0$ and $\varphi_i$ is some zero-mean stationary random field. Therefore, the random fluctuations in homogenization, measured in the strong sense, is a superposition of highly oscillatory random fields. Using the high order correctors obtained in Theroem \[t:mainTh\], the proof of Theorem \[t:2scale\] and Theorem \[t:highex\] mimics the periodic setting. It is well-known that in the *ideal* periodic setting (equation with smooth coefficients posed on the whole space), the formal two-scale expansion as in indeed approximates the solution up to arbitrary high order precision, so the message we want to convey here is that one can still hope the expansion to be valid in the random setting, provided that we have stationary correctors. We refer to [@gerard2011homogenization; @gerard2012homogenization] for a careful dealing with boundary layers in the periodic setting. The previous work on estimating the size of $u_{\varepsilon}-u_0$ includes e.g. [@yurinskii1986averaging; @conlon1; @conlon2], with non-optimal exponents or optimal exponents in the small ellipticity regime. With the study of the first order correctors, [@gloria2013fluctuation; @gloria2014optimal; @mourrat2012kantorovich] provides optimal estimates of the error size. The recent preprint [@BFFO] uses second order correctors to derive quantitative estimates in the weak spatial norms. The main contribution of Theorem \[t:2scale\] and Theorem \[t:highex\] is to identify the first and higher order fluctuations in the strong sense, which seems to be the first result of this type. The paper is organized as follows. In Section \[s:pr\], we present the proof of Theorem \[t:mainTh\]. The proofs of Theorem \[t:2scale\] and Theorem \[t:highex\] are in Section \[s:1stex\] and Section \[s:high\]. We leave some discussions to Section \[s:compa\]. Notations --------- - For any $e=(\underline{e},\bar{e})\in {\mathbb{B}}$ and $f:{\mathbb{Z}}^d\to {\mathbb{R}}$, we will write $$\nabla f(e)=f(\bar{e})-f(\underline{e}),$$ and for any $\xi\in{\mathbb{R}}^d$, $$\xi(e)=\xi\cdot e.$$ - For $x\in {\mathbb{Z}}^d$, we define $$|x|_*=|x|+2.$$ - We write $a\les b$ when $a{\leqslant}Cb$ with the constant $C$ independent of the spatial variables $y,z$ and the edge variables $b,e$. - For $x,y,c>0$, we write $$x\les \frac{1}{y^{c-}}$$ if for any $\delta>0$, there exists $C_\delta>0$ such that $$x{\leqslant}C_\delta \frac{1}{y^{c-\delta}}.$$ - For $f:{\mathbb{Z}}^d\times{\mathbb{Z}}^d\to {\mathbb{R}}$, $x\in {\mathbb{Z}}^d$ and $e,b\in{\mathbb{B}}$, we write $$\nabla f(x,e)=f(x,\bar{e})-f(x,{\underline}{e}),$$ and $$\nabla\nabla f(b,e)=\nabla f(\bar{b},e)-\nabla f({\underline}{b},e).$$ - We use $\|\cdot\|_p$ to denote the $L^p(\Omega)$ norm. Recall that $\|\cdot\|_{2,{\varepsilon}}$ denote the $L^2({\varepsilon}{\mathbb{Z}}^d\times \Omega)$ norm. - We use $G^\lambda,{\mathcal{G}}^\lambda$ to denote the Green’s function of $\lambda+\nabla^*a\nabla, \lambda+\nabla^*\nabla$ on ${\mathbb{Z}}^d$ and ${\mathscr{G}}^\lambda$ to denote the Green’s function of $\lambda-\Delta$ on ${\mathbb{R}}^d$. Moment bounds of high order correctors {#s:pr} ====================================== We first review the classical two-scale expansion of equations in divergence form, then present a proof of Theorem \[t:mainTh\]. Two-scale expansions {#s:2scale} -------------------- Consider the equation on ${\mathbb{R}}^d$: $$-\nabla\cdot a(\frac{x}{{\varepsilon}})\nabla u_{\varepsilon}(x)=f(x),$$ where $a:{\mathbb{R}}^d\to {\mathbb{R}}^{d\times d}$ is a stationary random field. We introduce the fast variable $y=\frac{x}{{\varepsilon}}$ and write $\nabla=\nabla_x+\frac{1}{{\varepsilon}}\nabla_y$. By further assuming $$u_{\varepsilon}(x)=u_0(x)+{\varepsilon}u_1(x,y)+{\varepsilon}^2 u_2(x,y)+\ldots, \label{eq:twoscale}$$ we obtain $$-(\nabla_x+\frac{1}{{\varepsilon}}\nabla_y)\cdot a(y)(\nabla_x+\frac{1}{{\varepsilon}}\nabla_y) (u_0+{\varepsilon}u_1+\ldots+{\varepsilon}^n u_n+\ldots)=f.$$ By matching the order of ${\varepsilon}$, we get equations satisfied by $u_n$. When $n=1$, $$\nabla_y\cdot a(y)\nabla_y u_1+\nabla_y\cdot a(y)\nabla_x u_0=0. \label{eq:1stcorEq}$$ When $n{\geqslant}2$, $$\begin{aligned} &\nabla_y \cdot a(y)\nabla_y u_n+\nabla_y\cdot a(y)\nabla_x u_{n-1}+\nabla_x \cdot a(y)\nabla_y u_{n-1}+\nabla_x \cdot a(y)\nabla_x u_{n-2}\\ =&{\langle}\nabla_x \cdot a(y)\nabla_y u_{n-1}+\nabla_x \cdot a(y)\nabla_x u_{n-2}{\rangle}. \label{eq:highcorEq} \end{aligned}$$ Note that in general, the corrector equations take the form $$\nabla_y\cdot a(y)\nabla_y u_n(x,y)=F$$ and a solvability condition requires that the source satisfies ${\langle}F{\rangle}=0$, so we added some expectations to the r.h.s. of . For , we can write $$u_1(x,y)=\sum_{k=1}^d \partial_k u_0(x)\phi_{e_k}(y)=\nabla u_0(x)\cdot \phi(y), \label{eq:u1}$$ with $\phi=[\phi_{e_1},\ldots,\phi_{e_d}]$ and $\phi_{e_k}$ the first order corrector solving $$\nabla_y\cdot a(y)(\nabla\phi_{e_k}+e_k)=0.$$ The equation satisfied by the second order corrector takes the form $$\begin{aligned} \nabla_y \cdot a(y)\nabla_y u_2 =&\nabla_x \cdot a_{\hom}\nabla_x u_0(x)-\nabla_x\cdot a(y)\nabla\phi(y)\nabla u_0(x)\\ &-\nabla_x\cdot a(y)\nabla_x u_0(x)-\nabla_y\cdot a(y)\nabla^2 u_0(x)\phi(y), \end{aligned} \label{eq:u20}$$ where $\nabla^2u_0$ is the Hessian matrix of $u_0$ and the matrix $\nabla\phi=[\nabla\phi_{e_1},\ldots,\nabla\phi_{e_d}]$. The r.h.s. of the above expression has mean zero since the homogenization matrix $a_{\hom}$ is given by $$a_{\hom}={\langle}a(y)(\mathrm{I}_d+\nabla\phi(y)){\rangle}.$$ By induction, it is clear that $u_n$ is a linear combination of $\partial_\alpha u_0$, where $\alpha$ is a multi-index with $|\alpha|=n$, so we write $$u_n(x,y)=\sum_\alpha \phi_\alpha(y)\partial_\alpha u_0(x),$$ and those $\phi_\alpha$ are the $n-$th order correctors. By , we further observe that $\phi_\alpha$ with $|\alpha|=n$ is a linear combination of $\psi_n$, which solves the following three types of equations: $$\nabla\cdot a(y)\nabla \psi_n(y)=\left\{\begin{array}{l} \partial_i (a_{ij}(y)\psi_{n-1}(y)),\\ a_{ij}(y) \partial_j \psi_{n-1}(y)-{\langle}a_{ij}(y)\partial_j \psi_{n-1}(y){\rangle},\\ a_{ij}(y)\psi_{n-2}(y)-{\langle}a_{ij}(y)\psi_{n-2}(y){\rangle}. \end{array} \right. \label{eq:corn}$$ Here $i,j=1,\ldots,d$, and we have the following correspondence between the r.h.s. of and : $$\begin{aligned} \partial_i (a_{ij}(y)\psi_{n-1}(y)) &\longleftrightarrow \nabla_y\cdot a(y)\nabla_x u_{n-1}(x,y), \\ a_{ij}(y) \partial_j \psi_{n-1}(y)-{\langle}a_{ij}(y)\partial_j \psi_{n-1}(y){\rangle}&\longleftrightarrow \nabla_x\cdot a(y)\nabla_y u_{n-1}(x,y)-{\langle}\nabla_x\cdot a(y)\nabla_y u_{n-1}(x,y){\rangle},\\ a_{ij}(y)\psi_{n-2}(y)-{\langle}a_{ij}(y)\psi_{n-2}(y){\rangle}&\longleftrightarrow \nabla_x \cdot a(y)\nabla_x u_{n-2}(x,y)-{\langle}\nabla_x \cdot a(y)\nabla_x u_{n-2}(x,y){\rangle}. \end{aligned}$$ In other words, since is a linear equation, we have decomposed it into finitely many “small equations” written in the generic form of . A proof of Theorem \[t:mainTh\] by induction {#s:proof} -------------------------------------------- Recalling that the coefficient $a(x)=\mathrm{diag}(a_1(x),\ldots,a_d(x))$ is a diagonal matrix for $x\in {\mathbb{Z}}^d$, so given in the continuous setting, we consider corrector equations on ${\mathbb{Z}}^d$ of the following form: $$\nabla^*a(y)\nabla \psi_n(y)=\left\{\begin{array}{l} \nabla_i^* (a_i(y)\psi_{n-1}(y)),\\ a_i(y) \nabla_i\psi_{n-1}(y)-{\langle}a_i(y)\nabla_i \psi_{n-1}(y){\rangle},\\ a_{i}(y)\psi_{n-2}(y)-{\langle}a_{i}(y)\psi_{n-2}(y){\rangle}. \end{array} \right. \label{eq:corndis}$$ For any $\lambda>0$, we regularize by adding a massive term and define $\psi_n^\lambda$ as the unique solution to $$(\lambda+\nabla^*a(y)\nabla) \psi_n^\lambda(y)=\left\{\begin{array}{l} \nabla_i^* (a_i(y)\psi_{n-1}(y)),\\ a_i(y) \nabla_i\psi_{n-1}(y)-{\langle}a_i(y)\nabla_i \psi_{n-1}(y){\rangle},\\ a_{i}(y)\psi_{n-2}(y)-{\langle}a_{i}(y)\psi_{n-2}(y){\rangle}. \end{array} \right. \label{eq:corndislambda}$$ To prove Theorem \[t:mainTh\], we only analyze when $d{\geqslant}2n-1$. Since our proof is based on induction, we do not worry about the meaning of the r.h.s. of for the moment. Let $\psi_1$ denote the stationary first order correctors $\phi_{e_i}$ when $d{\geqslant}3$. We make the statement $\mathscr{I}_n$ for $n{\geqslant}1$: **Statement $\mathscr{I}_1$**: when $d{\geqslant}3$, for any $p{\geqslant}1$, we have $$\|\partial_e \psi_1(y)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{d-1}}, \label{eq:eq01}$$ and $$\|\partial_e \nabla\psi_1(b)\|_p\les \frac{1}{|{\underline}{b}-{\underline}{e}|_*^{d}}, \label{eq:eq02}$$ with the proportional constant independent of $y\in {\mathbb{Z}}^d$ and $b,e\in{\mathbb{B}}$. **Statement $\mathscr{I}_n, n{\geqslant}2$**: when $d{\geqslant}2n-1$, for any $p{\geqslant}1$, we have $$\|\partial_e \psi_n^\lambda(y)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{(d-n)-}}, \label{eq:eq1}$$ and $$\|\partial_e \nabla\psi_n^\lambda(b)\|_p\les \frac{1}{|{\underline}{b}-{\underline}{e}|_*^{(d-n+1)-}}, \label{eq:eq2}$$ with the proportional constant independent of $\lambda>0$, $y\in {\mathbb{Z}}^d$ and $b,e\in{\mathbb{B}}$. The following result is the main result in this section. For any $n{\geqslant}2$, $\mathscr{I}_{n-2}+\mathscr{I}_{n-1}$ implies $\mathscr{I}_{n}$, where we take $\mathscr{I}_0$ as an empty statement. \[p:induction\] First, we note that $\mathscr{I}_1$ holds. For the first order correctors $\psi_1\in\{\phi_{e_k}\}_{k=1,\ldots,d}$, $$\partial_e \phi_{e_k}^\lambda(y)=-\nabla G^\lambda(y,e)(\nabla\phi_{e_k}+e_k)(e),$$ and $$\partial_e \nabla \phi_{e_k}^\lambda(b)=-\nabla\nabla G^\lambda(b,e)(\nabla\phi_{e_k}+e_k)(e),$$ where $G^\lambda$ is the Green’s function of $\lambda+\nabla^*a\nabla$. By [@marahrens2013annealed Theorem 1], we have $$\|\nabla G^\lambda(y,e)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{d-1}} \mbox{ and } \|\nabla\nabla G^\lambda(b,e)\|_p\les \frac{1}{|{\underline}{b}-{\underline}{e}|_*^{d}}, \label{eq:bdgr}$$ which implies $$\|\partial_e \phi_{e_k}(y)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{d-1}} \mbox{ and } \|\partial_e \nabla\phi_{e_k}(b)\|_p\les \frac{1}{|{\underline}{b}-{\underline}{e}|_*^d}. \label{eq:j1}$$ By Proposition \[p:induction\], $\mathscr{I}_n$ holds for $n{\geqslant}2$, and this implies Theorem \[t:mainTh\]. First, the log-Soblev inequality implies the Spectral-Gap inequality: $${\langle}|\zeta|^2{\rangle}\les {\left\langle}\sum_e |\partial_e \zeta|^2{\right\rangle}\label{eq:sgin}$$ for zero-mean $\zeta$, which can be generalized to control the $2p-$th moment for any $p>1$ [@gloria2013quantification Lemma 2]: $${\langle}|\zeta|^{2p}{\rangle}\les {\left\langle}\left(\sum_e |\partial_e\zeta|^2\right)^p{\right\rangle}. \label{eq:psg}$$ The result presented in [@gloria2013quantification Lemma 2] is for a different type of vertical derivative, but the same proof applies in our case. Now we apply to $\psi_n^\lambda$, and use $\mathscr{I}_n$ to derive $$\begin{aligned} \|\psi_n^\lambda(0)\|_{2p}\les {\left\langle}\left(\sum_e |\partial_e\psi_n^\lambda(0)|^2\right)^p{\right\rangle}^{\frac{1}{2p}} \les &\left(\sum_e \|\partial_e \psi_n^\lambda(0)\|_{2p}^2\right)^{\frac12}\\ \les & \left(\sum_e \frac{1}{|{\underline}{e}|_*^{(2d-2n)-}}\right)^{\frac12}\les 1 \end{aligned}$$ when $d{\geqslant}2n+1$. Similarly, we apply to $\nabla\psi_n^\lambda$, and use $\mathscr{I}_{n}$ to derive $$\begin{aligned} \|\nabla\psi_n^\lambda(b)\|_{2p}\les {\left\langle}\left(\sum_e |\partial_e\nabla \psi_n^\lambda(b)|^2\right)^p{\right\rangle}^{\frac{1}{2p}} \les &\left(\sum_e \|\partial_e \nabla\psi_n^\lambda(b)\|_{2p}^2\right)^{\frac12}\\ \les & \left(\sum_e \frac{1}{|{\underline}{b}-{\underline}{e}|_*^{(2d-2n+2)-}}\right)^{\frac12}\les 1 \end{aligned}$$ when $d{\geqslant}2n-1$. By the uniform bound on $\|\psi_n^\lambda\|_{2p}$ and $\|\nabla\psi_n^\lambda(b)\|_{2p}$, it is standard to extract a subsequence as $\lambda\to 0$ to get the existence of a stationary corrector $\psi_n$ when $d{\geqslant}2n+1$ and the existence of a stationary gradient $\nabla\psi_n$ when $d{\geqslant}2n-1$. For the convenience of reader, we present the details in Appendix \[s:exst\]. It is clear that if $\psi_n^\lambda(0)$ converges to $\psi_n(0)$ weakly in $L^p(\Omega)$, then $\partial_e\psi_n^\lambda(0)$ converges to $\partial_e\psi_n(0)$ weakly in $L^p(\Omega)$, thus the same estimates in $\mathscr{I}_n$ holds for $\psi_n$ when $d{\geqslant}2n+1$. The same discussion applies to $\nabla\psi_n$.\[r:bscor\] The proof of Proposition \[p:induction\] is a straightforward calculation once we assume the high moment bounds for $\nabla\psi_{n}^\lambda$: $$\mathscr{I}_{n-2}+\mathscr{I}_{n-1}\Rightarrow \mbox{ when } d{\geqslant}2n-1, \mbox{ for any } p{\geqslant}1, \|\nabla \psi_{n}^\lambda(b)\|_p\les 1.\label{eq:mmgra}$$ We will leave the proof of $\eqref{eq:mmgra}$ to the next section. *Proof of Proposition \[p:induction\].* By the previous discussion, when $d{\geqslant}2n-1$, we have stationary $(n-2)-$th and $(n-1)-$th order corrector, so the equation of $\psi_n^\lambda$ given by is well-defined:$$(\lambda+\nabla^*a(y)\nabla) \psi_n^\lambda(y)=\left\{\begin{array}{l} \nabla_i^* (a_i(y)\psi_{n-1}(y)),\\ a_i(y) \nabla_i\psi_{n-1}(y)-{\langle}a_i(y)\nabla_i \psi_{n-1}(y){\rangle},\\ a_{i}(y)\psi_{n-2}(y)-{\langle}a_{i}(y)\psi_{n-2}(y){\rangle}. \end{array} \right.$$ We take $\partial_e$ on both sides to obtain $$(\lambda+\nabla^* a(y)\nabla) \partial_e \psi_n^\lambda(y)=(1_{y={\underline}{e}}-1_{y=\bar{e}})\nabla\psi_n^\lambda(e)+\left\{\begin{array}{l} \nabla_i^* [\partial_e(a_i(y)\psi_{n-1}(y))],\\ \partial_e(a_i(y) \nabla_i\psi_{n-1}(y)),\\ \partial_e(a_{i}(y)\psi_{n-2}(y)). \end{array} \right.$$ By the Green’s function representation, we write $$\partial_e \psi_n^\lambda(y)=-\nabla G^\lambda(y,e)\nabla\psi_n^\lambda(e)+\left\{\begin{array}{l} \sum_{z}\nabla_i G^\lambda(y,z) [\partial_e(a_i(z)\psi_{n-1}(z))],\\ \sum_z G^\lambda(y,z)\partial_e(a_i(z) \nabla_i\psi_{n-1}(z)),\\ \sum_z G^\lambda(y,z)\partial_e(a_{i}(z)\psi_{n-2}(z)), \end{array} \right. \label{eq:depsin}$$ which implies that for any $p{\geqslant}1$: $$\|\partial_e \psi_n^\lambda(y)\|_p{\leqslant}\|\nabla G^\lambda(y,e)\nabla\psi_n^\lambda(e)\|_p+\left\{\begin{array}{l} \sum_{z}\|\nabla_i G^\lambda(y,z) [\partial_e(a_i(z)\psi_{n-1}(z))]\|_p,\\ \sum_z \|G^\lambda(y,z)\partial_e(a_i(z) \nabla_i\psi_{n-1}(z))\|_p,\\ \sum_z \|G^\lambda(y,z)\partial_e(a_{i}(z)\psi_{n-2}(z))\|_p. \end{array} \right.$$ Since $\partial_e a_i(z)=0$ for $e\neq (z,z+e_i)$, it is clear by $\mathscr{I}_{n-2},\mathscr{I}_{n-1}$ and Remark \[r:bscor\] that $$\begin{aligned} \|\partial_e(a_i(z)\psi_{n-1}(z))\|_p &\les \frac{1}{|z-{\underline}{e}|_*^{(d-n+1)-}},\\ \|\partial_e(a_i(z)\nabla_j\psi_{n-1}(z))\|_p &\les \frac{1}{|z-{\underline}{e}|_*^{(d-n+2)-}},\\ \|\partial_e(a_i(z)\psi_{n-2}(z))\|_p &\les \frac{1}{|z-{\underline}{e}|_*^{(d-n+2)-}}. \end{aligned}$$ Note that when $n=2$, $\psi_{n-2}(z)$ is a constant as can be seen from , so the above estimates still holds. An application of Hölder inequality together with and leads to $$\|\partial_e \psi_n^\lambda(y)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{d-1}}+\left\{\begin{array}{l} \sum_{z}\frac{1}{|y-z|_*^{d-1}}\frac{1}{|z-{\underline}{e}|_*^{(d-n+1)-}},\\ \sum_z \frac{1}{|y-z|_*^{d-2}}\frac{1}{|z-{\underline}{e}|_*^{(d-n+2)-}},\\ \sum_z\frac{1}{|y-z|_*^{d-2}}\frac{1}{|z-{\underline}{e}|_*^{(d-n+2)-}}. \end{array} \right.$$ By the discrete convolution inequality [@gu-mourratmms Lemma A.6], we obtain $$\|\partial_e \psi_n^\lambda(y)\|_p\les \frac{1}{|y-{\underline}{e}|_*^{(d-n)-}}. \label{eq:psin1}$$ By , we also have $$\partial_e \nabla \psi_n^\lambda (b)=-\nabla \nabla G^\lambda (b,e)\nabla\psi_n^\lambda(e)+\left\{\begin{array}{l} \sum_{z}\nabla \nabla_i G^\lambda(b,z) [\partial_e(a_i(z)\psi_{n-1}(z))],\\ \sum_z \nabla G^\lambda(b,z)\partial_e(a_i(z) \nabla_i\psi_{n-1}(z)),\\ \sum_z \nabla G^\lambda(b,z)\partial_e(a_{i}(z)\psi_{n-2}(z)), \end{array} \right. \label{eq:degrpsin}$$ so the same discussion as above gives $$\|\partial_e \nabla \psi_n^\lambda(b)\|_p\les \frac{1}{|{\underline}{b}-{\underline}{e}|_*^{(d-n+1)-}}. \label{eq:psin2}$$ The statement $\mathscr{I}_n$ consists of and , so the proof of Proposition \[p:induction\] is complete. The reason we choose to have $(d-n)-$ and $(d-n+1)-$ as the exponents in and is due to the following discrete convolution inequality: for any $\alpha\in (0,d)$, $$\sum_z \frac{1}{|z|_*^\alpha}\frac{1}{|y-z|_*^d} \les \frac{\log |y|_*}{|y|_*^\alpha}\les \frac{1}{|y|_*^{\alpha-}}. \label{eq:dcon}$$ As can be seen from , the term $\nabla \nabla_i G^\lambda (b,z)$ produces a factor of $|{\underline}{b}-z|_*^{-d}$, and we use “$-$” to absorb the logarithmic factor when applying . Estimates on gradient of correctors: proof of {#s:boundg} ---------------------------------------------- To prove $$\|\nabla\psi_n^\lambda(b)\|_p\les 1$$ for any $p{\geqslant}1$, we first consider $p=2$, then derive it for any $p>2$. ### $p=2$. {#s:q2} We write the equation as $$(\lambda+\nabla^* a(y)\nabla)\psi_n^\lambda(y)=F(y)=\left\{\begin{array}{l} \nabla_i^* (a_i(y)\psi_{n-1}(y)),\\ a_i(y) \nabla_i\psi_{n-1}(y)-{\langle}a_i(y)\nabla_i \psi_{n-1}(y){\rangle},\\ a_{i}(y)\psi_{n-2}(y)-{\langle}a_{i}(y)\psi_{n-2}(y){\rangle}. \end{array} \right. \label{eq:corgene}$$ By $\mathscr{I}_{n-2}+\mathscr{I}_{n-1}$, when $d{\geqslant}2n-1$, $F$ is a zero-mean stationary random field, and $$\|\psi_{n-1}\|_q+\|\nabla \psi_{n-1}\|_q+\|\psi_{n-2}\|_q\les1$$ for any $q{\geqslant}1$. We claim that there exists a stationary random field $\Psi$ such that $\Psi(0)\in L^q(\Omega)$ for any $q{\geqslant}1$ and $$\nabla^* \Psi=F. \label{eq:dipsif}$$ Using the fact $a_i> \delta>0$, we conclude from that $$\|\nabla\psi_n^\lambda\|_2\les \|\Psi\|_2\les 1.$$ For $F=\nabla_i^* (a_i(y)\psi_{n-1}(y))$, we only need to choose $$\Psi=\Psi_ie_i \mbox{ with }\Psi_i=a_i\psi_{n-1}.$$ For the other two cases, we consider the equation with $k=1,\ldots,d$ $$(\lambda+\nabla^*\nabla)\Psi_k^\lambda=\nabla_k F, \label{eq:plala}$$ and by $\mathscr{I}_{n-2}+\mathscr{I}_{n-1}$, we have $$\begin{aligned} \|\partial_e \Psi_k^\lambda(y)\|_q{\leqslant}\sum_z \|\nabla_k^* {\overline}{G^\lambda}(y,z)\partial_eF(z)\|_q\les& \sum_z \frac{1}{|y-z|_*^{d-1}}\|\partial_e F(z)\|_q\\ \les &\sum_z \frac{1}{|y-z|_*^{d-1}}\frac{1}{|z-{\underline}{e}|_*^{(d-n+2)-}}\\ \les &\frac{1}{|y-{\underline}{e}|_*^{(d-n+1)-}}, \end{aligned}$$ where we used ${\overline}{G^\lambda}$ to denote the Green’s function of $\lambda+\nabla^*\nabla$. Thus, by the same application of the spectral gap estimate , we have $$\|\Psi_k^\lambda\|_q\les 1$$ when $d{\geqslant}2n-1$. By Lemma \[l:stacor\], there exists a stationary random field $\Psi_k(y)$ such that $\Psi_k(0)\in L^q(\Omega)$ and $\Psi_k^\lambda(0) \to \Psi_k(0)$ weakly in $L^q(\Omega)$. Let $\Psi=(\Psi_1,\ldots,\Psi_d)$, and we claim $$\nabla^*\Psi=F.$$ From , we have $$(\lambda+\nabla^*\nabla)\nabla_k^*\Psi_k^\lambda=\nabla_k^*\nabla_kF,$$ so $$(\lambda+\nabla^*\nabla)\nabla^*\Psi^\lambda=\nabla^*\nabla F$$ by summing over $k$. Let $\lambda\to 0$, it is clear that $$\nabla^*\nabla (\nabla^*\Psi-F)=0.$$ By ergodicity, we have $\nabla^*\Psi-F$ is a constant, but since it has mean zero, we conclude $\nabla^*\Psi=F$. The proof is complete for $p=2$. ### $p>2$. {#s:qg2} By [@marahrens2013annealed Lemma 4], the log-Soblev inequality implies a type of inverse Hölder inequality: fix any $q>1$, for any $\delta>0$, there exists $C_\delta>0$ such that $${\langle}|\xi|^{2q}\rangle^{\frac{1}{2q}}{\leqslant}C_\delta\langle |\xi|\rangle+\delta{\left\langle}\left(\sum_e |\partial_e \xi|^2\right)^{q}{\right\rangle}^{\frac{1}{2q}}. \label{eq:inho}$$ The result in [@marahrens2013annealed Lemma 4] is for a different type of derivative, but the same proof applies in our case. In view of and the bound on $\|\nabla\psi_n^\lambda\|_2$, in order to bound $\|\nabla\psi_n^\lambda\|_q$ for any $q>1$, it suffices to prove $${\left\langle}\left(\sum_e |\partial_e \nabla\psi_n^\lambda|^2\right)^q{\right\rangle}^{\frac{1}{2q}}\les {\langle}|\nabla\psi_n^\lambda|^{2q}\rangle^{\frac{1}{2q}}+1, \label{eq:deto}$$ and choose $\delta$ sufficiently small. By recalling , we have $$|\partial_e \nabla\psi_n^\lambda(b)|^2\les |\nabla\nabla G^\lambda(b,e)\nabla\psi_n^\lambda(e)|^2+I_e^2, \label{eq:decom}$$ where $$I_e:=\left\{\begin{array}{l} \sum_{z}\nabla \nabla_i G^\lambda(b,z) [\partial_e(a_i(z)\psi_{n-1}(z))],\\ \sum_z \nabla G^\lambda(b,z)\partial_e(a_i(z) \nabla_i\psi_{n-1}(z)),\\ \sum_z \nabla G^\lambda(b,z)\partial_e(a_{i}(z)\psi_{n-2}(z)). \end{array} \right.$$ We first have $${\left\langle}\left(\sum_e I_e^2\right)^q{\right\rangle}\les \left(\sum_e \|I_e\|_{2q}^2\right)^q\les 1$$ by $\mathscr{I}_{n-2}+\mathscr{I}_{n-1}$ and the fact that $d{\geqslant}2n-1$. For the first term on the r.h.s. of , we have $$\begin{aligned} \left(\sum_e|\nabla\nabla G^\lambda(b,e)\nabla\psi_n^\lambda(e)|^2\right)^q \les &\left(\sum_e |\nabla\nabla G^\lambda(b,e)|^2\right)^{q-1} \sum_e |\nabla\nabla G^\lambda(b,e)|^2|\nabla\psi_n^\lambda(e)|^{2q}\\ \les &\sum_e |\nabla\nabla G^\lambda(b,e)|^2|\nabla\psi_n^\lambda(e)|^{2q}, \end{aligned}$$ where we used the quenched bound [@marahrens2013annealed Equation (39)] $$\sum_e |\nabla\nabla G^\lambda(b,e)|^2\les 1. \label{eq:qubd}$$ Thus we have $${\left\langle}\left(\sum_e|\nabla\nabla G^\lambda(b,e)\nabla\psi_n^\lambda(e)|^2\right)^q {\right\rangle}\les\sum_e{\langle}|\nabla\nabla G^\lambda(b,e)|^2|\nabla\psi_n^\lambda(e)|^{2q}{\rangle}.$$ By stationarity we can write $${\langle}|\nabla\nabla G^\lambda(b,e)|^2|\nabla\psi_n^\lambda(e)|^{2q}{\rangle}={\langle}|\nabla\nabla G^\lambda(b-{\underline}{e},e-{\underline}{e})|^2|\nabla\psi_n^\lambda(e-{\underline}{e})|^{2q}{\rangle},$$ so by considering the summation over edges of the same direction, $\nabla\psi_n^\lambda(e-{\underline}{e})$ is unchanged over $e$, and using again we have $$\sum_e{\langle}|\nabla\nabla G^\lambda(b,e)|^2|\nabla\psi_n^\lambda(e)|^{2q}{\rangle}\les {\langle}|\nabla\psi_n^\lambda|^{2q}{\rangle}.$$ To summarize, we have proved that $${\left\langle}\left(\sum_e |\partial_e \nabla\psi_n^\lambda|^2\right)^q{\right\rangle}\les \langle |\nabla\psi_n^\lambda|^{2q}{\rangle}+1,$$ which leads to and completes the proof. First order fluctuations when $d{\geqslant}3$ {#s:1stex} ============================================= The proof of Theorem \[t:2scale\] follows the ideas of [@papanicolaou1979boundary Section 6], which itself is analogous to the periodic case. Since the first order correctors are used to prove the convergence of $u_{\varepsilon}\to u_0$, it is natural to consider using the second order correctors to obtain the first order fluctuations. We define the remainder in the expansion as $$z_{\varepsilon}(x)=u_{\varepsilon}(x)-u_0(x)-{\varepsilon}u_1(x,y)-{\varepsilon}^2 u_2(x,y), \label{eq:remainder}$$ where $u_1,u_2$ will be constructed by the first and second order correctors later. Since $d{\geqslant}3$, we have stationary first order correctors, so ${\varepsilon}u_1\sim O({\varepsilon})$. We do not necessarily have stationary second order correctors (by Theorem \[t:mainTh\], we need at least $d=5$), but they have zero-mean stationary gradients, so we can expect the second order correctors to grow sublinearly, which implies $|{\varepsilon}^2u_2(x,y)|\ll {\varepsilon}$. The rest is to analyze $z_{\varepsilon}$. We first construct $u_1$ and $u_2$ for our purpose. Then we prove ${\varepsilon}^2 u_2$ and $z_{\varepsilon}$ are both of order $o({\varepsilon})$. Construction of $u_1,u_2$ {#s:conu1u2} ------------------------- Similar to the continuous setting, we introduce the fast variable $y=x/{\varepsilon}\in {\mathbb{Z}}^d$ when $x\in {\varepsilon}{\mathbb{Z}}^d$. In the discrete setting, the Leibniz rule is different, and for any function $f:{\varepsilon}{\mathbb{Z}}^d\to {\mathbb{R}}$ and $g:{\mathbb{Z}}^d\to {\mathbb{R}}$, we have $$\begin{aligned} \nabla_{{\varepsilon},i}(f(x)g(y))=&[f(x+{\varepsilon}e_i)g(y+e_i)-f(x)g(y)]/{\varepsilon}\\ =&\left\{\begin{array}{l} \nabla_{{\varepsilon},i}f(x) g(y)+f(x+{\varepsilon}e_i)\frac{1}{{\varepsilon}}\nabla_ig(y),\\ \nabla_{{\varepsilon},i}f(x)g(y+e_i)+f(x)\frac{1}{{\varepsilon}}\nabla_ig(y), \end{array} \right. \end{aligned}$$ and $$\begin{aligned} \nabla_{{\varepsilon},i}^*(f(x)g(y))=&[f(x-{\varepsilon}e_i)g(y-e_i)-f(x)g(y)]/{\varepsilon}\\ =&\left\{\begin{array}{l} \nabla_{{\varepsilon},i}^*f(x) g(y)+f(x-{\varepsilon}e_i)\frac{1}{{\varepsilon}}\nabla_i^*g(y),\\ \nabla_{{\varepsilon},i}^*f(x)g(y-e_i)+f(x)\frac{1}{{\varepsilon}}\nabla_i^*g(y). \end{array} \right. \end{aligned}$$ There are two different expressions for the Leibniz rule, and for our purpose, we choose the one that does not change the microscopic variable $y$:$$\begin{aligned} \nabla_{{\varepsilon},i}(f(x)g(y))&=&\nabla_{{\varepsilon},i}f(x) g(y)+f(x+{\varepsilon}e_i)\frac{1}{{\varepsilon}}\nabla_ig(y),\\ \nabla_{{\varepsilon},i}^*(f(x)g(y))&=&\nabla_{{\varepsilon},i}^*f(x) g(y)+f(x-{\varepsilon}e_i)\frac{1}{{\varepsilon}}\nabla_i^*g(y). \label{eq:disle} \end{aligned}$$ We construct $u_1,u_2$ by expanding $\nabla_{\varepsilon}^*a(y)\nabla_{\varepsilon}(u_0+{\varepsilon}u_1+{\varepsilon}^2u_2)$ with and equating the like power of ${\varepsilon}$, which is similar to what we presented in Section \[s:2scale\]. Due to the lattice effect, the expressions of $u_1,u_2$ will be different from the continuous setting, so we present the details of the calculations here. We first have $$\nabla_{\varepsilon}^*a(y) \nabla_{\varepsilon}u_0(x)=\sum_{i=1}^d a_i(y)\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i} u_0(x)+\sum_{i=1}^d\frac{1}{{\varepsilon}}\nabla_i^*a_i(y)\nabla_{{\varepsilon},i}u_0(x-{\varepsilon}e_i), \label{eq:nau0}$$ and we define $$u_1(x,y)=\sum_{j=1}^d \nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j) \phi_{e_j}(y), \label{eq:u1}$$ where we recall that $\phi_{e_j}$ is the first order corrector in the direction of $e_j$, which is a mean-zero stationary random field when $d{\geqslant}3$ and satisfies $$\nabla^*a(y)(\nabla\phi_{e_j}(y)+e_j)=0.$$ Next, we have $$\begin{aligned} \nabla_{\varepsilon}^*a(y)\nabla_{\varepsilon}({\varepsilon}u_1(x,y))=&\sum_{i,j=1}^d {\varepsilon}\nabla_{{\varepsilon},i}^* a_i(y)\nabla_{{\varepsilon},i}(\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)\phi_{e_j}(y))\\ =&\sum_{i,j=1}^d {\varepsilon}\nabla_{{\varepsilon},i}^*a_i(y)(\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j) \phi_{e_j}(y))\\ +&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^* a_i(y)(\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)\nabla_{i}\phi_{e_j}(y)):=I_1+I_2. \end{aligned}$$ For $I_1$, we have $$\begin{aligned} I_1=&\sum_{i,j=1}^d {\varepsilon}\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)a_i(y)\phi_{e_j}(y)\\ +&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i)\nabla_i^*(a_i(y)\phi_{e_j}(y)). \end{aligned}$$ For $I_2$, we have $$\begin{aligned} I_2=&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)a_i(y)\nabla_i\phi_{e_j}(y)\\ +&\sum_{i,j=1}^d \frac{1}{{\varepsilon}}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j) \nabla_i^*(a_i(y)\nabla_i\phi_{e_j}(y)). \end{aligned}$$ By the equation satisfied by $\phi_{e_j}$, we have the second term on the r.h.s. of the above display equals to $$\sum_{i,j=1}^d \frac{1}{{\varepsilon}}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j) \nabla_i^*(a_i(y)\nabla_i\phi_{e_j}(y))=-\frac{1}{{\varepsilon}}\sum_{j=1}^d \nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)\nabla^* a(y)e_j,$$ which cancels the second term on the r.h.s. of . Therefore, we have $$\begin{aligned} &\nabla_{\varepsilon}^* a(y)\nabla_{\varepsilon}(u_0(x)+{\varepsilon}u_1(x,y))-\nabla_{\varepsilon}^*a_{\hom}\nabla_{\varepsilon}u_0(x)\\ &=\sum_{i=1}^d (a_i(y)-\bar{a})\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i} u_0(x)\\ &+\sum_{i,j=1}^d {\varepsilon}\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)a_i(y)\phi_{e_j}(y)\\ &+\sum_{i,j=1}^d \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i)\nabla_i^*(a_i(y)\phi_{e_j}(y))\\ &+\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)a_i(y)\nabla_i\phi_{e_j}(y):=J_1+J_2+J_3, \end{aligned}$$ where $$J_1=\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)[a_i(y)(1_{i=j}+\nabla_i\phi_{e_j}(y))-\bar{a}1_{i=j}],$$ $$J_2=\sum_{i,j=1}^d \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i)\nabla_i^*(a_i(y)\phi_{e_j}(y)),$$ and $$J_3=\sum_{i,j=1}^d {\varepsilon}\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)a_i(y)\phi_{e_j}(y).$$ Since ${\langle}a(\mathrm{I}_d+\nabla\phi){\rangle}=\bar{a} \mathrm{I}_d$ and $\phi_{e_j}$ is stationary, we have ${\langle}J_1{\rangle}={\langle}J_2{\rangle}=0$. On one hand, we define the second correctors $\psi_{2,i,j}$ and $\tilde{\psi}_{2,i,j}$ by $$\nabla^* a\nabla\psi_{2,i,j}=-[a_i(1_{i=j}+\nabla_i\phi_{e_j})-\bar{a}1_{i=j}] \label{eq:2ndcor1}$$ and $$\nabla^* a\nabla\tilde{\psi}_{2,i,j}=-\nabla_i^*(a_i\phi_{e_j}), \label{eq:2ndcor2}$$ then $u_2(x,y)$ is given by $$\begin{aligned} u_2(x,y)=&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},j} u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)\psi_{2,i,j}(y)\\ +&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i) \tilde{\psi}_{2,i,j}(y). \end{aligned}$$ On the other hand, by the discussion in Section \[s:q2\], we can write $$a_i(y)(1_{i=j}+\nabla_i\phi_{e_j}(y))-\bar{a}1_{i=j}=\nabla^* \sigma_{ij}(y)$$ for some stationary random field $\sigma_{ij}(y)$ with $\sigma_{ij}(0)\in L^2(\Omega)$ (this was first proved in [@gloria2014regularity; @gu-mourratmms]). Thus $J_1+J_2$ is a linear combination of terms of the form $F_V(x)V(y)$, where $F_V$ is a second derivative of $u_0$, $V=\nabla_k^* \mathcal{V}$ for some $k=1,\ldots,d$ and stationary random field ${\mathcal{V}}(y)$ with ${\mathcal{V}}(0)\in L^2(\Omega)$. Thus we can write the equation for the second order corrector corresponding to $V$ as $$\nabla^*a(y)\nabla \psi_{2,V}(y)=-V(y)=-\nabla_k^* {\mathcal{V}}(y). \label{eq:u2}$$ By [@kunneman Theorem 3, Lemma 5], we can construct solutions to such that - $\psi_{2,V}(0)=0$, - $\nabla\psi_{2,V}(y)$ is a zero-mean stationary random field and $\nabla\psi_{2,V}(0)\in L^2(\Omega)$, - $\psi_{2,V}(y)$ grows sublinearly in the following sense: $${\langle}|{\varepsilon}\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}\to 0 \label{eq:sublinear}$$ as ${\varepsilon}\to 0$ for all $x\in {\mathbb{R}}^d$, where we extend $\psi_2$ from ${\mathbb{Z}}^d$ to ${\mathbb{R}}^d$ such that $\psi_{2,V}(x)=\psi_{2,V}([x])$ with $[x]$ the integer part of $x\in{\mathbb{R}}^d$. Furthermore, we have $${\langle}|\psi_{2,V}(x)|^2 {\rangle}\les 1+|x|^2. \label{eq:subes}$$ To simplify the notations, we write $$J_1+J_2=\sum_VF_V(x)V(y),$$ where the summation is over $$V\in \{a_i(1_{i=j}+\nabla_i\phi_{e_j})-\bar{a}1_{i=j}, \nabla_i^*(a_i\phi_{e_j})\}_{i,j=1,\ldots,d},$$ and write $$u_2(x,y)=\sum_V F_V(x)\psi_{2,V}(y). \label{eq:u2ex}$$ By a similar calculation as before, we have $$\begin{aligned} &\nabla_{\varepsilon}^* a(y)\nabla_{\varepsilon}({\varepsilon}^2\sum_VF_V(x)\psi_{2,V}(y))\\ &={\varepsilon}^2\sum_V \sum_{i=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},i}F_V(x) a_i(y)\psi_{2,V}(y)\\ &+{\varepsilon}\sum_V\sum_{i=1}^d\nabla_{{\varepsilon},i} F_V(x-{\varepsilon}e_i)\nabla_i^*(a_i(y)\psi_{2,V}(y))\\ &+{\varepsilon}\sum_V\sum_{i=1}^d \nabla_{{\varepsilon},i}^* F_V(x+{\varepsilon}e_i)a_i(y)\nabla_i\psi_{2,V}(y)\\ &+\sum_V\sum_{i=1}^d F_V(x) \nabla_i^*(a_i(y)\nabla_i \psi_{2,V}(y)):=K_1+K_2+K_3+K_4. \end{aligned} \label{eq:boca}$$ By , we have $J_1+J_2+ K_4=0$, which implies $$\begin{aligned} &\nabla_{\varepsilon}^* a(y)\nabla_{\varepsilon}(u_0(x)+{\varepsilon}u_1(x,y)+{\varepsilon}^2 u_2(x,y))-\nabla_{\varepsilon}^*a_{\hom}\nabla_{\varepsilon}u_0(x)\\ =&J_3+K_1+K_2+K_3. \end{aligned}$$ To summarize, with $u_1,u_2$ given by and , we can write the equation satisfied by $z_{\varepsilon}=u_{\varepsilon}-u_0-{\varepsilon}u_1-{\varepsilon}^2u_2$ as $$\begin{aligned} &(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})z_{\varepsilon}\\ =&f-\alpha u_0-{\varepsilon}\alpha u_1-{\varepsilon}^2\alpha u_2-\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon}(u_0+{\varepsilon}u_1+{\varepsilon}^2u_2)\\ =&-{\varepsilon}\alpha u_1-{\varepsilon}^2\alpha u_2-J_3-K_1-K_2-K_3, \end{aligned} \label{eq:zeps}$$ where we recall $$\begin{aligned} J_3&=&{\varepsilon}\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)a_i(y)\phi_{e_j}(y),\\ K_1&=&{\varepsilon}^2\sum_V \sum_{i=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},i}F_V(x) a_i(y)\psi_{2,V}(y),\\ K_2&=&{\varepsilon}\sum_V\sum_{i=1}^d\nabla_{{\varepsilon},i} F_V(x-{\varepsilon}e_i)\nabla_i^*(a_i(y)\psi_{2,V}(y)),\\ K_3&=&{\varepsilon}\sum_V\sum_{i=1}^d \nabla_{{\varepsilon},i}^* F_V(x+{\varepsilon}e_i)a_i(y)\nabla_i\psi_{2,V}(y).\end{aligned}$$ Proof of Theorem \[t:2scale\] {#s:p2scale} ----------------------------- The goal of this section is to analyze $z_{\varepsilon}$ through . Since the equation is linear, we can deal with different sources term separately. We write $$(\alpha+\nabla_{\varepsilon}^* a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon}) w_{\varepsilon}=W,$$ with $W$ belonging to one of the following four groups. - *Group I.* $W(x,y)={\varepsilon}^2 F_1(x)\psi_{2,V}(y)$. We have ${\varepsilon}^2\alpha u_2,K_1$ belonging to this group, and the key feature is the factor of ${\varepsilon}^2 \psi_{2,V}(y)$. - *Group II.* $W(x,y)={\varepsilon}F_1(x)\nabla_i^*(a_i(y)\psi_{2,V}(y))$, which includes $K_2$. - *Group III.* $W(x,y)={\varepsilon}F_1(x)F_2(y)$ with $F_2$ some stationary zero-mean random field and $F_2(0)\in L^2(\Omega)$. We have ${\varepsilon}\alpha u_1$, $J_3-{\langle}J_3{\rangle}$ and $K_3-{\langle}K_3{\rangle}$ belonging to group III. - *Group IV.* $W(x,y)={\varepsilon}F_1(x)$ for some deterministic function $F_1$, which includes ${\langle}J_3{\rangle}$ and ${\langle}K_3{\rangle}$. Recalling that $F_V$ are second order derivatives of $u_0$ taking the form of $$\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},j} u_0(x-{\varepsilon}e_j+{\varepsilon}e_i) \mbox{ or } \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i),$$ by Lemma \[l:de\], the function $F_1(x)$ appearing above in all groups can be bounded by $e^{-c|x|}$ for some $c>0$. ### Group I We consider $w_{\varepsilon}$ satisfying $$(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})w_{\varepsilon}={\varepsilon}^2 F_1(x)\psi_{2,V}(y),$$ with $|F_1(x)|^2\les e^{-c|x|}$, and a simple energy estimate together Cauchy-Schwartz inequality gives $$\begin{aligned} \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|w_{\varepsilon}(x)|^2{\rangle}\les& {\varepsilon}^2 \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|F_1(x)\psi_{2,V}(\frac{x}{{\varepsilon}}) w_{\varepsilon}(x)|{\rangle}\\ \les &{\varepsilon}^2 \sqrt{ \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|F_1(x)\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}}\sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}{\langle}|w_{\varepsilon}(x)|^2{\rangle}}, \end{aligned}$$ so we have $$\begin{aligned} {\varepsilon}^d \sum_{x\in{\varepsilon}{\mathbb{Z}}^d}{\langle}|w_{\varepsilon}(x)|^2{\rangle}\les& {\varepsilon}^2{\varepsilon}^d \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|F_1(x){\varepsilon}\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}\\ \les &{\varepsilon}^2{\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} e^{-c|x|} {\langle}|{\varepsilon}\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}={\varepsilon}^2o(1), \end{aligned}$$ where the last step comes from and . Thus we have $\|w_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon})$. ### Group II We consider $w_{\varepsilon}$ satisfying $$(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})w_{\varepsilon}={\varepsilon}F_1(x)\nabla_i^*(a_i(y)\psi_{2,V}(y)),$$ and use energy estimate to obtain $$\begin{aligned} \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} ( {\langle}|w_{\varepsilon}(x)|^2{\rangle}+ {\langle}|\nabla_{\varepsilon}w_{\varepsilon}|^2{\rangle})\les& {\varepsilon}|{\langle}\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} F_1(x)\nabla_i^*(a_i(y)\psi_{2,V}(y))w_{\varepsilon}(x){\rangle}|\\ =&{\varepsilon}^2 |{\langle}\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} F_1(x)\nabla_{{\varepsilon},i}^*(a_i(\frac{x}{{\varepsilon}})\psi_{2,V}(\frac{x}{{\varepsilon}}))w_{\varepsilon}(x) {\rangle}|. \end{aligned}$$ An integration by parts leads to $$\begin{aligned} \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} ( {\langle}|w_{\varepsilon}(x)|^2{\rangle}+ {\langle}|\nabla_{\varepsilon}w_{\varepsilon}|^2{\rangle})\les &{\varepsilon}^2\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}{\langle}|\nabla_{{\varepsilon},i}(F_1(x)w_{\varepsilon}(x))a_i(\frac{x}{{\varepsilon}})\psi_{2,V}(\frac{x}{{\varepsilon}})|{\rangle}\\ \les &{\varepsilon}^2 \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|\nabla_{{\varepsilon},i}F_1(x) w_{\varepsilon}(x)a_i(\frac{x}{{\varepsilon}})\psi_{2,V}(\frac{x}{{\varepsilon}})|{\rangle}\\ +& {\varepsilon}^2 \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} {\langle}|F_1(x+{\varepsilon}e_i) \nabla_{{\varepsilon},i}w_{\varepsilon}(x)a_i(\frac{x}{{\varepsilon}})\psi_{2,V}(\frac{x}{{\varepsilon}})|{\rangle}. \end{aligned}$$ By Lemma \[l:de\], $|\nabla_{{\varepsilon},i}F_1(x)|^2+|F_1(x+{\varepsilon}e_i)|^2\les e^{-c|x|}$, so by Cauchy-Schwartz inequality we have $$\begin{aligned} &\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} ( {\langle}|w_{\varepsilon}(x)|^2{\rangle}+ {\langle}|\nabla_{\varepsilon}w_{\varepsilon}|^2{\rangle})\\ \les& {\varepsilon}^2\sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}e^{-c|x|}{\langle}|\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}} \left(\sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}{\langle}|w_{\varepsilon}(x)|^2{\rangle}}+\sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}{\langle}|\nabla_{{\varepsilon},i}w_{\varepsilon}(x)|^2{\rangle}} \right)\\ \les &{\varepsilon}^2\sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}e^{-c|x|}{\langle}|\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}} \sqrt{\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} ( {\langle}|w_{\varepsilon}(x)|^2{\rangle}+ {\langle}|\nabla_{\varepsilon}w_{\varepsilon}|^2{\rangle})}, \end{aligned}$$ thus $${\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} ( {\langle}|w_{\varepsilon}(x)|^2{\rangle}+ {\langle}|\nabla_{\varepsilon}w_{\varepsilon}|^2{\rangle})\les {\varepsilon}^2 {\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}e^{-c|x|}{\langle}|{\varepsilon}\psi_{2,V}(\frac{x}{{\varepsilon}})|^2{\rangle}={\varepsilon}^2o(1),$$ and we conclude $ \|w_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon})$. ### Group III We consider $z_{\varepsilon}$ satisfying $$(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})w_{\varepsilon}={\varepsilon}F_1(x)F_2(y),$$ where $|F_1(x)|\les e^{-c|x|}$ and $F_2$ is a zero-mean stationary random field with $F_2(0)\in L^2(\Omega)$. The proof borrows the following result from [@kunneman Equation (4.23i) – (4.23iii)], which itself comes from [@papanicolaou1979boundary] in the continuous setting: there exists a random field $H(y)=(H_1(y),\ldots,H_d(y))$ on ${\mathbb{Z}}^d$ that satisfies the following properties: - $H_j(0)=0$ and $\sum_{j=1}^d \nabla_j H_j(y)=F_2(y), y\in{\mathbb{Z}}^d$, - by extending $H(y)$ from ${\mathbb{Z}}^d$ to ${\mathbb{R}}^d$ such that $H(x)=H([x])$ for any $x\in{\mathbb{R}}^d$ with $[x]$ the integer part of $x$, we have $${\langle}|{\varepsilon}H(\frac{x}{{\varepsilon}})|^2{\rangle}\to 0 \label{eq:sub1}$$ as ${\varepsilon}\to 0$, and $${\langle}|H(x)|^2{\rangle}\les 1+|x|^2. \label{eq:sub2}$$ Now we only need to apply the same proof as for *Group II* to conclude $$\|w_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon}).$$ ### Group IV This is the key part where we show the deterministic bias vanishes in the order of ${\varepsilon}$. We consider $w_{\varepsilon}$ satisfying $$(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})w_{\varepsilon}={\varepsilon}F_1(x),$$ where $F_1:{\varepsilon}{\mathbb{Z}}^d\to {\mathbb{R}}$ is some fast decaying deterministic function. If we define $z_0$ satisfying $$(\alpha+\nabla_{\varepsilon}^*a_{\hom}\nabla_{\varepsilon})w_0={\varepsilon}F_1(x),$$ it is clear by the standard homogenization result that $ \|w_{\varepsilon}-w_0\|_{2,{\varepsilon}} =o({\varepsilon})$. Now we claim that in our specific case the deterministic bias $\|w_0\|_{2,{\varepsilon}}=o({\varepsilon})$, which implies $\|w_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon})$. First we note that $$\begin{aligned} (\alpha+\nabla_{\varepsilon}^* a_{\hom}\nabla_{\varepsilon}) w_0=&-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}\\ =&-{\varepsilon}\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j){\langle}a_i\phi_{e_j}{\rangle}\\ &-{\varepsilon}\sum_V\sum_{k=1}^d \nabla_{{\varepsilon},k}^* F_V(x+{\varepsilon}e_k){\langle}a_k\nabla_k\psi_{2,V}{\rangle}, \end{aligned} \label{eq:w0}$$ and by recalling $u_2(x,y)$ given by $$\begin{aligned} u_2(x,y)=\sum_VF_V(x)\psi_{2,V}(y)=&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},j} u_0(x-{\varepsilon}e_j+{\varepsilon}e_i)\psi_{2,i,j}(y)\\ +&\sum_{i,j=1}^d \nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i) \tilde{\psi}_{2,i,j}(y), \end{aligned}$$ we can write the second term on the r.h.s. of as $$\begin{aligned} &\sum_V\sum_{k=1}^d \nabla_{{\varepsilon},k}^* F_V(x+{\varepsilon}e_k){\langle}a_k\nabla_k\psi_{2,V}{\rangle}\\ =&\sum_{k,i,j=1}^d \nabla_{{\varepsilon},k}^*\nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},j} u_0(x-{\varepsilon}e_j+{\varepsilon}e_i+{\varepsilon}e_k) {\langle}a_k\nabla_k\psi_{2,i,j}{\rangle}\\ +&\sum_{k,i,j=1}^d \nabla_{{\varepsilon},k}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j-{\varepsilon}e_i+{\varepsilon}e_k){\langle}a_k\nabla_k\tilde{\psi}_{2,i,j}{\rangle}. \end{aligned}$$ By Lemma \[l:fide\], we have $\|w_0-\bar{w}_0\|_{2,{\varepsilon}}=o({\varepsilon})$ with $\bar{w}_0$ solving the corresponding equation in ${\mathbb{R}}^d$: $$\begin{aligned} \frac{1}{{\varepsilon}}(\alpha-\nabla\cdot a_{\hom}\nabla)\bar{w}_0 &= \sum_{i,j=1}^d \partial_{x_ix_ix_j} \bar{u}_0(x){\langle}a_i\phi_{e_j}{\rangle}\\ &- \sum_{k,i,j=1}^d \partial_{x_kx_ix_j}\bar{u}_0(x){\langle}a_k\nabla_k\psi_{2,i,j}{\rangle}+ \sum_{k,i,j=1}^d \partial_{x_kx_ix_j}\bar{u}_0(x){\langle}a_k\nabla_k\tilde{\psi}_{2,i,j}{\rangle}, \end{aligned} \label{eq:z0}$$ where $\bar{u}_0$ satisfies $$(\alpha-\nabla\cdot a_{\hom} \nabla)\bar{u}_0=f$$ on ${\mathbb{R}}^d$. Now we show that r.h.s. of is zero so $\bar{w}_0\equiv 0$. For the second term on the r.h.s. of , we have $${\langle}a_k\nabla_k\psi_{2,i,j}{\rangle}={\langle}\nabla_k^* a_k\psi_{2,i,j}{\rangle}=-{\langle}\nabla^* a\nabla \phi_{e_k}\psi_{2,i,j}{\rangle}=-{\langle}\phi_{e_k}\nabla^*a\nabla\psi_{2,i,j}{\rangle}.$$ By the equation satisfied by $\psi_{2,i,j}$ , we have $${\langle}a_k\nabla_k\psi_{2,i,j}{\rangle}={\langle}\phi_{e_k}a_i(1_{i=j}+\nabla_i\phi_{e_j}){\rangle}={\langle}\phi_{e_k}a_i{\rangle}1_{i=j}+{\langle}\phi_{e_k}a_i\nabla_i\phi_{e_j}{\rangle}.$$ Similarly, for the third term on the r.h.s. of , we have $${\langle}a_k\nabla_k\tilde{\psi}_{2,i,j}{\rangle}=-{\langle}\phi_{e_k}\nabla^*a\nabla\tilde{\psi}_{2,i,j}{\rangle}={\langle}\phi_{e_k}\nabla_i^*(a_i\phi_{e_j}){\rangle}={\langle}\nabla_i\phi_{e_k}a_i\phi_{e_j}{\rangle}.$$ Therefore, we have $$\begin{aligned} \mbox{r.h.s. of \eqref{eq:z0}}=&\sum_{i,j=1}^d \partial_{x_ix_ix_j}\bar{u}_0(x){\langle}a_i\phi_{e_j}{\rangle}-\sum_{k,i,j=1}^d \partial_{x_kx_ix_j}\bar{u}_0(x){\langle}\phi_{e_k} a_i{\rangle}1_{i=j}\\ -&\sum_{k,i,j=1}^d \partial_{x_kx_ix_j} \bar{u}_0(x){\langle}\phi_{e_k}a_i\nabla_i\phi_{e_j}{\rangle}+\sum_{k,i,j=1}^d \partial_{x_kx_ix_j}\bar{u}_0(x){\langle}\nabla_i\phi_{e_k}a_i\phi_{e_j}{\rangle}\equiv0. \end{aligned}$$ To summarize, we have $\|w_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon})$. ### Summary Recall that $z_{\varepsilon}(x)=u_{\varepsilon}(x)-u_0(x)-{\varepsilon}u_1(x,y)-{\varepsilon}^2 u_2(x,y)$ and $$\begin{aligned} &(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})z_{\varepsilon}\\ =&-{\varepsilon}\alpha u_1-{\varepsilon}^2\alpha u_2-(J_3-{\langle}J_3{\rangle})-K_1-K_2-(K_3-{\langle}K_3{\rangle})-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}. \end{aligned}$$ By the previous discussion on *Group I,II,III,IV*, we have $\|z_{\varepsilon}\|_{2,{\varepsilon}}=o({\varepsilon})$ and the discussion for *Group I* already shows that $$\|{\varepsilon}^2u_2\|_{2,{\varepsilon}}=o({\varepsilon}),$$ and this gives $$\|u_{\varepsilon}-u_0-{\varepsilon}u_1\|_{2,{\varepsilon}}=o({\varepsilon}).$$ Now we only need to note that $\|u_1-\tilde{u}_1\|_{2,{\varepsilon}}=O({\varepsilon})$ with $$\tilde{u}_1(x,\frac{x}{{\varepsilon}})=\sum_{j=1}^d \nabla_{{\varepsilon},j}u_0(x)\phi_{e_j}(\frac{x}{{\varepsilon}})$$ to complete the proof of Theorem \[t:2scale\]. Higher order fluctuations {#s:high} ========================= With the proof of Theorem \[t:2scale\], it is clear that a similar approach will lead to a higher order expansions of $u_{\varepsilon}$ in higher dimensions, e.g., when $d{\geqslant}5$, we can include in the expansion the second order correctors (which are stationary by Theorem \[t:mainTh\]) and prove the remainder is of order $o({\varepsilon}^2)$. The calculation is more involved but the idea is the same. We first illustrate it in the continuous setting, then we present a proof of Theorem \[t:highex\] in the discrete setting. High order errors in the continuous setting {#s:excon} ------------------------------------------- For fixed $n{\geqslant}2$ and equations $$(\alpha-\nabla\cdot a(\frac{x}{{\varepsilon}})\nabla) u_{\varepsilon}(x)= f(x),\ \ (\alpha-\nabla\cdot a_{\hom}\nabla) u_0(x)=f(x),$$ with $x\in{\mathbb{R}}^d,\alpha>0,f\in{\mathcal{C}}_c^\infty({\mathbb{R}}^d)$, the goal is to obtain an expansion of $u_{\varepsilon}-u_0$ up to order ${\varepsilon}^n$ in dimension $d{\geqslant}2n+1$. When $d{\geqslant}2n+1$, the $n-$th order correctors are stationary by Theorem \[t:mainTh\]. To show the remainder is small, we need to construct up to $(n+1)$-th order correctors. It can be done as in [@papanicolaou1979boundary Theorem 2] since we have zero-mean stationary gradients. Now we consider $$\mathscr{J}_n:=[\alpha-(\nabla_x+\frac{1}{{\varepsilon}}\nabla_y)\cdot a(y) (\nabla_x+\frac{1}{{\varepsilon}}\nabla_y)] (u_0+{\varepsilon}u_1+\ldots+{\varepsilon}^n u_n+{\varepsilon}^{n+1}u_{n+1}),$$ and construct $u_1,u_2, \{u_k\}_{3{\leqslant}k{\leqslant}n+1}$ satisfy the following equations: $$\nabla_y\cdot a(y)\nabla_y u_1+\nabla_y\cdot a(y)\nabla_x u_0=0,$$ $$\begin{aligned} &\nabla_y\cdot a(y)\nabla_yu_2+\nabla_y\cdot a(y)\nabla_x u_1+\nabla_x\cdot a(y)\nabla_y u_1+\nabla_x \cdot a(y)\nabla_x u_0\\ =&{\langle}\nabla_x\cdot a(y)\nabla_y u_1+\nabla_x \cdot a(y)\nabla_x u_0{\rangle}, \end{aligned}$$ and $$\begin{aligned} &\nabla_y \cdot a(y)\nabla_y u_k+\nabla_y\cdot a(y)\nabla_x u_{k-1}+\nabla_x \cdot a(y)\nabla_y u_{k-1}+\nabla_x \cdot a(y)\nabla_x u_{k-2}-\alpha u_{k-2}\\ =&{\langle}\nabla_x \cdot a(y)\nabla_y u_{k-1}+\nabla_x \cdot a(y)\nabla_x u_{k-2}{\rangle}. \end{aligned} \label{eq:highcor1}$$ Our definition of higher order correctors $u_k,k{\geqslant}3$ in is different from due to the extra term $\alpha u_{k-2}$, but it is clear that Theorem \[t:mainTh\] still applies in the corresponding discrete setting. By the above construction, it is clear that $$\begin{aligned} \mathscr{J}_n=&\alpha u_0-\sum_{k=2}^{n+1}{\varepsilon}^{k-2}{\langle}\nabla_x \cdot a(y)\nabla_y u_{k-1}+\nabla_x \cdot a(y)\nabla_x u_{k-2}{\rangle}\\ -&{\varepsilon}^n\nabla_x\cdot a(y)\nabla_x u_n-{\varepsilon}^{n}\nabla_x\cdot a(y)\nabla_y u_{n+1}-{\varepsilon}^{n}\nabla_y \cdot a(y)\nabla_x u_{n+1}-{\varepsilon}^{n+1}\nabla_x\cdot a(y)\nabla_x u_{n+1}\\ +& \alpha {\varepsilon}^n u_n+\alpha {\varepsilon}^{n+1} u_{n+1}. \end{aligned}$$ By defining $$z_{\varepsilon}=u_{\varepsilon}-u_0-{\varepsilon}u_1-\ldots-{\varepsilon}^{n+1}u_{n+1}, \label{eq:ex1}$$ and the fact that $\alpha u_0-{\langle}\nabla_x \cdot a(y)\nabla_y u_1+\nabla_x \cdot a(y)\nabla_x u_0{\rangle}=f$, we derive $$\begin{aligned} (\alpha-\nabla\cdot a(\frac{x}{{\varepsilon}})\nabla)z_{\varepsilon}=&f-\mathscr{J}_n\\ =&\sum_{k=3}^{n+2}{\varepsilon}^{k-2}{\langle}\nabla_x \cdot a(y)\nabla_y u_{k-1}+\nabla_x \cdot a(y)\nabla_x u_{k-2}{\rangle}\\ &+{\varepsilon}^n\nabla_x\cdot a(y)\nabla_x u_n-{\langle}{\varepsilon}^n\nabla_x\cdot a(y)\nabla_x u_n{\rangle}\\ &+{\varepsilon}^{n}\nabla_x\cdot a(y)\nabla_y u_{n+1}-{\langle}{\varepsilon}^{n}\nabla_x\cdot a(y)\nabla_y u_{n+1}{\rangle}\\ &+{\varepsilon}^{n}\nabla_y \cdot a(y)\nabla_x u_{n+1}+{\varepsilon}^{n+1}\nabla_x\cdot a(y)\nabla_x u_{n+1}\\ &-\alpha {\varepsilon}^n u_n-\alpha {\varepsilon}^{n+1} u_{n+1}, \label{eq:exhigh} \end{aligned}$$ and the goal reduces to refine $z_{\varepsilon}$ up to the order ${\varepsilon}^n$. There are two types of sources on the r.h.s. of : - *random sources*: by the same proof for *Group I,II,III* in Section \[s:p2scale\], we have the terms in the last four lines of the above display contribute $o({\varepsilon}^n)$ to $z_{\varepsilon}$. Here we need a stationary $n-$th order corrector, and a sublinear $(n+1)$-th order corrector with a zero-mean stationary gradient. - *deterministic sources*: we can write $$\sum_{k=3}^{n+2}{\varepsilon}^{k-2}{\langle}\nabla_x \cdot a(y)\nabla_y u_{k-1}+\nabla_x \cdot a(y)\nabla_x u_{k-2}{\rangle}=\sum_{k=1}^n {\varepsilon}^k f_k(x) \label{eq:desou}$$ where $f_k$ is some linear combination of derivatives of $u_0$. For equations of the form $$(\alpha-\nabla\cdot a(\frac{x}{{\varepsilon}})\nabla)w_{\varepsilon}(x)=f_k(x),$$ we can apply the expansion again, and derive equations of the form to obtain an expansion of $w_{\varepsilon}$. By iteration, we can go up to the order of ${\varepsilon}^n$ in finite steps. To summarize, when $d{\geqslant}2n+1$, we expect there exists $\{u_k\}_{k=1}^n$ and $\{v_k\}_{k=1}^n$ such that ${\langle}u_i{\rangle}=0$, $v_i$ are deterministic, and $$\left(\int_{{\mathbb{R}}^d}\bigg{\langle}\bigg|u_{\varepsilon}-u_0-\sum_{k=1}^n {\varepsilon}^k u_k-\sum_{k=1}^n {\varepsilon}^kv_k\bigg|^2\bigg{\rangle}\, \d x\right)^{\frac12}=o({\varepsilon}^n). \label{eq:highexre}$$ It is worth mentioning that the $\{u_k\}_{k=1}^n$ appearing in is different from the ones in , since the deterministic source in also contributes to the random error. Proof of Theorem \[t:highex\] ----------------------------- For any $n{\geqslant}2$, to expand up to $o({\varepsilon}^n)$, we need to construct the $(n+1)$-th correctors $\psi_{n+1}$. By Theorem \[t:mainTh\] and [@kunneman Theorem 3, Lemma 5], this can be done when $d{\geqslant}2n+1$, and we have a sublinear random field $\psi_{n+1}(y)$ with $\psi_{n+1}(0)=0$ and a zero-mean stationary gradient $\nabla\psi_{n+1}$. We can continue the calculation in Section \[s:conu1u2\] and construct higher order correctors $\{u_k\}_{3{\leqslant}k{\leqslant}n+1}$ as in Section \[s:excon\]. In the following we only discuss the case $n=2$ since it already includes all the ingredients of the proof. To construct the third order correctors, we recall $$\begin{aligned} &(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})(u_{\varepsilon}-u_0-{\varepsilon}u_1-{\varepsilon}^2u_2)\\ =&-{\varepsilon}\alpha u_1-{\varepsilon}^2\alpha u_2-(J_3-{\langle}J_3{\rangle})-K_1-K_2-(K_3-{\langle}K_3{\rangle})-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}, \label{eq:ex2nd} \end{aligned}$$ with $$\begin{aligned} J_3&=&{\varepsilon}\sum_{i,j=1}^d \nabla_{{\varepsilon},i}^*\nabla_{{\varepsilon},i}\nabla_{{\varepsilon},j}u_0(x-{\varepsilon}e_j)a_i(y)\phi_{e_j}(y),\\ K_1&=&{\varepsilon}^2\sum_V \sum_{i=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},i}F_V(x) a_i(y)\psi_{2,V}(y),\\ K_2&=&{\varepsilon}\sum_V\sum_{i=1}^d\nabla_{{\varepsilon},i} F_V(x-{\varepsilon}e_i)\nabla_i^*(a_i(y)\psi_{2,V}(y)),\\ K_3&=&{\varepsilon}\sum_V\sum_{i=1}^d \nabla_{{\varepsilon},i}^* F_V(x+{\varepsilon}e_i)a_i(y)\nabla_i\psi_{2,V}(y).\end{aligned}$$ Similar to $u_2$, we define $u_3$ to get rid of the zero-mean terms on the r.h.s. of of order ${\varepsilon}$, which is written as $${\varepsilon}\sum_U G_U(x)U(y)=-{\varepsilon}\alpha u_1 -(J_3-{\langle}J_3{\rangle}) -K_2-(K_3-{\langle}K_3{\rangle}).$$ Here $G_U$ is some derivative of $u_0$ (note that due to the term $-{\varepsilon}\alpha u_1$, $G_U$ is not necessarily a *third order* derivative of $u_0$) and the summation is over zero-mean stationary random fields $$U\in \{\phi_{e_i},a_i\phi_{e_j}-{\langle}a_i\phi_{e_j}{\rangle},\nabla_i^*(a_i\psi_{2,V}),a_i\nabla_i\psi_{2,V}-{\langle}a_i\nabla\psi_{2,V}{\rangle}\}_{i,j=1,\ldots,d}.$$ We define $$u_3(x,y)=\sum_U G_U(x)\psi_{3,U}(y),$$ with $\psi_{3,U}$ the corrector corresponding to $U$, i.e., $$\nabla^*a(y)\nabla \psi_{3,U}(y)=U(y).$$ By the same calculation as in , we have $$\begin{aligned} \nabla_{\varepsilon}^* a(y)\nabla_{\varepsilon}({\varepsilon}^3u_3)=&{\varepsilon}^3\sum_U \sum_{i=1}^d \nabla_{{\varepsilon},i}^* \nabla_{{\varepsilon},i}G_U(x) a_i(y)\psi_{3,U}(y)\\ +&{\varepsilon}^2 \sum_U\sum_{i=1}^d\nabla_{{\varepsilon},i} G_U(x-{\varepsilon}e_i)\nabla_i^*(a_i(y)\psi_{3,U}(y))\\ +&{\varepsilon}^2 \sum_U\sum_{i=1}^d \nabla_{{\varepsilon},i}^* G_U(x+{\varepsilon}e_i)a_i(y)\nabla_i\psi_{3,U}(y)\\ +&{\varepsilon}\sum_U\sum_{i=1}^d G_U(x) \nabla_i^*(a_i(y)\nabla_i \psi_{3,U}(y)):=L_1+L_2+L_3+L_4. \end{aligned}$$ It is clear that $L_4={\varepsilon}\sum_U G_U(x)U(y)$, and by , we have $$z_{\varepsilon}=u_{\varepsilon}-u_0-{\varepsilon}u_1-{\varepsilon}^2u_2-{\varepsilon}^3u_3$$ satisfies $$\begin{aligned} &(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})z_{\varepsilon}\\ =&-{\varepsilon}^2\alpha u_2-{\varepsilon}^3\alpha u_3-K_1-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}-L_1-L_2-L_3. \end{aligned}$$ Similar to , we can write $$\begin{aligned} (\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})z_{\varepsilon}=&-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}-{\langle}K_1{\rangle}-{\langle}L_3{\rangle}\\ &-K_1+{\langle}K_1{\rangle}-L_3+{\langle}L_3{\rangle}\\ &-L_1-L_2\\ &-{\varepsilon}^2\alpha u_2-{\varepsilon}^3\alpha u_3. \end{aligned}$$ For the random source in the last three lines of the above display, by the same proof as in Section \[s:p2scale\], their contributions to $z_{\varepsilon}$ is $o({\varepsilon}^2)$. For the deterministic source $-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}-{\langle}K_1{\rangle}-{\langle}L_3{\rangle}$, it is of the form ${\varepsilon}F_1(x)$, where $F_1(x),x\in{\varepsilon}{\mathbb{Z}}^d$ is deterministic and fast decaying, so we can apply the same expansion again to refine the solution to $$(\alpha+\nabla_{\varepsilon}^*a(\frac{x}{{\varepsilon}})\nabla_{\varepsilon})w_{\varepsilon}=-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}-{\langle}K_1{\rangle}-{\langle}L_3{\rangle}$$ up to $o({\varepsilon}^2)$. More precisely, for the source $-{\langle}K_1{\rangle}-{\langle}L_3{\rangle}$, it is of order $O({\varepsilon}^2)$, so we can define $v_2$ as the solution to $$(\alpha+\nabla_{\varepsilon}^*a_{\hom}\nabla_{\varepsilon})v_2=-({\langle}K_1{\rangle}+{\langle}L_3{\rangle})/{\varepsilon}^2.$$ For the source $-{\langle}J_3{\rangle}-{\langle}K_3{\rangle}$, it is of order $O({\varepsilon})$. If we define $v_{\varepsilon}$ as the solution to $$(\alpha+\nabla_{\varepsilon}^* a(\frac{x}{{\varepsilon}}) \nabla_{\varepsilon})v_{\varepsilon}= -({\langle}J_3{\rangle}+{\langle}K_3{\rangle})/{\varepsilon},$$ then by Theorem \[t:2scale\], $v_{\varepsilon}$ can be refined up to order $o({\varepsilon})$, i.e., let $v_1$ solve $$(\alpha+\nabla_{\varepsilon}^* a_{\hom} \nabla_{\varepsilon})v_1= -({\langle}J_3{\rangle}+{\langle}K_3{\rangle})/{\varepsilon},$$ which is the same as , and there exists $v_{11}$ such that ${\langle}v_{11}{\rangle}=0$ and $$\|v_{\varepsilon}-v_1-{\varepsilon}v_{11}\|_{2,{\varepsilon}}=o({\varepsilon}).$$ Now it is clear that $$\|w_{\varepsilon}-{\varepsilon}v_1-{\varepsilon}^2v_{11}-{\varepsilon}^2 v_2\|_{2,{\varepsilon}}=o({\varepsilon}^2).$$ To summarize, we have shown that $$\|u_{\varepsilon}-u_0-{\varepsilon}u_1-{\varepsilon}^2u_2-{\varepsilon}v_1-{\varepsilon}^2v_{11}-{\varepsilon}^2v_2\|_{2,{\varepsilon}}=o({\varepsilon}^2),$$ where $u_1,u_2,v_{11}$ are zero-mean random fluctuations and $v_1,v_2$ are deterministic bias. This completes the proof of Theorem \[t:highex\] for $n=2$. Discussion: local vs global or strong vs weak random fluctuations {#s:compa} ================================================================= Theorem \[t:2scale\] shows that when $d{\geqslant}3$, $\|u_{\varepsilon}-u_0-{\varepsilon}u_1\|_{2,{\varepsilon}}=o({\varepsilon})$ with the first order random fluctuations represented by $${\varepsilon}u_1(x,\frac{x}{{\varepsilon}})={\varepsilon}\sum_{j=1}^d \nabla_{{\varepsilon},j}u_0(x)\phi_{e_j}(\frac{x}{{\varepsilon}}).$$ It is consistent with [@gu-mourrat Theorem 2.5] where a pointwise version is proved in the continuous setting: for fixed $x\in {\mathbb{R}}^d$, $$u_{\varepsilon}(x)=u_0(x)+{\varepsilon}\sum_{j=1}^d \partial_{x_j}u_0(x)\phi_{e_j}(\frac{x}{{\varepsilon}})+o({\varepsilon}), \label{eq:point}$$ with $o({\varepsilon})/{\varepsilon}\to0$ in $L^1(\Omega)$. Neither of the two results implies the other. On one hand, it is not clear how to obtain the pointwise estimates by the analytic approach, in particular the energy estimate described in Section \[s:p2scale\]; on the other hand, the probabilistic approach used in [@gu-mourrat] loses track of the dependence of $o({\varepsilon})$ in on $x\in{\mathbb{R}}^d$, so does not easily extend to an $L^2({\mathbb{R}}^d\times \Omega)$ version. Both results indicate that $u_1(x,x/{\varepsilon})$ represents the first order random fluctuation measured in a strong sense. It is however not the case after a spatial average with respect to a test function as pointed out in [@gu-mourratmms]. Central limit theorems were derived for large scale fluctuations of ${\varepsilon}u_1(x,x/{\varepsilon})$ and $u_{\varepsilon}-{\langle}u_{\varepsilon}{\rangle}$ in [@mourrat2014correlation; @MN; @gu-mourratmms; @gloria2016random; @armstrong2016scaling]. When $d{\geqslant}3$, the result shows for any $g\in{\mathcal{C}}_c^\infty({\mathbb{R}}^d)$ $$\frac{1}{{\varepsilon}^{d/2}}\int_{{\mathbb{R}}^d} {\varepsilon}u_1(x,\frac{x}{{\varepsilon}})g(x)\,\d x\Rightarrow N(0,\sigma^2), \label{eq:cltcor}$$ $$\frac{1}{{\varepsilon}^{d/2}}\int_{{\mathbb{R}}^d} (u_{\varepsilon}(x)-{\langle}u_{\varepsilon}(x){\rangle})g(x)\,\d x \Rightarrow N(0,\tilde{\sigma}^2), \label{eq:cltso}$$ and $\sigma^2\neq \tilde{\sigma}^2$. The mismatch between the two variances in and suggests that the higher order correctors become visible in . When $d{\geqslant}2n+1$, the $n-$th order correctors $\psi_n$ are stationary, and a covariance estimate gives $$\begin{aligned} |{\langle}\psi_n(0)\psi_n(x){\rangle}| \les& \sum_e \|\partial_e \psi_n(0)\|_2\|\partial_e \psi_n(x)\|_2\\ \les &\sum_e \frac{1}{|{\underline}{e}|_*^{(d-n)-}}\frac{1}{|x-{\underline}{e}|_*^{(d-n)-}}\les \frac{1}{|x|_*^{(d-2n)-}}. \end{aligned}$$ The scaling indicates that as a stationary random field, $\psi_n(x)$ decorrelates almost at the rate of $|x|^{-(d-2n)}$ (it was proved rigorously in [@mourrat2014correlation] for $n=1$). In other words, the higher order correctors have stronger correlations and the decay of correlation functions slows down as the order increases! It partly explains why we have contributions from high order correctors weakly in space. If we abuse the notation and consider $$I_n:=\frac{1}{{\varepsilon}^{d/2}}\int_{{\mathbb{R}}^d} {\varepsilon}^n \psi_n(\frac{x}{{\varepsilon}}) g(x)dx,$$ we have $${\langle}|I_n|^2{\rangle}=\int_{{\mathbb{R}}^{2d}} {\varepsilon}^{2n-d} {\langle}\psi_n(\frac{x}{{\varepsilon}})\psi_n(\frac{y}{{\varepsilon}}){\rangle}g(x)g(y)dxdy.$$ If we assume ${\langle}\psi_n(0)\psi_n(x){\rangle}\sim |x|^{-(d-2n)}$ as $|x|\to\infty$, the scaling indicates that ${\langle}|I_n|^2{\rangle}$ is of order $O(1)$, i.e., the $n-$th order corrector contributes weakly in space. In the recent preprint [@gloria2016random], the authors provided another way of understanding the mismatch between and in terms of the so-called homogenization commutator. It turns out that if the formal expansion applies to the gradient $\nabla u_{\varepsilon}$, it leads to the correct large scale random fluctuation; see [@gloria2016random Page 6]. The expansion on the gradient of the Green’s function in [@gu-mourratmms Proposition 4.2] shares the same spirit. It is worth mentioning that a nice non-local expansion was proved in [@gloria2016random Corollary 2] which identifies a single term that captures the large scale fluctuation. The existence of stationary correctors: removing the massive term {#s:exst} ================================================================= The goal in this section is to show that for the corrector equation defined when $d{\geqslant}2n-1$ $$(\lambda+\nabla^*a(y)\nabla) \psi_n^\lambda(y)=F(y)=\left\{\begin{array}{l} \nabla_i^* (a_i(y)\psi_{n-1}(y)),\\ a_i(y) \nabla_i\psi_{n-1}(y)-{\langle}a_i(y)\nabla_i \psi_{n-1}(y){\rangle},\\ a_{i}(y)\psi_{n-2}(y)-{\langle}a_{i}(y)\psi_{n-2}(y){\rangle}, \end{array} \right.$$ the uniform estimates $\|\psi_n^\lambda\|_p\les 1$ and $\|\nabla\psi_n^\lambda\|_p\les 1$ implies the existence of a stationary corrector and a stationary gradient, respectively. In Section \[s:q2\], we proved that when $d{\geqslant}2n-1$, there exists a stationary random field $\Psi$ such that $\Psi(0)\in L^2(\Omega)$ and $$F=\nabla^* \Psi.$$ The rest is standard, and we present it for the convenience of the reader. We first introduce some notations. Let $\omega=(\omega_e)_{e\in{\mathbb{B}}}\in \Omega$ denote the sample point, and for $x\in {\mathbb{Z}}^d$, we define the shift operator $\tau_x$ on $\Omega$ by $(\tau_x \omega)_e=\omega_{x+e}$, where $x+e:=(x+\underline{e},x+\bar{e})$ is the edge obtained by shifting $e$ by $x$. Since $\{\omega_e\}_{e\in\mathbb{B}}$ are i.i.d., $\{\tau_x\}_{x\in {\mathbb{Z}}^d}$ is a group of measure-preserving transformations. We can define the operator $$T_x f(\zeta)=f(\tau_x\zeta)$$ for any measurable function $f$ on $\Omega$, and the generators of $T_x$, denoted by $\{D_i\}_{i=1}^d$, are defined by $D_if:=T_{e_i}f-f$. The adjoint $D_i^*$ is defined by $D_i^* f:=T_{-e_i}f-f$. We denote the gradient on $\Omega$ by $D=(D_1,\ldots,D_d)$ and the divergence $D^*F:=\sum_{i=1}^d D_i^*F_i$ for $F:\Omega\to {\mathbb{R}}^d$. Let $\phi_\lambda$ solve $(\lambda+\nabla^* a\nabla)\phi_\lambda=\nabla^*\Psi$ with $\Psi$ a stationary random field on ${\mathbb{Z}}^d$ such that $\Psi(0)\in L^2(\Omega)$. For the equation $$\nabla^*a\nabla\phi=\nabla^*\Psi, \label{eq:corap}$$ \(i) there exists a random field $\phi$ solving such that $\nabla\phi$ is stationary and $\nabla\phi(0)\in L^2(\Omega)$. \(ii) if ${\langle}|\phi_\lambda|^2{\rangle}\les1$, then there exists a stationary random field $\phi$ solving such that $\phi(0)\in L^2(\Omega)$. \[l:stacor\] Part (i) comes from [@kunneman Theorem 3]. For Part (ii), let $\tilde{a}(\omega)=(\omega_{e_1},\ldots,\omega_{e_d})$ and $\tilde{\Psi}\in L^2(\Omega)$ so that $\Psi(x)=\tilde{\Psi}(\tau_x\omega)$, we lift the equation to the probability space $$(\lambda+D^*\tilde{a} D)\tilde{\phi}_\lambda=D^*\tilde{\Psi},$$ so it is clear that $\phi_\lambda(x)=\tilde{\phi}_\lambda(\tau_x\omega)$. Since ${\langle}|\tilde{\phi}_\lambda|^2{\rangle}\les 1$, we can extract a subsequence $\tilde{\phi}_\lambda\to \tilde{\phi}$ weakly in $L^2$, which implies $D\tilde{\phi}_\lambda\to D\tilde{\phi}$ weakly in $L^2(\Omega)$. For any $G\in L^2(\Omega)$, we have $$\lambda{\langle}\tilde{\phi}_\lambda G{\rangle}+{\langle}\tilde{a}D\tilde{\phi}_\lambda DG{\rangle}={\langle}D^*\tilde{\Psi}G{\rangle},$$ and sending $\lambda\to 0$ leads to $${\langle}\tilde{a}D\tilde{\phi}DG{\rangle}={\langle}D^*\tilde{\Psi}G{\rangle},$$ so $D^*\tilde{a}D\tilde{\phi}=D^*\tilde{\Psi}$. Now we define $\phi(x)=\tilde{\phi}(\tau_x\omega)$, and it is clear that $$\nabla^*a\nabla\phi=\nabla^*\Psi.$$ The proof is complete. Finite difference approximation =============================== The following are some classical results and we present it for the convenience of the reader. We recall the classical Green’s function estimates: $${\mathcal{G}}^\alpha(x,y)\les \frac{e^{-c\sqrt{\alpha}|x-y|}}{|x-y|_*^{d-2}} \label{eq:gres}$$ for some $c>0$, with ${\mathcal{G}}^\alpha$ the Green’s function of $\alpha+\nabla^*\nabla$ on ${\mathbb{Z}}^d$. Let ${\mathcal{G}}_{\varepsilon}^\alpha(x,y)$ be the Green’s function of $\alpha+\nabla_{\varepsilon}^*\nabla_{\varepsilon}$ on ${\varepsilon}{\mathbb{Z}}^d$, it is clear that $${\mathcal{G}}_{\varepsilon}^\alpha(x,y)={\varepsilon}^2{\mathcal{G}}^{{\varepsilon}^2 \alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}}).$$ Let $u_{\varepsilon}$ solve $$(\alpha+\nabla_{\varepsilon}^*\nabla_{\varepsilon})u_{\varepsilon}=f_{\varepsilon},$$ on ${\varepsilon}{\mathbb{Z}}^d$, with $f_{\varepsilon}$ satisfying (i) there exists $\lambda>0$ s.t. $|f_{\varepsilon}(x)|\les e^{-\lambda|x|}$, (ii) all discrete derivatives of $f_{\varepsilon}$ satisfies (i), then for $v_{\varepsilon}=u_{\varepsilon}$ or any derivative of $u_{\varepsilon}$, we have $$|v_{\varepsilon}(x)|\les e^{-c|x|}$$ for constant $c>0$ independent of ${\varepsilon}>0,x\in{\varepsilon}{\mathbb{Z}}^d$. \[l:de\] It suffices to consider the case $v_{\varepsilon}=u_{\varepsilon}$. For derivatives, e.g., $v_{\varepsilon}=\nabla_{{\varepsilon},i} u_{\varepsilon}$, using the fact that $\nabla_{{\varepsilon},i}$ commutes with $\nabla_{{\varepsilon},j}$ and $\nabla_{{\varepsilon},j}^*$, we have $$(\alpha+\nabla_{\varepsilon}^*\nabla_{\varepsilon})v_{\varepsilon}=\nabla_{{\varepsilon},i}f_{\varepsilon},$$ so we only need to apply the result when $v_{\varepsilon}=u_{\varepsilon}$. For $u_{\varepsilon}(x)$, we use the Green’s function representation $$u_{\varepsilon}(x)=\sum_{y\in{\varepsilon}{\mathbb{Z}}^d} {\mathcal{G}}^\alpha_{\varepsilon}(x,y)f(y)=\sum_{y\in{\varepsilon}{\mathbb{Z}}^d}{\varepsilon}^2 {\mathcal{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})f(y),$$ and by , we have $${\varepsilon}^2 {\mathcal{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})\les {\varepsilon}^2 \frac{e^{-\rho|x-y|}}{|\frac{x-y}{{\varepsilon}}|_*^{d-2}}$$ for some constant $\rho>0$. Thus $$|u_{\varepsilon}(x)|\les {\varepsilon}^d\sum_{y\in {\varepsilon}{\mathbb{Z}}^d, y\neq x} \frac{e^{-\rho |x-y|}}{|x-y|^{d-2}}e^{-\lambda|y|}+{\varepsilon}^2e^{-\lambda|x|}.$$ For the first term on the r.h.s. of the above expression, we only need to decompose the summation into $\sum_{|y|<|x|/2}$ and $\sum_{|y|{\geqslant}|x|/2}$ to complete the proof. Let $u_{\varepsilon}$ solve $$(\alpha+\nabla_{\varepsilon}^* \nabla_{\varepsilon})u_{\varepsilon}=f_{\varepsilon}$$ on ${\varepsilon}{\mathbb{Z}}^d$, with $f_{\varepsilon}$ satisfying (i) there exists $\lambda>0$ s.t. $|f_{\varepsilon}(x)|\les e^{-\lambda |x|}$, (ii) there exists a continuous function $\bar{f}:{\mathbb{R}}^d\to {\mathbb{R}}$ such that $\|f_{\varepsilon}-\bar{f}\|_{2,{\varepsilon}}\to0$ as ${\varepsilon}\to 0$, then for $\bar{u}$ solving $$(\alpha-\Delta)\bar{u}=\bar{f}$$ on ${\mathbb{R}}^d$, we have $\|u_{\varepsilon}-\bar{u}\|_{2,{\varepsilon}}\to0$ as ${\varepsilon}\to 0$. \[l:fide\] We first note that $|\bar{f}(x)|\les e^{-\lambda|x|}$ for $x\in{\mathbb{R}}^d$. The rest of the proof is decomposed into three steps. *Step 1.* We write $$u_{\varepsilon}(x)=\sum_{y\in{\varepsilon}{\mathbb{Z}}^d} {\varepsilon}^2 {\mathcal{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})f_{\varepsilon}(y),$$ and define $$\bar{u}_{\varepsilon}(x):=\sum_{y\in{\varepsilon}{\mathbb{Z}}^d} {\varepsilon}^2 {\mathcal{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})\bar{f}(y)$$ on ${\varepsilon}{\mathbb{Z}}^d$. The goal is to show $\|u_{\varepsilon}-\bar{u}_{\varepsilon}\|_{2,{\varepsilon}}\to 0$ as ${\varepsilon}\to 0$. By , we deduce $$\begin{aligned} |u_{\varepsilon}(x)-\bar{u}_{\varepsilon}(x)|\les& \sum_{y\in{\varepsilon}{\mathbb{Z}}^d} {\varepsilon}^2 \frac{e^{-\rho|x-y|}}{|\frac{x-y}{{\varepsilon}}|_*^{d-2}}|f_{\varepsilon}(y)-\bar{f}(y)|\\ \les&{\varepsilon}^d \sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x} \frac{e^{-\rho|x-y|}}{|x-y|^{d-2}}|f_{\varepsilon}(y)-\bar{f}(y)|+{\varepsilon}^2|f_{\varepsilon}(x)-\bar{f}(x)|, \end{aligned}$$ thus $$\begin{aligned} &{\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} |u_{\varepsilon}(x)-\bar{u}_{\varepsilon}(x)|^2 \\ \les &{\varepsilon}^d \sum_{x\in{\varepsilon}{\mathbb{Z}}^d} \left( \sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x} {\varepsilon}^d\frac{e^{-\rho|x-y|}}{|x-y|^{d-2}} \sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x}{\varepsilon}^d\frac{e^{-\rho|x-y|}}{|x-y|^{d-2}}|f_{\varepsilon}(y)-\bar{f}(y)|^2\right)\\ +&{\varepsilon}^d{\varepsilon}^4\sum_{x\in{\varepsilon}{\mathbb{Z}}^d}|f_{\varepsilon}(x)-\bar{f}(x)|^2:=I_1+I_2 \end{aligned}$$ It is clear by assumption that $I_2\to 0$, and $$\begin{aligned} I_1\les& {\varepsilon}^d \sum_{x\in{\varepsilon}{\mathbb{Z}}^d}\sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x}{\varepsilon}^d\frac{e^{-\rho|x-y|}}{|x-y|^{d-2}}|f_{\varepsilon}(y)-\bar{f}(y)|^2\\ \les &{\varepsilon}^d \sum_{y\in{\varepsilon}{\mathbb{Z}}^d}|f_{\varepsilon}(y)-\bar{f}(y)|^2\to 0 \end{aligned}$$ as ${\varepsilon}\to 0$. *Step 2.* Define $$\tilde{u}_{\varepsilon}(x)={\varepsilon}^d\sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x} {\mathscr{G}}^\alpha(x,y)\bar{f}(y)$$ on ${\varepsilon}{\mathbb{Z}}^d$, with ${\mathscr{G}}^\alpha$ the continuous Green’s function of $\alpha-\Delta$ in ${\mathbb{R}}^d$. The goal is to show $\|\bar{u}_{\varepsilon}-\tilde{u}_{\varepsilon}\|_{2,{\varepsilon}}\to 0$ as ${\varepsilon}\to 0$. We first have $$|\bar{u}_{\varepsilon}(x)-\tilde{u}_{\varepsilon}(x)|\les \sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x} |{\varepsilon}^2 {\mathcal{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})-{\varepsilon}^2{\mathscr{G}}^{{\varepsilon}^2\alpha}(\frac{x}{{\varepsilon}},\frac{y}{{\varepsilon}})||\bar{f}(y)|+{\varepsilon}^2|\bar{f}(x)|.$$ where we used the scaling property ${\mathscr{G}}^{{\varepsilon}^2\alpha}(x/{\varepsilon},y/{\varepsilon})={\varepsilon}^{d-2}{\mathscr{G}}^\alpha(x,y)$. By [@conlon2 Lemma 3.1], $$|{\mathcal{G}}^\lambda(x,y)-{\mathscr{G}}^\lambda(x,y)|\les \frac{e^{-c \sqrt{\lambda}|x-y|}}{|x-y|_*^{d-1}}$$ for some $c>0$, so $$|\bar{u}_{\varepsilon}(x)-\tilde{u}_{\varepsilon}(x)|\les \sum_{y\in{\varepsilon}{\mathbb{Z}}^d,y\neq x} {\varepsilon}^{d+1} \frac{e^{-c\sqrt{\alpha}|x-y|}}{|x-y|^{d-1}}|\bar{f}(y)|+{\varepsilon}^2|\bar{f}(x)| \les {\varepsilon}e^{-c|x|}$$ for some $c>0$, which implies $\|\bar{u}_{\varepsilon}-\tilde{u}_{\varepsilon}\|_{2,{\varepsilon}}\to 0$. *Step 3.* We first extend $\tilde{u}_{\varepsilon}(x)$ from ${\varepsilon}{\mathbb{Z}}^d$ to ${\mathbb{R}}^d$ such that $\tilde{u}_{\varepsilon}(x)=\tilde{u}_{\varepsilon}([x]_{\varepsilon})$ with $[x]_{\varepsilon}$ the ${\varepsilon}-$integer part of $x$. Then we consider $$\|\tilde{u}_{\varepsilon}-\bar{u}\|_{2,{\varepsilon}}^2={\varepsilon}^d\sum_{x\in{\varepsilon}{\mathbb{Z}}^d} |\tilde{u}_{\varepsilon}(x)-\bar{u}(x)|^2 =\int_{{\mathbb{R}}^d} |\tilde{u}_{\varepsilon}(x)-\bar{u}([x]_{\varepsilon})|^2dx.$$ It is clear that $\int_{{\mathbb{R}}^d} |\bar{u}([x]_{\varepsilon})-\bar{u}(x)|^2dx\to 0$ as ${\varepsilon}\to 0$, so we only need to show $$\int_{{\mathbb{R}}^d} |\tilde{u}_{\varepsilon}(x)-\bar{u}(x)|^2dx\to 0.$$ Since $\bar{u}(x)=\int_{{\mathbb{R}}^d} {\mathscr{G}}^\alpha(x,y)\bar{f}(y)dy$, we have $\tilde{u}_{\varepsilon}(x)\to \bar{u}(x)$ for $x\in{\mathbb{R}}^d$. Now by Lemma \[l:de\], $|\tilde{u}_{\varepsilon}(x)|+|\bar{u}(x)|\les e^{-c|x|}$ for some $c>0$, so by dominated convergence theorem, the proof is complete. Acknowledgments {#acknowledgments .unnumbered} --------------- We would like to thank the anonymous referee for several helpful suggestions. The author’s research is partially funded by grant DMS-1613301 from the US National Science Foundation. [10]{} , [*Quantitative stochastic homogenization of convex integral functionals*]{}, Annales scientifiques de l’Ecole normale supérieure, 48 (2016), pp. 423–481. , [*Mesoscopic higher regularity and subadditivity in elliptic homogenization*]{}, Comm. Math. Phys., to appear. height 2pt depth -1.6pt width 23pt, [*Scaling limits of energies and correctors*]{}, arXiv preprint arXiv:1603.03388, (2016). height 2pt depth -1.6pt width 23pt, [*The additive structure of elliptic homogenization*]{}, preprint, arXiv:1602.00512 (2016). , [*Lipschitz regularity for elliptic equations with random coefficients*]{}, Arch. Ration. Mech. Anal., 219 (2016), pp. 255–348. , [*Stochastic homogenization of linear elliptic equations: High-order error estimates in weak norms via second-order correctors*]{}, preprint, arXiv:1609.01528 (2016). , [*A central limit theorem for the effective conductance: Linear boundary data and small ellipticity contrasts*]{}, Communications in Mathematical Physics, 328 (2014), pp. 701–731. , [*Estimates in probability of the residual between the random and the homogenized solutions of one-dimensional second-order operator*]{}, Asymptot. Anal., 21 (1999), pp. 303–315. , [*Rates of convergence for the homogenization of fully nonlinear uniformly elliptic pde in random media*]{}, Inventiones Mathematicae, 180 (2010), pp. 301–360. , [*On homogenization of elliptic equations with random coefficients*]{}, Electronic Journal of Probability, 5 (2000), pp. 1–58. , [*Strong convergence to the homogenized limit of elliptic equations with random coefficients*]{}, Trans. Amer. Math. Soc., 366 (2014), pp. 1257–1288. , [*On estimating the derivatives of symmetric diffusions in stationary random environments, with applications to the $\nabla\phi$ interface model*]{}, Probability Theory and Related Fields 133 (2005), pp. 358–390. , [*Homogenization in polygonal domains*]{}, J. Eur. Math. Soc.(JEMS), 13 (2011), pp. 1477–1503. height 2pt depth -1.6pt width 23pt, [*Homogenization and boundary layers*]{}, Acta mathematica, (2012), pp. 1–46. , [*The structure of fluctuations in stochastic homogenization*]{}, preprint, arXiv: 1602.01717 (2016). , [*A higher-order large-scale regularity theory for random elliptic operators*]{}, arXiv:1503.07578 (2015). , [*Fluctuation of solutions to linear elliptic equations with noisy diffusion coefficients*]{}, Comm. Partial Differential Equations, 38 (2013), pp. 304–338. , [*Annealed estimates on the Green functions and uncertainty quantification*]{}, Ann. Inst. H. Poincaré Anal. Non Linéaire, to appear. , [*Quantification of ergodicity in stochastic homogenization: optimal bounds via spectral gap on glauber dynamics*]{}, Inventiones mathematicae 199 (2015), pp. 455–515. height 2pt depth -1.6pt width 23pt, [*An optimal quantitative two-scale expansion in stochastic homogenization of discrete elliptic equations*]{}, ESAIM: Mathematical Modelling and Numerical Analysis, 48 (2014), pp. 325–346. height 2pt depth -1.6pt width 23pt, [*A regularity theory for random elliptic operators*]{}, preprint, arXiv:1409.2678 (2014). , [*A quantitative central limit theorem for the effective conductance on the discrete torus*]{}, Communications on Pure and Applied Mathematics, to appear. , [*An optimal variance estimate in stochastic homogenization of discrete elliptic equations*]{}, Annals of Probability, 39 (2011), pp. 779–856. height 2pt depth -1.6pt width 23pt, [*An optimal error estimate in stochastic homogenization of discrete elliptic equations*]{}, Annals of Applied Probability, 22 (2012), pp. 1–28. height 2pt depth -1.6pt width 23pt, [*Quantitative results on the corrector equation in stochastic homogenization*]{}, Journal of the European Mathematical Society, to appear. height 2pt depth -1.6pt width 23pt, [*The corrector in stochastic homogenization: Near-optimal rates with optimal stochastic integrability*]{}, preprint, arXiv: 1510.08290 (2015). , [*A central limit theorem for fluctuations in 1d stochastic homogenization*]{}, , onlinefirst, 2016. , [*Pointwise two-scale expansion for parabolic equations with random coefficients*]{}, , onlinefirst, 2015. height 2pt depth -1.6pt width 23pt, [*Scaling limit of fluctuations in stochastic homogenization*]{}, Multiscale Model. Simul., 14 (2016), pp. 452–481. , [*Averaging of random operators*]{}, Matematicheskii Sbornik, 151 (1979), pp. 188–202. , [*The diffusion limit for reversible jump processes on ${\mathbb{Z}}^d$ with ergodic random bond conductivities*]{}, Communications in Mathematical Physics, 90 (1983), pp. 27–68. , [*Annealed estimates on the green’s function*]{}, Probability Theory and Related Fields, 163 (2015), pp. 527–573. , [*Variance decay for functionals of the environment viewed by the particle*]{}, Ann. Inst. H. Poincar[é]{} Probab. Statist, 47 (2011), pp. 294–327. height 2pt depth -1.6pt width 23pt, [*Kantorovich distance in the martingale clt and quantitative homogenization of parabolic equations with random coefficients*]{}, Probability Theory and Related Fields 160 (2014), pp. 279–314. , [*A Scaling limit of the corrector in stochastic homogenization*]{}, Annals of Applied Probability, to appear. , [*Correlation structure of the corrector in stochastic homogenization*]{}, Annals of Probability, to appear. , [*Estimates on the variance of some homogenization problems*]{}, Preprint, (1998). , [*Normal approximation for a random elliptic equation*]{}, Probability Theory and Related Fields 159 (2014), pp. 661–700. height 2pt depth -1.6pt width 23pt, [*Normal approximation for the net flux through a random conductor*]{}, SPDE: Analysis and Computations, 4 (2016), pp. 439–476. , [*Boundary value problems with rapidly oscillating random coefficients*]{}, in Random fields, Vol. I, II (Esztergom, 1979), Colloq. Math. Soc. J[á]{}nos Bolyai, 27, North Holland, Amsterdam, New York, 1981, pp. 835–873. , [*Noise-stability and central limit theorems for effective resistance of random electric networks*]{}, Annals of Probability, to appear. , [*Averaging of symmetric diffusion in random medium*]{}, Siberian Mathematical Journal, 27 (1986), pp. 603–613.

--- abstract: 'A method of formation of the tightly confined distortion-free fs pulses with the step-like decreasing of intensity under the finite-length propagation in free space is described. Such pulses are formed by the Fresnel source of a high refraction-index waveguide. The source reproduces in free space a propagation-invariant (distortion-free) pulse confined by the waveguide. Converse to the case of material waveguides, when the pulse goes out from the Fresnel (virtual) waveguide its shape is not changed, but the intensity immediately drops down to the near-zero level. OCIS numbers: 320.0320, 320.2250, 320.5540, 320.5550.' author: - 'G. Nyitray and S. V. Kukhlevsky' date: | *Department of Experimental Physics, Institute of Physics,\ University of Pecs, Ifjusag u. 6, H-7624 Pecs, Hungary* title: 'Distortion-free tight confinement and step-like decay of fs pulses in free space' --- Tight transverse confinement of a light wave in free space together with fast decreasing of its intensity after the finite-length propagation are required in many fields of optics, such as non-destroying light-matter interaction, control of light penetration-depth and surface processing. The effect is usually achieved by strong focusing a light beam by short-curvature-radius lenses or spherical mirrors. In the case of fs pulses, the control of the transverse and longitudinal intensity distributions is a particularly difficult problem, because the ultrashort pulses are distorted in the space and time domains by such optical elements (see, for example Refs. [@Kemp; @Bor; @Jack; @Bert; @Niso; @Ferm]). Recently, it was shown that the tightly confined propagation-invariant (distortion-free) continuous waves with the fast decreasing of the intensity after the finite-length propagation in free space could be formed by the Fresnel source of a high refractive-index hollow waveguide having total-reflection walls [@Kuk1; @Cann]. If such an effect exists for the ultrashort pulses, this could be very important for the fs optics and applications. In the present article, the formation of tightly confined distortion-free fs pulses with the step-like decreasing of intensity under finite-length propagation in free space is described. Let us describe the Fresnel source that reproduces in free space a propagation-invariant (distortion-free) pulse tightly confined by a high refractive-index hollow waveguide with total-reflection walls. The plane-parallel guide having the length $L_w$ and width $2a$ is considered. A propagation-invariant time-harmonic wave $E'(x',z,t,{\omega})$ is supported in the free-space by the constructive interference of multiple beams $E'_m(x',z,t,{\omega})$ launching from the Fresnel zones $[x_{m}^{min},x_m^{max}]$ of the virtual source of the guide [@Kuk2]: $$\begin{aligned} E'(x',z,t,{\omega})=\sum_{m=-M}^{M}E'_m(x',z,t,{\omega}),\end{aligned}$$ with $$\begin{aligned} E'_m(x',z,t,{\omega})={\frac{1}{\sqrt{2{i}\lambda}}}\int_{x_m^{min}}^{x_m^{max}} {\frac{\exp{[ik{r(x',x)}]}}{r(x',x)}}\nonumber \\ {(1+\cos\Theta_m){E_{m}(x,0,t,{\omega})}dx},\end{aligned}$$ where $2M+1$ is the number of zones (beams) of the Fresnel source that contribute the energy into the field $E'(x',z,t,{\omega})$; ${\mid}x_m^{max}{\mid}=(2m+1)a$ and ${\mid}x_m^{min}{\mid}=(2m-1)a$; ${\omega}=kc$ is the wave frequency; $r(x',x)$ is the distance between points $x'$ and $x$; $E_m(x,0,t,{\omega})=E_0([x-2am](-1)^m,0,t,{\omega})$, where $E_0(x,0,t,{\omega})$ is the field at the guide entrance $(z=0)$. The value $M=M(z,d_M)$ is determined by the transverse dimension $d_M(k,a,z)$ of the beam $E'_M(x',z,t,{\omega})$ at the guide exit $(z=L_w)$. By analogy with the Fresnel lens, the virtual source of the waveguide is called the Fresnel waveguide [@Cann]. The source reproduces in free space a diffraction-free beam [@Brit; @Ziol; @Durn; @Gori; @Lu] confined by the waveguide. In the case of an ultrashort pulse $E'(x',z,t)$ guided by the waveguide, the field of the pulse at the guide entrance is represented in the form of a Fourier integral $F[E(x,z=0,{\omega})]$. Using the Fresnel-waveguide representation for the Fourier components and substituting the result into the Fourier integral we get the propagation-invariant pulse $E'(x',z,t)$ in free space [@Kuk2]. The pulse is supported in free space by the constructive interference of the $2M(z)+1$ ultrashort-pulses launching from the Fresnel source of the waveguide. The Fresnel source reproduces in free space a distortion-free (propagation-invariant) pulse confined by the waveguide. Converse to the case of material waveguides [@Marc; @Oka], when the pulse goes out from the Fresnel (virtual) waveguide its shape is not changed, but the intensity immediately drops down to the near-zero level. The effect, which looks like the instant decay or annihilation of a tightly confined pulse, is illustrated by the numerical examples in Figs. 1(a) and 1(b). Figure 1(a) shows the intensity distribution ${\mid}E(x,z){\mid}^2$ on the axis $(x=0)$ of the virtual waveguide computed for the pulse having the two different durations $\tau$: 100 and 150 fs. The intensity distribution of the continuous wave $({\tau}=\infty)$ is shown in the figure for the comparison. Figure 1(b) demonstrates the intensity distributions of the pulse (${\tau}$=150 fs) at the points B, C and D. We notice that the pulse propagates in the region of the virtual waveguide as the superposition of many pulses that diffract in the off-axis direction and interfere with each other. Analysis of the intensity distributions indicates existence of the two main stages of the evolution of the pulse under the free-space propagation: the propagation-invariant AC $(z<L_w)$ and collapse CD $(z>L_w)$ stages. In the region AC $(z<L_w)$, the intensity distribution of the pulse is practically unchanged that demonstrates the propagation-invariant properties of the pulse. The pulse amplitude is constant in the region AB and shows some oscillations in the part BC of the region AC. The amplitude of these oscillations, which increases with increasing the pulse duration, reaches maximum value before the pulse decay. In the collapse region CD $(z>L_w)$, the number of pulses $2M(z=Lw)+1$ is not sufficient for supporting of the propagation-invariant field distribution. In this region, the pulse shape is not changed, but the intensity immediately drops down to the near-zero level. The pulse decay is caused by the destructive interference between the pulses launched from the finite-width Fresnel source of the finite-length waveguide. The distance $AC=L_w$ strongly correlates with the width $(2M(z=L_w)+1)2a$ of the Fresnel source. This presents a method of formation of the tightly confined distortion-free fs pulses with the step-like decreasing of intensity under the finite-length propagation in free space. The propagation length AC could be controlled by variation of the width of the Fresnel source. The above-described theoretical principle of the Fresnel source can be realized experimentally by using the $2M(z)+1$ fs-pulses. In the simplest case, the source can be formed by the interference of only two fs-pulses propagating in the off-axis direction, as shown in Fig. 2. Notice, that the two-pulse technique is similar to the method of formation of the dynamical gratings (see, for example Refs. [@Mazn; @Goodno]). In conclusion, a method of formation of the tightly confined distortion-free fs pulses with the step-like decreasing of intensity under the finite-length propagation in free space was described. The pulses are formed by the Fresnel source of a high refraction-index waveguide. The length of the distortion-free propagation and the time of the pulse decay are controlled by variation of the width of the Fresnel-waveguide source. It should be noted that the method can be extended to the guiding of fs pulses in free-space by the 3-dimensional virtual waveguides. In this case, one could use the Fresnel sources of 3-dimensional high-refraction-index waveguides [@Kuk3]. This study was supported by the Fifth Framework of the European Commission (Financial support from the EC for shared-cost RTD actions: research and technological development projects, demonstration projects and combined projects. Contract NG6RD-CT-2001-00602). The authors thank the Computing Services Center, Faculty of Science, University of Pecs, for providing computational resources. [99]{} M. Kempe, U. Stamm, B. Wilhelmi, Opt. Commun. [**59**]{}, 119 (1986). Z. Bor, Opt. Lett. [**[14]{}**]{}, 119 (1989). S. Jackel, R. Burris, J. Grun, A. Ting, C. Manaka, K. Evans, J. Kosakowskii, Opt. Lett. [**20**]{}, 1086 (1995). M. Bertolotti, A. Ferrari, L. Sereda, J. Opt. Soc. Am. B, 1519 (1995). M. Nisoli, S. De Silvestri, O. Svelto, R. Szipocs, K. Ferencz, Ch. Spielmann, S. Sartania, Opt. Lett. [**22**]{}, 522 (1997). M.E. Ferman, Opt. Letts. [**23**]{}, 52 (1998). S.V. Kukhlevsky, G. Nyitray, Opt. Commun. [**218**]{}, 213 (2003). J. Canning, E. Buckley, K. Lytikainen, Opt. Lett., 230 (2003). S.V. Kukhlevsky, G. Nyitray, V.L. Kantsyrev, Phys. Rev. E [**64**]{}, 026603 (2001). J.N. Brittingham, Appl. Phys. [**54**]{}, 1179 (1983). R.W. Ziolkowski, J. Math. Phys. [**26**]{}, 861 (1985). J. Durnin, et al, Phys. Rev. Lett. [**58**]{}, 1499 (1987). F. Gori, et al, Opt. Commun. [**64**]{}, 491 (1987). J. Y. Lu, J. G. Greenleaf, IEEE Trans. Ultrason.Ferroelec. Freq. Contr. [**39**]{}, 19 (1992). D. Marcuse, Theory of Dielectric OpticalWaveguides (Academic Publishers, New York, 1974). K. Okamoto, Fundamentals of OpticalWaveguides (Academic Press, New York, 2000). A. A. Maznev, T. F. Crimmins, K. A. Nelson Opt. Lett. [**23**]{}, 1378 (1998). G. D. Goodno, G. Dadusc, R. J. Miller J. Opt. Soc. Am. B, 1791 (1998). S.V. Kukhlevsky, G. Nyitray, J. Mod. Opt. [**50**]{}, 2043 (2003). List of Figure Captions {#list-of-figure-captions .unnumbered} ======================= Fig. 1. (a) The step-like evolution of the normalized intensity of the Gaussian-shaped ultrashort light pulse on the virtual-waveguide axis under the free-space propagation computed for the two different pulse durations $\tau$ : (1) - 100 fs and (2) - 150 fs. The intensity distribution of the continuous wave (3) and the step-like function (step(z)) are shown for the comparison. (b) The normalized intensity distributions of the propagation invariant (distortion-free) pulse ($\tau$ = 150 fs) at the points B, C and D. The white dashed line indicates the region of the virtual waveguide. The guided pulse is supported in the free space by the diffraction and interference of $2M+1$ pulses of the Fresnel-waveguide source. The intensity distributions were calculated using the following parameters: ${\lambda}_0$ = 500 nm, $z{\in}$\[0.05 m, 4.05 m\], $x{\in}$\[-750 ${\mu}$m, 750 ${\mu}$m\], ${\tau}{\in}$\[-750 fs, 750 fs\] and $M=71$, where ${\lambda}_0$ is the he central wavelength of the pulse (wave-packet). Fig. 2. The schematic diagram of the Fresnel-wavegiude source formed by the interference of the two fs-pulses ($p_1$ and $p_2$) propagating in the off-axis direction. The box indicated by the white dashed line shows the region of the virtual waveguide.

--- abstract: 'The aim of this note is to show that every subset of a given topological space is the intersection of a preopen and a preclosed set, therefore [$\beta$-locally]{} closed, and that every topological space is [$\beta$-submaximal]{}.' author: - 'Julian [Dontchev]{} and Maximilian [Ganster]{}' title: 'A remark on $\beta$-locally closed sets' --- plus 1pt minus 1pt Introduction {#s1} ============ In a recent paper, Gnanambal and Balachandran [@GB1] introduced the classes of [$\beta$-locally]{} closed sets, [$\beta$-submaximal]{} spaces and $\beta$-LC-continuous functions. The purpose of our note is to show that every subset of any topological space is the intersection of a preopen set and a preclosed set, hence [$\beta$-locally]{}closed, and therefore every function [$f \colon (X,\tau) \rightarrow (Y,\sigma)$]{} is $\beta$-LC-continuous. We have felt the need to point out explicitly this observation since over the years several papers have investigated concepts like “pre-locally closed sets” or “[$\beta$-locally]{} closed sets” which do not have any nontrivial meaning. In addition, we will show that every space is [$\beta$-submaximal]{} and we will point out that most results of [@GB1] are either trivial or false. Let $A$ be a subset of a topological space $(X,\tau)$. Following Kronheimer [@K1], we call the interior of the closure of $A$, denoted by $A^+$, the [*consolidation*]{} of $A$. Sets included in their consolidation are called [*preopen*]{} or [*locally dense*]{}. Complements of preopen sets are called [*preclosed*]{} and the preclosure of a set $A$, denoted by ${\rm pcl} (A)$, is the intersection of all preclosed supersets of $A$. Since union of preopen sets is also preopen, the preclosure of every set is in fact a preclosed set. If $A$ is included in the closure of its consolidation, then $A$ is called [*$\beta$-open*]{} or [*semi-preopen*]{}. Complements of $\beta$-open sets are called [*$\beta$-closed*]{}. The $\beta$-closure of $A$, denoted by ${\rm cl}_{\beta} (A)$ is the intersection of all $\beta$-closed supersets of $A$. In [@GB1], Gnanambal and Balachandran called a set $A$ [*[$\beta$-locally]{} closed*]{} if $A$ is intersection of a $\beta$-open and a $\beta$-closed set. They defined a set $A$ to be [*$\beta$-dense*]{} [@GB1] if ${\rm cl}_{\beta} (A) = X$ and called a space $X$ [*[$\beta$-submaximal]{} *]{} [@GB1] if every $\beta$-dense subset is $\beta$-open. A function [$f \colon (X,\tau) \rightarrow (Y,\sigma)$]{} is called [*$\beta$-LC-continuous*]{} [@GB1] if the preimage of every open subset of $Y$ is [$\beta$-locally]{} closed in $X$. The following implications hold and none of them is reversible: dense $\Rightarrow$ preopen $\Rightarrow$ $\beta$-open $\Rightarrow$ [$\beta$-locally]{} closed Every set is [$\beta$-locally]{} closed {#s2} ======================================= \[p1\] Every subset $A$ of a topological space $(X,\tau)$ is the intersection of a preopen and a preclosed set, hence pre-locally closed. [*Proof.*]{} Let $A \subseteq (X,\tau)$. Set $A_1 = A \cup (X \setminus {\rm cl} (A))$. Since $A_1$ is dense in $X$, it is also preopen. Let $A_2$ be the preclosure of $A$, i.e., $A_2 = A \cup {\rm cl} ({\rm int} (A))$. Clearly, $A_2$ is a preclosed set. Note now that $A = A_1 \cap A_2$. $\Box$ \[c1\] [(i)]{} Every set is [$\beta$-locally]{} closed and every function is $\beta$-LC-continuous. [(i)]{} Every topological space is [$\beta$-submaximal]{}. [*Proof.*]{} (i) Every preopen (resp. preclosed) set is $\beta$-open (resp. $\beta$-closed). \(ii) By [@GB1 Corollary 3.24] a topological space is [$\beta$-submaximal]{}if and only if every set is [$\beta$-locally]{} closed. \[r1\] *(i) Corollary \[c1\] makes [@GB1] trivial.* \(ii) Example 3.4 from [@GB1] is wrong as the subset $A = \{ \frac{1}{n} \colon n = 1,2,\ldots \} \cup (2,3) \cup (3,4) \cup \{ 4 \} \cup (5,6) \cup \{ x \colon x$ is irrational and $7 \leq x < 8 \}$ of the real line $\mathbb R$ is indeed [$\beta$-locally]{} closed. \(iii) Proposition 3.6 from [@GB1] is wrong as every proper nonempty subset of the real line $\mathbb R$ with the indiscrete topology is $\beta$-open and preclosed but not semi-open. \(iv) Example 4.11 from [@GB1] is wrong, since the space $(X,\tau)$, where $X = \{ a,b,c,d \}$, $\tau = \{ \emptyset, \{ a,b \}, \{ c,d \}, X \}$ is [*not*]{} an ${\alpha}{\beta}$-space. Note that $\{ a \}$ is $\beta$-open but not $\alpha$-open (an [*$\alpha$-open set*]{} is a set which is the difference of an open and a nowhere dense set). \(v) An ${\alpha}{\beta}$-space [@GB1] is in fact a strongly irresolvable, extremally disconnected space. \(vi) An $\alpha$-locally closed set ([@GB1 Definition 2.1 (x)]) is nothing else but a simly-open set. [2]{} , [$\beta$-locally closed sets and $\beta$-LC-continuous functions]{}, [*Mem. Fac. Sci. Kochi Univ. Ser. A Math.*]{}, [**19**]{} (1998), 35–44. , [The topology of digital images]{}, [*Topology Appl.*]{}, [**46**]{} (3) (1992), 279–303. Department of Mathematics\ University of Helsinki\ PL 4, Yliopistonkatu 5\ 00014 Helsinki\ Finland\ e-mail: [dontchev@cc.helsinki.fi]{} Department of Mathematics\ Graz University of Technology\ Steyrergasse 30\ A-8010 Graz\ Austria\ e-mail: [ganster@weyl.math.tu-graz.ac.at]{}

--- author: - | Bonavera L., Suarez Gomez S. L., Gonz[á]{}lez-Nuevo J., Santos J.D., Sanchez M.L.,\ Muñiz R. , de Cos F.J. bibliography: - './SDNN.bib' title: 'Point Source Detection with Fully-Convolutional Networks: Performance in Realistic Simulations' --- [Point Sources are one of the main contaminants to the recovery of Cosmic Microwave Background signal at small scales, and their careful detection will be important for the next generation of Cosmic Microwave Background experiments like CORE.]{} [We want to develop a method based on Fully Convolutional Networks to detect sources in realistic simulations and compare its performance against one of the most used point source detection method in this context, the Mexican Hat wavelet 2 (MHW2). The frequencies for our analysis are the 143, 217 and 353 GHz Planck channels.]{} [We produce realistic simulations of Point Sources at each frequency taking into account potential contaminating signals as the Cosmic Microwave Background, the Cosmic Infrared Background, the Galactic thermal emission and the instrumental noise. We first produce a set of training simulations a 217 GHz to train the network. Then we apply both the neural network and the wavelet to recover the point sources in the validating simulations at all the frequencies, comparing the results by estimating the reliability, completeness and flux density estimation accuracy.]{} [In the extra-galactic region with a 30$^\circ$ galactic cut, the neural network successfully recover point sources with 90% of completeness corresponding to 300, 139 and 227 mJy for 143, 217 and 353 GHz respectively. On the same validation simulations, the wavelet with a 3$\sigma$ flux density detection limit, recover point sources till 224, 124 and 154 mJy at the 90% completeness. To reduce the amount of spurious sources, we also apply a safer 4$\sigma$ flux density detection limit increasing the 90% of completeness levels: 298, 173 and 227 mJy. In all cases the neural network produce a much lower number of spurious sources with respect the MHW2. As expected, the results on spurious sources for both techniques worsen when increasing the frequency or reducing the galactic cut to 10$^\circ$.]{} [Our results suggests that the neural networks are a very promising approach to detect point sources using data from Cosmic Microwave Background experiments, providing overall better results with respect to the more usual filtering approaches.]{} Introduction ============ The importance of compact sources (galaxy clusters and extra-galactic sources) for ground- and space-based Cosmic Microwave Background (CMB) experiments has been clear since the conception of the WMAP [@BEN13] and Planck [@PLA18_I] missions. Point sources (hereafter PS) in the microwave regime are mainly blazars (i.e. AGNs with the relativistic jets aligned along the line of sight) and dusty galaxies. At such frequencies, PS are one of the contaminants to the recovery of the CMB anisotropies signal whose effect is more important at small angular scales. For this reason PS are even more important for the next generation of CMB experiments with higher resolution than *Planck*, such as the Cosmic Origins Explorer [CORE, @Del18], the Probe of Inflation and Cosmic Origins [PICO, @Han19] or the Lite satellite for the studies of B-mode polarization and Inflation from cosmic background Radiation Detection [LiteBird, @Mat14]. Generally, they are planned to keep the PS contamination low, but precise CMB measurements will still be affected by PS. For this reason, it is quite important to develop highly performing methods for PS detection. The standard single-frequency approach for PS detection in the CMB and far IR frequencies rely on the Mexican Hat Wavelet [MHW; @VIE03; @GN06] or on the matched filter techniques [@TEG98; @BAR03; @LOP06; @Her02]. Matched filter is theoretically the optimal filter when the PS shape is known providing the maximum signal-to-noise amplification. However, as concluded by @LOP06, the second member of the MHW family [MHW2; @GN06] provides a similar performance as the Matched Filter one, but it is easier to implement and more robust. Such wavelet has been successfully applied to *Planck* realistic simulation [@GN06; @LOP06; @LEA08] as well as to WMAP [@LOP07; @MAS09] and *Planck* real data: the Early Release Compact Source Catalogue [ERCSC, @ERCSC], the *Planck* Catalogue of Compact Sources [PCCS, @PCCS] and the Second *Planck* Catalogue of Compact Source [PCCS2, @PCCS2]. This is why we decide to compare our results against such method. Although there was always a tight relationships between Machine Learning techniques and astrophysics/cosmology, in the recent years the particular usage of neural networks has become a mainstream technique to derive new results. Artificial Neural Networks are Artificial Intelligence (AI) techniques involving numerical mathematical models which can be trained to represent complex physical systems by supervised or unsupervised learning [@SUA19a; @SUA17]. This characteristics are perfect to provide further results in Cosmology. Some example of recent interesting applications of Artificial Neural Network in cosmology are the identification of galaxy mergers [@Pea19] and strongly gravitational lenses [@Pet17] in astronomical images, a better estimation of cosmological constrains from weak lensing maps [@Flu19] and high fidelity generation of weak lensing convergence maps [@Mus19] and cosmological structure formation simulations under different assumptions [@Mat18; @He19; @Per19; @Gui19]. Some AI approaches, such as Multi-Layer Perceptron (MLP) [@JUE12] or Convolutional Neural Networks (CNNs) [@SUA19b; @KRI12], have been successfully applied to image processing (and related fields) for modelling and forecasting [@GRA13; @GIU13]. In this work, we propose the use of Fully-Convolutional Networks (FCN) [@LON15; @DAI16] as a very promising tool for PS detection. They are usually applied in image recognition and make use of different layers in order to get various image features (e.g. shapes, smoothness and borders). Important features of the input images are generally obtained by pairing convolution and merging layers. After that, layers are applied to get the output (image or numerical). In this work we present an application of FCN to the detection of PS in realistic simulations, the Point Source Image Detection Network () that can be summarised in the search of spheroids in a noisy background (i.e. the rest of the components in the microwave sky). The outline of the paper is the following. Section \[sec:simulations\] describes how the simulated maps are generated and Section \[sec:methodology\] reviews our methodology. The results are presented in Section \[sec:results\] and our conclusion are in Section \[sec:conclusions\]. Simulations {#sec:simulations} =========== {width="19cm"} {width="19cm"} {width="19cm"} In this work, we make use of realistic simulated maps of the microwave sky. The simulations correspond to sky patches at 143, 217 and 353 GHz, the central channels of the *Planck* mission, with $pixsize=90$ arcsec (a round number close to the 1.72 arcmin used in the *Planck* maps [npix=2048 in the `HEALPIX` all-sky pixelization schema, @GOR05]. For memory and speediness reasons, we use patches of $128\times128$ pixels, a trade of between density of bright sources per patch and size. We tested anyway that using bigger patches ($256\times256$) does not alter our statistical results or our conclusions. First, a catalogue of radio PS is simulated at each frequency independently by following the model by @TUC11. The flux density limit is $1$ mJy at all the frequencies. From the simulated catalogue we then create the simulated PS map and convolve it with the FWHM of the instrument [7.22, 4.90 and 4.92 arcmin at 143, 217 and 353 GHz respectively; @PLA18_I]. In order for the simulations to be realistic at these frequencies, we need to take into account fluctuations due to high redshift infrarred PS [massive proto-spheroidal galaxies in the process of forming most of their stellar mass, @Gra04; @Lap06; @LAP11; @CAI13] too faint to be detected one by one. Such contamination [@Bla98; @Lag03; @Dol04] is dominant at few arcmin resolution and it is called the Cosmic Infrared Background [CIB; @Pug96; @Hau01; @Dol06]. We use the software `CORRSKY` [@GN05] to simulate a sample of galaxies with a particular clustering properties, described by their angular power spectrum *P(k)*. We adopt the power spectrum and the source number counts (different at each frequency) given by the [@LAP11] and [@CAI13] models. As the main idea of this work is to compare detection methodologies, we do not simulate the late-type infrared galaxies [@Tof98; @PEPXIII; @PIPVII; @PCCS], that dominate the bright part of the source number counts above 217 GHz. Radio and late-type galaxies are only distinguishable for their different spectral emission, not for their shape and/or size. Compared with the *Planck* beam they are both point like sources. Their introduction will have supposed simply a higher density of brighter PS *per* patch in the highest simulated frequency without appreciably modifying the statistical properties of the background. On larger angular scales, we must include in our simulation the contamination due to diffuse emission by our Galaxy and the CMB. Such contaminants are introduced in our simulated maps by randomly select patches in *Planck* 143, 217 and 353 GHz official CMB maps [from the last release described in @PLA18_I]. The CMB maps are the one by the SEVEM method [@LEA08; @FER12], that are provided at all *Planck* frequencies. For the Galaxy emission we use the *Planck* `FFP10` simulations[^1], available for all *Planck* channels. The *Planck* maps are at nside=2048, that corresponds to a pixel size of 1.72 arcmin, and the selected sky patches are projected into flat patches with pixel size of 1.5 arcmin using the `gnomview` function provided with the `HEALPIX` framework [@GOR05]. Finally, we add the instrumental noise to the simulations. The noise maps are produced by simulating white noise accordingly to the Planck values: 0.55, 0.78 and 2.56 $\mu K_{CMB}$ deg, respectively [Noise rms computed after smoothing to $1^\circ$; @PLA18_I]. In this work, we study the performance of two detection methods, PoSeIDoN and the MHW2, especially focusing on their dependence with increasing intensity of Galactic emission by applying two different homogeneous galactic cuts (at 10$^\circ$ and 30$^\circ$ galactic latitudes). Moreover, such intensity increase also arises with higher frequencies due to Galactic emission spectral behaviour [@PEPXIX; @PLA14_11; @PLA16_10; @PLA18_IV]. Examples of random simulated patches are shown in the first two columns of Fig. \[Fig:patches\] for 143, 217 and 353 GHz (top, middle and bottom panels, respectively). The first column is the total input simulated map, including CMB, Galactic emission, CIB, PS and instrumental noise, whereas the second column is the input PS only map. Methodology {#sec:methodology} =========== PoSeIDoN {#sec:FCNN} -------- {width="17cm"} Neurons, sorted in layers, are the basic computing elements of an artificial network model. Their responses are modelled by weights that represents the influence of the neuron response on the neurons of the subsequent layer. In particular, for some models (such as CNNs) the weights correspond with kernel values [@LEC15]. The response is finally given after the process is completed along each computation units. In supervised learning, the implementation of the training procedure is performed via estimation of a loss function, usually a Mean Square Error (MSE) function, computed over the data from a training set (i.e. the network responses to certain inputs compared with their corresponding labels). Back-propagation algorithms are then employed to correct weights and kernel values and thus minimise the loss function with methods as the Stochastic Gradient Descent (SGD) [@RUM88; @CHA13]. FCNs allow us to perform dense predictions over the data used as input [@LON15; @DAI16]. In this case, the most relevant characteristics are first extracted using a convolutional block where each convolutional layer allows the extraction of several feature maps from the image obtained using kernels, frequently modulated by an activation function and processed by a down-sampling in terms of pooling. In addition to the typical convolutional process, an FCN has a second block where deconvolutions are performed, allowing the recovery of a dense response, also by means of layers with the correspondent kernels. Moreover, during the deconvolution process, information on the convolutional segment is included through the addition of fine-grained features in specific steps. In this work, the FCN parameters and hyperparameters are selected through a grid search. The selected topology is detailed as follows (see Figure \[Fig:FCN\]): - Convolutional block: the network has six convolutional layers, with 8, 2, 4, 2, 2 and 2 kernels respectively. Their correspondent kernel sizes are of 9, 9, 7, 7, 5 and 3 values of side. The activation function is leaky ReLU [@NAI10] in all the layers. Strides are of pixels both horizontally and vertically and padding has been added. - Deconvolutional block: the feature maps obtained after the convolutions are connected to a block of six deconvolutional layers. These layers have 2, 2, 2, 4, 2 and 8 kernels respectively. Their correspondent kernel sizes are of 3, 5, 7, 7, 9 and 9 values of side. The activation function is leaky ReLU in all the layers. Strides are of pixels both horizontally and vertically and padding has been added. Moreover, feature maps resulted from the five last convolutions are added, as fine-grained features, to the results of the five first deconvolutions. The training procedure is performed using an MSE loss function, with a training set of 50000 samples and a validation set of 5000 samples. The test sets for performance assessment consists of 5000 samples too. We produce 50000 simulations at 217 GHz to train the network. For each simulation we randomly chose a position of the available sky with the selected cut in latitude ($10^\circ$ or $30^\circ$) for both the CMB and the galactic emission. Moreover, the positions and fluxes for the input PS are also different in each realisation. At this stage, for each patch, two images are provided to PoSeIDoN: the total image (the simulated patch including all the components, the “Input Total” column in Fig. \[Fig:patches\]) and the PS image (the image containing only the input PS that should be detected; the “Input PS” column in Fig. \[Fig:patches\]). Mind that just for the training purpose, the sources flux density in the simulated catalogue are amplified by a “training factor” of 10, before being added to the other components. The reason is simply to increase the density of possible bright PS inside the patch without modifying the source number count shape, i.e. without altering the statistical properties of the PS sample, just their normalisation. Please note that PoSeIDoN is trained just at 217 GHz and for the $30^\circ$ galactic cut. Such trained FCN is then applied to all the cases studied in this work (i.e. 143, 217 and 353 GHz with a $30^\circ$ Galactic cut and 217 GHz with a $10^\circ$ Galactic cut). Better results are expected, although probably modest ones, if PoSeIDoN can be trained at each case individually. On this respect, the detection of PS in regions with intense Galactic emission, as the Galactic plane, is probably the most interesting case and also the one that can be improved the most by a dedicated FCN training. However, this is beyond the main scope of the current work. On the other hand, in the validation process, the simulated sources flux densities are the realistic ones (no additional training factor is applied), that also allow us to compare our results with the Planck catalogues. The validation simulation is built using realistic PS flux densities and realistic contaminants simulated in the same way as for the training ones (although the sky positions are always randomly chosen). Each validation patch is then provided to the trained network that returns an output map of recovered PS. An example of the PoSeIDoN output patch at the studied frequencies is shown in Fig. \[Fig:patches\], last column. Such output is then compared with the input PS only map for a performance analysis: estimation of the completeness, reliability and flux density accuracy. Mexican Hat Wavelet 2 --------------------- To assess PoSeIDoN performance, we also compare it against the MHW2 filter. The Mexican Hat Wavelet Family in the plane is derived by a applying the Laplacian operator iteratively to the 2D Gaussian [@GN06]. Any member of the family can be written in Fourier space as: $$\psi_n(k)=\frac{k^{2n}e^{-k^2/2}}{2^nn!}$$ The first member of the family, $\psi_1$, is the traditional MHW. It is one of the first wavelets applied successfully to the detection of PS in flat CMB maps [@CAY00; @VIE01]. The MHW2 is therefore the second member of the family, $\psi_2$, and it was demonstrated even more suited to the task than it predecessor [@GN06] or the theoretical optimal Matched Filter [@TEG98]. In fact, it was successfully applied to the WMAP data [@GN08; @MAS09] and it became the standard filtering technique for the production of the PS catalogues for the *Planck* mission [@ERCSC; @PCCS; @PCCS2]. The wavelets coefficients, $w_n(\textbf{b}, R)$ can be obtained for each member of the family as: $$w_n(\textbf{b}, R)=\int{d\textbf{k} e^{-i\textbf{k·b}}f(\textbf{k})\psi_n(kR)},$$ with $\textbf{b}$ being the location and $R$ the wavelet scale. By definition, a PS adopt the beam profile or point spread function, usually approximated by a Gaussian: $$\tau(x)= \frac{1}{2\pi\sigma_b^2} e^{-(\frac{x}{2\sigma_b})^2},$$ where $\sigma_b$ is the instrumental Gaussian beam dispersion. Therefore, the intensity of each source can be written as: $$I(x)=I_0 e^{-(\frac{x}{2\sigma_b})^2}.$$ Then, the scale R of the wavelet can be optimised by finding the maximum signal-to-noise ratio (S/N) of the sources in the filtered patch, i.e. maximising the amplification factor $\lambda_n=\frac{w_n/\sigma_{wn}}{I_0/\sigma}$, with $\sigma$ and $\sigma_{wn}$ the rms deviation of the background before and after filtering, respectively. The optimal scale is determined for each patch independently and it is always near to unity. The wavelets can be used for blind source detection (no prior information on the sources’ positions) and in non-blind mode, usually to get the estimated flux densities of PS at known positions. In our case, we apply the filter blindly to each total input validation simulation to produce the filtered image. Examples of the MHW2 output image for our frequencies are the patches shown in Fig. \[Fig:patches\], third column. Catalogue production and statistical comparison ----------------------------------------------- Both PoSeIDoN and the MWH2 methods provide just an output image, not a list of detections. In this section we describe the catalogue production process and the statistical quantities that we use for the performance comparison. The catalogue extraction consists simply in searching peaks, i.e. local maxima, above a certain intensity threshold, separated by at least a given minimum distance. This distance is 1.5 times the instrumental Gaussian beam dispersion or $\sigma_b$, that can be different for each channel. Taking into account that the MHW2 is the most used technique to detect PS in these kind of images, we use it as our reference for the comparison. In particular, we use the standard deviation, $\sigma_{MHW2}$, of the MHW2 output map to set up the thresholds for catalogue production. In the case of the MHW2 we set a $3\sigma_{MHW2}$ threshold to build the catalogue of detected PS (positions, flux densities and uncertainties). To reduce the number of spurious sources we also apply the MHW2 with the flux density threshold set to 4$\sigma_{MHW2}$, reducing the completeness, as shown in section \[sec:results\]. This last threshold level is more or less the one used for the *Planck* official PS catalogues [@ERCSC; @PCCS; @PCCS2]. The input catalogue is built from the input PS only image. Taking into account that this image does not have noise or any kind of background we use a lower PS threshold for the input catalogue. We choose one $\sigma_{MHW2}$ in order to be sure to have fainter PS. In the PoSeIDoN case, the output map is an attempt to mimic the input PS image. It contains just PS candidates without any background or instrumental noise residuals. Therefore, we apply to the FCN output map the same procedure followed for the input catalogue, i.e. the one $\sigma_{MHW2}$ threshold. A well known issue of any filtering technique is the border effects. As can be seen in the MHW2 patches from Fig. \[Fig:patches\] (third column), the filtering procedure produces artefacts near the patch border that can introduce spurious PS detections. In order to make a fair comparison with the MHW2 results, we exclude the detections within 5 pixels ($\sim 1$ FWHM) from the patches’ borders on every side for all the cases (input, MHW2 and PoSeIDoN). To describe the performance of the two techniques we focus on three statistical quantities: completeness, number of spurious detections and flux density estimation. These statistical quantities are commonly used to validate a detection technique or a produced catalogue [see e.g., @LOP07; @ERCSC; @PCCS; @PCCS2; @PIPLIV; @Hop15]. Completeness is estimated by cross-matching the detected PS against the input catalogue. It is a function of the intrinsic flux density, the detection threshold and sky location (not analysed in this work because the location dependency is common to both techniques by using the same simulations). Completeness provides information about the cumulative number of input sources that are missed at fainter flux densities: $C(>S_0)=\frac{N_{det}(>S_0)}{N_{input}(>S_0)}$, with $S_0$ the input flux density. All the detection techniques misidentify background fluctuations as PS at faint flux densities or around positions with strong Galactic emissions as the Galactic plane. Those wrongly detected PS, that are not in the input catalogue, are called spurious sources. Depending on the background characteristics and intensity, in many occasions it is the number of spurious sources, and not the completeness, that put a lower limit to the minimum flux density achievable with a given detection method. Therefore, the number of spurious sources is another important statistical quantity for the performance assessment of a detection technique. Finally, the third statistical quantity is the flux density estimation. For those input PS detected, we can compare their flux densities. This comparison can provide useful information about potential flux density bias or to identify spurious sources that were detected by chance on the same positions of faint input PS. Results {#sec:results} ======= {width="9cm"} {width="9cm"}\ {width="9cm"} {width="9cm"}\ {width="9cm"} {width="9cm"}\ {width="9cm"} {width="9cm"} As mentioned above, Fig. \[Fig:patches\] shows examples of output maps provided by the MHW2 and PoSeIDoN techniques when applied to the simulations. The third column corresponds to the MHW2, where it can be appreciated the typical granulate background after filtering. Moreover, these patches clearly show the border effects produced by the filtering approach. The fourth column are instead the PoSeIDoN output maps: no background fluctuations are present here, as the FCN only provides the best guess for input PS. It must be stressed that no border effects are present in the PoSeIDoN output images. As explained in section \[sec:FCNN\], we train the FCN only at 217 GHz for a Galactic mask of $|b|>30$º and we apply it to all the other studied cases. The performance of both techniques at this frequency are compared in the middle panel of Fig. \[Fig:compl\_rel\] for the the 30$^\circ$ galactic cut. The completeness (top sub-panel) and the percentage of spurious sources with respect to the input ones (bottom sub-panel) are shown on the left. The relative errors of the flux density estimation ($\Delta S / S_0= (S_{det} - S_0)/ S_0$) are shown on the right. With the MHW2 we obtain the expected results: we have general agreement with previous applications of the MHW2 and in particular with the *Planck* catalogues. Using a $3\sigma$ threshold (blue dashed line) the MHW2 provides good completeness results with a 90% completeness level at 124 mJy. However, such aggressive threshold implies a spurious PS detection problem already at $\sim400$ mJy with more than 20% of the total detected sources being spurious ones. In fact, using a more conservative and traditional $4\sigma$ threshold (cyan dot-dashed line), the spurious problem is highly reduced at least until $\sim200$ mJy. The price to be paid for this improvement is a reduction of the 90% completeness level that increases to 173 mJy. On the other hand, PoSeIDoN (red solid line) has a similar completeness performance with an intermediate 90% completeness level at 139 mJy. The clear advantage of the FCN is in the much lower number of spurious PS: in this case, it starts to be an issue only below $\sim100$ mJy (a flux density level below the 90% completeness level). The spurious PS issue is strongly related with high intensity regions of the background (mainly the Galactic emission). Therefore, the fact that we have a lower number of spurious PS implies that PoSeIDoN is most robust in distinguish a PS from a background local maxima. The flux density accuracy, right column in Fig. \[Fig:compl\_rel\] (central panel for 217 GHz), provides additional information in the understanding of the better PoSeIDoN performance when dealing with spurious PS. On the one hand, in the MHW2 case, most of the flux densities are correctly recovered within a 10% relative error. For the $3\sigma$ case the relative flux density error distribution shows a strong tail toward positive error, i.e. there are many sources whose flux density is overestimated by factors $>50\%$. Usually, those cases correspond to spurious PS, caused by strong Galactic emission, that by chance are near the position of a very faint input PS that should not be detectable. As expected, this issue almost disappeared with the more conservative $4\sigma$ threshold. On the other hand, PoSeIDoN behaviour on the recovery of the flux densities side is completely the opposite. The FCN recovers correctly the flux density of the most bright sources, but tends to under-estimate the flux density of fainter sources. A way to understand this specific behaviour is to consider that the final flux density recovered by PoSeIDoN is multiplied by a ‘confidence’ factor: $S_{est}\propto p_{conf} S_0$. For bright input PS or those in low background fluctuations areas, $p_{conf}\sim 1$. On the contrary, for faint input PS or those near high background fluctuations areas, $p_{conf}< 1$. This ‘confidence’ factor has the advantage to put the most dubious detected PS at fainter flux densities (see the steep increase of spurious sources below $\sim 100$ mJy). But it also means that the FCN recovered flux densities are not reliable. Although this is not ideal, it is not a limitation at all in the application of this novel technique: it is not unusual to firstly apply one technique for detection and then a second different one on the detected positions to estimate the flux density with better accuracy. One simple pipeline might consist in estimating the flux densities with the aperture flux, or with the MHW2 in non-blind mode, on the PS positions provided by PoSeIDoN. Another interesting possibility is to train a second neural network to get an accurate flux density estimation in known PS positions. To test the robustness of PoSeIDoN, we apply it to slightly different situations without additional training. On the one hand, we apply it at 143 (top panel) and 353 GHz (bottom panel) with the same galactic cut ( $|b|>30^\circ$). The first channel has a lower Galactic emission but higher instrumental noise and bigger beam. The second one has the same beam as 217 GHz but higher instrumental noise and the Galactic thermal emission is stronger. On the other hand, the FCN is applied again at 217 GHz but allowing patches at lower galactic latitudes, $|b|>10^\circ$, that implies stronger Galactic emission (see Fig. \[Fig:b10\]). At 143 GHz (Fig. \[Fig:compl\_rel\], top panel) the performance of PoSeIDoN with respect to the completeness is almost the same as the MHW2 with a $4\sigma$ threshold. The 90% completeness level in this channel are 224 mJy , 298 mJy and 300 mJy for the 3$\sigma$ MHWF, the 4$\sigma$ MHWF and PoSeIDoN, respectively. By comparing with the 217 GHz case, the FCN performance has worsen with respect to the MHW2 ones. The most probable reason is the change in the instrumental beam that will produce PS slightly bigger that the ones used to train (and thus expected by) the FCN at 217 GHz. This issue also explain PoSeIDoN under-estimation of the flux densities ($p_{conf}< 1$ for almost all detected PS). On the other hand, PoSeIDoN is still the most robust technique as for spurious PS. While both MHW2 results have already more than 20% spurious PS at the 90% completeness level, PoSeIDoN detects lower number of spurious PS well below $\sim 200$ mJy. At higher frequencies the Galactic emission is stronger and the spurious PS issue is much worse. This is clearly shown in the spurious results of all the techniques at 353 GHz (Fig. \[Fig:compl\_rel\], lower panel). The level of spurious PS is always above 50% for both MHW2 cases. PoSeIDoN is performing slightly better, although such issue is still present. At this frequency, a more conservative Galactic masking is needed or additional steps are required to decrease the number of spurious PS [see @PCCS; @PCCS2]. As for the completeness, the levels are similar to the 217 GHz case: 154, 227 and 227 mJy for the 3$\sigma$ MHWF, the 4$\sigma$ MHWF and PoSeIDoN, respectively. Again, as in the 143 GHz case, PoSeIDoN completeness results are equal to the 4$\sigma$ MHW2. More or less the flux densities estimation results are also similar to the 217 GHz case. So, for all three cases the source detection worsen with higher frequencies due to the increase of the foregrounds contribution to the total map, being PoSeIDoN the overall best performing method. To complete the analysis of the robustness of PoSeIDoN, we perform an additional test at 217 GHz. Without any additional training, we apply the FCN to a new set of validation simulations at 217 GHz but using a less aggressive Galactic mask of $|b|>10^\circ$ (see Fig. \[Fig:b10\]). The completeness and flux density estimation results remains more or less the same as in the 217 GHz case with $|b|>30^\circ$ for all the techniques. The 90% completeness levels in this case are 135, 195 and 149 mJy for the 3$\sigma$ MHWF, the 4$\sigma$ MHWF and PoSeIDoN, respectively. However, the spurious PS numbers increase dramatically, similarly to the 353 GHz case. Again, PoSeIDoN gives better results with this issue and it shows a more linear increase of spurious PS with respect to the 353 GHz case. This difference can be an indication that, as expected, the more the situation resembles the simulations training set, the better is the performance of the FCN. Therefore, by training PoSeIDoN in each particular situation the results can be slightly improved, although probably not much comparing with the MHW2. The detection of Galactic sources inside the complicate Galactic plane is most likely the most interesting case where the re-training would significantly improve the results. However, we have demonstrated that even without specialised training, a FCN is able to compete with the MHW2 filtering schema, when applied to typical CMB experiment observed patches. Finally, in the completeness panels of Figures \[Fig:compl\_rel\] and \[Fig:b10\], we also point out, with the grey dotted line, the PCCS2 90% completeness flux density limit [@PCCS2]: 177, 152 and 304 mJy at 143, 217, 353 GHz, respectively. Such values are in fair agreement with our findings. However, it should be stressed that such information has been added to guide the reader and it is not meant as a direct comparison with our results. First of all it must be taken into account that the PCCS2 is built to ensure at least 80% reliability. Then, the different masking must also be considered. As for the percentage of masked sky, the PCCS2 excludes the 15%, 35.1% and 52.4% for 143, 217 and 353 GHz [@PCCS2], whereas our $30^\circ$ Galactic cut corresponds approximately to a 50% of the sky. Moreover, the PCCS2 masks are tailored to avoid the most contaminating Galactic areas and to maximise the sky coverage of the catalogue. We could have used a more effective masking, but for our comparison-between-techniques purposes a simple galactic cut is enough. This point should just be taken into account when comparing with PCC2 numbers. Conclusions {#sec:conclusions} =========== In this work, we successfully apply PoSeIDoN to the detection of sources in a realistic situation: we include simulated PS and CIB at the Planck frequencies of 143, 217 and 353 GHz; we add CMB and Galactic emission by randomly choosing the patches in the real *Planck* CMB (provided by the `SEVEM` method) and the Galactic simulated (provided by the `FFP10` simulations) maps; finally we also add the instrumental noise, according to the Planck characteristics. The network was trained at 217 GHz with a Galactic cut of 30$^{\circ}$, using 50000 simulations. Then it was applied to the validation simulations at 143, 217 and 353 GHz and Galactic cut of 30$^{\circ}$. At 217 GHz the network was also tested with a Galactic cut of 10$^{\circ}$. Such results were then compared with those coming from the application of the MHW2 technique: in the overall, is performing better, providing more reliable results at lower flux densities. In should be stressed that in the MHW2 case, in order to get rid of the many spurious sources detected at low fluxes, we need to increase the flux density detection limit from 3$\sigma$ to 4$\sigma$. On the contrary, PoSeIDoN application is straightforward, well performing even at $1\sigma$ (i.e. the results given in this work). Another advantage of with respect to MHW2 is that it doesn’t have border effects like any filtering approach. In the MHW2 analysis we need to remove those pixels near the patch border, subsequently missing those sources falling in that regions [they can be recovered by selecting overlapping patches as done in @ERCSC; @PCCS; @PCCS2]. is not affected by such problem, being able to detect sources placed near the patches limits. As expected, both methods worsen their performance with increasing frequencies, i.e. with the increase of the relative importance of the contaminants (mainly due to the Galactic thermal emission in our set of simulations). Moreover, they also get worse with a smaller galactic cut because, as expected, the Galactic contamination is higher. As a reference, we also indicate the flux density limit at 90% completeness for the PCCS2, which is in fair agreement with our results. However it should be kept in mind that the PCCS2 is built imposing an overall 80% of reliability and by using tailored sky mask to better avoid Galactic contamination and preserving as much sky coverage as possible. Finally, as a limit of PoSeIDoN, it must be said that the flux density estimation of the FCN method is not optimal, at least with respect to the MHW2: the network behaviour in flux density estimation is to give lower flux densities with respect to the true ones. We notice a trend to assign lower flux densities to less reliable sources. So, our advice when building a catalogue (which is beyond the scope of this work) is to first blindly detect sources in a map with PoSeIDoN and then estimate the flux density of the retrieved sources by non-blindly applying some flux density estimation methods (e.g. non-blind MHW2) in the obtained PoSeIDoN positions. A future development of the current work would be to train a second neural network to derive accurate flux density estimations in known PS positions. LB and JGN acknowledge financial support from the PGC 2018 project PGC2018-101948-B-I00 (MICINN, FEDER) and PAPI-19-EMERG-11 (Universidad de Oviedo). JGN acknowledges financial support from the Spanish MINECO for the ‘Ramon y Cajal’ fellowship (RYC-2013-13256). JDCJ, MLS, SLSG, JDS acknowledge financial support from the I+D 2017 project AYA2017-89121-P and support from the European Union’s Horizon 2020 research and innovation programme under the H2020-INFRAIA-2018-2020 grant agreement No 210489629.\ This research has made use of the python packages `ipython` [@ipython], `matplotlib` [@matplotlib] and `Scipy` [@scipy]. [^1]: available at http://pla.esac.esa.int/pla/

--- abstract: 'Let $Y$ be a smooth rational surface and let $D$ be a cycle of rational curves on $Y$ which is an anticanonical divisor, i.e.  an element of $|-K_Y|$. Looijenga studied the geometry of such surfaces $Y$ in case $D$ has at most five components and identified a geometrically significant subset $R$ of the divisor classes of square $-2$ orthogonal to the components of $D$. Motivated by recent work of Gross, Hacking, and Keel on the global Torelli theorem for pairs $(Y,D)$, we attempt to generalize some of Looijenga’s results in case $D$ has more than five components. In particular, given an integral isometry $f$ of $H^2(Y)$ which preserves the classes of the components of $D$, we investigate the relationship between the condition that $f$ preserves the “generic" ample cone of $Y$ and the condition that $f$ preserves the set $R$.' address: 'Department of Mathematics, Columbia University, New York, NY 10027' author: - Robert Friedman title: | On the ample cone of a rational surface\ with an anticanonical cycle --- Introduction {#introduction .unnumbered} ============ The ample cone of a del Pezzo surface $Y$ (or rather the associated dual polyhedron) was studied classically by, among others, Gosset, Schoute, Kantor, Coble, Todd, Coxeter, and Du Val. For a brief historical discussion, one can consult the remarks in §11.x of [@Coxeter]. From this point of view, the lines on $Y$ are the main object of geometric interest, as they are the walls of the ample cone or the vertices of the dual polyhedron. The corresponding root system (in case $K_Y^2 \leq 6$) only manifests itself geometrically by allowing del Pezzo surfaces with rational double points, or equivalently smooth surfaces $Y$ with $-K_Y$ nef and big but not ample. This is explicitly worked out in Part II of Du Val’s series of papers [@duV]. On the other hand, the root system, or rather its Weyl group, appears for a smooth del Pezzo surface as a group of symmetries of the ample cone, a fact which (in a somewhat different guise) was already known to Cartan. Perhaps the culmination of the classical side of the story is Du Val’s 1937 paper [@duV2], where he also systematically considers the blowup of $\Pee^2$ at $n\geq 9$ points. In modern times, Manin explained the appearance of the Weyl group by noting that the orthogonal complement to $K_Y$ in $H^2(Y;\Zee)$ is a root lattice $\Lambda$. Moreover, given any root of $\Lambda$, in other words an element $\beta$ of square $-2$, there exists a deformation of $Y$ for which $\beta = \pm[C]$, where $C$ is a smooth rational curve of self-intersection $-2$. For modern expositions of the theory, see for example Manin’s book [@Manin] or Demazure’s account in [@777]. In general, it seems hard to study an arbitrary rational surface $Y$ without imposing some extra conditions. One very natural condition is that $-K_Y$ is effective, i.e. that $-K_Y = D$ for an effective divisor $D$. In case the intersection matrix of $D$ is negative definite, such pairs $(Y,D)$ arise naturally in the study of minimally elliptic singularities: the case where $D$ is a smooth elliptic curve corresponds to the case of simple elliptic singularities, the case where $D$ is a nodal curve or a cycle of smooth rational curves meeting transversally corresponds to the case of cusp singularities, and the case where $D$ is reduced but has one component with a cusp, two components with a tacnode or three components meeting at a point, corresponds to triangle singularities. From this point of view, the case where $D$ is a cycle of rational curves is the most plentiful. The systematic study of such surfaces in case the intersection matrix of $D$ is negative definite dates back to Looijenga’s seminal paper [@Looij]. However, for various technical reasons, most of the results of that paper are proved under the assumption that the number of components in the cycle is at most $5$. Some of the main points of [@Looij] are as follows: Denote by $R$ the set of elements in $H^2(Y; \Zee)$ of square $-2$ which are orthogonal to the components of $D$ and which are of the form $\pm [C]$, where $C$ is a smooth rational curve disjoint from $D$, for some deformation of the pair $(Y,D)$. In terms of deformations of singularities, the set $R$ is related to the possible rational double point singularities which can arise as deformations of the dual cusp to the cusp singularity corresponding to $D$. Looijenga noted that, in general, there exist elements in $H^2(Y; \Zee)$ of square $-2$ which are orthogonal to the components of $D$ but which do not lie in $R$. Moreover, reflections in elements of the set $R$ give symmetries of the “generic" ample cone (which is the same as the ample cone in case there are no smooth rational curves on $Y$ disjoint from $D$). Finally, still under the assumption of at most 5 components, any isometry of $H^2(Y; \Zee)$ which preserves the positive cone, the classes $[D_i]$ and the set $R$, preserves the generic ample cone. This paper, which is an attempt to see how much of [@Looij] can be generalized to the case of arbitrarily many components, is motivated by a question raised by the recent work of Gross, Hacking and Keel [@GHK] on, among matters, the global Torelli theorem for pairs $(Y,D)$ where $D$ is an anticanonical cycle on the rational surface $Y$. In order to formulate this theorem in a fairly general way, one would like to characterize the isometries $f$ of $H^2(Y, \Zee)$, preserving the positive cone and fixing the classes $[D_i]$, which preserve the ample cone of $Y$. It is natural to ask if, at least in the generic case, the condition that $f(R)=R$ is sufficient. In this paper, we give various criteria on $R$ which insure that, if an isometry $f$ of $H^2(Y; \Zee)$ preserves the positive cone, the classes $[D_i]$ and the set $R$, then $f$ preserves the generic ample cone. Typically, one needs a hypothesis which says that $R$ is large. For example, one such hypothesis is that there is a subset of $R$ which spans a negative definite codimension one subspace of the orthogonal complement to the components of $D$. In theory, at least under various extra hypotheses, such a result gives a necessary and sufficient condition for an isometry to preserve the generic ample cone. In practice, however, the determination of the set $R$ in general is a difficult problem, which seems close in its complexity to the problem of describing the generic ample cone of $Y$. Finally, we show that some assumptions on $(Y,D)$ are necessary, by giving examples where $R=\emptyset$, so that the condition that an isometry $f$ preserves $R$ is automatic, and of isometries $f$ such that $f$ preserves the positive cone, the classes $[D_i]$ and (vacuously) the set $R$, but $f$ does not preserve the generic ample cone. We do not yet have a good understanding of the relationship between preserving the ample cone and preserving the set $R$. An outline of this paper is as follows. The preliminary Section 1 reviews standard methods for constructing nef classes on algebraic surfaces and applies this to the study of when the normal surface obtained by contracting a negative definite anticanonical cycle on a rational surface is projective. In Section 2, we analyze the ample cone and generic ample cone of a pair $(Y,D)$ and show that the set $R$ defined by Looijenga is exactly the set of elements $\beta$ in $H^2(Y; \Zee)$ of square $-2$ which are orthogonal to the components of $D$ such that reflection about $\beta$ preserves the generic ample cone. Much of the material of §2 overlaps with results in [@GHK], proved there by somewhat different methods. Section 3 is devoted to giving various sufficient conditions for an isometry $f$ of $H^2(Y; \Zee)$ to preserve the generic ample cone, including the one described above. Section 4 gives examples of pairs $(Y,D)$ satisfying the sufficient conditions of §3 where the number of components of $D$ and the multiplicity $-D^2$ are arbitrarily large, as well as examples showing that some hypotheses on $(Y,D)$ are necessary. **Acknowledgements.** It is a pleasure to thank Mark Gross, Paul Hacking and Sean Keel for access to their manuscript [@GHK] and for extremely stimulating correspondence and conversations about these and other matters, and Radu Laza for many helpful discussions. **Notation and conventions.** We work over $\Cee$. If $X$ is a smooth projective surface with $h^1(\scrO_X) = h^2(\scrO_X) = 0$ and $\alpha \in H^2(X; \Zee)$, we denote by $L_\alpha$ the corresponding holomorphic line bundle, i.e. $c_1(L_\alpha) = \alpha$. Given a curve $C$ or divisor class $G$ on $X$, we denote by $[C]$ or $[G]$ the corresponding element of $H^2(X; \Zee)$. Intersection pairing on curves or divisors, or on elements in the second cohomology of a smooth surface (viewed as a canonically oriented $4$-manifold) is denoted by multiplication. Preliminaries ============= Throughout this paper, $Y$ denotes a smooth rational surface with $-K_Y = D= \sum_{i=1}^rD_i$ a (reduced) cycle of rational curves, i.e. each $D_i$ is a smooth rational curve and $D_i$ meets $D_{i\pm 1}$ transversally, where $i$ is taken mod $r$, except for $r=1$, in which case $D_1=D$ is an irreducible nodal curve. We note, however, that many of the results in this paper can be generalized to the case where $D\in |-K_Y|$ is not assumed to be a cycle. The integer $r=r(D)$ is called the *length* of $D$. An *orientation* of $D$ is an orientation of the dual graph (with appropriate modifications in case $r=1$). We shall abbreviate the data of the surface $Y$ and the oriented cycle $D$ by $(Y,D)$ and refer to it as a *anticanonical pair*. If the intersection matrix $(D_i\cdot D_j)$ is negative definite, we say that $(Y,D)$ is a *negative definite anticanonical pair*. \[defcurves\] An irreducible curve $E$ on $Y$ is an *exceptional curve* if $E\cong \Pee^1$, $E^2 = -1$, and $E \neq D_i$ for any $i$. An irreducible curve $C$ on $Y$ is a *$-2$-curve* if $C\cong \Pee^1$, $C^2 = -2$, and $C \neq D_i$ for any $i$. Let $\Delta_Y$ be the set of all $-2$-curves on $Y$, and let $\mathsf{W}({\Delta_Y})$ be the group of integral isometries of $H^2(Y; \Ar)$ generated by the reflections in the classes in the set $\Delta_Y$. Let $\Lambda = \Lambda(Y,D) \subseteq H^2(Y; \Zee)$ be the orthogonal complement of the lattice spanned by the classes $[D_i]$. Fixing the identification $\Pic^0D \cong \mathbb{G}_m$ defined by the orientation of the cycle $D$, we define the *period map* $\varphi_Y \colon \Lambda \to \mathbb{G}_m$ via: if $\alpha \in \Lambda$ and $L_\alpha$ is the corresponding line bundle, then $\varphi_Y(\alpha) \in \mathbb{G}_m$ is the image of the line bundle of multi-degree $0$ on $D$ defined by $L_\alpha|D$. Clearly $\varphi_Y$ is a homomorphism. By [@Looij], [@FriedmanScattone], [@Fried2], we have: \[surjper\] The period map is surjective. More precisely, given $Y$ as above and given an arbitrary homomorphism $\varphi \colon \Lambda \to\mathbb{G}_m$, there exists a deformation of the pair $(Y,D)$ over a smooth connected base, which we can take to be a product of $\mathbb{G}_m$’s, such that the monodromy of the family is trivial and there exists a fiber of the deformation, say $(Y', D')$ such that, under the induced identification of $\Lambda(Y',D')$ with $\Lambda$, $\varphi_{Y'} = \varphi$. For future reference, we recall some standard facts about negative definite curves on a surface: \[negdef\] Let $X$ be a smooth projective surface and let $G_1, \dots , G_n$ be irreducible curves on $X$ such that the intersection matrix $(G_i\cdot G_j)$ is negative definite. Let $F$ be an effective divisor on $X$, not necessarily reduced or irreducible, and such that, for all $i$, $G_i$ is not a component of $F$. 1. Given $r_i \in \Ar$, if $(F + \sum_ir_iG_i) \cdot G_j = 0$ for all $j$, then $r_i \geq 0$ for all $i$, and, for every subset $I$ of $\{1, \dots, n\}$, if $\bigcup_{i\in I}G_i$ is a connected curve such that $F\cdot G_j \neq 0$ for some $j\in I$, then $r_i > 0$ for $i\in I$. 2. Given $s_i, t_i \in \Ar$, if $[F] + \sum_is_i[G_i] = \sum_it_i[G_i]$, then $F=0$ and $s_i = t_i$ for all $i$. The following general result is also well-known: \[constructnef\] Let $X$ be a smooth projective surface and let $G_1, \dots , G_n$ be irreducible curves on $X$ such that the intersection matrix $(G_i\cdot G_j)$ is negative definite. (We do not, however, assume that $\bigcup_iG_i$ is connected.) Then there exists a nef and big divisor $H$ on $X$ such that $H\cdot G_j = 0$ for all $j$ and, if $C$ is an irreducible curve such that $C \neq G_j$ for any $j$, then $H\cdot C >0$. In fact, the set of nef and big $\Ar$-divisors which are orthogonal to $\{G_1, \dots, G_n\}$ is a nonempty open subset of $\{G_1, \dots, G_n\}^\perp \otimes \Ar$. Fix an ample divisor $H_0$ on $X$. Since $(G_i\cdot G_j)$ is negative definite, there exist $r_i\in \Q$ such that $(\sum_ir_iG_i) \cdot G_j = -(H_0\cdot G_j)$ for every $j$, and hence $(H_0 + \sum_ir_iG_i) \cdot G_j = 0$. By Lemma \[negdef\], $r_i > 0$ for every $i$. There exists an $N > 0$ such that $Nr_i \in \Zee$ for all $i$. Then $H = N(H_0 + \sum_ir_iG_i)$ is an effective divisor satisfying $H\cdot G_j = 0$ for all $j$. If $C$ is an irreducible curve such that $C \neq G_j$ for any $j$, then $H_0 \cdot C > 0$ and $G_i \cdot C \geq 0$ for all $i$, hence $H\cdot C >0$. In particular $H$ is nef. Finally $H$ is big since $H^2 = NH\cdot(H_0 + \sum_ir_iG_i) = N(H\cdot H_0) > 0$, as $H_0$ is ample. To see the final statement, we apply the above argument to an ample $\Ar$-divisor $x$ (i.e. an element in the interior of the ample cone) to see that $x + \sum_ir_iG_i$ is a nef and big $\Ar$-divisor orthogonal to $\{G_1, \dots, G_n\}$. Since $x + \sum_ir_iG_i$ is simply the orthogonal projection $p$ of $x$ onto $\{G_1, \dots, G_n\}^\perp \otimes \Ar$, and $p\colon H^2(X; \Ar) \to \{G_1, \dots, G_n\}^\perp \otimes \Ar$ is an open map, the image of the interior of the ample cone of $X$ is then a nonempty open subset of $\{G_1, \dots, G_n\}^\perp \otimes \Ar$ consisting of nef and big $\Ar$-divisors orthogonal to $\{G_1, \dots, G_n\}$. Applying the above construction to $X=Y$ and $D_1, \dots, D_r$, we can find a nef and big divisor $H$ such that $H\cdot D_j = 0$ for all $j$ and such that, if $C$ is an irreducible curve such that $C \neq D_j$ for any $j$, then $H\cdot C >0$. Let $(Y,D)$ be a negative definite anticanonical pair and let $H$ be a nef and big divisor such that $H\cdot D_j = 0$ for all $j$ and such that, if $C$ is an irreducible curve such that $C \neq D_j$ for any $j$, then $H\cdot C >0$. Suppose in addition that $\scrO_Y(H)|D = \scrO_D$, i.e. that $\varphi_Y([H]) =1$. Then the $D_i$ are not fixed components of $|H|$. Hence, if $\overline{Y}$ denotes the normal complex surface obtained by contracting the $D_i$, then $H$ induces an ample divisor $\overline{H}$ on $\overline{Y}$ and $|3\overline{H}|$ defines an embedding of $\overline{Y}$ in $\Pee^N$ for some $N$. Consider the exact sequence $$0 \to \scrO_Y(H-D) \to \scrO_Y(H) \to \scrO_D \to 0.$$ Looking at the long exact cohomology sequence, as $$H^1(Y; \scrO_Y(H-D)) = H^1(Y;\scrO_Y(H) \otimes K_Y)$$ is Serre dual to $H^1(Y; \scrO_Y(-H)) = 0$, by Mumford vanishing, there exists a section of $\scrO_Y(H)$ which is nowhere vanishing on $D$, proving the first statement. The second follows from Nakai-Moishezon and the third from general results on linear series on anticanonical pairs [@Fried1]. By the surjectivity of the period map \[surjper\], for any $(Y,D)$ a negative definite anticanonical pair and $H$ a nef and big divisor on $Y$ such that $H\cdot D_j = 0$ for all $j$ and $H\cdot C > 0$ for all curves $C\neq D_i$, there exists a deformation of the pair $(Y,D)$ such that the divisor corresponding to $H$ has trivial restriction to $D$. More generally, one can consider deformations such that $\varphi_Y([H])$ is a torsion point of $\mathbb{G}_m$. In this case, if $\overline{Y}$ is the normal surface obtained by contracting $D$, then $\overline{Y}$ is projective. Note that this implies that the set of pairs $(Y,D)$ such that $\overline{Y}$ is projective is Zariski dense in the moduli space. However, as the set of torsion points is not dense in $\mathbb{G}_m$ in the classical topology, the set of projective surfaces $\overline{Y}$ will not be dense in the classical topology. Roots and nodal classes ======================= Let $\mathcal{C}=\mathcal{C}(Y)$ be the positive cone of $Y$, i.e.  $$\mathcal{C} = \{x\in H^2(Y; \Ar): x^2 >0\}.$$ Then $\mathcal{C}$ has two components, and exactly one of them, say $\mathcal{C}^+=\mathcal{C}^+(Y)$, contains the classes of ample divisors. We also define $$\mathcal{C}^+_D = \mathcal{C}^+_D(Y) = \{x\in \mathcal{C}^+: x \cdot [D_i] \geq 0 \text{ for all $i$ }\}.$$ Let $\overline{\mathcal{A}}(Y)\subseteq \mathcal{C}^+ \subseteq H^2(Y; \Ar)$ be the (closure of) the ample (nef, Kähler) cone of $Y$ in $\mathcal{C}^+$. By definition, $\overline{\mathcal{A}}(Y)$ is closed in $\mathcal{C}^+$ but not in general in $H^2(Y; \Ar)$. Let $\alpha \in H^2(Y; \Zee), \alpha \neq 0$. The *oriented wall $W^\alpha$ associated to $\alpha$* is the set $\{x\in \mathcal{C}^+: x\cdot \alpha =0\}$, i.e. the intersection of $\mathcal{C}^+$ with the orthogonal space to $\alpha$, together with the preferred half space defined by $x\cdot \alpha \geq 0$. If $C$ is a curve on $Y$, we write $W^C$ for $W^{[C]}$. A standard result (see for example [@FriedmanMorgan], II (1.8)) shows that, if $I$ is a subset of $H^2(Y; \Zee)$ and there exists an $N\in \Zee^+$ such that $-N \leq \alpha^2 < 0$ for all $\alpha \in I$, then the collection of walls $\{W^\alpha: \alpha \in I\}$ is locally finite on $\mathcal{C}^+$. Finally, we say that $W^\alpha$ is a *face* of $\overline{\mathcal{A}}(Y)$ if $\partial \overline{\mathcal{A}}(Y)\cap W^\alpha$ contains an open subset of $W^\alpha$ and $x\cdot \alpha \geq 0$ for all $x\in \overline{\mathcal{A}}(Y)$. $\overline{\mathcal{A}}(Y)$ is the set of all $x\in \mathcal{C}^+$ such that $x\cdot [D_i]\geq 0$, $x\cdot [E] \geq 0$ for all exceptional curves $E$ and $x\cdot [C] \geq 0$ for all $-2$-curves $C$. Moreover, if $\alpha$ is the class associated to an exceptional or $-2$-curve, or $\alpha =[D_i]$ for some $i$ such that $D_i^2< 0$ then $W^\alpha$ is a face of $\overline{\mathcal{A}}(Y)$. If $\alpha, \beta$ are two such classes, $W^\alpha = W^\beta$ $\iff$ $\alpha =\beta$. For the first claim, it is enough to show that, if $G$ is an irreducible curve on $Y$ with $G^2 < 0$, then $G$ is either $D_i$ for some $i$, an exceptional curve or a $-2$-curve. This follows immediately from adjunction since, if $G\neq D_i$ for any $i$, then $G\cdot D \geq 0$ and $-2 \leq 2p_a(G) -2 = G^2 - G \cdot D < 0$, hence $p_a(G) =0$ and either $G^2 = -2$, $G \cdot D =0$, or $G^2 = G \cdot D =-1$. The last two statements follow from the openness statement in Proposition \[constructnef\] and the fact that no two distinct classes of the types listed above are multiples of each other. As an alternate characterization of the classes in the previous lemma, we have: \[numeric\] Let $H$ be a nef divisor such that $H\cdot D >0$. 1. If $\alpha \in H^2(Y; \Zee)$ with $\alpha ^2 = \alpha \cdot [K_Y] = -1$, then $\alpha \cdot [H] \geq 0$ $\iff$ $\alpha$ is the class of an effective curve. In particular, the wall $W^\alpha$ does not pass through the interior of $\overline{\mathcal{A}}(Y)$ (cf. [@FriedmanMorgan], p. 332 for a more general statement). 2. If $\beta \in H^2(Y; \Zee)$ with $\beta ^2= -2$, $\beta\cdot [D_i] = 0$ for all $i$, $\beta \cdot [H] \geq 0$, and $\varphi_Y(\beta) =1$, then $\pm \beta$ is the class of an effective curve, and $\beta$ is effective if $\beta \cdot [H] > 0$. Hence the ample cone $\overline{\mathcal{A}}(Y)$ is the set of all $x\in \mathcal{C}^+$ such that $x\cdot [D_i]\geq 0$ and $x\cdot\alpha \geq 0$ for all classes $\alpha$ and $\beta$ as described in [(i)]{} and [(ii)]{} above, where in case [(ii)]{} we assume in addition that $\beta$ is effective, or equivalently that $\beta \cdot [H] > 0$ for some nef divisor $H$. \(i) Clearly, if $\alpha$ is the class of an effective curve, then $\alpha \cdot [H] \geq 0$ since $H$ is nef. Conversely, assume that $\alpha ^2 = \alpha \cdot [K_Y] = -1$ and that $\alpha \cdot [H] \geq 0$. By Riemann-Roch, $\chi(L_\alpha) = 1$. Hence either $h^0(L_\alpha) > 0$ or $h^2(L_\alpha) > 0$. But $h^2(L_\alpha) = h^0(L_\alpha^{-1}\otimes K_Y)$ and $[H] \cdot (-\alpha - [D] ) < 0$, by assumption. Thus $h^0(L_\alpha) > 0$ and hence $\alpha$ is the class of an effective curve. \(ii) As in (i), $H \cdot (-\beta - [D] ) < 0$, and hence $h^0(L_\beta^{-1}\otimes K_Y) = 0$. Thus $h^2(L_\beta) =0$. Suppose that $h^0(L_\beta) = 0$. Then, by Riemann-Roch, $\chi(L_\beta) = 0$ and hence $h^1(L_\beta) = 0$. Hence $h^1(L_\beta^{-1}\otimes K_Y) = 0$. Since $\varphi_Y(\beta) =1$, $L_\beta^{\pm 1}|D = \scrO_D$. Thus there is an exact sequence $$0 \to L_\beta^{-1}\otimes \scrO_Y(-D) \to L_\beta^{-1}\to\scrO_D \to 0.$$ Since $H^1(L_\beta^{-1}\otimes K_Y) =H^1(L_\beta^{-1}\otimes \scrO_Y(-D)) =0$, the map $H^0(L_\beta^{-1})\to H^0(\scrO_D)$ is surjective and hence $-\beta$ is the class of an effective curve. It is natural to make the following definition: Let $\alpha \in H^2(Y; \Zee)$. Then $\alpha$ is a *numerical exceptional curve* if $\alpha ^2 = \alpha \cdot [K_Y] = -1$. The numerical exceptional curve $\alpha$ is *effective* if $h^0(L_\alpha) > 0$, i.e. if $\alpha = [G]$, where $G$ is an effective curve. A minor variation of the proof of Lemma \[numeric\] shows: \[remarkafternumeric\] Let $H$ be a nef and big divisor such that $H\cdot G > 0$ for all $G$ an irreducible curve not equal to $D_i$ for some $i$, and let $\alpha$ be a numerical exceptional curve. 1. Suppose that $[H]\cdot\alpha \geq 0$. Then either $[H]\cdot\alpha > 0$ and $\alpha$ is effective or $H\cdot D = [H]\cdot\alpha =0$ and $\alpha$ is an integral linear combination of the $[D_i]$. 2. If $(Y,D)$ is negative definite and $\alpha$ is an integral linear combination of the $[D_i]$, then either some component $D_i$ is a smooth rational curve of self-intersection $-1$ or $K_Y^2=-1$, $\alpha = K_Y$ and hence $\alpha$ is not effective. 3. If no component $D_i$ is a smooth rational curve of self-intersection $-1$, then $\alpha$ is effective $\iff$ $[H] \cdot \alpha > 0$. \(i) As in the proof of Lemma \[numeric\], either $\alpha$ or $-\alpha -[D]$ is the class of an effective divisor. If $-\alpha -[D]$ is the class of an effective divisor, then $0\leq [H]\cdot (-\alpha -[D]) \leq 0$, so that $[H]\cdot \alpha=H \cdot D = 0$. In particular $(Y,D)$ is negative definite. Moreover, if $G$ is an effective divisor with $[G] = -\alpha -[D]$, then every component of $G$ is equal to some $D_i$, hence $[G]$ and therefore $\alpha = -[G]-[D]$ are integral linear combinations of the $[D_i]$. \(ii) Suppose that $\alpha$ is an integral linear combination of the $[D_i]$ but that no $D_i$ is a smooth rational curve of self-intersection $-1$. We shall show that $K_Y^2=-1$ and $\alpha = K_Y$. First suppose that $K_Y^2=-1$. Then $\bigoplus_i\Zee\cdot [D_i] = \Zee\cdot [K_Y] \oplus L$, where $L$, the orthogonal complement of $[K_Y]$ in $\bigoplus_i\Zee\cdot [D_i]$, is even and negative definite. Thus $\alpha = a[K_Y] + \beta$, with either $\beta = 0$ or $\beta^2 \leq -2$, and $\alpha ^2 = -a^2 + \beta^2$. Hence, if $\alpha ^2 = \alpha \cdot [K_Y] = -1$, the only possibility is $\beta = 0$ and $a=1$. In case $K_Y^2 < -1$, $D$ is reducible, and no $D_i$ is a smooth rational curve of self-intersection $-1$, then $D_i^2\leq -2$ for all $i$ and either $D_i^2 \leq -4$ for some $i$ or there exist $i\neq j$ such that $D_i^2 = D_j^2=-3$. In this case, it is easy to check that, for all integers $a_i$ such that $a_i \neq 0$ for some $i$, $(\sum_ia_iD_i)^2 < -1$. This contradicts $\alpha^2 =-1$. \(iii) If $[H]\cdot \alpha > 0$, then $\alpha$ is effective by (i). If $[H]\cdot \alpha< 0$, then clearly $\alpha$ is not effective. Suppose that $[H]\cdot \alpha =0$; we must show that, again, $\alpha$ is not effective. Suppose that $\alpha=[G]$ is effective. By the hypothesis on $H$, every component of $G$ is a $D_i$ for some $i$, so that $\alpha = \sum_ia_i[D_i]$ for some $a_i \in \Zee$, $a_i \geq 0$. Let $I\subseteq \{1, \dots, r\}$ be the set of $i$ such that $a_i > 0$. Then $H\cdot D_i = 0$ for all $i\in I$. If $I = \{1, \dots, r\}$, then $(Y,D)$ is negative definite and we are done by (ii). Otherwise, $\bigcup_{i\in I}D_i$ is a union of chains of curves whose components $D_i$ satisfy $D_i ^2 \leq -2$. It is then easy to check that $\alpha^2 < -1$ in this case, a contradiction. Hence $\alpha$ is not effective. Let $Y_t$ be a generic small deformation of $Y$, and identify $H^2( Y_t ;\Ar)$ with $H^2( Y ;\Ar)$. Define $\overline{\mathcal{A}}_{\text{\rm{gen}}}= \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ to be the ample cone $\overline{\mathcal{A}}(Y_t)$ of $Y_t$, viewed as a subset of $H^2( Y ;\Ar)$. \[describeA\] With notation as above, 1. If there do not exist any $-2$-curves on $Y$, then $\overline{\mathcal{A}}(Y) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$. More generally, $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ is the set of all $x\in \mathcal{C}^+$ such that $x\cdot [D_i] \geq 0$ and $x\cdot \alpha \geq 0$ for all effective numerical exceptional curves. In particular, $$\overline{\mathcal{A}}(Y) \subseteq \overline{\mathcal{A}}_{\text{\rm{gen}}}.$$ 2. $\overline{\mathcal{A}}(Y) = \{ x\in \overline{\mathcal{A}}_{\text{\rm{gen}}}: x\cdot [C] \geq 0 \text{ for all $-2$-curves $C$} \}$. Let $Y$ be a surface with no $-2$-curves (such surfaces exist and are generic by the surjectivity of the period map, Theorem \[surjper\]). Fix a nef divisor $H$ on $Y$ with $H\cdot D >0$. Then $\overline{\mathcal{A}}(Y)$ is the set of all $x\in \mathcal{C}^+$ such that $x\cdot [D_i] \geq 0$ and $x\cdot [E] \geq 0$ for all exceptional curves $E$, and this last condition is equivalent to $x\cdot \alpha \geq 0$ for all $\alpha\in H^2(Y; \Zee)$ such that $\alpha^2 =\alpha\cdot [K_Y] =-1$ and $\alpha \cdot [H] \geq 0$, by Lemma \[numeric\]. Since this condition is independent of the choice of $Y$, because we can choose the divisor $H$ to be ample and to vary in a small deformation, the first part of (i) follows, and the remaining statements are clear. In fact, the argument above shows: \[definv\] The set of effective numerical exceptional curves and the set $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ are locally constant, and hence are invariant in a global deformation with trivial monodromy under the induced identifications. \[reflect\] If $C$ is a $-2$-curve on $Y$, then the wall $W^C$ meets the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, and in fact $r_C( \overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$, where $r_C\colon H^2(Y; \Ar) \to H^2(Y; \Ar)$ is reflection in the class $[C]$. Hence $\overline{\mathcal{A}}(Y)$ is a fundamental domain for the action of the group $\mathsf{W}({\Delta_Y})$ on $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, where $\mathsf{W}({\Delta_Y})$ is the group generated by the reflections in the classes in the set $\Delta_Y$ of $-2$-curves on $Y$. Clearly, if $r_C( \overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$, then $W^C$ meets the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$. To see that $r_C( \overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$, assume first more generally that $\beta\in \Lambda$ is any class with $\beta^2 = -2$, and let $r_\beta$ be the corresponding reflection. Then $r_\beta$ permutes the set of $\alpha \in H^2(Y; \Zee)$ such that $\alpha^2 =\alpha\cdot [K_Y] =-1$, but does not necessarily preserve the condition that $\alpha$ is effective, i.e. that $\alpha \cdot [H] \geq 0$ for some nef divisor $H$ on $Y$ with $H\cdot D >0$. However, for $\beta = [C]$, there exists by Proposition \[constructnef\] a nef and big divisor $H_0$ such that $H_0 \cdot C = 0$ and $H\cdot D > 0$. Hence $[H_0]$ is invariant under $r_C$, and so $r_C$ permutes the set of $\alpha \in H^2(Y; \Zee)$ such that $\alpha^2 =\alpha\cdot [K_Y] =-1$ and $\alpha \cdot [H_0] \geq 0$. Thus $r_C$ permutes the set of effective numerical exceptional curves and hence the faces of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, so that $r_C( \overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Since $\overline{\mathcal{A}}(Y) \subseteq \overline{\mathcal{A}}_{\text{\rm{gen}}}$ is given by (ii) of Lemma \[describeA\], the final statement is then a general result in the theory of reflection groups (cf. [@Bour], V §3). \[monodromyinvar\] (i) The argument for the first part of Lemma \[reflect\] essentially boils down to the following: let $\overline{Y}$ be the normal surface obtained by contracting $C$. Then the reflection $r_C$ is the monodromy associated to a generic smoothing of the singular surface $\overline{Y}$, and the cone $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ is invariant under monodromy. \(ii) If $E$ is an exceptional curve, then $W^E$ is a face of $\overline{\mathcal{A}}(Y)$. For a generic $Y$ (i.e. no $-2$-curves), Lemma \[reflect\] then says that the set of exceptional curves on $Y$ is invariant under the reflection group generated by all classes of square $-2$ which become the classes of a $-2$-curve under some specialization. A somewhat more involved statement holds in the nongeneric case. \[permapinvar\] With $\mathsf{W}({\Delta_Y})$ as in Definition \[defcurves\], for all $w\in \mathsf{W}({\Delta_Y})$ and all $\beta \in \Lambda$, $\varphi_{Y}(w(\alpha)) = \varphi_{Y}(\alpha)$. This is clear since $\varphi_Y([C]) =1$, hence $\varphi_{Y}(r_C(\alpha)) = \varphi_{Y}(\alpha)$ for all $\alpha \in \Lambda$. \[Weyltrans\] Suppose that $C=\sum_ia_iC_i$, where the $C_i$ are $-2$-curves, $a_i\in \Zee$, $C^2 = -2$, the support of $C$ is connected, and $(C_i\cdot C_j)$ is negative definite. Then there exists an element $w$ in the group generated by reflections in the $[C_i]$ such that $w([C]) = [C_i]$ for some $i$. This follows from the well known fact that, if $R$ is an irreducible root system such that all roots have the same length, then the Weyl group $\mathsf{W}(R)$ acts transitively on the set of roots. \[mainprop\] Let $\beta \in \Lambda$ with $\beta^2 = -2$. Then the following are equivalent: 1. Let $Y_1$ be a deformation of $Y$ with trivial monodromy such that $\varphi_{Y_1}(\beta) = 1$. Then, with $\mathsf{W}({\Delta_{Y_1}})$ as in Definition \[defcurves\], there exists $w\in \mathsf{W}({\Delta_{Y_1}})$ such that $w(\beta)=[C]$, where $C$ is a $-2$-curve on $Y_1$. In particular, if $Y_1$ is generic subject to the condition that $\varphi_{Y_1}(\beta) = 1$ (i.e. if $\Ker \varphi_{Y_1} = \Zee \cdot \beta$), then $\pm \beta = [C]$ for a $-2$-curve $C$. 2. The wall $W^\beta$ meets the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$. 3. If $r_\beta$ is reflection in the class $\beta$, then $r_\beta(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Lemma \[reflect\] implies that (i) $\implies$ (iii) in case $Y=Y_1$ and $\beta = [C]$ where $C$ is a $-2$-curve. The case where $w(\beta) = [C]$ follows easily from this since, for all $w\in \mathsf{W}({\Delta_{Y_1}})$, $w\circ r_\beta \circ w^{-1} = r_{w(\beta)}$. Lemma \[definv\] then handles the case where $Y_1$ is replaced by a general deformation $Y$. Also, clearly (iii) $\implies$ (ii). So it is enough to show that (ii) $\implies$ (i). In fact, by Lemma \[Weyltrans\], it is enough to show that, if $Y$ is any surface such that $\varphi_Y(\beta) = 1$ and $W^\beta$ meets the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, then there exists a $w\in \mathsf{W}({\Delta_Y})$ such that $w(\beta) = [\sum_ia_iC_i]$ where $a_i\in \Zee^+$, the $C_i$ are curves disjoint from $D$, and $\bigcup_iC_i$ is connected. By hypothesis, there exists an $x$ in the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ such that $x\cdot \beta =0$. In particular, $x\cdot [D_i] >0$ for all $i$. We can assume that $x=[H]$ is the class of a divisor $H$. After replacing $x$ by $w(x)$ and $\beta$ by $w(\beta)$ for some $w\in \mathsf{W}({\Delta_Y})$, we can assume that $x$ (and hence $H$) lies in $\overline{\mathcal{A}}(Y)$, so that $H$ is a nef and big divisor with $H\cdot D_i > 0$ for all $i$, and we still have $\varphi_Y(\beta) = 1$ by Lemma \[permapinvar\]. By Lemma \[numeric\], possibly after replacing $\beta$ by $-\beta$, $\beta = [\sum_ia_iC_i]$ where the $C_i$ are irreducible curves and $a_i \in \Zee^+$. Since $\beta\cdot [H] =\sum_ia_i(C_i\cdot H)=0$, $C_i\cdot H \geq 0$, and $D_j\cdot H > 0$, $C_i\cdot H = 0$ for all $i$ and no $C_i$ is equal to $D_j$ for any $j$. Hence the $C_i$ are curves meeting each $D_j$ in at most finitely many points and $\sum_ia_i(C_i\cdot D_j)=0$, so that $C_i\cap D_j =\emptyset$. Finally each $(C_i)^2 < 0$ by Hodge index, and so each $C_i$ is a $-2$-curve. Moreover the $C_i$ span a negative definite lattice, and in particular their classes are independent. From this, the statement about the connectedness of $\bigcup_iC_i$ is clear. Let $R=R_Y$ be the set of all $\beta \in \Lambda$ such that $\beta ^2 = -2$ and such that there exists some deformation of $Y$ for which $\beta$ becomes the class of a $-2$-curve. Following [@GHK], we call $R$ the set of *Looijenga roots* (or briefly *roots*) of $Y$. Note that $R$ only depends on the deformation type of $Y$. The definition of $R$ is slightly ill-posed, since we have not specified an identification of the cohomologies of the fibers along the deformation. In particular, if $\beta = [C]$ is a $-2$-curve on $Y$, then by (i) of Remark \[monodromyinvar\], if $Y'$ is a nearby deformation of $Y$, then a general smoothing of the ordinary double point on the contraction of $C$ on $Y$ has monodromy which sends $[C]$ to $-[C]$, and hence $-\beta \in R$ as well. To avoid this issue, it is simpler to define $R$ to be the set of $\beta \in \Lambda$, $\beta^2=-2$, which satisfy either of the equivalent conditions (ii), (iii) of Theorem \[mainprop\]. Given $Y$, let $\Delta_Y$ be the set of classes of $-2$-curves on $Y$ and $\mathsf{W}({\Delta_Y})$ the reflection group generated by $\Delta_Y$. Finally set $R^{\text{\rm{nod}}}$, the set of *nodal classes*, to be $\mathsf{W}({\Delta_Y})\cdot \Delta_Y$. Then $R^{\text{\rm{nod}}} \subseteq R$. \[preserveamp\] If $f\colon H^2(Y; \Zee) \to H^2(Y; \Zee)$ is an integral isometry preserving the classes $[D_i]$ such that $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$, then $f(R) = R$. If $\mathsf{W}(R)$ is the reflection group generated by reflections in the elements of $R$, then $\mathsf{W}(R) \cdot R = R$ and $w(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$ for all $w\in \mathsf{W}(R)$. A result similar to Theorem \[mainprop\] classifies the elements of $H^2(Y;\Zee)$ which are represented by the class of a smoothly embedded $2$-sphere of self-intersection $-2$ in terms of the “super $P$-cell" of [@FriedmanMorgan]. In [@Looij], for the case where the length $r(D) \leq 5$, Looijenga defines a subset $R_L$ of $\Lambda$ by starting with a particular configuration $B$ of elements of square $-2$ (a *root basis* in the terminology of [@Looij]), and setting $R_L = \mathsf{W}(B)\cdot B$, where $\mathsf{W}(B)$ is the reflection group generated by $B$. In fact, the set $R_L$ is just the set $R$ of Looijenga roots: In the above notation, $R_L = R$. It is easy to see from the construction of [@Looij I §2] that $B \subseteq R$. Hence $R_L \subseteq R$. Conversely, if $\alpha \in R$, then, by (ii) of Corollary \[preserveamp\], $r_\alpha(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$. It then follows from [@Looij Proposition I (4.7)] that $r_\alpha \in \mathsf{W}(B)$. By a general result in the theory of reflection groups [@Bour V §3.2, Thm. 1(iv)], $r_\alpha = r_\beta$ for some $\beta \in R_L$. Thus $\alpha =\pm \beta$, so that $\alpha \in R_L$. Hence $R\subseteq R_L$, and therefore $R_L = R$. \[irredex\] Let $(Y,D)$ be the blowup of $\Pee^2$ at $N \geq 10$ points on an irreducible nodal cubic curve. We let $h$ be the pullback of the class of a line on $\Pee^2$ and $e_1, \dots, e_N$ be the classes of the exceptional curves. \(i) Let $\alpha = -3h + \sum_{i=1}^{10}e_i$. Then $\alpha^2 = \alpha \cdot [K_Y] = -1$, so that $\alpha$ is a numerical exceptional curve. But there exists a nef and big divisor $H$ (for example $h$) such that $\alpha \cdot [H] < 0$, so that $\alpha$ is not effective. Hence, $\alpha \cdot x \leq 0$ for all $x\in \overline{\mathcal{A}}(Y)=\overline{\mathcal{A}}_{\text{\rm{gen}}}$, since $W^\alpha$ does not pass through the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$. Note that $W^\alpha$ is never a face of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$. For $N=10$, $W^{-\alpha}$ is a face of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, but this is no longer the case for $N \geq 11$. Thus the condition $\alpha \cdot [H] \geq 0$ for some $H$ such that $H\cdot D > 0$ is necessary for $\alpha$ to be effective. More generally, let $f = 3h -\sum_{i=1}^9e_i$ and set $\alpha = kf + e_{10}$ (the case above corresponds to $k=-1$). As above, $\alpha$ is a numerical exceptional curve. For $k\leq -1$, $h\cdot \alpha < 0$, and hence $\alpha$ is not effective. For $k\geq 1$, $\alpha$ is effective but it is not the class of an exceptional curve: for all $x\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$, $x\cdot f > 0$, and $x\cdot e_{10}\geq 0$. Hence $x\cdot \alpha > 0$ for all $x \in \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Thus $W^\alpha$ is not a face of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ and so $\alpha$ is not the class of an exceptional curve. \(ii) With $\alpha$ any of the classes as above, suppose that $N \geq 11$ and $k\neq 0$ and set $\beta = \alpha -e_{11}$. Then $\beta^2 =-2$ and $\beta \cdot [K_Y] = 0$. However, $$r_\beta(e_{11}) = e_{11} + (e_{11}\cdot \beta)\beta = \alpha.$$ Since $W^{e_{11}}$ is a face of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ and $W^\alpha$ is not a face of $\overline{\mathcal{A}}_{\text{\rm{gen}}}$, $r_\beta( \overline{\mathcal{A}}_{\text{\rm{gen}}})\neq \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Hence $\beta$ does not satisfy any of the equivalent conditions of Theorem \[mainprop\], so that $\beta \notin R$. \[bestposs\] In the situation of the example above, it is well-known that if $D$ is irreducible, $N \leq 9$ (i.e. $D^2\geq 0$), and there are no $-2$-curves on $Y$, then every numerical exceptional curve is the class of an exceptional curve, so (i) above is best possible. A generalization is given in Proposition \[nonneg\] below. We shall show in Proposition \[minusone\] that the example in (ii) is best possible as well. The numerical exceptional curves given in (i) of Example \[irredex\] were known to Du Val. In fact, he showed that they are essentially the only numerical curves in case $Y$ is the blowup of $\Pee^2$ at $10$ points ([@duV2], pp. 46–47): Suppose that $(Y,D)$ is the blowup of $\Pee^2$ at $10$ points lying on an irreducible cubic, that $Y$ is generic in the sense that there are no $-2$-curves on $Y$, and that $\alpha$ is a numerical exceptional curve. Then there exists an exceptional curve $E$ on $Y$ and an integer $k$ such that $\alpha$ is the class of $k(D + E) + E$. Suppose that $\alpha$ is a numerical exceptional curve on $Y$. Then, since $K_Y^2=-1$, $\lambda = \alpha +[D] = \alpha -[K_Y]$ satisfies: $\lambda^2 = \lambda \cdot \alpha = \lambda \cdot [K_Y] = 0$. In particular, $\lambda \in \Lambda$. Conversely, given an isotropic vector $\lambda \in \Lambda$, if we set $\alpha = \lambda + [K_Y]$, then $\alpha$ is a numerical exceptional curve. Any isotropic vector $\lambda \in \Lambda$ can be uniquely written as $n\lambda_0$, where $n \in \Zee$ and $\lambda_0$ is primitive and lies in $\overline{\mathcal{C}^+}$. Note that $H^2(Y; \Zee) = \Zee[K_Y] \oplus \Lambda$ and that $\Lambda = U \oplus (-E_8)$ (both sums orthogonal). An easy exercise shows that, if $\Aut^+(\Lambda)$ is the group of integral isometries $A$ of $\Lambda$ such that $A(\mathcal{C}^+\cap \Lambda) = \mathcal{C}^+\cap \Lambda$, i.e. $A$ has real spinor norm equal to $1$, then every $A \in \Aut^+(\Lambda)$ extends uniquely to an integral isometry of $H^2(Y; \Zee)$ fixing $[K_Y]$ and hence $[D]$, and moreover that $\Aut^+(\Lambda)$ acts transitively on the set of (nonzero) primitive isotropic vectors in $\overline{\mathcal{C}^+} \cap \Lambda$. Hence there exists an $A \in \Aut^+(\Lambda)$ such that $A(\lambda_0) = f$, in the notation of Example \[irredex\]. If we continue to denote by $A$ the extension of $A$ to an isometry of $H^2(Y; \Zee)$, then $A(\alpha) = nf + [K_Y] = (n-1) f + e_{10}$, since $f = -[K_Y] + e_{10}$. It follows that $\alpha = (n-1)\lambda_0 + A^{-1}(e_{10})$. Using Proposition \[minusone\] below, $A^{-1}$ preserves the walls of the ample cone of $Y$, and thus $A^{-1}(e_{10}) =e$ is the class of an exceptional curve $E$, and $\lambda_0 =A^{-1}(f) = A^{-1}([D] + e_{10}) = [D] + E$. Hence, setting $k=n-1$, $\alpha$ is the class of $k(D + E) + E$ as claimed. The proof above shows the following: Let $(Y,D)$ be the blowup of $\Pee^2$ at $10$ points lying on an irreducible cubic and such that there are no $-2$-curves on $Y$, let $\alpha$ be a numerical exceptional curve on $Y$, and let $\lambda = \alpha -[K_Y]$. Then: 1. $\alpha$ is effective $\iff$ $\lambda \in (\overline{\mathcal{C}^+}-\{0\}) \cap \Lambda$. 2. $\alpha$ is not effective $\iff$ $\lambda \in (-\overline{\mathcal{C}^+}) \cap \Lambda$. 3. $\alpha$ is the class of an exceptional curve $\iff$ $\lambda$ is a primitive isotropic vector in $\overline{\mathcal{C}^+} \cap \Lambda$. Thus there is a bijection from the set of exceptional curves on $Y$ to the set of primitive isotropic vectors in $\overline{\mathcal{C}^+} \cap \Lambda$. In the above situation, let $\mathsf{W}$ be the group generated by the reflections in the classes $e_1-e_2, \dots, e_9-e_{10}, h-e_2-e_2-e_3$, which are easily seen to be Looijenga roots. A classical argument (usually called Noether’s inequality) shows that, if $\lambda_0$ is a primitive integral isotropic vector in $\Lambda$ lying in $\overline{\mathcal{C}^+}$, then there exists $w\in \mathsf{W}$ such that $w(\lambda_0) = f = 3h -\sum_{i=1}^9 e_i$, in the notation of Example \[irredex\]. Thus, $\mathsf{W}$ acts transitively on the set of such vectors. Using standard results about the affine Weyl group of $E_8$, it is then easy to see that $\mathsf{W} = \Aut^+(\Lambda)$. This was already noted by Du Val in [@duV2]. Roots and the ample cone ======================== By Corollary \[preserveamp\], if $f\colon H^2(Y; \Zee) \to H^2(Y; \Zee)$ is an integral isometry preserving the classes $[D_i]$ such that $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$, then $f(R) = R$. In this section, we find criteria for when the converse holds. We begin with the following: \[containopen\] Let $f\colon H^2(Y; \Zee) \to H^2(Y; \Zee)$ be an integral isometry preserving $\mathcal{C}^+$ and the classes $[D_i]$. If $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}) \cap \overline{\mathcal{A}}_{\text{\rm{gen}}}$ contains an open set, then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}) = \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Choosing $x\in f(\overline{\mathcal{A}}_{\text{\rm{gen}}}) \cap \overline{\mathcal{A}}_{\text{\rm{gen}}}$ corresponding to an ample divisor, it is easy to see that $f(\overline{\mathcal{A}}_{\text{\rm{gen}}})$ and $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ have the same set of walls, hence are equal. Next we deal with the case where one component of $D$ is a smooth rational curve of self-intersection $-1$. \[excepcomp\] Suppose that $D$ is reducible and that $D_r^2=-1$. Let $(\overline{Y}, \overline{D})$ be the anticanonical pair obtained by contracting $D_r$. Then any isometry $f$ of $H^2(Y;\Zee)$ preserving the classes $[D_i]$, $1\leq i\leq r$, defines an isometry $\bar{f}$ of $H^2(\overline{Y};\Zee)$ preserving the classes $[\overline{D}_i]$, $1\leq i\leq r-1$, and conversely. Moreover, $f$ preserves $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ $\iff$ $\bar{f}$ preserves $\overline{\mathcal{A}}_{\text{\rm{gen}}}(\overline{Y})$, and $R_Y$ is naturally identified with the roots $R_{\overline{Y}}$ of $\overline{Y}$. The first statement is clear. Identifying $H^2(\overline{Y}, \Zee)$ with $[D_r]^\perp \subseteq H^2(Y; \Zee)$, it is clear that $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y) \cap [D_r]^\perp =\overline{\mathcal{A}}_{\text{\rm{gen}}}(\overline{Y})$. Hence, if $f$ preserves $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$, then $\bar{f}$ preserves $\overline{\mathcal{A}}_{\text{\rm{gen}}}(\overline{Y})$. Since a divisor $\overline{H}$ on $\overline{Y}$ is ample $\iff$ $N \overline{H} - D_r$ is ample for all $N \gg 0$, it follows that, if $\bar{f}$ preserves $\overline{\mathcal{A}}_{\text{\rm{gen}}}(\overline{Y})$, then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) \cap \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ contains an open set, and hence $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ by Lemma \[containopen\]. It follows from this and from Theorem \[mainprop\] that $R_Y$ is naturally identified with $R_{\overline{Y}}$ (or directly from the definition by noting that there is a bijection from the set of deformations of $(Y,D)$ to those of $(\overline{Y}, \overline{D})$). Henceforth, then, we shall always assume if need be that no component of $D$ is a smooth rational curve of self-intersection $-1$. We turn to the straightforward case where $(Y,D)$ is not negative definite: \[nonneg\] Suppose that $(Y,D)$ and $(Y', D')$ are two anticanonical pairs with $r(D) = r(D')$ and such that neither pair is negative definite. If $f\colon H^2(Y; \Zee) \to H^2(Y'; \Zee)$ is an integral isometry with $f([D_i]) = [D_i']$ for all $i$, then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$ and hence $f(R_Y) = R_{Y'}$. Moreover, $$R_Y = \{\beta \in \Lambda(Y,D): \beta^2 =-2\}.$$ By Lemma \[excepcomp\], we may assume that no $D_i$ has self-intersection $-1$. The statement that the cycle is not negative definite is then equivalent to the statement that either $D_j^2 \geq 0$ for some $j$ or $D_i^2 = -2$ for all $i$ and $r\geq 2$. In the first case, $D_j$ is nef and $D_j \cdot D > 0$. Hence, if $\alpha$ is a numerical exceptional curve such that $\alpha\cdot [D_i] \geq 0$, then $\alpha$ is effective by Lemma \[numeric\]. Thus $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ is the set of all $x\in \mathcal{C}^+_D(Y)$ such that $x\cdot \alpha \geq 0$ for all numerical exceptional curves $\alpha$ such that $\alpha\cdot [D_i] \geq 0$. Since $f(\alpha)^2 = \alpha^2$, $f([D_i]) = [D_i']$, and $f(\alpha) \cdot [K_{Y'}] = \alpha \cdot [K_Y]$, it follows that $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. Applying this to reflection in a class $\beta$ of square $-2$ in $\Lambda(Y,D)$ then implies that $\beta \in R_Y$. The case where $D_i^2 = -2$ for every $i$ is similar, using the nef divisor $D = \sum_iD_i$ with $D^2 = 0$. If $\alpha$ is a numerical exceptional curve, then $\alpha$ is effective since $(-\alpha + [K_Y]) \cdot [D] = \alpha \cdot [K_Y] = -1$. The rest of the argument proceeds as before. If $D$ is irreducible and not negative definite (i.e. $D^2 \geq 0$) and there are no $-2$-curves on $Y$, then, as is well-known and noted in Remark \[bestposs\], every numerical exceptional curve is the class of an exceptional curve. However, if $D$ is reducible but not negative definite, then, even if there are no $-2$-curves on $Y$, there may well exist numerical exceptional curves which are not effective, and effective numerical exceptional curves which are not the class of an exceptional curve. From now on we assume that $D$ is negative definite. The case $K_Y^2=-1$ can also be handled by straightforward methods, as noted in [@Looij]. (See also [@FriedmanMorgan], II(2.7)(c) in case $D$ is irreducible.) \[minusone\] Let $(Y,D)$ and $(Y', D')$ be two negative definite anticanonical pairs with $r(D) = r(D')$ and $K_Y^2 = K_{Y'}^2 = -1$. Let $f\colon H^2(Y; \Zee) \to H^2(Y'; \Zee)$ be an isometry such that $f([D_i]) = [D_i']$ for all $i$ and $f (\mathcal{C}^+(Y)) = \mathcal{C}^+(Y')$. Then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. Moreover, $$R_Y = \{\beta \in \Lambda(Y,D): \beta^2 = -2\},$$ and hence $f(R_Y) = R_{Y'}'$. Since $(Y,D)$ is negative definite, no component of $D$ is a smooth rational curve of self-intersection $-1$. Fix a nef and big divisor $H$ such that $H\cdot D_i =0$ for all $i$ and $H\cdot G > 0$ for every irreducible curve $G \neq D_i$. If $\alpha$ is a numerical exceptional curve, $(\alpha -[K_Y])^2 = (\alpha + [D]) ^2 = 0$. By Lemma \[remarkafternumeric\], $\alpha$ is effective $\iff$ $[H] \cdot \alpha > 0$ $\iff$ $[H] \cdot (\alpha + [D]) > 0$. By the Light Cone Lemma (cf. [@FriedmanMorgan], p. 320), this last condition is equivalent to: $\alpha + [D] \in \overline{\mathcal{C}^+}-\{0\}$. Since this condition is clearly preserved by an isometry $f$ as in the statement of the proposition, we see that $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. The final statement then follows as in the proof of Proposition \[nonneg\]. The hypothesis $K_Y^2 =-1$ implies that $r(D) \leq 10$, so there are only finitely many examples of the above type. For $r(D) = 10$, there is essentially just one combinatorial possibility for $(Y,D)$ neglecting the orientation (cf. [@FriedmanMiranda], (4.7), where it is easy to check that this is the only possibility). For $r(D) = 9$, however, there are two different possibilities for the combinatorial type of $(Y,D)$ (again ignoring the orientation). Begin with an anticanonical pair $(\overline{Y}, \overline{D})$, where $\overline{Y}$ is a rational elliptic surface and $\overline{D}=\overline{D}_0 + \cdots + \overline{D}_8$ is a fiber of type $\widetilde{A}_8$ (or $I_9$ in Kodaira’s notation). There is a unique such rational elliptic surface $\overline{Y}$ and its Mordell-Weil group has order $3$ (see for example [@MirandaPersson]). In particular, possibly after relabeling the components, there is an exceptional curve meeting $\overline{D}_i$ $\iff$ $i=0,3,6$. It is easy to see that blowing up a point on a component $\overline{D}_i$ meeting an exceptional curve leads to a different combinatorial possibility for an anticanonical pair $(Y,D)$ with $K_Y^2 =-1$ and $r(D) = 9$ than blowing up a point on a component $\overline{D}_i$ which does not meet an exceptional curve. We turn now to the case where $(Y,D)$ is negative definite but with no assumption on $K_Y^2$. A point $x\in \mathcal{C}^+ \cap \Lambda$ is *$R$-distinguished* if there exists a codimension one negative definite subspace $V$ of $\Lambda \otimes \Ar$ spanned by elements of $R$ such that $x\in V^\perp$. Note that the definition only depends on the deformation type of the pair $(Y,D)$. Clearly, if $V$ is a codimension one negative definite subspace of $\Lambda \otimes \Ar$ spanned by elements of $R$, then $V$ is defined over $\Q$ and $V^\perp \cap (\Lambda \otimes \Ar)$ is a one-dimensional subspace of $H^2(Y; \Ar)$ defined over $\Q$ and spanned by an $h\in H^2(Y; \Zee)$ with $h^2 > 0$, $h\cdot [D_i] =0$, and $h\cdot \beta = 0$ for all $\beta \in R\cap V$. Hence, if $h\in \mathcal{C}^+\cap \Lambda$, then $h$ is $R$-distinguished. Also, if the rank of $\Lambda$ is one, then $\{0\}$ is a codimension one negative definite subspace of $\Lambda \otimes \Ar$, and hence every point of $\mathcal{C}^+ \cap \Lambda$ is $R$-distinguished. However, as we shall see, there exist deformation types $(Y,D)$ with no $R$-distinguished points. The following is also clear: Let $(Y,D)$ and $(Y', D')$ be two anticanonical pairs with $r(D) = r(D')$ and let $f\colon H^2(Y; \Zee) \to H^2(Y'; \Zee)$ be an isometry such that $f([D_i]) = [D_i']$ for all $i$, $f (\mathcal{C}^+(Y)) = \mathcal{C}^+(Y')$, and $f(R_Y) = R_{Y'}$. Then, if $x$ is a $R_Y$-distinguished point of $\mathcal{C}^+(Y) \cap \Lambda(Y,D)$, $f(x)$ is a $R_{Y'}$-distinguished point of $ \mathcal{C}^+(Y') \cap \Lambda(Y',D')$. Our goal now is to prove: \[disttheorem\] Let $(Y,D)$ and $(Y', D')$ be two anticanonical pairs with $r(D) = r(D')$ and let $f\colon H^2(Y; \Zee) \to H^2(Y'; \Zee)$ be an isometry such that $f([D_i]) = [D_i']$ for all $i$, $f (\mathcal{C}^+(Y)) = \mathcal{C}^+(Y')$, and $f(R_Y) = R_{Y'}$. If there exists a $R$-distinguished point of $\mathcal{C}^+ \cap \Lambda$, then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. We begin by showing: \[aprop\] Let $x$ be a $R$-distinguished point of $\mathcal{C}^+ \cap \Lambda$. Then $x\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Moreover, if $\alpha$ is a numerical exceptional curve and $\alpha$ is not in the span of the $[D_j]$, then $\alpha$ is effective $\iff$ $\alpha \cdot x \geq 0$. It is enough by Lemma \[definv\] to check this on some (global) deformation of $(Y,D)$ with trivial monodromy. By Theorem \[surjper\], we can assume that $$\Ker \varphi_Y = V \cap \Lambda,$$ where $V$ is as in the definition of $R$-distinguished. In particular, if $C \in \Delta_Y$, i.e. $C$ is a $-2$-curve on $Y$, then $[C] \in V$. It follows from (i) of Theorem \[mainprop\] that every $\beta \in V\cap R$ is a sum of elements of $\Delta_Y$, so that $\Delta_Y$ spans $V$ over $\Q$. Thus there exist $-2$-curves $C_1, \dots, C_k$ such that $V$ is spanned by the classes $[C_i]$, and the intersection matrix $(C_i\cdot C_j)$ is negative definite. The classes $[C_1], \dots, [C_k], [D_1], \dots, [D_r]$ span a negative definite sublattice of $H^2(Y; \Zee)$. By Lemma \[constructnef\] there exists a nef and big divisor $H$ such that $H$ is perpendicular to the curves $C_1, \dots, C_k, D_1, \dots, D_r$. Clearly, then, $[H] \in \overline{\mathcal{A}}(Y) \subseteq \overline{\mathcal{A}}_{\text{\rm{gen}}}$ and $[H] = tx$ for some $t\in \Ar^+$. Hence $x\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$ as well. Note that $[H]^\perp$ is spanned over $\Q$ by $[C_1], \dots, [C_k], [D_1], \dots, [D_r]$. Since $x\in \overline{\mathcal{A}}(Y)$, if $\alpha$ is effective, $x\cdot \alpha \geq 0$. Conversely, suppose that $\alpha$ is a numerical exceptional curve with $x\cdot \alpha \geq 0$ and that $\alpha$ is not effective. Then $-\alpha + [K_Y]=[G]$, where $G$ is effective, and $H\cdot (-\alpha + [K_Y]) = -\alpha \cdot [H] \leq 0$. Hence $(-\alpha + [K_Y]) \cdot [H] = 0$. $-\alpha + [K_Y] = \sum_ia_i[C_i] + \sum_jb_j[D_j]$ where the $a_i, b_j \in \Zee$. In any case, since $-\alpha + [K_Y]$ is perpendicular to $[H]$, there exist $a_i, b_j \in \Q$ such that $-\alpha + [K_Y] = \sum_ia_i[C_i] + \sum_jb_j[D_j]$. Write $-\alpha + [K_Y]= [G] = \sum_in_i[C_i] + \sum_j m_j[D_j] +[F]$, where $n_i, m_j \in \Zee$ and $F$ is an effective curve not containing $C_i$ or $D_j$ in its support for any $i,j$. By (ii) of Lemma \[negdef\], $F=0$, $a_i = n_i$ and $b_j= m_j$ for all $i,j$. Hence $a_i, b_j \in \Zee$. Since $-\alpha + [K_Y]$ is an integral linear combination of the $[C_i]$ and $[D_j]$, the same holds for $\alpha$. Then $\alpha = \sum_ic_i[C_i] + \sum_jd_j[D_j]$ with $c_i, d_j\in \Zee$. But $\alpha ^2 =-1 = (\sum_ic_iC_i)^2 + (\sum_jd_jD_j)^2$. Both terms are non-positive, and so $(\sum_ic_iC_i)^2 \geq -1$. But if $\sum_ic_iC_i \neq 0$, then $(\sum_ic_iC_i)^2 \leq -2$. Thus $\sum_ic_iC_i =0$ and $\alpha$ lies in the span of the $[D_j]$. Conversely, if $\alpha$ is not in the span of the $[D_j]$ and $\alpha \cdot x \geq 0$, then $\alpha$ is the class of an effective curve. It follows from Proposition \[aprop\] that, if $x\in \mathcal{C}^+(Y) \cap \Lambda(Y,D)$ is $R_Y$-distinguished, then $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)$ is the set of all $y\in \mathcal{C}^+_D(Y)$ such that $\alpha \cdot y \geq 0$ for all $\alpha$ a numerical exceptional curve on $Y$, not in the span of the $[D_i]$, such that $\alpha \cdot x \geq 0$. Let $f$ be an isometry satisfying the conditions of the theorem. Then $f(x)$ is $R_{Y'}$-distinguished, and $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y))$ is clearly the set of all $y\in \mathcal{C}^+_{D'}(Y')$ such that $\alpha \cdot y \geq 0$ for all $\alpha$ a numerical exceptional curve on $Y'$, not in the span of the $[D_i']$, such that $\alpha \cdot f(x) \geq 0$. Again by Proposition \[aprop\], this set is exactly $\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. Theorem \[disttheorem\] covers all of the cases in [@Looij] except for the case of $5$ components: By inspection of the root diagrams on pp. 275–277 of [@Looij], the complement of any trivalent vertex spans a negative definite codimension one subspace, except in the case of $5$ components. To give a direct argument along the above lines which also handles this case (and all of the other cases in [@Looij]), we recall the basic setup there: There exists a subset $B= \{\beta_1, \dots, \beta_n\} \subseteq R$ such that $B$ is a basis for $\Lambda\otimes \Ar$, and there exist $n_i\in \Zee^+$ such that $(\sum_in_i\beta_i) \cdot \beta_j > 0$ for all $j$ (compare also [@Looijpre] (1.18)). In particular, note that the intersection matrix $(\beta_i\cdot \beta_j)$ is non-singular. Finally, by the classification of Theorem (1.1) in [@Looij], there exists a deformation of $(Y,D)$ for which $\beta_i = [C_i]$ is the class of a $-2$-curve for all $i$. (With some care, this explicit argument could be avoided by appealing to the surjectivity of the period map and (i) of Theorem \[mainprop\].) Let $(Y,D)$ and $(Y', D')$ be two anticanonical pairs satisfying the hypotheses of the preceding paragraph, both negative definite, with $r(D) = r(D')$, and let $f\colon H^2(Y; \Zee) \to H^2(Y'; \Zee)$ be an isometry such that $f([D_i]) = [D_i']$ for all $i$, $f (\mathcal{C}^+(Y)) = \mathcal{C}^+(Y')$, and $f(R_Y) = R_{Y'}$. Then $f(\overline{\mathcal{A}}_{\text{\rm{gen}}}(Y)) = \overline{\mathcal{A}}_{\text{\rm{gen}}}(Y')$. (Sketch) With notation as in the paragraph preceding the statement of the theorem, let $h =\sum_in_i\beta_i$ have the property that $h\cdot \beta_i > 0$. By the arguments used in the proof of Theorem \[disttheorem\], it is enough to show that $h\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$ and that, if $\alpha$ is a numerical exceptional curve and $\alpha$ is not in the span of the $[D_j]$, then $\alpha$ is effective $\iff$ $\alpha \cdot h \geq 0$. Also, it is enough to prove this for some deformation of $(Y,D)$, so we can assume $\beta_i = [C_i]$ is the class of a $-2$-curve for all $i$, hence that $h$ is the class of $H=\sum_in_iC_i$. By construction, $H\cdot C_j > 0$ for every $j$, hence $H$ is nef and big. By Lemma \[remarkafternumeric\], it is enough to show that, if $G$ is an irreducible curve not equal to $D_i$ for any $i$, then $H\cdot G > 0$. Since $H$ is nef, it suffices to rule out the case $H\cdot G =0$, in which case $G^2 < 0$. As $G\neq D_j$ for any $j$, then $G$ is either a $-2$-curve or an exceptional curve. The case where $G$ is a $-2$-curve is impossible since then $G$ is orthogonal to the span of the $[C_i]$, but the $[C_i]$ span $\Lambda$ over $\Q$ and the intersection form is nondegenerate. So $G=E$ is an exceptional curve disjoint from the $C_i$. If $(\overline{Y}, \overline{D})$ is the anticanonical pair obtained by contracting $E$, then the $[C_i]$ define classes in $\overline{\Lambda} = \Lambda(\overline{Y}, \overline{D})$. Since the intersection form $(C_i\cdot C_j)$ is nondegenerate, the rank of $\overline{\Lambda}$ is at least that of the rank of $\Lambda$. It is easy to check that the classes of $\overline{D}_1, \dots, \overline{D}_r$ are linearly independent: if say $E$ meets $D_1$, then the intersection matrix of $\overline{D}_2, \dots, \overline{D}_r$ is still negative definite, and then (ii) of Lemma \[negdef\] (with $F = \overline{D}_1$ and $G_1, \dots, G_n = \overline{D}_2, \dots, \overline{D}_r$) shows that the classes of $\overline{D}_1, \dots, \overline{D}_r$ are linearly independent. Hence the rank of $H^2(\overline{Y}; \Zee)$ is greater than or equal to the rank of $H^2(Y; \Zee)$, which contradicts the fact that $\overline{Y}$ is obtained from $Y$ by contracting an exceptional curve. Some examples ============= We give a series of examples satisfying the hypotheses of Theorem \[disttheorem\] where the number of components and the multiplicities are arbitrarily large. Let $(\overline{Y}, \overline{D})$ be the anticanonical pair obtained by making $k+6$ infinitely near blowups starting with the double point of a nodal cubic. Thus $\overline{D} = \overline{D}_0 + \cdots + \overline{D}_{k+6}$, where $\overline{D}_0^2 = -k$, $\overline{D}_i^2 = -2$, $1\leq i\leq k+5$, and $\overline{D}_{k+6}^2 = -1$. Now blow up $N \geq 1$ points $p_1, \dots, p_N$ on $\overline{D}_{k+6}$, and let $(Y,D)$ be the resulting anticanonical pair. Note that $(Y,D)$ is negative definite as long as $k\geq 3$ or $k=2$ and $N\geq 2$. Clearly $r(D) = k+7$ and $K_Y^2 = 3-k-N$. It follows that $\Lambda = \Lambda(Y,D)$ has rank $N$. If $E_1, \dots, E_N$ are the exceptional curves corresponding to $p_1, \dots, p_N$, then the classes $[E_i] - [E_{i+1}]$ span a negative definite root lattice of type $A_{N-1}$ in $\Lambda$. By making all of the blowups infinitely near to the first point, we see that all of the classes $[E_i] - [E_{i+1}]$ lie in $R$. Hence $(Y,D)$ satisfies the hypotheses of Theorem \[disttheorem\]. Next we turn to examples where the rank of $\Lambda$ is small. The case where the rank of $\Lambda$ is $1$ is covered by Theorem \[disttheorem\], as well as the case where the rank of $\Lambda$ is $2$ and $R\neq \emptyset$. Note that, conjecturally at least, the case where $R\neq \emptyset$ should be related to the question of whether the dual cusp singularity deforms to an ordinary double point. It is easy to construct examples where the rank of $\Lambda$ is $2$ and with $R\neq \emptyset$: begin with an anticanonical pair $(\hat{Y}, \hat{D})$ where the rank of $\Lambda(\hat{Y}, \hat{D})$ is $1$, locate a component $\hat{D}_i$ such that there exists an exceptional curve $E$ on $\hat{Y}$ with $E \cdot \hat{D}_i = 1$, and blow up a point of $\hat{D}_i$ to obtain a new anticanonical pair $(Y,D)$ together with exceptional curves $E,E'$ (where we continue to denote by $E$ the pullback to $Y$ and by $E'$ the new exceptional curve), such that $[E] -[E'] \in R$. So our interest is in finding examples where $R=\emptyset$. In case the rank of $\Lambda$ is $2$ and $R\neq \emptyset$, it is easy to see that either $(\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda)/\Ar^+$ is a closed (compact) interval or $\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda =\mathcal{C}^+\cap \Lambda$ (and in fact both cases arise). In either case, there is at most one wall $W^\beta$ with $\beta \in R$ passing through the interior of $\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda$, and hence either $R=\emptyset$ or $R =\{\pm \beta\}$. We give an example where the rank of $\Lambda$ is $2$ and there are no $\beta \in \Lambda$ such that $\beta^2 = -2$, in particular $R=\emptyset$, hence the condition $f(R) = R$ is automatic for every isometry $f$, and of an isometry $f$ which preserves $\mathcal{C}^+$ and the classes $[D_i]$ but not the generic ample cone. Let $(\overline{Y}, \overline{D})$ be the anticanonical pair obtained by making $9$ infinitely near blowups starting with the double point of a nodal cubic. Thus $\overline{D} = \overline{D}_0 + \cdots + \overline{D}_9$, where $\overline{D}_0 = 3H -2E_1 -\sum_{i=2}^9E_i$, $\overline{D}_i = E_i - E_{i+1}$, $1\leq i \leq 8$, and $\overline{D}_9 = E_9$. Make two more blowups, one at a point $p_{10}$ on $\overline{D}_9$, and one at a point $p_{11}$ on $\overline{D}_4$. This yields an anticanonical pair $(Y,D)$ with $D_0 = 3H -2E_1 -\sum_{i=2}^9E_i$, $D_i = E_i - E_{i+1}$, $i>0$ and $i\neq 4$, and $D_4 = E_4 - E_5 - E_{11}$. Thus $$(-d_0, \dots, -d_9) = (3,2,2,2,3,2,2,2,2,2),$$ i.e.  $D$ is of type $\displaystyle \begin{pmatrix} 3&3\\3&5\end{pmatrix}$, with dual cycle $\displaystyle \begin{pmatrix} 6&8\\0&0\end{pmatrix}$ in the notation of [@FriedmanMiranda]. Set $$\begin{aligned} G_1 &= 5H - 2\sum_{i=1}^4E_i - \sum_{i=5}^{10}E_i -E_{11};\\ G_2 &= 10H-5\sum_{i=1}^4E_i-\sum_{i=5}^{10}E_i -4E_{11}.\end{aligned}$$ It is straightforward to check that $(G_i\cdot D_j) = 0$ for $i=1,2$ and $0\leq j \leq 9$. Hence $G_1, G_2 \in \Lambda$. Also, $$G_1^2 = 2; \qquad G_2^2 = -22; \qquad G_1\cdot G_2 = 0.$$ The corresponding quadratic form $$q(n,m) = (nG_1 + mG_2)^2 = 2n^2 - 22m^2$$ has discriminant $-44=-2^2\cdot 11$. Note that this is consistent with the fact that the discriminant of the dual cycle is $$\det \begin{pmatrix} -6&2\\2&-8\end{pmatrix} = 44.$$ It is easy to see that $G_1$ and $G_2$ are linearly independent mod $2$ and hence span a primitive lattice, which must therefore equal $\Lambda$. First we claim that there is no element of $\Lambda$ of square $-2$. This is equivalent to the statement that there is no solution in integers to the equation $n^2 - 11m^2 = -1$, i.e. that the fundamental unit in $\Zee[\sqrt{11}]$ has norm $1$. But clearly if there were an integral solution to $n^2 - 11m^2 = -1$, then since $-11\equiv 1 \bmod 4$, we could write $-1$ as a sum of squares mod $4$, which is impossible. In fact, the fundamental unit in $\Zee[\sqrt{11}]$ is $10 + 3\sqrt{11}$. Thus, if $R$ is the set of roots for $(Y,D)$, then $R=\emptyset$. In particular, any isometry $f$ trivially satisfies: $f(R) = R$. Finally, we claim that there is an isometry $f$ of $H^2(Y; \Zee)$ such that $f([D_i]) = [D_i]$ for all $i$ and $f(\mathcal{C}^+) = \mathcal{C}^+$, but such that $f$ does not preserve the generic ample cone. Note that the unit group $U$ of $\Zee[\sqrt{11}]$ acts as a group of isometries on $\Lambda$, and hence acts as a group of isometries (with $\Q$-coefficients) of the lattice $H^2(Y; \Q) = (\Lambda \otimes \Q) \oplus \bigoplus_i\Q[D_i]$, fixing the classes $[D_i]$. Also, any isometry of $\Lambda$ which is trivial on the discriminant group $\Lambda\spcheck/\Lambda$ extends to an integral isometry of $H^2(Y; \Zee)$ fixing the $[D_i]$. Concretely, the discriminant form $\Lambda\spcheck/\Lambda \cong \Zee/2\Zee \oplus \Zee/22\Zee$. If $\mu = 10 + 3\sqrt{11}$, then it is easy to check that the automorphism of $\Lambda$ corresponding to $\mu^2 = 199+ 60\sqrt{11}$ acts trivially on $\Lambda\spcheck/\Lambda$ and hence defines an isometry $f$ of $H^2(Y; \Zee)$ fixing the $[D_i]$. Then $f$ acts freely on $(\mathcal{C}^+\cap \Lambda)/\Ar^+$, which is just a copy of $\Ar$ (and $f$ acts on it via translation). But the intersection of the generic ample cone with $\Lambda$ has the nontrivial wall $W^{E_{11}}$, so that the intersection cannot be all of $\mathcal{C}^+\cap \Lambda$. It then follows that $f^{\pm1}$ does not preserve the generic ample cone. Explicitly, let $(\hat{Y}, \hat{D})$ be the surface obtained by contracting $E_{11}$ and let $\hat{G}_1 = 4G_1-G_2 = 10H - 3\sum_{i=1}^{10}E_i$ be the pullback of the positive generator of $\Lambda(\hat{Y}, \hat{D})$. Thus $\hat{G}_1$ is nef and big, so that $\hat{G}_1\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$. Clearly $\hat{G}_1\in W^{E_{11}}$. If $A = \displaystyle \begin{pmatrix} a & 11b\\b&a\end{pmatrix}$ is the isometry of $\Lambda$ corresponding to multiplication by the unit $a + b\sqrt{11}$, then $A(G_1) = aG_1 +bG_2$, $A(G_2) = 11bG_1 + aG_2$, and $A( \hat{G}_1) = (4a-11b)G_1 + (4b-a)G_2$. Thus $$E_{11} \cdot A( \hat{G}_1) = (4a-11b) + 4(4b-a) = 5b,$$ hence $E_{11} \cdot A( \hat{G}_1) < 0$ if $b< 0$. Taking $f^{-1}$, which corresponds to $199- 60\sqrt{11}$, we see that $f^{-1}(\hat{G}_1)\notin \overline{\mathcal{A}}_{\text{\rm{gen}}}$. In this example, the rank of $\Lambda$ is $2$ and $R=\emptyset$, but there exist infinitely many $\beta \in \Lambda$ such that $\beta^2 = -2$. The condition $f(R) = R$ is again automatic for every isometry $f$, and reflection about every $\beta \in \Lambda$ with $\beta^2=-2$ is an isometry which preserves $\mathcal{C}^+$ and the classes $[D_i]$ but not the generic ample cone. As in the previous example, let $(\overline{Y}, \overline{D})$ be the anticanonical pair obtained by making $9$ infinitely near blowups starting with the double point of a nodal cubic. Thus $\overline{D} = \overline{D}_0 + \cdots + \overline{D}_9$, where $\overline{D}_0 = 3H -2E_1 -\sum_{i=2}^9E_i$, $\overline{D}_i = E_i - E_{i+1}$, $1\leq i \leq 8$, and $\overline{D}_9 = E_9$. Make two more blowups, one at a point $p_{10}$ on $\overline{D}_9$, and one at a point $p_{11}$ on $\overline{D}_0$. This yields an anticanonical pair $(Y,D)$ with $D_0 = 3H -2E_1 -\sum_{i=2}^9E_i-E_{11}$ and $D_i = E_i - E_{i+1}$, $1\leq i\leq 9$. Thus $$(-d_0, \dots, -d_9) = (4,2,2,2,2,2,2,2,2,2),$$ i.e.  $D$ is of type $\displaystyle \begin{pmatrix} 4\\9\end{pmatrix}$, with dual cycle $\displaystyle \begin{pmatrix} 12\\1\end{pmatrix}$ in the notation of [@FriedmanMiranda]. Set $$\begin{aligned} G_1 &= 10H - 3\sum_{i=1}^{10}E_i ;\\ G_2 &= 3H- \sum_{i=1}^{10}E_i +E_{11}.\end{aligned}$$ It is straightforward to check that $(G_i\cdot D_j) = 0$ for $i=1,2$ and $0\leq j \leq 9$. Hence $G_1, G_2 \in \Lambda$. Also, $$G_1^2 = 10; \qquad G_2^2 = -2 ; \qquad G_1\cdot G_2 = 0.$$ The corresponding quadratic form $$q(n,m) = (nG_1 + mG_2)^2 = 10n^2 - 2m^2$$ has discriminant $-20=-2^2\cdot 5$. Note that this is consistent with the fact that the discriminant of the dual cycle is $$\det \begin{pmatrix} -12&2\\2&-2\end{pmatrix} = 20.$$ It is easy to see that $G_1$ and $G_2$ are linearly independent mod $2$ and hence span a primitive lattice, which must therefore equal $\Lambda$. To give a partial description of $\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda$, note that (as for $\hat{G}_1$ in the previous example) $G_1$ is the pullback to $Y$ of a positive generator for $\Lambda(\hat{Y}, \hat{D})$, where $\hat{Y}$ denotes the surface obtained by contracting $E_{11}$. Thus $G_1$ is nef and big, so that $G_1\in \overline{\mathcal{A}}_{\text{\rm{gen}}}$ and also $G_1\in W^{E_{11}}$. Hence $$\mathcal{C}^+\cap \Lambda = \{nG_1+ mG_2: 5n^2 - m^2 > 0, n>0\},$$ i.e. $n>0$ and $-n\sqrt{5} < m < n\sqrt{5}$. The condition $E_{11} \cdot (nG_1+ mG_2) \geq 0$ gives $m\leq 0$. To get a second inequality on $n$ and $m$, let $$E' = 5H - 4E_{11} - \sum_{i=1}^{10}E_i.$$ Then $(E')^2 = E' \cdot K_Y = -1$, and $H\cdot E' > 0$. Hence $E'$ is effective. (In fact one can show that $E'$ is generically the class of an exceptional curve.) Thus, for all $nG_1+ mG_2 \in \overline{\mathcal{A}}_{\text{\rm{gen}}}$, $$E' \cdot (nG_1+ mG_2) = 20n + 9m \geq 0,$$ hence $$\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda \subseteq \{nG_1 + mG_2: n > 0, -{\textstyle\frac{20}{9}}n \leq m \leq 0.\}.$$ Next we describe the classes $\beta\in \Lambda$ with $\beta^2 = -2$. The element $\beta = aG_1 + bG_2\in \Lambda$ satisfies $\beta^2 = -2$ $\iff$ $5a^2 -b^2 =-1$, i.e. $\iff$ $b+ a\sqrt{5}$ is a unit in the (non-integrally closed) ring $\Zee[\sqrt{5}]$. For example, the class $G_2$ corresponds to $1$; as we have seen, the wall $W^{G_2} = W^{E_{11}}$. The fundamental unit in $\Zee[\sqrt{5}]$ is easily checked to be $9 + 4\sqrt{5}$. However, since we are only concerned with walls which are rays in the fourth quadrant $\{(nG_1+ mG_2): n > 0, m< 0\}$, we shall consider instead $\pm(9-4\sqrt{5})$, and shall choose the sign corresponding to $\beta = 4G_1 - 9G_2$. Note that $$\beta\cdot (nG_1+ mG_2) = 40n + 18m = 0 \iff E'\cdot (nG_1+ mG_2) = 0.$$ Hence $W^\beta = W^{E'}$. Moreover, for every $\gamma\in \Lambda$ such that $\gamma^2 = -2$ and such that the wall $W^\gamma$ passes through the fourth quadrant, either $W^\gamma =W^\beta$ or the corresponding ray $W^\gamma$ lies below $W^\beta$. Thus, for every $\gamma \in \Lambda$ with $\gamma^2 = -2$, $r_\gamma$ does not preserve $\overline{\mathcal{A}}_{\text{\rm{gen}}} \cap \Lambda$. Hence $R=\emptyset$. Note that, aside from the isometries $r_\beta$, where $\beta^2 = -2$, one can also construct isometries of infinite order preserving $\mathcal{C}^+$ and the classes $[D_i]$ which do not fix preserve $\overline{\mathcal{A}}_{\text{\rm{gen}}}$ using multiplication by fundamental units in $\Zee[\sqrt{5}]$, as in the previous example. The exceptional curve $E'$ used in the above example is part of a general series of such. For $n\geq 0$, let $Y$ be the blowup of $\Pee^2$ at $2n+1$ points $p_0, \dots, p_{2n}$, with corresponding exceptional curves $E_0, \dots, E_{2n}$, and consider the divisor $$A= nH - (n-1)E_0 - \sum_{i=1}^{2n}E_i.$$ Then $A^2 = A\cdot K_Y = -1$, and it is easy to see that there exist $p_0, \dots, p_{2n}$ such that $A$ is the class of an exceptional curve. In fact, if $\mathbb{F}_1$ is the blowup of $\Pee^2$ at $p_0$, then $\Sigma = nH - (n-1)E_0$ is very ample on $\mathbb{F}_1$ and, for an anticanonical divisor $D\in |-K_{\mathbb{F}_1}| = |3H-E_0|$, $\Sigma \cdot D = 2n+1$. From this it is easy to see that we can choose the points $p_1, \dots, p_{2n}$ to lie on the image of $D$ in $\Pee^2$, and hence we can arrange the blowup $Y$ to have (for example) an irreducible anticanonical nodal curve. [99]{} N. Bourbaki, *Groupes et Algèbres de Lie*, Chap. 4, 5, et 6, Masson, Paris, 1981. H.S.M. Coxeter, *Regular Polytopes*, Third Edition, New York, Dover, 1973. M. Demazure, *Surfaces de Del Pezzo II, III, IV, V*, in *Séminaire sur les Singularités des Surfaces*, M. Demazure, H. Pinkham and B. Teissier (eds.), Lecture Note in Mathematics **777**, Berlin Heidelberg New York, Springer-Verlag, 1980, 23–69. P. Du Val, *On isolated singularities of surfaces which do not affect the conditions of adjunction*, Math. Proc. Cambridge Philos. Soc. **30** (1934), Part I: 453–459, Part II: 460-465, Part III: 483–491. P. Du Val, *On the Kantor group of a set of points in a plane*, Proc. Lond. Math. Soc. (2) **42** (1937), 18–51. R. Friedman, *Linear systems on anticanonical pairs*, Appendix to a paper of N.I. Shepherd-Barron, in *The Birational Geometry of Degenerations*, R. Friedman and D. Morrison (eds.), Progress in Mathematics **29**, Birkhäuser, Boston, 1983, 162–171. R. Friedman, *The mixed Hodge structure of an open variety*, manuscript (1984). R. Friedman and R. Miranda, *Smoothing cusp singularities of small length*, Math. Ann. **263** (1983), 185–212. R. Friedman and J.W. Morgan, *On the diffeomorphism types of certain algebraic surfaces I*, J. Differential Geom. **27** (1988), 297–369. R. Friedman and F. Scattone, *Type III degenerations of $K3$ surfaces*, Invent. Math. **83** (1986), 1–39. M. Gross, P. Hacking, and S. Keel, *Moduli of surfaces with an anti-canonical cycle*, arXiv:1211.6367. E. Looijenga, *Invariant theory for generalized root system*, Invent. Math. **61** (1980), 1–32. E. Looijenga, *Rational surfaces with an anti-canonical cycle*, Ann. of Math. (2) **114** (1981), 267–322. Y. Manin, *Cubic forms. Algebra, geometry, arithmetic* (Second edition), Translated from the Russian by M. Hazewinkel. North-Holland, Amsterdam, 1986. R. Miranda and U. Persson, *On extremal rational elliptic surfaces*, Math. Z. **193** (1986), 537–558.

--- abstract: 'Secure communication over a wiretap channel is investigated, in which an active adversary modifies the state of the channel and the legitimate transmitter has the opportunity to sense and learn the adversary’s actions. The adversary has the ability to switch the channel state and observe the corresponding output at every channel use while the encoder has *causal* access to observations that depend on the adversary’s actions. A joint learning/transmission scheme is developed in which the legitimate users learn and adapt to the adversary’s actions. For some channel models, it is shown that the achievable rates, defined precisely for the problem, are arbitrarily close to those obtained *with hindsight*, had the transmitter known the actions ahead of time. This initial study suggests that there is much to exploit and gain in physical-layer security by learning the adversary, e.g., monitoring the environment.' author: - 'Mehrdad Tahmasbi, Matthieu R. Bloch, and Aylin Yener[^1]' bibliography: - 'active-wiretap.bib' title: 'Learning an Adversary’s Actions for Secret Communication' --- Introduction {#sec:introduction} ============ The seminal papers of Wyner [@Wyner1975] and Csiszár and Körner [@Csiszar1978] on the wiretap channel have provided the foundation for advances in information-theoretic secrecy. In more recent past, primarily motivated by the advent of wireless networks, a number of new secure communication models have been studied leading to new design insights [@yener2015wireless]. An example is the realization that secrecy capacity in wireless channels with fading could be positive even with an eavesdropper obtaining a higher than the legitimate receiver [@Bloch2008c; @Liang2008a]. Another example is the merits of introducing judicious interference, i.e., cooperative jamming [@Tekin2008], and its ability to increase secrecy rates in multi-terminal settings [@Tekin2008; @he2014providing; @He2008a; @Pierrot2011a; @ElGamal2013]. Yet another example is the ability to network with entities even if they are untrusted [@He2010b]. In all these initial studies, one critical assumption is that the eavesdropper’s channel is completely or partially known to the legitimate parties. Another crucial assumption is that it is a purely passive observer, is unable to make strategic decisions, or tamper with the channels in any way. More recently, several studies have aimed at removing these assumptions or introducing new models that account for more powerful attacks. To this end, a first direction has addressed the wiretap model in which the eavesdropper’s channel is completely unknown and can be varying in each channel use. In this set up, it has been shown in [@He2014a] that utilizing multiple antennas are useful in providing secrecy, albeit with reduction in degrees of freedom as compared to the other extreme of completely known channels. A model in which the adversary can modify its channel based on overheard signals has been addressed in the specific setting of a two-way wiretap channel, utilizing cooperative jamming to counter the attack [@He2011]. Additionally, the extensions of models to active adversaries that can also jam the channel has been captured with arbitrarily-varying wiretap channel models, in which both main and eavesdropper’s channels depend on states under complete control of the adversary. Secrecy rates for arbitrarily-varying wiretap channels have been studied for point-to-point channels [@MolavianJazi2009; @Schaefer2015; @Goldfeld2016a] and multiple-access channels [@Chou2017], leading to characterizations of situations in which secure communication is possible. Another line of work has been towards addressing passive adversaries with strategic capabilities in their monitoring of signals. These efforts build on the model known as wiretap channel Type II [@ozarow1984wire], in which the main channel is noiseless and the eavesdropper has the ability to observe only a subset of the transmitted codeword bits with known size, but can choose the subset it taps. This model imposes a more stringent secrecy constraint, requiring a universal guarantee againt any choice of observed subset. Like [@ozarow1984wire], follow-up works that have generalized the wiretap II model beyond a noiseless main channel [@Nafea2015; @Goldfeld2016c] have demonstrated that, the impact of the adversary’s strategic ability to choose the observed subset is no worse than random erasures of a subset of the same size. More recently, the secrecy capacity of a model that unifies the wiretap and wiretap II models has been established in [@nafea2016new]. Some multi-terminal extensions of this model have been also been studied [@nafea2016newmult; @nafea2017newmod; @Nafea2019]. Despite these successes, application of information-theoretic security in practical systems is yet to take place. This is in large part because the models to date include assumptions that are either extremely optimistic or perhaps overly pessimistic in regards to what can be known about the adversary. The rationale behind the present paper is that there might be a middle ground to develop adversarial yet realistic models. More precisely, we suggest that, although an adversary may potentially control communication channels, its actions are likely to come at a cost, i.e., the modification may induce some physical effect in the environment that can be detected by other parties. Thus, legitimate parties may have the ability to *learn* the adversary’s actions and accordingly adapt their coding scheme. This idea is also motivated by several studies that have investigated the role of feedback for reliable communication over arbitrary-varying channels [@shulman2003communication; @lomnitz2013universal; @eswaran2010zero] and have shown that the empirical capacity of an arbitrary-varying channel can be achieved with negligible feedback. As a first step towards integrating learning into information-theoretic secrecy models, we study here a wiretap channel model in which an active adversary is able to attack the signals on the main channel by selecting one of two main channels at each channel use. Unlike previously studied models, we allow the legitimate transmitter to monitor the channel and *causally* receive a signal correlated with the adversary’s observations. This consequently allows the legitimate parties to simultaneously “explore” the adversary’s behavior and “exploit” it for providing secrecy. More concretely, our coding scheme chains the transmission of successive sub-blocks to learn the adversary’s actions in past sub-blocks and causally generate secret keys from past observations. Secure communication is achieved by superposing the transmission of uniform random bits and one-time-padded message bits through a suitably generalized layered-secrecy coding scheme [@zou2013layered]. To ensure reliability in the presence of an attacker actively tampering with the main channel, which was not considered in our preliminary results [@tahmasbi2017learning], our coding scheme also leverages a universal list-decoder chained over sub-blocks. We emphasize that key generation is the crucial building block that enables learning for secrecy, by allowing our coding scheme to *defer* the decisions as to which bits are secret until *after* the adversary’s actions have been learned. We also point out that the greater flexibility offered by key generation compared to direct wiretap coding can be traced back to earlier works but for completely different models and without any connection to learning. For instance, key generation has been used [@Bloch2008c; @Gungor2013] in fading channels as a means to buffer secret keys and *provision* secrecy. Perhaps surprisingly, we show that the legitimate parties achieve the secrecy rates that they would have obtained *with hindsight*, had they known the attacker’s actions non-causally. This result is conceptually similar to those that exist in the context of multi-arm bandit problems [@Bubeck2012]: without knowing the adversary’s actions a priori, one can simultaneously exploit and explore to develop an asymptotically optimal strategy. This connection is not fortuitous, as our proof extends ideas laid out in the context of universal channel and source coding [@merhav1998universal; @lomnitz2013universal] that make explicit use of reinforcement learning. In particular, our definition of rate is similar to [@lomnitz2013universal], in which the number of bits required to be decoded correctly is only specified at the end of the transmission, and our converse proof follows the same approach as in [@lomnitz2013universal]. [[A direct comparison of our result with previous works is not entirely fair since our model is different. Nevertheless, it is perhaps useful to explicitly highlight the similarities and differences with the most related prior works. In particular, two major characteristics of our model are to consider some form of channel state information and an adversarial state. The usefulness of channel state information at the transmitter to increase secrecy rates over wiretap channels has been extensively analyzed both in the non-causal [@Chen2008; @Khisti2011] and causal case [@Chia2012] but only under the assumption that the channel state follows a known distribution. In contrast, in our model, the state is adversarial and need not have a well-defined distribution. Adversarial arbitrarily-varying wiretap channel models have also been explored [@MolavianJazi2009; @Schaefer2015; @Goldfeld2016a] but in situations without state information. Consequently, by nature, known results establish worst case rates that are only positive under certain assumptions, such as the type constraint analyzed in [@Goldfeld2016a].]{}]{} The remainder of the paper is organized as follows. In Section \[sec:probl-form-main\], we introduce the model under investigation. In Section \[sec:achievability-proof\], which constitutes the core of our contribution, we develop an achievability proof based on the coding scheme outlined earlier. In Section \[sec:converse-proof\], we establish a converse that matches our achievability. In Section \[sec:discussion\], we conclude with a discussion of natural extensions of our model. Problem Formulation and Main Results {#sec:probl-form-main} ==================================== Notation -------- We denote random variables by uppercase letters (e.g., $X$), their realizations by lowercase letters (e.g., $x$), sets by calligraphic letters (e.g., ${{\mathcal{X}}}$), and vectors by bold face letters (e.g., $\mathbf{x}$). For $\mathbf{x} = (x_1, \cdots, x_n) \in {{\mathcal{X}}}^n$ and $a\in {{\mathcal{X}}}$, let $N(\mathbf{x}|a){\ensuremath{\triangleq}}|\{i: x_i = a\}|$. For $\mathbf{x}\in\{0, 1\}^n$, let ${\ensuremath{\textnormal{wt}(\mathbf{x})}} {\ensuremath{\triangleq}}N(\mathbf{x}|1)$ and $\alpha(\mathbf{x}){\ensuremath{\triangleq}}\frac{{\ensuremath{\textnormal{wt}(\mathbf{x})}}}{n}$. If $P_X$ is a over ${{\mathcal{X}}}$, let ${{\mathcal{T}}}_{P_X} {\ensuremath{\triangleq}}\{\mathbf{x} \in{{\mathcal{X}}}^n: \text{for all } a \in {{\mathcal{X}}}:~N(\mathbf{x}|a) = P(a) n\}$. [^2] We denote by ${{\mathcal{P}}}_n({{\mathcal{X}}})$ the set of all $P_X$ for which ${{\mathcal{T}}}_{P_X} \neq \emptyset$ and by ${{\mathcal{P}}}_n({{\mathcal{X}}}|{{\mathcal{Y}}})$ the set of all conditional $P_{X|Y}$ for which there exists a joint $P_{XY}$ such that $P_{X|Y}= \frac{P_{XY}}{P_Y}$ and ${{\mathcal{T}}}_{P_{XY}} \neq 0$. For $\mathbf{x} \in {{\mathcal{X}}}^n$ and a conditional $P_{Y|X}$, we also define ${{\mathcal{T}}}_{P_{Y|X}}(\mathbf{x}){\ensuremath{\triangleq}}\{\mathbf{y}\in {{\mathcal{Y}}}^n:\text{for all }a\in {{\mathcal{X}}}, b\in {{\mathcal{Y}}}: N(\mathbf{x}, \mathbf{y}|a, b) = P_{Y|X}(b|a) N(\mathbf{x}|a)\}$. For three discrete random variables $(X, Y, Z)$ with joint $P_{XYZ}$, we define $$\begin{aligned} P_{X|YZ} \circ P_{Z} &{\ensuremath{\triangleq}}\sum_z P_{X|YZ=z}P_Z(z) {\ensuremath{\triangleq}}P_{X|Y}\\ P_{Z|Y} \times P_{X|YZ} &{\ensuremath{\triangleq}}P_{Z|Y} P_{X|YZ} {\ensuremath{\triangleq}}P_{XZ|Y}\\ I(P_X, P_{Y|X}) {\ensuremath{\triangleq}}I(P_{XY}) &{\ensuremath{\triangleq}}{{{\mathbb{I}}\!\left(X;Y\right)}}.\end{aligned}$$ For two sequences $\mathbf{x} \in {{\mathcal{X}}}^n$ and $\mathbf{y} \in {{\mathcal{Y}}}^n$ such that $(\mathbf{x}, \mathbf{y}) \in {{\mathcal{T}}}_{P_{XY}}$, we define $$\begin{aligned} I(\mathbf{x}\wedge \mathbf{y}) {\ensuremath{\triangleq}}I(P_{XY}).\end{aligned}$$ For two integers $a$ and $b$ such that $a{\leqslant}b$, we denote the set $\{a, a+1, \cdots, b - 1, b\}$ by ${\ensuremath{\llbracket{a},{b}\rrbracket}}$. If $a>b$, then ${\ensuremath{\llbracket{a},{b}\rrbracket}}{\ensuremath{\triangleq}}\emptyset$. Throughout the paper, we measure the information in bits and $\log(\cdot)$ should be understood to be base $2$; we use $\ln(\cdot)$ for the logarithm base $e$. \[sec:problem\_form\] ![Problem setup[]{data-label="fig:problem-setup"}](channelmodel.pdf) Problem Formulation and Main Results {#problem-formulation-and-main-results} ------------------------------------ We consider the channel model illustrated in Figure \[fig:problem-setup\], in which a transmitter wishes to communicate securely to a receiver over $N$ channel uses of an arbitrarily varying wiretap $({{\mathcal{X}}}\times {{\mathcal{S}}}, W_{YZ|XS}, {{\mathcal{Y}}}, {{\mathcal{Z}}})$. We assume that the capacity *with* common randomness of the arbitrarily varying channel $({{\mathcal{X}}}\times {{\mathcal{S}}}, W_{Y|XS}, {{\mathcal{Y}}})$ is positive, otherwise it will be impossible to send any information reliably over this channel. The terminals corresponding to $Y$ and $Z$ are in control of the legitimate receiver and the adversary, respectively. The adversary is allowed to actively and arbitrarily choose the state of the channel $S$ at each channel use. The transmitter and the receiver may share secret randomness prior to the transmission whose amount will be made precise. For simplicity, we suppose ${{\mathcal{S}}}=\{0, 1\}$; however, generalizing our results to arbitrary finite ${{\mathcal{S}}}$ is fairly direct, and we discuss this extension in Section \[sec:discussion\]. We assume that the transmitter is able to monitor the effect of the adversary’s actions, which we model as *strictly causal* observations at the output of an arbitrarily varying $(\mathcal{X}\times {{\mathcal{S}}}, W_{\overline{X}|XS}, \overline{\mathcal{X}})$ controlled by the same states as the main wiretap channel. Our only assumption for this channel is that $W_{\overline{X}|XS=0}\neq W_{\overline{X}|XS=1}$, as otherwise the model reduces to an arbitrarily varying wiretap channel [@Bjelakovic2012]. We assume that all channel outputs are conditionally independent given the input and that the statistical description of the channels are known to all parties. As discussed in Remark \[rm:assumptions\] below, we rely on these assumptions for our converse proof, but they play no crucial role in the achievability. The transmitted sequence is denoted by $\mathbf{X} {\ensuremath{\triangleq}}(X_1, \cdots, X_N)$, while the corresponding observations of the receiver and the adversary are denoted by $\mathbf{Y} {\ensuremath{\triangleq}}(Y_1, \cdots, Y_N)$ and $\mathbf{Z} {\ensuremath{\triangleq}}(Z_1, \cdots, Z_N)$, respectively. The monitored sequence is $\mathbf{\overline{X}} {\ensuremath{\triangleq}}(\overline{X}_1, \cdots, \overline{X}_N)$. Formally, a code for this channel model operates as follows. Unlike traditional wiretap channel models, the number of message bits is unknown at the beginning of transmission and potentially depends on the adversary’s actions. Therefore, it is convenient to assume that the transmitter has access to $K$ uniformly distributed bits $\mathbf{W} {\ensuremath{\triangleq}}(W_1, \cdots, W_K)$, and that only the first $\psi$ bits will be transmitted[^3]. Both the encoder and the decoder also have access to a secret common randomness source $Q$ distributed according to $P_{Q}$ over ${{\mathcal{Q}}}$. The encoder consists of $N$ possibly stochastic functions $\mathbf{f} = (f_1, \cdots, f_N)$ where $f_i:\overline{\mathcal{X}}^{i-1} \times\{0,1\}^K \times {{\mathcal{Q}}}\to \mathcal{X}$ outputs a symbol for the transmission over the channel. The total number of transmitted bits $\psi:\overline{\mathcal{X}}^N\times {{\mathcal{Q}}}\to{\ensuremath{\llbracket{0},{K}\rrbracket}}$ is a function of the transmitter’s observations and is determined after the $N^{\text{th}}$ transmission. The decoder is a function $\phi: \mathcal{Y}^N\times {{\mathcal{Q}}}\to \{0,1\}^K$, which allows the receiver to form an estimate $(\widehat{W}_1, \cdots, \widehat{W}_K) {\ensuremath{\triangleq}}\phi(\mathbf{Y}, Q)$ of the transmitted bits. Since the channel is varying according to the adversary’s actions, the receiver is *not* required to reliably decode all bits. We assume that there exists a function $\widehat{\psi}: \mathcal{Y}^N\times {{\mathcal{Q}}}\to {\ensuremath{\llbracket{0},{K}\rrbracket}}$ that estimates the number of bits actually transmitted. The quintuple $(\mathbf{f}, \phi, \psi, \widehat{\psi}, Q)$ defines an $(N, K)$ code ${{\mathcal{C}}}$, and the functions $\mathbf{f}$, $\mathbf{\phi}$, $\psi$, and $\widehat{\psi}$ are assumed to be publicly known. For all $\mathbf{s}\in{{\mathcal{S}}}^N$, reliability is measured with a probability of error defined as $$\begin{aligned} P_e({{\mathcal{C}}}|\mathbf{s}){\ensuremath{\triangleq}}{{\mathbb{P}_{}}{\left(\widehat{\psi}(\mathbf{Y}, Q)\neq \psi(\overline{\mathbf{X}}, Q)\text{ or }\exists k \in {\ensuremath{\llbracket{1},{\widehat{\psi}(\mathbf{Y}, Q)}\rrbracket}}:~\widehat{W}_k \neq W_k \big |\mathbf{s}\right)}}.\end{aligned}$$ Secrecy is measured in terms of the average mutual information between the message bits $\mathbf{W}$ and the observations $\mathbf{Z}$ given $\mathbf{s}$ as $$\begin{aligned} \mathtt{S}({{\mathcal{C}}}|\mathbf{s}){\ensuremath{\triangleq}}{{{\mathbb{I}}\!\left(\mathbf{W};\mathbf{Z}|\mathbf{s}\right)}}.\end{aligned}$$ The rate of the code is a function of the adversary’s actions and is a random variable defined as $ \frac{\psi(\overline{\mathbf{X}}, Q)}{n}$. Furthermore, for a fixed value of common randomness $q\in{{\mathcal{Q}}}$, we define $P_e({{\mathcal{C}}}|\mathbf{s}, q)$ and $\mathtt{S}({{\mathcal{C}}}|\mathbf{s}, q)$ analogously using probability distributions conditioned on $Q=q$. For a fixed sequence $\{\mathbf{s}_N\in{{\mathcal{S}}}^N\}_{N{\geqslant}1}$ of adversarial actions, a sequence of $(N, K_N)$ codes $\{\mathcal{C}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ achieves a rate $R$, if and only if, $$\begin{aligned} \lim_{N\to\infty} \frac{{{\mathbb{H}}\!\left(Q_N\right)}}{N} &= 0,\\ \lim_{N\to \infty}\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N)&= 0,\displaybreak[0]\\ \lim_{N\to \infty} P_e({{\mathcal{C}}}_N|\mathbf{s}_N) &= 0,\end{aligned}$$ and $$\begin{aligned} &\lim_{N\to \infty} {{\mathbb{P}_{}}{\left(\frac{\psi_N(\mathbf{\overline{X}}, Q_N)}{N} {\leqslant}R\right)}} = 0.\label{eq:rate-constraint}\end{aligned}$$ \[rm:rate-details\] The number of secret bits transmitted depends on the noisy observations $\overline{\mathbf{X}}$ of the adversary’s actions and on the common randomness $Q_N$, both of which are random variables. Consequently, the number of secret bits is itself a random variable and our notion of achievable rate in  only requires a rate $R$ to be achieved with high probability. \[rm:assumptions\] The technical assumptions behind our model have concrete operational significance. Since $\psi$ is publicly known, no secrecy is conveyed through the *number* of secret bits. Since only the transmitter monitors the environment, the receiver does not benefit from another channel observation that could potentially increase its reliability. Finally, since channel outputs are conditionally independent given the input, the transmitter only obtains information about the adversary’s actions and not about the receiver or adversary’s [[observations]{}]{}. These assumptions are not crucial in our achievability proof but they are needed in the converse. For $\alpha\in[0, 1]$, let $P_S$ be Bernoulli$(\alpha)$ over ${{\mathcal{S}}}$ and $C(\alpha) {\ensuremath{\triangleq}}\sup_{P_{XU}} {{\left({{{\mathbb{I}}\!\left(U;Y\right)}} - {{{\mathbb{I}}\!\left(U;Z|S\right)}}\right)}}$ where the random variables $(U, X,S, Y, Z)$ have joint $P_{UX}P_SW_{YZ|XS}$ . [[Since $\sup_{P_{XU}} {{\left({{{\mathbb{I}}\!\left(U;Y\right)}} - {{{\mathbb{I}}\!\left(U;Z|S\right)}}\right)}}$ is a special case of the right hand side of [@Goldfeld2016a Eq. (50)], by [@Goldfeld2016a Theorem 1], it is enough to consider $U$ with the support of size at most $|{{\mathcal{X}}}|$.]{}]{} Our main results are as follows. \[thm:achv\] For any $\zeta>0$, there exists a sequence of $(N, K_N) $ codes $\{\mathcal{C}_N\}_{N{\geqslant}1}$ that achieves the rate $C(\alpha) - \zeta$ for all $\alpha\in]0,1[$ and every sequence $\{\mathbf{s}_N \in \{0, 1\}^N\}_{N{\geqslant}1}$ with $\lim_{N\to\infty}\alpha(\mathbf{s}_N) = \alpha$. \[thm:converse\] If a sequence of $(N, K_N)$ codes $\{\mathcal{C}_N\}_{N{\geqslant}1}$ achieves a rate $R$ for *all* sequences $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{n\to \infty}\alpha(\mathbf{s}_N) = \alpha$, then $R {\leqslant}C(\alpha)$. Before we prove Theorem \[thm:achv\] in Section \[sec:achievability-proof\] and Theorem \[thm:converse\] in Section \[sec:converse-proof\], we first make a few important remarks regarding our results. Note that the main contribution of Theorem \[thm:achv\] is in guaranteeing the existence of codes having good performance for *all* choices of $\mathbf{s}$. If the weight of the actions sequence ${\ensuremath{\textnormal{wt}(\mathbf{s}_N)}}$ were known, our model would reduce to a channel with type-constrained states as studied in [@Goldfeld2016a; @nafea2016new]. In this case, the transmitter and the receiver would know ahead of time the optimal number of secret bits that could be transmitted, and proofs would follow from more standard techniques. For instance, reliability in [@Goldfeld2016a] is established using the random coding technique for channels (e.g., [@csiszar2011information Chapter 12]), which guarantees low probability of error at the decoder for a class of channel states. Similarly, secrecy is derived in [@Goldfeld2016a; @nafea2016new] by proving the existence of a universal scheme for all adversary’s actions with a certain type. While [@Goldfeld2016a] and [@nafea2016new] have different approaches to prove universal secrecy, both papers show that a random code fails to provide secrecy for a specific state with doubly-exponentially decreasing probability, which then allows on to use the union bound and show the existence of a code secure for all states with a fixed type. Without constraints on the adversary’s actions, one might wonder why the adversary would not always choose the “best" action over all $N$ channel uses, so that the problem would reduce to a traditional wiretap channel. Our modeling allows us to remove all assumptions regarding the rationality or possible limitations of the adversary that the legitimate parties could be *unaware* of. Our scheme performs optimally as if these constraints were known a priori. In both achievability and converse, if $\lim_{N\to \infty}\alpha(\mathbf{s}_N)$ does not exist, we need to consider the sub-sequence of $\{\mathbf{s}_N\}_{N{\geqslant}1}$ such as $\{\mathbf{s}_{N_k}\}_{k{\geqslant}1}$ so that $\alpha({\mathbf{s}_{N_k}})$ is convergent, which we know always exists, and $$\begin{aligned} C\left(\lim_{k\to \infty}{\alpha({\mathbf{s}_{N_k}})}\right)\end{aligned}$$ is minimized. This subtlety is a consequence of our asymptotic formulation of the rate, and one should note that our achievability scheme provides guarantees at finite length. Although we assume that the action sequence $\mathbf{s}$ is arbitrary, it cannot depend on the adversary’s observations $\mathbf{Z}$. In such case, the adversary could potentially choose strategies that introduce memory in the channel, which would require different coding schemes. We conclude this section by illustrating the results for an example. Suppose that when $s=0$, the main channel and the adversary’s channel are with flipping probabilities $0.1$ and $0.4$, respectively, and when $s=1$, the main and the adversary’s channels are with flipping probabilities $0.4$ and $0.1$. Fig. \[fig:example\] is the plot of $C(\alpha)$ in terms of $\alpha$ for this channel. We next compare our results with [@Goldfeld2016a] in Fig. \[fig:example2\] when there is a constraint $\alpha {\leqslant}0.2$. Using the scheme in [@Goldfeld2016a], we have to operate at the worst case, which corresponds to $\alpha = 0.2$. However, in our proposed scheme, we can operate at $C(\alpha)$ for all $\alpha$. ![$C(\alpha)$ in terms of $\alpha$ for []{data-label="fig:example"}](learning-example.pdf) ![ Comparing our results with those of [@Goldfeld2016a] for []{data-label="fig:example2"}](learning-example2.pdf) Proof of Theorem \[thm:achv\]: Achievability Scheme {#sec:achievability-proof} =================================================== ![Illustration of oracle-assisted coding scheme. In every sub-block $b$, uniformly random bits $\mathbf{A}^b$ (represented as vertical bars) are transmitted using a layered secrecy code. At the end of sub-block $b$ transmission, an oracle indicates to the legitimate parties which bits were unreliable (white stripes) and secret (gray stripes). Unreliable bits are repeated in the next sub-block while reliable secret bits are used to one-time pad a message $\mathbf{W}^{b+1}$ in sub-block $b+1$. Additional uniformly random bits $\mathbf{L}^{b+1}$ are appended to form $\mathbf{A}^{b+1}$.[]{data-label="fig:coding-scheme"}](coding-scheme){width="0.6\linewidth"} We begin the achievability proof by sketching our proposed coding scheme. The transmitter and the receiver first agree on an input distribution $P_X$ before the transmission. Our coding scheme, which is illustrated in Figure \[fig:coding-scheme\], is an $(N, K)$ coding scheme that consists of a transmission over $B$ sub-blocks of length $n$ and one terminal sub-block whose length will be specified later. We denote all quantities corresponding to the sub-block $b$ by a superscript $b$; specifically, $\mathbf{s}^b$ denotes the channel states sequence in sub-block $b$. We initially assume that “an oracle” shall causally provide partial information $\alpha(\mathbf{s}^b)$ of the channel states $\mathbf{s}^b$ to both the transmitter and the receiver at the *end* of sub-block $b$. In every sub-block $b$, the transmitter prepares $k$ uniform random bits $\mathbf{A}^b = (A_1^b, \cdots, A_k^b)$ that do not convey information on their own. It encodes $\mathbf{A}^b$ to a sequence $\mathbf{X}^b$ through a random encoder generated using $P_X$ from the common randomness shared between the transmitter and the receiver. By layer secrecy results [@zou2013layered], the adversary obtains negligible information about the first $\approx k - (1-\alpha(\mathbf{s}^b))I(P_X, W_{Z|XS=0}) + \alpha(\mathbf{s}^b) I(P_X, W_{Z|XS=1})$ bits of $\mathbf{A}^b$. In the next sub-block, the number of secure bits is evaluated using the oracle information, and those bits are used as a key to one-time-pad information bits in the next sub-block $b+1$. The use of uniform random bits and a layered transmission scheme is crucial to enable the extraction of secrecy *with hindsight*. Since the capacity of the main channel depends on the channel states, the decoder outputs a list of possible values for $\mathbf{A}^b$ with a list size determined by $\alpha({\mathbf{s}^b})$. To be able to decode the messages uniquely, the transmitter adds further structure to the transmitted bits by repeating some of the bits of the sub-block $b$ in the sub-block $b+1$. We call such a scheme an $(N, K, n)$ oracle assisted scheme. We then need the following three steps to complete the proof. - In Section \[sec:estimation\], we show how to replace the oracle by an actual noisy estimator of the attacker’s past action sequence without affecting asymptotic performance (Corollary \[cor:no\_oracle\_scheme\]). In particular, if the adversary’s actions change the statistics of any feedback symbol, based on the frequency of the observation of that symbol, we can construct an estimator of $\alpha(\mathbf{s}^b)$. - In Section \[sec:input-dist\], we show how the input distribution underlying the construction of the layered-secrecy code may be adapted from one block to another. In more details, the optimal input distribution depends on the adversary’s action. By resorting to results for adverserial multi-arm bandits, we prove that the transmitter and the receiver could adaptively choose the input distribution for random coding such that its effect on the rate is negligible. - Finally, in Section \[sec:red\_com\_rand\], we finally show how to reduce the randomness in the randomized scheme (Lemma \[lm:cr-reduction\]) to achieve similar performance with a negligible rate of common randomness (Lemma \[lm:cr-reduction\]). We use the standard Ahlswede’s robustification technique to do so. Formal Description and Analysis of the Oracle-Assisted Coding Scheme with Arbitrary Common Randomness {#sec:simp_coding_scheme} ----------------------------------------------------------------------------------------------------- In this section, we first formally describe our *oracle-assisted* coding scheme and then separately analyze its reliability, secrecy, and rate. For all $b\in {\ensuremath{\llbracket{1},{B}\rrbracket}}$ and $\kappa\in {\ensuremath{\llbracket{1},{k}\rrbracket}}$, let $W_\kappa^b$ and $L_\kappa^b$ be random bits uniformly distributed on $\{0, 1\}$; all these random variables are mutually independent. Furthermore, for all $(\kappa, b) \in {\ensuremath{\llbracket{1},{k}\rrbracket}} \times {\ensuremath{\llbracket{1},{B}\rrbracket}}$, we assume that $W_{\kappa}^b$ contains useful information but $L_{\kappa}^b$ is an auxiliary bit. ### Encoding Fix a distribution over ${{\mathcal{X}}}$, $P_X$, and $\zeta > 0$. For $\alpha \in [0, 1]$, define $$\begin{aligned} \mathbb{I}_Y^\alpha &=I{{\left(P_X, (1-\alpha)W_{Y|XS=0} + \alpha W_{Y|XS=1}\right)}} \\ \mathbb{I}_Z^\alpha &= (1-\alpha)I(P_X, W_{Z|XS=0}) + \alpha I(P_X, W_{Z|XS=1})\\ \mathbb{I}_Y^{\max} &= \max_{ \alpha \in[0, 1]}\mathbb{I}_Y^{\alpha}\\ \mathbb{I}_Z^{\max} &= \max(\mathbb{I}_Z^0, \mathbb{I}_Z^1)\\ \mathbb{I}^{\max} &= \max(\mathbb{I}^{\max}_Y, \mathbb{I}^{\max}_Z).\end{aligned}$$ Let both the encoder and the decoder have access to a common sequence of mutually independent random functions $F^1, \cdots, F^B$ from $\{0,1\}^k$ to ${{\mathcal{X}}}^n$ where for all $\mathbf{a}\in\{0,1\}^k$ and $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, $F^b(\mathbf{a})$ is distributed according to $P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}$. In the sub-block $b$, the transmitter prepares $k {\ensuremath{\triangleq}}\lceil {{\left(\mathbb{I}^{\max} + 2\zeta\right)}}n + 1\rceil$ bits $\mathbf{A}^b = (A_1^b, \cdots, A_k^b)$ defined as follows. Let $A_1^0 = A_2^0 = \cdots = A_k^0 {\ensuremath{\triangleq}}0$, $\alpha^b{\ensuremath{\triangleq}}\alpha(\mathbf{s}^b)$, and $$\begin{aligned} m^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b = 1\\ k - \left\lceil{{\left(\mathbb{I}_Z^{\alpha^{b-1}}+\zeta\right)}}n\right\rceil\quad &b\in{\ensuremath{\llbracket{2},{B}\rrbracket}}\end{cases}\label{eq:mb-def}\\ u^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b=1\\k - \left\lfloor{{\left(\mathbb{I}_Y^{\alpha^{b-1}}-2\zeta\right)}}n\right\rfloor\quad &b\in{\ensuremath{\llbracket{2},{B+1}\rrbracket}} \end{cases}.\end{aligned}$$ By Lemma \[lm:universal\_secrecy\_weight\], $m^b$ represents the number of secure bits in the sub-block $b-1$ that can be used as a key in sub-block $b$; by Lemma \[lm:universal\_list\_decode\], $2^{u^b}$ is almost the size of the list required for correct list-decoding in the sub-block $b-1$. We repeat $u^b$ bits of $\mathbf{A}^{b-1}$ in $\mathbf{A}^b$ to form $\mathbf{A}^b$ as $$\begin{aligned} \label{eq:def_s} A_\kappa^b {\ensuremath{\triangleq}}\begin{cases} A_\kappa^{b-1} \quad& \kappa \in {\ensuremath{\llbracket{1},{u^b}\rrbracket}}\\ A_\kappa^{b-1}\oplus W_\kappa^b\quad& \kappa \in {\ensuremath{\llbracket{u^b+1},{m^b}\rrbracket}}\\ L_\kappa^b\quad & \kappa\in{\ensuremath{\llbracket{\max(m^b, u^b)+1},{k}\rrbracket}} \end{cases}.\end{aligned}$$ Note that if $m^b<u^b +1$, the set ${\ensuremath{\llbracket{u^b+1},{m^b}\rrbracket}}$ is empty and no message bit is transmitted in the sub-block $b$. Finally, after transmission of sub-block $B$, we use an code to transmit $\mathbf{A}^{B+1} {\ensuremath{\triangleq}}(A_{1}^{B}, \cdots, A_{u^{B+1}}^B)$. The existence of codes is established in the following lemma. \[lm:avc-code\] Suppose for the $({{\mathcal{X}}}\times \cal S, W_{Y|XS}, {{\mathcal{Y}}})$, we have $\min_{\alpha\in[0, 1]}\mathbb{I}_Y^{\alpha} > 0$. Then, there exists $R>0$, $\xi>0$, and a sequence of codes with common randomness $\{{{\mathcal{C}}}_m^{\text{AVC}}\}_{m{\geqslant}1}$ such that ${{\mathcal{C}}}_m$ transmits a message in ${\ensuremath{\llbracket{1},{\lfloor 2^{mR}\rfloor}\rrbracket}}$ over $m$ channel uses with vanishing probability of error $\epsilon_m^{\text{AVC}} {\leqslant}2^{-\xi m}$. See [@csiszar2011information Lemma 12.10]. In particular, if $m = \Theta(n)$, we can transmit $\mathbf{A}^{B+1}$ by ${{\mathcal{C}}}_m^{\text{AVC}}$. ### Decoding The decoder operates recursively to decode $\mathbf{A}^{B+1}, \mathbf{A}^B, \cdots, \mathbf{A}^1$ in this order. Using the decoder of ${{\mathcal{C}}}_m^{\text{AVC}}$, the receiver first recovers $\mathbf{A}^{B+1}$ as $\widehat{\mathbf{A}}^{B+1}$. Subsequently, for the sub-block $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, we assume that an estimate $\widehat{\mathbf{A}}^{b+1}$ of $\mathbf{A}^{b+1}$ is available at the receiver. The receiver forms a list ${{\mathcal{L}}}^b$ of $\ell^b$ messages $\mathbf{w}$ with highest $I(F^b(\mathbf{w})\wedge \mathbf{Y}^b)$ with $$\begin{aligned} \label{eq:list-size} \log \ell^b {\ensuremath{\triangleq}}k - \left\lceil{{\left(\mathbb{I}_Y^{\alpha^b}-\zeta\right)}}n\right\rceil.\end{aligned}$$ The receiver then seeks a sequence $\mathbf{a} \in {{\mathcal{L}}}^b$ such that for all $\kappa\in{\ensuremath{\llbracket{1},{u^{b+1}}\rrbracket}}$ we have $a_\kappa = \widehat{A}_\kappa^{b+1}$. If there is a unique such sequence, the decoder sets $\widehat{\mathbf{A}}^b {\ensuremath{\triangleq}}\mathbf{a}$; otherwise, it declares an error. The message bits $W_{\kappa}^b$ can finally be decoded from $\widehat{\mathbf{A}}^1, \cdots, \widehat{\mathbf{A}}^B$. ### Reliability Analysis The probability of error is upper-bounded by $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbf{A}^1 \neq \widehat{\mathbf{A}}^1 \text{ or } \cdots \text{ or }\mathbf{A}^{B+1} \neq \widehat{\mathbf{A}}^{B+1}\right)}} &= \sum_{b=1}^{B+1} {{\mathbb{P}_{}}{\left(\mathbf{A}^b \neq \widehat{\mathbf{A}}^b, {{\left(\mathbf{A}^{b+1} = \widehat{\mathbf{A}}^{b+1}, \cdots, \mathbf{A}^{B+1} = \widehat{\mathbf{A}}^{B+1}\right)}} \right)}}\\ &{\leqslant}\sum_{b=1}^{B+1} {{\mathbb{P}_{}}{\left(\mathbf{A}^b \neq \widehat{\mathbf{A}}^b \big |{\mathbf{A}^{b+1} = \widehat{\mathbf{A}}^{b+1} ,\cdots,\mathbf{A}^{B+1} = \widehat{\mathbf{A}}^{B+1}} \right)}}.\end{aligned}$$ For the last term in the sum, we have ${{\mathbb{P}_{}}{\left(\mathbf{A}^{B+1} \neq \widehat{\mathbf{A}}^{B+1}\right)}} {\leqslant}\epsilon_{m}^{\text{AVC}}$ by construction. For the sub-block $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, let $$\begin{aligned} {{\mathcal{A}}}^b &{\ensuremath{\triangleq}}\{\mathbf{A}^b \},\\ {{\mathcal{B}}}^b &{\ensuremath{\triangleq}}\{\mathbf{a} \in \{0,1\}^k: \text{ for all } \kappa\in{\ensuremath{\llbracket{1},{u^{b+1}}\rrbracket}} \text{ we have } a_\kappa= A_\kappa^{b+1}\},\end{aligned}$$ which are the set of the transmitted message in the sub-block $b$, and the set of all messages matching with the bits of the next sub-block, respectively. With these notations, we can write the probability of decoding error in the sub-block $b$ as $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbf{A}^b \neq \widehat{\mathbf{A}}^b \big |{\mathbf{A}^{b+1} = \widehat{\mathbf{A}}^{b+1},\cdots, \mathbf{A}^{B+1} = \widehat{\mathbf{A}}^{B+1}} \right)}} &= {{\mathbb{P}_{}}{\left({{\mathcal{L}}}^b \cap {{\mathcal{B}}}^b \neq {{\mathcal{A}}}^b\right)}}\\\displaybreak[0] &{\leqslant}{{\mathbb{P}_{}}{\left({{\mathcal{A}}}^b \nsubseteq {{\mathcal{L}}}^b \right)}} + {{\mathbb{P}_{}}{\left(({{\mathcal{L}}}^b \cap {{\mathcal{B}}}^b) \setminus {{\mathcal{A}}}^b \neq \emptyset\right)}}\label{eq:dec-error}\end{aligned}$$ The following lemma that we prove in Appendix \[sec:list-dec\] upper-bounds the first term in the above expression by $2^{-\xi n}$ for some $\xi > 0$. \[lm:universal\_list\_decode\] Let $\ell^b$ be a positive integer and $(X, Y, S)$ be three random variables distributed according to $P_{XYS}(x, y, s) = P_X(x)P_S(s)W_{Y|XS}(y|xs)$, in which $P_S = \text{Bernoulli}(\alpha(\mathbf{s}^b))$. Then, $$\begin{aligned} \label{eq:list_bound_avc} {{\mathbb{P}_{}}{\left({{\mathcal{A}}}^b \nsubseteq {{\mathcal{L}}}^b \right)}} {\leqslant}2^{-n\min_{V_{XY|S}} \left[{{{\mathbb{D}}\!\left({V_{XY|S}\VertW_{Y|XS}\times P_X|P_S}\right)}} + {{\color{black}\ell^b}}\left[I(V_{XY|S}\circ P_S) - \frac{1}{n} {{\left(k + \log \frac{e}{{{\color{black}\ell^b}}} \right)}}+ O{{\left(\frac{\log n}{n}\right)}}\right]^+\right]}.\end{aligned}$$ In particular, if holds, there exists $\xi > 0$ independent of $\mathbf{s}^b$ such that $$\begin{aligned} \label{eq:avc_list_asymptotic} {{\mathbb{P}_{}}{\left({{\mathcal{A}}}^b \nsubseteq {{\mathcal{L}}}^b \right)}}{\leqslant}2^{-\xi n}.\end{aligned}$$ To upper-bound the second term on the RHS of , let $\mathbf{0}$ denote the all-zero vector, and note that $$\begin{aligned} {{\mathbb{P}_{}}{\left({{\left({{\mathcal{L}}}^b\cap{{\mathcal{B}}}^b\right)}}\setminus {{\mathcal{A}}}^b\neq \emptyset\right)}} &= \sum_{\mathbf{a}}{{\mathbb{P}_{}}{\left(\mathbf{A}^{b} = \mathbf{a}\right)}}{{\mathbb{P}_{}}{\left({{\left({{\mathcal{L}}}^b\cap {{\mathcal{B}}}^b\right)}} \setminus {{\mathcal{A}}}^b \neq \emptyset \big| \mathbf{A}^{b} = \mathbf{a}\right)}}\\ &\stackrel{(a)}{=} {{\mathbb{P}_{}}{\left({{\left({{\mathcal{L}}}^b\cap {{\mathcal{B}}}^b\right)}} \setminus {{\mathcal{A}}}^b \neq \emptyset \big| \mathbf{A}^{b} = \mathbf{0}\right)}},\end{aligned}$$ where $(a)$ follows since our encoding and decoding processes are symmetric with respect to all messages. Furthermore, when $\mathbf{A}^b = \mathbf{0}$, we have ${{\mathcal{A}}}^{b} = \{\mathbf{0}\}$ and ${{\mathcal{B}}}^b = \{\mathbf{a}:\text{ for all } \kappa\in{\ensuremath{\llbracket{1},{u^{b+1}}\rrbracket}}~a_\kappa = 0\}$. Therefore, by the union bound, we have $$\begin{aligned} {{\mathbb{P}_{}}{\left({{\left({{\mathcal{L}}}^b\cap {{\mathcal{B}}}^b\right)}} \setminus {{\mathcal{A}}}^b \neq \emptyset \big| \mathbf{A}^{b} = \mathbf{0}\right)}} &{\leqslant}\sum_{\widetilde{\mathbf{a}}:\text{ for all } \kappa\in{\ensuremath{\llbracket{1},{u^{b+1}}\rrbracket}}~a_\kappa = 0, \widetilde{\mathbf{a}} \neq 0} {{\mathbb{P}_{}}{\left(\tilde{\mathbf{a}}\in {{\mathcal{L}}}^b\big | \mathbf{A}^b = \mathbf{0}\right)}}\displaybreak[0]\\ &\stackrel{(a)}{{\leqslant}} \frac{\ell^b}{2^{u^{b+1}}}\displaybreak[0]\\ &{\leqslant}\frac{2^{k - \left\lceil{{\left(\mathbb{I}_Y^{\alpha^{b-1}}-\zeta\right)}}n\right\rceil}}{2^{k - \left\lfloor{{\left(\mathbb{I}_Y^{\alpha^{b-1}}-2\zeta\right)}}n\right\rfloor}}\displaybreak[0]\\ &{\leqslant}2^{-\zeta n},\end{aligned}$$ where $(a)$ follows since all $\widetilde{\mathbf{a}}\neq \mathbf{0}$ have an equal probability to be in ${{\mathcal{L}}}^b$. As a result, the total probability of error is upper-bounded by $B{{\left(2^{-n\zeta} + 2^{-n \xi}\right)}} + \epsilon_{m}^{\text{AVC}}$. ### Secrecy Analysis Let $\mathbf{Z}^b$ denote the adversary’s observation in the sub-block $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, $\mathbf{Z}^{B+1}$ denote what the adversary observes during the transmission of $\mathbf{A}^{B+1}$, and $$\begin{aligned} \mathbf{W}^b &{\ensuremath{\triangleq}}(W_\kappa^b:k\in {\ensuremath{\llbracket{1},{k}\rrbracket}}),\displaybreak[0]\label{eq:W}\\ \mathbf{L}^b &{\ensuremath{\triangleq}}(L_k^b:\kappa\in {\ensuremath{\llbracket{\max(m^b, u^b)+1},{k}\rrbracket}})\displaybreak[0]\label{eq:L}\\ \overline{\mathbf{A}}^b &{\ensuremath{\triangleq}}(A_\kappa^b:\kappa\in {\ensuremath{\llbracket{1},{m^{b+1}}\rrbracket}})\displaybreak[0]\label{eq:barA}\\ \widetilde{\mathbf{A}}^b &{\ensuremath{\triangleq}}(A_\kappa^b:\kappa\in {\ensuremath{\llbracket{m^{b+1} + 1},{k}\rrbracket}}).\label{eq:tildeA}\end{aligned}$$ The functional dependence graph illustrating the dependencies introduced by the chaining in  and with the notation in - is shown in Fig. \[fig:fdg\]. ![Functional dependence graph illustrating chaining in  for the variables in -[]{data-label="fig:fdg"}](fdg){width="\linewidth"} We first state a layered secrecy result, which we prove in Appendix \[sec:layer-sec\]. \[lm:universal\_secrecy\_weight\] Let $k{\geqslant}{{\left(\mathbb{I}^{\max} _Z + 2\zeta\right)}}n + 1$. There exists $\xi > 0$ such that for large $n$, $$\begin{aligned} \label{eq:universal_sec} {{\mathbb{P}_{F^b}}{\left(\text{for all }\mathbf{s^b}\in{{\mathcal{S}}}^n,~{{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b\right)}} {\leqslant}2^{-\xi n}\right)}} {\geqslant}1 - 2^{-2^{n\xi}}.\end{aligned}$$ Our next objective is to upper-bound $$\begin{aligned} {{{\mathbb{I}}\!\left(\mathbf{W}^1,\cdots,\mathbf{W}^B;\mathbf{Z}^1,\cdots, \mathbf{Z}^{B+1}|F^1=f^1, \cdots, F^B=f^B\right)}}\end{aligned}$$ when encoders $f^1, \cdots, f^B$ are such that for all $b$, ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b|F^b=f^b\right)}} {\leqslant}2^{-\xi n}$ which holds with high probability by construction of the layered scheme and the definition of $\overline{\mathbf{A}}^b$. From now on, all expressions should be interpreted as conditioned on $F^1=f^1, \cdots, F^B=f^B$ for such $f^1, \cdots, f^B$; we will omit $F^1=f^1, \cdots, F^B=f^B$ from our notation for the sake of simplicity. We prove the following auxiliary lemma. \[lm:s\_z\_bound\] For all $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, we have ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^b\right)}} {\leqslant}b 2^{-\xi n}.$ It follows from our construction using layered coding that for all $b$, ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b\right)}} {\leqslant}2^{-\xi n}$. We use induction on $b$ to prove the result, i.e., assuming ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^{b-1}; \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}\right)}} {\leqslant}(b-1) 2^{-\xi n}$, we show that ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^b\right)}} {\leqslant}b 2^{-\xi n}$. For $b>1$, note that $$\begin{aligned} {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^b\right)}} &= {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b\right)}} + {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}|\mathbf{Z}^b\right)}}\displaybreak[0]\\ &{\leqslant}2^{-\xi n} + {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}|\mathbf{Z}^b\right)}}\displaybreak[0]\\ &{\leqslant}2^{-\xi n} + {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b,\mathbf{Z}^b; \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}\right)}}\displaybreak[0]\\ &\stackrel{(a)}{{\leqslant}} 2^{-\xi n} + {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^{b-1}; \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}\right)}}\displaybreak[0]\\ &\stackrel{(b)}{{\leqslant}} b 2^{-\xi n},\end{aligned}$$ where $(a)$ follows from the Markov chain $(\mathbf{Z}^b, \overline{\mathbf{A}}^b)-\overline{\mathbf{A}}^{b-1}-(\mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1})$ as seen in Fig. \[fig:fdg\], and $(b)$ follows from the induction hypothesis. We now resume the proof of the secrecy of the scheme. By the chain rule, we have $$\begin{aligned} \label{eq:chain_rule} {{{\mathbb{I}}\!\left(\mathbf{W}^1,\cdots,\mathbf{W}^B;\mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}\right)}} = \sum_{b=1}^B {{{\mathbb{I}}\!\left(\mathbf{W}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{B+1}|\mathbf{W}^{b+1},\cdots,\mathbf{W}^B\right)}}.\end{aligned}$$ Considering each term in the above expression separately, we have $$\begin{aligned} {{{\mathbb{I}}\!\left(\mathbf{W}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{B+1}|\mathbf{W}^{b+1},\cdots,\mathbf{W}^B\right)}} &\stackrel{(a)}{{\leqslant}} {{{\mathbb{I}}\!\left(\mathbf{W}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{B+1}|\mathbf{W}^{b+1},\cdots,\mathbf{W}^B, \mathbf{A}^b\right)}}\\\displaybreak[0] &\stackrel{(b)}{=} {{{\mathbb{I}}\!\left(\mathbf{W}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{b-1}|\mathbf{A}^b\right)}}\\\displaybreak[0] &{\leqslant}{{{\mathbb{I}}\!\left(\mathbf{W}^b, \mathbf{A}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{b-1}\right)}}\\\displaybreak[0] &\stackrel{(c)}{=}{{{\mathbb{I}}\!\left(\mathbf{W}^b, \overline{\mathbf{A}}^{b-1}, \mathbf{L}^b; \mathbf{Z}^1,\cdots, \mathbf{Z}^{b-1}\right)}}\\\displaybreak[0] &\stackrel{(d)}{=} {{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^{b-1}; \mathbf{Z}^1,\cdots, \mathbf{Z}^{b-1}\right)}}\\ &\stackrel{(e)}{{\leqslant}} b2^{-\xi n} \label{eq:w_bound},\end{aligned}$$ where $(a)$ follows from the independence of $(\mathbf{W}^{b+1},\cdots,\mathbf{W}^B, \mathbf{A}^b)$ and $\mathbf{W}^b$ (it holds because the bits of $\mathbf{A}^b$ that depend on $\mathbf{W}^b$ are XOR of $\mathbf{W}^b$ and bits of $\mathbf{A}^{b-1}$ which are independent of $\mathbf{W}^b$), $(b)$ follows form the Markov chain $$\begin{aligned} (\mathbf{W}^{b+1},\cdots, \mathbf{W}^B, \mathbf{Z}^b, \cdots, \mathbf{Z}^{B+1})-\mathbf{A}^b-(\mathbf{W}^b, \mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}),\end{aligned}$$ $(c)$ follows since there is a one-to-one mapping between $(\mathbf{W}^b, \mathbf{A}^b)$ and $(\mathbf{W}^b, \overline{\mathbf{A}}^{b-1}, \mathbf{L}^b)$, $(d)$ follows from the Markov chain $$\begin{aligned} (\mathbf{W}^b,\mathbf{L}^b)-\overline{\mathbf{A}}^{b-1}-(\mathbf{Z}^1, \cdots, \mathbf{Z}^{b-1}),\end{aligned}$$ and $(e)$ follows from Lemma \[lm:s\_z\_bound\]. Combining and , we obtain $$\begin{aligned} {{{\mathbb{I}}\!\left(\mathbf{W}^1,\cdots,\mathbf{W}^B;\mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}\right)}} {\leqslant}B^22^{-\xi n}.\end{aligned}$$ ### Rate Analysis Note that the rate is not random here since the number of information bits is fixed. We know that the bits $\{W_k^b: b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}, \kappa\in{\ensuremath{\llbracket{u^b+1},{m^b}\rrbracket}}\}$ are supposed to be transmitted. Hence, if we define $$\begin{aligned} \alpha {\ensuremath{\triangleq}}\frac{1}{B}\sum_{b=1}^B \alpha(\mathbf{s}^b),\end{aligned}$$ the rate would be $$\begin{aligned} \sum_{b=1}^{B} \frac{[m^b-u^b]^+}{N} &{\geqslant}\sum_{b=2}^{B-1} \frac{k - \left\lceil{{\left(\mathbb{I}_Z^{\alpha^b}+\zeta\right)}}n\right\rceil- k + \left\lceil{{\left(\mathbb{I}_Y^{\alpha^b}-2\zeta\right)}}n\right\rceil}{nB + m}\\ &{\geqslant}\sum_{b=2}^{B-1} \frac{ - \mathbb{I}_Z^{\alpha^b}n + \mathbb{I}_Y^{\alpha^b}n - 3\zeta n- 2}{nB + m}\\ &\stackrel{(a)}{{\geqslant}} \mathbb{I}^{\alpha}_Y - \mathbb{I}^\alpha_Z -O\left(\zeta + \frac{1}{n} + \frac{1}{B}\right),\end{aligned}$$ where $(a)$ follows from the convexity of $\mathbb{I}^\alpha_Y$ in $\alpha$ and the linearity of $\mathbb{I}^\alpha_Z$ in $\alpha$. With an appropriate choice of the parameters of the oracle-assisted scheme, we obtain the following. \[cor:oracle-cr-result\] Fix $\zeta>0$ and $P_X$. There exists $\xi>0$ and a sequence of $(N, K_N, \lceil N^{\frac{2}{3}} \rceil)$ oracle-assisted coding schemes $\{{{\mathcal{C}}}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ such that for all $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{N\to \infty}\alpha(\mathbf{s}_N)= \alpha$ and $N$ large enough $$\begin{aligned} \frac{\psi_N}{N} &{\geqslant}\mathbb{I}^\alpha_Y - \mathbb{I}_Z^\alpha - O{{\left(\zeta + \frac{1}{N^{\frac{1}{3}}}\right)}},\\ \label{eq:rel_oracle} P_e({{\mathcal{C}}}_N|\mathbf{s}_N) &{\leqslant}2^{-\xi N^{\frac{2}{3}}},\\ {{\mathbb{P}_{Q_N}}{\left(\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, Q_N) {\geqslant}2^{-\xi N^{\frac{2}{3}}}\right)}} &{\leqslant}2^{-2^{\xi N^{\frac{2}{3}}}}.\end{aligned}$$ For a fixed $N$, consider an oracle-assisted scheme, ${{\mathcal{C}}}_N$, with $n=\lceil N^{\frac{2}{3}}\rceil$ and $B=\lfloor N^{\frac{1}{3}}\rfloor - O(1)$ which operates as described above. Note that the common randomness $Q_N$ consists of $F^1, \cdots, F^B$ and the common randomness required for ${{\mathcal{C}}}_m^{\text{AVC}}$. Since the probability of error is upper-bounded by $B(2^{-n\zeta} + 2^{-n\xi} + \epsilon_m^{AVC}$, $m= \Theta(n)$, and $n{\geqslant}N^{\frac{2}{3}}$, for $\xi$ small enough, we have $P_e({{\mathcal{C}}}_N|\mathbf{s}_N) {\leqslant}2^{-\xi N^{\frac{2}{3}}}$. Additionally, Lemma \[lm:universal\_secrecy\_weight\] together with union bound imply that with probability at least $1-B2^{-\xi N^{\frac{2}{3}}}$ for all $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, we have ${{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b\right)}} {\leqslant}2^{-\xi N^{\frac{2}{3}}}$. Therefore, by our secrecy analysis in this section, we have $$\begin{aligned} {{\mathbb{P}_{Q_N}}{\left(\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, Q_N) {\geqslant}2^{-\xi N^{\frac{2}{3}}}\right)}} &{\leqslant}2^{-2^{\xi N^{\frac{2}{3}}}}.\end{aligned}$$ Finally, our rate analysis yields that $\frac{\psi_N}{N} {\geqslant}\mathbb{I}^\alpha_Y - \mathbb{I}_Z^\alpha - O{{\left(\zeta + \frac{1}{N^{\frac{1}{3}}}\right)}}$. Note that in our analysis, we did not take into the account the portion of $\mathbf{s}_N$ which corresponds to the use of ${{\mathcal{C}}}^{\text{AVC}}_m$, since it does not change $\lim_{N\to \infty}\alpha(\mathbf{s}_N)$; we will use this simplification several times in the later proofs. Estimation of Adversary’s Actions {#sec:estimation} --------------------------------- In this section, we consider an $(N, K, n)$ oracle-assisted coding scheme introduced in Section \[sec:simp\_coding\_scheme\] and modify it to construct a regular coding scheme which does not require the oracle. The main idea is to select some positions in each sub-block at random and transmit a fixed symbol in those positions. Using the feedback channel $W_{\overline{X}|XS}$, the transmitter then estimates the weight of the channel states provided that the length of the sub-block is large enough. Formally, let the new sub-block length be $n'=n + t$ for $t$ specified later; At the beginning of every sub-block $b \in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, the transmitter selects $t$ distinct positions $\mathbf{J}^b {\ensuremath{\triangleq}}(J_1^b,\cdots, J_t^b)$ with $J_1^b<\cdots < J_t^b$ uniformly at random out of $n'$ positions, in which the transmitter sends a fixed symbol $x_0\in{{\mathcal{X}}}$ with $W_{\overline{X}|X=x_0S=0} \neq W_{\overline{X}|X=x_0S=1}$, which always exists by our assumption that $W_{\overline{X}|XS=0} \neq W_{\overline{X}|XS=1}$. The encoder operates according to the sub-block $b$ of the oracle-assisted scheme in the remaining $n$ positions. Since $W_{\overline{X}|X=x_0S=0} \neq W_{\overline{X}|X=x_0S=1}$, there exists $\overline{x}_0\in\overline{{{\mathcal{X}}}}$ with $W_{\overline{X}|XS}(\overline{x}_0|x_0, 1) \neq W_{\overline{X}|XS}(\overline{x}_0|x_0, 0)$. For simplicity, let $p_0 {\ensuremath{\triangleq}}W_{\overline{X}|XS}(\overline{x}_0|x_0, 0)\text{ and } p_1 {\ensuremath{\triangleq}}W_{\overline{X}|XS}(\overline{x}_0|x_0, 1)$. We also define $$\begin{aligned} \overline{\alpha}^b &{\ensuremath{\triangleq}}\alpha(\mathbf{s}^b)\\ \alpha^b &{\ensuremath{\triangleq}}\frac{{\ensuremath{\textnormal{wt}(\mathbf{s}^b)}} - \sum_{i=1}^t s_{J^b_i}^b}{n}\\ \widehat{\alpha}^b &{\ensuremath{\triangleq}}\frac{1}{t}\sum_{i=1}^t\frac{{\ensuremath{\mathds{1}\!\left\{\overline{X}_{J_i^{b}}^{b}=\overline{x}_0\right\}}} - p_0 }{p_1-p_0 }.\end{aligned}$$ In other words, $\overline{\alpha}^b$ is the overall fraction of ones in the state sequence of the sub-block $b$, $\alpha^b$ is the fraction of ones in the state sequence corresponding to the positions used for transmission, and $\widehat{\alpha}^b$ is our estimation of $\alpha^b$. Each $\widehat{\alpha}^b$ takes $t+1$ possible values by definition. Note that in the protocol described in Section \[sec:simp\_coding\_scheme\], the parameters depending on $\alpha^1, \cdots, \alpha^B$ are $m^1, \cdots, m^B$, $u^1, \cdots, u^B$, and $\ell^1, \cdots, \ell^B$. In the modified protocol, we substitute them with the approximations $$\begin{aligned} \widehat{m}^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b = 1\\ k - \left\lceil{{\left(\mathbb{I}_Z^{\widehat{\alpha}^{b-1}}+2\zeta\right)}}n\right\rceil\quad &b\in{\ensuremath{\llbracket{2},{B}\rrbracket}}\end{cases}\\ \widehat{u}^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b=1\\k - \left\lfloor{{\left(\mathbb{I}_Y^{\widehat{\alpha}^{b-1}}-3\zeta\right)}}n\right \rfloor\quad &b\in{\ensuremath{\llbracket{2},{B+1}\rrbracket}} \end{cases}\\ \log \widehat{\ell}^b &{\ensuremath{\triangleq}}k - \left\lceil{{\left(\mathbb{I}_Y^{\widehat{\alpha}^b}-2\zeta\right)}}n\right\rceil\end{aligned}$$ After the $B^{\text{th}}$ sub-block, the transmitter sends the positions $\mathbf{J}^1, \cdots, \mathbf{J}^B$ together with $\widehat{\alpha}^1, \cdots, \widehat{\alpha}^B$ using the code ${{\mathcal{C}}}_m^{\text{AVC}}$ for $m = O{{\left(Bt \log (Bt) + B \log t\right)}}$. To analyze the secrecy and reliability of the modified coding scheme, we first prove that, with high probability, $\widehat{m}^b$, $\widehat{u}^b$, and $\widehat{\ell}^b$ properly estimate $m^b$, $u^b$, and $\ell^b$, where we redefined $m^b$, $u^b$, $\ell^b$ with the modified version of $\alpha^b$ as $$\begin{aligned} \label{eq:true_mb} {m}^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b = 1\\ k - \left\lceil{{\left(\mathbb{I}_Z^{{\alpha}^{b-1}}+\zeta\right)}}n\right\rceil\quad &b\in{\ensuremath{\llbracket{2},{B}\rrbracket}}\end{cases}\\ \label{eq:true_ub} {u}^{b} &{\ensuremath{\triangleq}}\begin{cases}0\quad&b=1\\k - \left\lfloor{{\left(\mathbb{I}_Y^{{\alpha}^{b-1}}-2\zeta\right)}}n\right \rfloor\quad &b\in{\ensuremath{\llbracket{2},{B+1}\rrbracket}} \end{cases}\\ \label{eq:true_lb} \log {\ell}^b &{\ensuremath{\triangleq}}k - \left\lceil{{\left(\mathbb{I}_Y^{{\alpha}^b}-\zeta\right)}}n\right\rceil.\end{aligned}$$ We recall that all $m^b$, $\widehat{m}^b$, $u^b$, $\widehat{u}^b$, $\ell^b$, and $\widehat{\ell}^b$ are random variables, but for simplicity and with slightly abusing our notation, we use small letters to represent them. Notice that we should be careful not to overestimate $m^b$ to remain secure and not to underestimate $u^b$ to remain reliable. \[lm:estimation\_performance\] For $\frac{t}{n}$ small enough, there exists some $\xi>0$ such that $$\begin{aligned} {{\mathbb{P}_{}}{\left(\widehat{m}^b {\leqslant}m^b \text{ and } \widehat{u}^b {\geqslant}u^b \text{ and } \widehat{\ell}^b {\geqslant}\ell^b\right)}} {\geqslant}1- 2^{-t\xi}.\end{aligned}$$ See Appendix \[sec:estimation-lem\]. Now for all $b\in{\ensuremath{\llbracket{1},{B}\rrbracket}}$, we fix some values for $\widehat{m}^b$, $\widehat{u}^b$, $\widehat{\ell}^b$, $\mathbf{J}^b$ denoted by $\widetilde{m}^b$, $\widetilde{u}^b$, $\widetilde{\ell}^b$, $\widetilde{\mathbf{j}}^b$. Conditioned on these fixed values, we can omit the adversary’s observations in the positions used for estimation from secrecy analysis because they only depend on the noise of the channel. Since conditioned on the positions selected for estimation, the estimated values are independent from all other sources of randomness, the probability of error and mutual information between the message and adversary’s observations are the same as those of the coding scheme of Section \[sec:simp\_coding\_scheme\] when it operates on the positions selected for transmission and $\tilde{m}^b$, $\tilde{u}^b$, $\tilde{\ell}^b$ are used in the scheme. Note that after conditioning, the variables $m^b$, $u^b$, $\ell^b$ defined in - are fixed. Thus, we can define the set $$\begin{aligned} {{\mathcal{A}}}{\ensuremath{\triangleq}}\{(\widetilde{m}^1, \cdots, \widetilde{m}^B, \widetilde{u}^1, \cdots, \widetilde{u}^B, \widetilde{\ell}^1, \cdots, \widetilde{\ell}^B, \widetilde{\mathbf{j}}^1, \cdots, \widetilde{\mathbf{j}}^B): \text{for all }b,~{m}^b {\leqslant}\widetilde{m}^b, {u}^b {\geqslant}\widetilde{u}^b,\widehat{\ell}^b {\leqslant}\widetilde{\ell}^b\}.\end{aligned}$$ Let ${{\mathcal{E}}}$ be the event that a decoding error happens. Then, $$\begin{aligned} {{\mathbb{P}_{}}{\left({{\mathcal{E}}}\right)}} &= \sum {{\mathbb{P}_{}}{\left(\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}}\displaybreak[0]\\ &~~~~~~~~~~~~~~~{{\mathbb{P}_{}}{\left({{\mathcal{E}}}|\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}}\displaybreak[0]\\ &\stackrel{(a)}{{\leqslant}} \sum_{{{\mathcal{A}}}}{{\mathbb{P}_{}}{\left(\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b,\mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}}\\ &~~~~~~~~~~~~~~~{{\mathbb{P}_{}}{\left({{\mathcal{E}}}|\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}} + B 2^{-\xi t}\\ &\stackrel{(b)}{{\leqslant}} 2^{-\xi N^{\frac{2}{3}}} + B 2^{-\xi t},\end{aligned}$$ where the first summation is taken over all possible values of $\tilde{m}^b$, $\tilde{u}^b$, $\tilde{\ell}^b$, $\tilde{j}^b_1, \cdots \tilde{j}^b_t$ for all $b$, $(a)$ follows from Lemma \[lm:estimation\_performance\], and $(b)$ follows from  and our argument that the reliability of the modified scheme is the same as the reliability of oracle-assisted scheme under the conditioning. For secrecy analysis, we first fix a value common randomness $q$ such that conditioned on $Q=q$, the coding scheme in Section \[sec:simp\_coding\_scheme\] is secure. In particular, let $q$ be such that for all state sequences $\mathbf{s}$, in the coding scheme of Section \[sec:simp\_coding\_scheme\], $$\begin{aligned} {{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}|Q=q\right)}}{\leqslant}2^{-\xi N^{\frac{2}{3}}}.\end{aligned}$$ Then, we have $$\begin{gathered} {{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}|Q=q\right)}}\\\displaybreak[0] \begin{split} & {\leqslant}{{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1},\widehat{m}^1, \cdots, \widehat{m}^B, \widehat{u}^1, \cdots, \widehat{u}^B, \widehat{\ell}^1, \cdots,\widehat{\ell}^B, \mathbf{J}^1, \cdots, \mathbf{J}^B |Q=q\right)}}\\ &\stackrel{(a)}{=} {{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}|\widehat{m}^1, \cdots, \widehat{m}^B, \widehat{u}^1, \cdots, \widehat{u}^B, \widehat{\ell}^1, \cdots,\widehat{\ell}^B, \mathbf{J}^1, \cdots, \mathbf{J}^B,Q=q\right)}}\\ &= \sum{{\mathbb{P}_{}}{\left(\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}}\\ &~~~~~~~~~~~~~~~{{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}|\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b,Q=q\right)}}\\ &{\leqslant}\sum_{{{\mathcal{A}}}}{{\mathbb{P}_{}}{\left(\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b\right)}}\\ &~~~~~~~~~~~~~~~{{{\mathbb{I}}\!\left(\mathbf{W}^1, \cdots, \mathbf{W}^B; \mathbf{Z}^1, \cdots, \mathbf{Z}^{B+1}|\text{for all }b,~\widehat{m}^b=\widetilde{m}^b, \widehat{u}^b = \widetilde{u}^b,\widehat{\ell}^b = \widetilde{\ell}^b, \mathbf{J}^b = \widetilde{\mathbf{j}}^b,Q=q\right)}}+ K_N B2^{-\xi t}\\ &{\leqslant}2^{-\xi N^{\frac{2}{3}}}+ K_N B2^{-\xi t}, \end{split}\end{gathered}$$ where $(a)$ follows from the independence of $\mathbf{W}^1, \cdots, \mathbf{W}^B$ and $\widehat{m}^1, \cdots, \widehat{m}^B, \widehat{u}^1, \cdots, \widehat{u}^B, \widehat{\ell}^1, \cdots,\widehat{\ell}^B, \mathbf{J}^1, \cdots, \mathbf{J}^B$. Finally, following the same reasoning of the rate analysis in Section \[sec:simp\_coding\_scheme\], for $\widehat{\alpha} {\ensuremath{\triangleq}}\frac{1}{B} \sum_{b=1}^B \widehat{\alpha}^b$, the rate would be greater than $$\begin{aligned} \mathbb{I}^{\widehat{\alpha}}_Y - \mathbb{I}^{\widehat{\alpha}}_Z + O{{\left(\zeta + \frac{1}{B} + \frac{1}{n}\right)}}.\end{aligned}$$ Using the uniform continuity of $\mathbb{I}^\alpha_Y$ and $\mathbb{I}^\alpha_Z$ in $\alpha$ and , for some $\xi > 0$, we obtain $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbb{I}^{\widehat{\alpha}}_Y - \mathbb{I}^{\widehat{\alpha}}_Z + O{{\left(\zeta + \frac{1}{B} + \frac{1}{n}\right)}} {\geqslant}\mathbb{I}^\alpha_Y - \mathbb{I}_Z^\alpha \right)}} {\geqslant}1 - 2^{-\xi t}.\end{aligned}$$ \[cor:no\_oracle\_scheme\] For any $\zeta>0$, there exists $\xi > 0$ and a sequence of $(N, K_N)$ codes $\{{{\mathcal{C}}}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ such that for all sequences of states $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{N\to \infty}{\frac{{\ensuremath{\textnormal{wt}(\mathbf{s}_N)}}}{N}} = \alpha$ and for large enough $N$, $$\begin{aligned} {{\mathbb{P}_{}}{\left( \frac{\psi_N(\mathbf{\overline{X}}, Q_N)}{N} {\geqslant}\mathbb{I}^\alpha_Y - \mathbb{I}^\alpha_Z - \zeta \big |\mathbf{s}_N\right)}} &{\geqslant}1 - 2^{-\xi N^{\frac{1}{3}}},\\ P_e({{\mathcal{C}}}_N|\mathbf{s}_N) &{\leqslant}2^{-\xi N^{\frac{1}{3}}},\label{eq:estimaiton-rel}\\ {{\mathbb{P}_{Q_N}}{\left(\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, Q_N) {\geqslant}2^{-\xi N^{\frac{1}{3}}}\right)}} &{\leqslant}2^{-2^{\xi N^{\frac{2}{3}}}}. \label{eq:estimation-sec}\end{aligned}$$ For any $N$, let $\widetilde{{{\mathcal{C}}}}_N$ be the $(N, K_N)$ coding scheme introduced in Corollary \[cor:oracle-cr-result\], $n' = \lceil N^{\frac{2}{3}} \rceil$, $t = n'-n = \sqrt{n'} $, and $B=\lfloor N^{\frac{1}{3}}\rfloor - O(1)$. Then, for the modified coding scheme, ${{\mathcal{C}}}_{N}$, we have for large enough $N$ $$\begin{aligned} P_e({{\mathcal{C}}}_{N}|\mathbf{s}_N) {\leqslant}2^{-\xi N^{\frac{2}{3}}} + B2^{-\xi t},\end{aligned}$$ which is less than $2^{-\xi N^{\frac{1}{3}}}$ for $\xi$ small enough. Furthermore, for large enough $N$ and any value of common randomness $q$, the information leakage conditioned on $Q=q$ is upper-bounded by $$\begin{aligned} \mathtt{S}({{\mathcal{C}}}_{N}|\mathbf{s}_N, q) {\leqslant}\mathtt{S}(\widetilde{{{\mathcal{C}}}}_{N}|\mathbf{s}_N, q) + B 2^{-\xi t}.\end{aligned}$$ Substituting the value of $n$, $B$, and $t$ and choosing $\xi$ small enough, we have ${{\mathbb{P}_{Q_N}}{\left(\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, Q_N) {\geqslant}2^{-\xi N^{\frac{1}{3}}}\right)}}{\leqslant}2^{-2^{\xi N^{\frac{2}{3}}}} $. Input Distribution Selection {#sec:input-dist} ---------------------------- We have assumed so far that a fixed $P_X$ is used for random coding for all sub-blocks. To achieve the secrecy capacity of the wiretap channel, in general, channel prefixing $P_{X|U}$ is needed [@Csiszar1978]. Additionally, since the capacity achieving input distribution and the optimal prefix channel might vary for different adversary’s action, we should adapt $P_X$ according to the feedback from adversary’s action to achieve $C(\alpha)$. Therefore, in each sub-block, the transmitter selects a distribution $P_{XU}$ which depends on the causal information the transmitter obtained about channel states. The challenge here is that in general, ${{{\mathbb{I}}\!\left(U;Y\right)}} - {{{\mathbb{I}}\!\left(U;Z|S\right)}}$ is not concave in $P_{XU}$, and therefore, we follow a different approach than [@lomnitz2013universal] to select the input distribution, which is based on the results of adversarial multi-arm bandit problem [@Bubeck2012]. Since those results cannot be applied to an arbitrary set of “bandits", we first introduce a technical lemma that helps us reduce the set of possible input distributions to a finite set. \[lm:cont-input-approx\] Suppose ${{\mathcal{A}}}$ and ${{\mathcal{B}}}$ are compact metric spaces, and $f:{{\mathcal{A}}}\times {{\mathcal{B}}}\to \mathbb{R}$ is continuous. For all $\zeta>0$, there exists $a_1, \cdots, a_{\nu}\in{{\mathcal{A}}}$ such that for all $b\in{{\mathcal{B}}}$, $$\begin{aligned} \left|\max_{i\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}}f(a_i, b) - \max_{a\in{{\mathcal{A}}}} f(a, b)\right| {\leqslant}\zeta.\end{aligned}$$ Since $f$ is continuous and its domain is a product of two compact sets, it is uniformly continuous. Thus, there exists $\delta>0$ such that for all $(a, b)$ and $(a', b')$ in ${{\mathcal{A}}}\times {{\mathcal{B}}}$, if $d((a, b), (a', b')) {\leqslant}\delta$, then $|f(a, b)-f(a',b')|{\leqslant}\zeta$. Because ${{\mathcal{A}}}$ is compact, there exist $a_1, \cdots, a_{\nu}$ such that for all $a\in {{\mathcal{A}}}$ $$\begin{aligned} \label{eq:compact-cover} \min_{i\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}} d(a_i, a) {\leqslant}\delta.\end{aligned}$$ We now claim that $a_1, \cdots, a_{\nu}$ have the desired property. Consider $b\in{{\mathcal{B}}}$. Since, ${{\mathcal{A}}}$ is compact, there exists $a^* \in {{\mathcal{A}}}$ such that $f(a^*, b) = \max_{a\in{{\mathcal{A}}}} f(a, b)$. Moreover, by , there exists $i\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}$ such that $d(a_i, a^*){\leqslant}\delta$. We know that $d((a_i, b), (a^*, b)) {\leqslant}d(a_i, a^*) {\leqslant}\delta$, so $|f(a_i, b) -f(a^*, b)| {\leqslant}\zeta$. This completes the proof. We now provide a result from reinforcement learning regarding the problem of “adversarial bandit" described as follows. Suppose for each $t\in{\ensuremath{\llbracket{1},{T}\rrbracket}}$, we make a choice ${{\color{black}V}}^t\in {\ensuremath{\llbracket{1},{\nu}\rrbracket}}$. The reward for choice $v$ at time $t$ is denoted by $g_{{{\color{black}v}}, t} \in[0, 1]$ assigned by an adversary. Only after we made the choice at time $t$, we are informed about $g_{{{\color{black}V}}^t, t}$. For a specific strategy ${{\color{black}V}}^1, \cdots, {{\color{black}V}}^T$, we define the regret as $$\begin{aligned} R {\ensuremath{\triangleq}}\max_{{{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}} \sum_{t=1}^T g_{{{\color{black}v}}, t} - \sum_{t=1}^T g_{{{\color{black}V}}^t, t}.\end{aligned}$$ The following lemma from [@Bubeck2012] guarantees the existence of strategies with sub-linear regret with high probability. \[lm:bandit\] For all $\delta \in (0, 1)$, there exists ${{\color{black}V}}^1, \cdots, {{\color{black}V}}^T$, which are possibly randomized, such that for all choices of $\{g_{{{\color{black}v}}, t}, {{\color{black}v}}\in {\ensuremath{\llbracket{1},{\nu}\rrbracket}}, t\in{\ensuremath{\llbracket{1},{T}\rrbracket}}\}$, with probability at least $1-\delta$, we have $$\begin{aligned} R {\leqslant}5.15 \sqrt{T\nu \ln{{\left(\nu \delta^{-1}\right)}}}.\end{aligned}$$ See [@Bubeck2012 Theorem 3.3]. We are ready now to describe how the transmitter chooses $P_{XU}$ for each sub-block. For a joint $P_{XU}$ and $\alpha\in[0, 1]$, let $P_S$ be Bernoulli$(\alpha)$, $(S, U, X, Y, Z)$ be distributed according to $P_{SXUYZ} {\ensuremath{\triangleq}}P_S P_{XU}W_{YZ|XS}$, and $g(P_{XU}, \alpha) {\ensuremath{\triangleq}}{{{\mathbb{I}}\!\left(U;Y\right)}} - {{{\mathbb{I}}\!\left(U;Z|S\right)}}$ which is continuous in its arguments. By definition, we have $\max_{P_{XU}} g(P_{XU}, \alpha) = C(\alpha)$. According to Lemma \[lm:cont-input-approx\], if we fix $\zeta > 0$, there exist $P_{XU}^1, \cdots, P_{XU}^{\nu}$ such that for all $\alpha\in[0, 1]$, $$\begin{aligned} \label{eq:dist-approx} \left|\max_{{{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}} g(P_{XU}^{{{\color{black}v}}}, \alpha) - \max_{P_{XU}} g(P_{XU}, \alpha) \right| = \left|\max_{{{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}} g(P_{XU}^{{{\color{black}v}}}, \alpha) - C(\alpha)\right| {\leqslant}\zeta.\end{aligned}$$ In the sub-block $b$, we therefore allow the transmitter to select a distribution $P_{XU}^{{{\color{black}V}}^b}$ where ${{\color{black}V}}^b\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}$ may depend on $\widehat{\alpha}^1, \cdots, \widehat{\alpha}^{b-1}$ defined in Section \[sec:estimation\] and local randomness of the transmitter. By Lemma \[lm:bandit\], we know that for all $\zeta>0$, there exists a choice of ${{\color{black}V}}^1, \cdots, {{\color{black}V}}^B$ and $\xi > 0$ such that $$\begin{aligned} {{\mathbb{P}_{}}{\left(\max_{{{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}}\frac{1}{B} \sum_{b=1}^B g(P_{XU}^{{{\color{black}v}}}, \widehat{\alpha}^b) - \zeta{\leqslant}\frac{1}{B} \sum_{b=1}^B g(P_{XU}^{{{\color{black}V}}^b}, \widehat{\alpha}^b) \right)}} {\geqslant}1 - 2^{-\xi B}.\end{aligned}$$ Using convexity of $g(P_{XU}, \alpha)$ in $\alpha$, for all ${{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}$, we have $\frac{1}{B}\sum_{b=1}^Bg(P_{XU}^{{{\color{black}v}}}, \widehat{\alpha}^b) {\geqslant}g(P_{XU}^{{{\color{black}v}}}, \widehat{\alpha})$ where $\widehat{\alpha}{\ensuremath{\triangleq}}\frac{1}{B}\sum_{b=1}^B \widehat{\alpha}_b$ with probability one. Therefore, we have $$\begin{aligned} {{\mathbb{P}_{}}{\left(\max_{{{\color{black}v}}\in{\ensuremath{\llbracket{1},{\nu}\rrbracket}}}g(P_{XU}^{{{\color{black}v}}}, \widehat{\alpha}) - \zeta {\leqslant}\frac{1}{B}\sum_{b=1}^B g(P_{XU}^{{{\color{black}V}}^b}, \widehat{\alpha}^b) \right)}} {\geqslant}1 - 2^{-\xi B}.\end{aligned}$$ Furthermore, implies that $$\begin{aligned} {{\mathbb{P}_{}}{\left(C(\widehat{\alpha})- 2\zeta{\leqslant}\frac{1}{B}\sum_{b=1}^B g(P_{XU}^{{{\color{black}V}}^b}, \widehat{\alpha}^b) \right)}} {\geqslant}1 - 2^{-\xi B}.\end{aligned}$$ We now show that with high probability $C(\widehat{\alpha})$ is close to $C(\alpha)$ where $\alpha {\ensuremath{\triangleq}}\frac{1}{B}\sum_{b=1}^B\alpha(\mathbf{s}^b)$. Continuity of $C(\alpha)$ in $\alpha$ implies that there exits $\zeta'>0$ such that if $|\alpha_1 - \alpha_2| {\leqslant}\zeta'$, then $|C(\alpha_1) - C(\alpha_2)| {\leqslant}\zeta$. Since $\widehat{\alpha}^1, \cdots, \widehat{\alpha}^B$ are independent and ${{\mathbb{E}_{}}{\left(\widehat{\alpha}^b\right)}} = \alpha(\mathbf{s}^b)$, by Hoeffding’s inequality, we have $$\begin{aligned} {{\mathbb{P}_{}}{\left(\left|\frac{1}{B}\sum_{b=1}^B\widehat{\alpha}^b - \frac{1}{B}\sum_{b=1}^B\alpha(\mathbf{s}^b)\right| {\geqslant}\zeta'\right)}} &={{\mathbb{P}_{}}{\left(|\widehat{\alpha} - \alpha| {\geqslant}\zeta'\right)}}\\ & {\leqslant}2 e^{-2B\zeta'^2}.\end{aligned}$$ Hence, we obtain $$\begin{aligned} {{\mathbb{P}_{}}{\left(C({\alpha})- 3\zeta{\leqslant}\frac{1}{B}\sum_{b=1}^B g(P_{XU}^{{{\color{black}V}}^b}, \widehat{\alpha}^b) \right)}} {\geqslant}1 - 2^{-\xi B}.\end{aligned}$$ In other words, with probability at least $1 -2^{-\xi n}$, our achieved rate, $\frac{1}{B}\sum_{b=1}^B g(P_{XU}^{{{\color{black}V}}^b}, \widehat{\alpha}^b) - O{{\left(\zeta + \frac{1}{B} + \frac{1}{n}\right)}}$, would be at least $C(\widehat{\alpha}) - 3\zeta$. At the end of the transmission, the transmitter sends ${{\color{black}V}}^1, \cdots, {{\color{black}V}}^B$ through the main channel using ${{\mathcal{C}}}_m^{\text{AVC}}$ introduced in Lemma \[lm:avc-code\] for $m = O(B\log \nu)$. Note that ${{\color{black}V}}^1, \cdots, {{\color{black}V}}^B$ are independent of the message bits since they depend only on $\widehat{\alpha}^1, \cdots, \widehat{\alpha}^B$ and local randomness of the transmitter, and their transmission does not leak any information about the messages. We summarize the results in this section as a corollary. \[cor:input-dist-sel\] Fix $\zeta>0$. There exists $\xi>0$ and a sequence of $(N, K_N)$ codes $\{{{\mathcal{C}}}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ such that for all $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{N\to \infty} \alpha(\mathbf{s}_N)= \alpha$ and $N$ large enough $$\begin{aligned} {{\mathbb{P}_{}}{\left(\frac{\psi_N(\mathbf{\overline{X}}, Q_N)}{N} {\geqslant}C(\alpha)- \zeta \big |\mathbf{s}_N\right)}} &{\geqslant}1 - 2^{-\xi N^\frac{1}{3}},\\\displaybreak[0] P_e({{\mathcal{C}}}_N|\mathbf{s}_N) &{\leqslant}2^{-\xi N^{\frac{1}{3}}},\\ \displaybreak[0] {{\mathbb{P}_{Q_N}}{\left(\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, Q_N) {\geqslant}2^{-\xi N^{\frac{1}{3}}}\right)}} &{\leqslant}2^{-2^{\xi N^{\frac{2}{3}}}}.\end{aligned}$$ Reducing Common Randomness {#sec:red_com_rand} -------------------------- In this section, we reduce the common randomness required for the coding scheme with a standard robustification argument [@ahlswede1986arbitrarily]. Specifically, the following lemma is the adaptation of [@csiszar2011information Lemma 12.8] to our setting. \[lm:cr-reduction\] Consider an $(N, K)$ code ${{\mathcal{C}}}= (\mathbf{f}, \phi, \psi, \widehat{\psi}, Q)$. Let us assume that for some $\delta$, $\epsilon_1$, $\epsilon_2>0$, and a function $R:{{\mathcal{S}}}^N\to\mathbb{R}$, for all $\mathbf{s}$, we have $$\begin{aligned} {{\mathbb{E}_{P_Q}}{\left(P_e({{\mathcal{C}}}|\mathbf{s}, Q)\right)}} = P_e({{\mathcal{C}}}|\mathbf{s})&{\leqslant}\epsilon_1,\displaybreak[0]\\ {{\mathbb{E}_{P_{Q}}}{\left(P_{R(\mathbf{s})}({{\mathcal{C}}}|\mathbf{s}, Q)\right)}} &{\leqslant}\epsilon_1,\displaybreak[0]\\ {{\mathbb{P}_{P_Q}}{\left( \mathtt{S}({{\mathcal{C}}}|\mathbf{s}, Q) {\geqslant}\delta\right)}} &{\leqslant}\epsilon_2,\end{aligned}$$ where $P_{\overline{R}}({{\mathcal{C}}}|\mathbf{s}, q) {\ensuremath{\triangleq}}{{\mathbb{P}_{}}{\left(\frac{\psi(\mathbf{\overline{X}}, Q)}{N} < \overline{R}|Q = q, \mathbf{s}\right)}}$ for $\overline{R} > 0$. Then, for every $L$ and $\epsilon>0$ satisfying $N\log |{{\mathcal{S}}}| + 2^{\frac{1}{2}L\epsilon} L \epsilon_2 + 1 < \frac{1}{2}L\epsilon $ and $\epsilon > 2\log(1+\epsilon_1)$, there exist $q_1, \cdots, q_L \in {{\mathcal{Q}}}$ such that for all $\mathbf{s}$ $$\begin{aligned} \frac{1}{L} \sum_{\ell = 1}^L P_e({{\mathcal{C}}}|\mathbf{s}, q_\ell) &{\leqslant}\epsilon,\\ \frac{1}{L}\sum_{\ell = 1}^L P_{R(\mathbf{s})}({{\mathcal{C}}}|\mathbf{s}, q_\ell) &{\leqslant}\epsilon,\\ \text{ for all } \ell,~\mathtt{S}({{\mathcal{C}}}|\mathbf{s}, q_\ell) &{\leqslant}\delta.\end{aligned}$$ Let $Q_1, \cdots, Q_L$ be according to $P_{Q}$. Then, union bound yields that $$\begin{gathered} {{\mathbb{P}_{}}{\left(\text{there exists }\mathbf{s}:~\frac{1}{L}\sum_{\ell=1}^L P_e({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \epsilon \text{ or } \frac{1}{L}\sum_{\ell = 1}^L P_{R(\mathbf{s})}({{\mathcal{C}}}|\mathbf{s}, Q_\ell)> \epsilon\text{ or there exists } \ell:~ \mathtt{S}({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \delta \right)}}\\ {{\leqslant}} \sum_{\mathbf{s}}\left[{{{\mathbb{P}_{}}{\left(\frac{1}{L}\sum_{\ell=1}^L P_e({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \epsilon\right)}} + {{\mathbb{P}_{}}{\left(\frac{1}{L}\sum_{\ell=1}^L P_{R(\mathbf{s})}({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \epsilon\right)}}+ \sum_{\ell=1}^L{{\mathbb{P}_{}}{\left( \mathtt{S}({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \delta\right)}}}\right]. \label{eq:prob_robust}\end{gathered}$$ By [@csiszar2011information Equation (12.15)], we have $$\begin{aligned} {{\mathbb{P}_{}}{\left(\frac{1}{L}\sum_{\ell=1}^L P_e({{\mathcal{C}}}|\mathbf{s}, q_\ell) > \epsilon\right)}} &{\leqslant}2^{-L{{\left(\epsilon-\log (1+\epsilon_1)\right)}}},\end{aligned}$$ and $$\begin{aligned} {{\mathbb{P}_{}}{\left(\frac{1}{L}\sum_{\ell=1}^L P_{R(\mathbf{s})}({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \epsilon\right)}} &{\leqslant}2^{-L{{\left(\epsilon-\log (1+\epsilon_1)\right)}}}.\end{aligned}$$ Moreover, because ${{\mathbb{P}_{}}{\left( \mathtt{S}({{\mathcal{C}}}|\mathbf{s}, Q_\ell) > \delta\right)}} {\leqslant}\epsilon_2$, the right hand side of is upper-bounded by $$\begin{aligned} |{{\mathcal{S}}}|^N{{\left(2^{-L{{\left(\epsilon-\log(1+\epsilon_1)\right)}} + 1} + L \epsilon_2\right)}} \stackrel{(a)}{{\leqslant}} |{{\mathcal{S}}}|^N{{\left(2^{-\frac{1}{2}L\epsilon + 1} + L\epsilon_2\right)}},\end{aligned}$$ where $(a)$ follows from $\epsilon > 2\log(1+\epsilon_1)$. Thus, it is sufficient to prove $|{{\mathcal{S}}}|^N{{\left(2^{-\frac{1}{2}L\epsilon + 1} + L \epsilon_2\right)}} < 1$, which follows from $N\log |{{\mathcal{S}}}| + 2^{\frac{1}{2}L\epsilon} L \epsilon_2 + 1 < \frac{1}{2}L\epsilon$. Applying Lemma \[lm:cr-reduction\] on the coding schemes introduced in Corollary \[cor:input-dist-sel\], we obtain a bound on the required amount of common randomness . \[cor:com-rand\] For any $\zeta>0$, there exists $\xi>0$ and a sequence of $(N, K_N)$ codes $\{{{\mathcal{C}}}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ such that for all $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{N\to \infty}\alpha(\mathbf{s}_N) = \alpha$ and $N$ large enough $$\begin{aligned} \label{eq:cr-red-cr1} {{\mathbb{P}_{}}{\left(\frac{\psi_N(\mathbf{\overline{X}}, Q_N)}{N} {\geqslant}C(\alpha) - \zeta \big |\mathbf{s}_N\right)}} &{\geqslant}1 - \frac{1}{N},\displaybreak[0]\\ P_e({{\mathcal{C}}}_N|\mathbf{s}_N)& {\leqslant}\frac{1}{N},\displaybreak[0]\\ \text{for all } q,~\mathtt{S}({{\mathcal{C}}}_N|\mathbf{s}_N, q) &{\leqslant}2^{-\xi N^{\frac{1}{3}}},\displaybreak[0]\\ {{\mathbb{H}}\!\left(Q_N\right)} &= O{{\left(\log N\right)}}.\label{eq:cr-red-cr4}\\end{aligned}$$ Consider the sequence of codes $\{{{\mathcal{C}}}_N\}_{N{\geqslant}1}$ from Corollary \[cor:input-dist-sel\]. For $\epsilon_1 = 2^{-\xi N^{\frac{1}{3}}}$, $\epsilon_2 = 2^{-2^{\xi N^{\frac{1}{3}}}}$, $\epsilon = \frac{1}{N}$, $\delta = 2^{-\xi N^{\frac{2}{3}}}$, $L=N^2$, all assumptions in Lemma \[lm:cr-reduction\] hold. Thus, we obtain a sequence of codes satisfying -. Proof of Theorem \[thm:converse\]: Converse {#sec:converse-proof} =========================================== Suppose $\{{{\mathcal{C}}}_N = (\mathbf{f}_N, \phi_N, \psi_N, \widehat{\psi}_N, Q_N)\}_{N{\geqslant}1}$ achieves a rate $R$ for all sequences $\{\mathbf{s}_N\}_{N{\geqslant}1}$ with $\lim_{N\to\infty}\alpha(\mathbf{s}_N) = \alpha$. Let $\{\mathbf{s}_N\}_{N{\geqslant}1}$ be a particular such sequence, and $\Pi_N$ be a uniformly chosen random permutation on the set ${\ensuremath{\llbracket{1},{N}\rrbracket}}$ that is independent from all other sources of randomness. For a fixed $N$, suppose the legitimate parties use the code ${{\mathcal{C}}}_N$, and the adversary applies the state sequence $\widetilde{\mathbf{S}}_N {\ensuremath{\triangleq}}(s_{\Pi_N(1)}, \cdots, s_{\Pi_N(n)})$. Considering a random state sequence defined in such a way allows us to upper-bound the rate of a coding scheme that operates well for all state sequences with a fixed type, although we considered a fixed state sequence in our problem formulation in Section \[sec:probl-form-main\]. By conditioning on different values of $\Pi_N$, it follows that $$\begin{aligned} \lim_{N\to \infty} {{\mathbb{P}_{}}{\left(\psi_N(\mathbf{\overline{X}}, Q_N) \neq \widehat{\psi}_N(\mathbf{Y}, Q_N) \text{ or there exists } k\in{\ensuremath{\llbracket{1},{\psi_N(\mathbf{\overline{X}}, Q_N)}\rrbracket}}: W_k \neq \widehat{W}_k\right)}}&=0,\\ \lim_{N\to\infty} {{{\mathbb{I}}\!\left(\mathbf{W}; \Pi_N, \mathbf{Z}\right)}} = \lim_{N\to\infty}{{{\mathbb{I}}\!\left(\mathbf{W}; \mathbf{Z}|\Pi_N\right)}} &= 0,\\end{aligned}$$ and $$\begin{aligned} \lim_{N\to\infty} {{\mathbb{P}_{}}{\left(\frac{\psi_N(\mathbf{\overline{X}}, Q_N)}{N} < R \right)}} = 0.\end{aligned}$$ Moreover, we define $$\begin{aligned} E_N {\ensuremath{\triangleq}}{\ensuremath{\mathds{1}\!\left\{\psi_N(\mathbf{\overline{X}}, Q_N) {\geqslant}NR \text{ and for all } k\in{\ensuremath{\llbracket{1},{\psi_N(\mathbf{\overline{X}}, Q_N)}\rrbracket}}: W_k = \widehat{W}_k\right\}}},\end{aligned}$$ which indicates whether the transmission of the first $NR$ bits is successful, and $\widetilde{\mathbf{W}} {\ensuremath{\triangleq}}\left(W_1, \cdots, W_{\lfloor NR\rfloor}\right)$. Then, $$\begin{aligned} {{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ; \mathbf{Y}\right)}} &= {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} \right)} - {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y}\right)}\\\displaybreak[0] &= {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} \right)} - {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y}, E_N, Q_N\right)} - {{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ; E_N, Q_N|\mathbf{Y}\right)}}\\\displaybreak[0] &{\geqslant}{{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} \right)} - {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y},E_N\right)} - {{\mathbb{H}}\!\left(E_N\right)} -{{\mathbb{H}}\!\left(Q_N\right)}\\\displaybreak[0] &{\geqslant}(NR - 1) - {{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y},E_N\right)} - {{\mathbb{H}}\!\left(E_N\right)} - {{\mathbb{H}}\!\left(Q_N\right)}\\\displaybreak[0] &= (NR-1) - {{\mathbb{P}_{}}{\left(E_N=0\right)}}{{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y},E_N = 0\right)} -{{\mathbb{P}_{}}{\left(E_N=1\right)}}{{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y},E_N = 1\right)}- {{\mathbb{H}}\!\left(E_N\right)}-{{\mathbb{H}}\!\left(Q_N\right)}\\\displaybreak[0] &\stackrel{(a)}{=} (NR-1) - {{\mathbb{P}_{}}{\left(E_N=0\right)}}{{\mathbb{H}}\!\left(\widetilde{\mathbf{W}} |\mathbf{Y},E_N = 0\right)} - {{\mathbb{H}}\!\left(E_N\right)}-{{\mathbb{H}}\!\left(Q_N\right)}\\ &= (NR-1) - {{\mathbb{P}_{}}{\left(E_N=0\right)}}NR - {{\mathbb{H}}\!\left(E_N\right)}-{{\mathbb{H}}\!\left(Q_N\right)}\end{aligned}$$ where $(a)$ follows since for $E_N=1$, $\widetilde{\mathbf{W}}$ is a function of $\mathbf{Y}$ and $Q_N$. Moreover, if we define $\widetilde{Z}_i {\ensuremath{\triangleq}}(\widetilde{S}_i, Z_i)$ and $\widetilde{\mathbf{Z}} {\ensuremath{\triangleq}}(\widetilde{Z}_1, \cdots, \widetilde{Z}_N)$, we have ${{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}}; \widetilde{\mathbf{Z}}\right)}} {\leqslant}{{{\mathbb{I}}\!\left(\mathbf{W}; \widetilde{\mathbf{Z}}\right)}} {\leqslant}{{{\mathbb{I}}\!\left(\mathbf{W}; \Pi_N, \mathbf{Z}\right)}}$, which is vanishing. Thus, applying [@csiszar2011information Lemma 17.12], we have $$\begin{aligned} \label{eq:rate-converse} &\frac{(NR-1) - {{\mathbb{P}_{}}{\left(E_N=0\right)}}NR - {{\mathbb{H}}\!\left(E_N\right)}-{{\mathbb{H}}\!\left(Q_N\right)} - {{{\mathbb{I}}\!\left(\mathbf{W}; \Pi_N, \mathbf{Z}\right)}}}{N} \\ &\phantom{====}{\leqslant}\frac{1}{N}\left({{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ; \mathbf{Y}\right)}}-{{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ;\widetilde{\mathbf{Z}}\right)}}\right)\displaybreak[0]\\ &\phantom{====}= \frac{1}{N}\sum_{i=1}^N \left({{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ;Y_i|Y_1, \cdots, Y_{i-1},\widetilde{Z}_{i+1}, \cdots, \widetilde{Z}_{N}\right)}}-{{{\mathbb{I}}\!\left(\widetilde{\mathbf{W}} ;\widetilde{Z}_i|Y_1, \cdots, Y_{i-1},\widetilde{Z}_{i+1}, \cdots, \widetilde{Z}_{N}\right)}}\right)\displaybreak[0]\\ &\phantom{====}= {{{\mathbb{I}}\!\left(V_N;Y_{J_N}|U_N\right)}} - {{{\mathbb{I}}\!\left(V_N;\widetilde{Z}_{J_N}|U_N\right)}},\end{aligned}$$ where $J_N$ is a random variable with uniform distribution on ${\ensuremath{\llbracket{1},{N}\rrbracket}}$ and independent of all other random variables, $U_N{\ensuremath{\triangleq}}(J_N, Y_1, \cdots Y_{J_N-1}, \widetilde{Z}_{J_N+1}, \cdots \widetilde{Z}_{N})$, and $V_N{\ensuremath{\triangleq}}(\widetilde{\mathbf{W}}, U_N)$. One can check that $U_N-V_N-X_{J_N}-Y_{J_N}\widetilde{Z}_{J_N}$ holds; thus, we can write the joint as $$\begin{aligned} P_{U_NV_NX_{J_N}Y_{J_N}\widetilde{Z}_{J_N}} &= P_{U_NV_NX_{J_N}Y_{J_N}{Z}_{J_N}S_{J_N}}\\ &= P_{U_NV_N|X_{J_N}Y_{J_N}{Z}_{J_N}S_{J_N}}P_{Y_{J_N}{Z}_{J_N}|X_{J_N}S_{J_N}}P_{X_{J_N}|S_{J_N} }P_{S_{J_N}}\\ &= P_{U_N|V_N}P_{V_N|X_{J_N}}P_{Y_{J_N}{Z}_{J_N}|X_{J_N}S_{J_N}}P_{X_{J_N}|S_{J_N} }P_{S_{J_N}}\\ &= P_{U_N|V_N}P_{V_N|X_{J_N}}P_{Y_{J_N}{Z}_{J_N}|X_{J_N}S_{J_N}}P_{X_{J_N} }P_{S_{J_N}}\\ &=P_{U_N|V_N}P_{V_NX_{J_N}}P_{Y_{J_N}{Z}_{J_N}|X_{J_N}S_{J_N}}P_{S_{J_N}}\\ &=P_{U_N|V_N}P_{V_NX_{J_N}}W_{YZ|XS}P_{S_{J_N}}.\end{aligned}$$ Thus, ${{{\mathbb{I}}\!\left(V_N;Y_{J_N}|U_N\right)}} - {{{\mathbb{I}}\!\left(V_N;\widetilde{Z}_{J_N}|U_N\right)}} {\leqslant}C{{\left(\alpha(\mathbf{s}_N)\right)}}$. Furthermore, we know $\lim_{N\to \infty} {{\mathbb{P}_{}}{\left(E_N=0\right)}} = 0$, and $$\begin{aligned} \lim_{N\to\infty}\frac{(NR-1) - {{\mathbb{P}_{}}{\left(E_N=0\right)}}NR - {{\mathbb{H}}\!\left(E_N\right)}-{{\mathbb{H}}\!\left(Q_N\right)} - {{{\mathbb{I}}\!\left(\mathbf{W}; \Pi_N, \mathbf{Z}\right)}}}{N} = R.\end{aligned}$$ Therefore, by taking the limit of both sides of and using the continuity of $C(\alpha)$ in $\alpha$, we obtain $$\begin{aligned} R {\leqslant}C(\alpha).\end{aligned}$$ Discussion and Conclusion {#sec:discussion} ========================= In the presence of a causal feedback channel controlled by the same states as the wiretap channel, our results show that secrecy and reliability can be achieved as if the state sequence had been known with hindsight. We have considered binary-state channels to simplify our notation and proofs, but all results extend to finite-state channels. Specifically, for general finite alphabets, one should substitute the weight of states with the *type* of the states. Our proof then carries over nearly unchanged, with perhaps the exception of the type estimation of the states based on the output of the feedback channel in Section \[sec:estimation\], which requires some care. The result follows, for instance, if there exists a symbol for which all channel output distributions corresponding to all states are linearly independent, so that by solving a system of linear equations, the transmitter can estimate the type of the states. While our model is still far from capturing the full range of active attacks that one could envision in realistic situations, it casts a more optimistic light onto what information-theoretic security may offer in the presence of active adversaries. Proof of Lemma \[lm:universal\_list\_decode\] {#sec:list-dec} ============================================= We first define a one-shot code with list decoder and introduce a generic universal list decoder. \[def:one\_shot\_list\_code\] For a channel $({{\mathcal{X}}}, W_{Y|X}, {{\mathcal{Y}}})$, an $(M, {{\color{black}\ell}}, \epsilon)$ list code $\mathcal{C}$ is a pair of encoder/decoder $(f, \phi)$ with $f:{\ensuremath{\llbracket{1},{M}\rrbracket}}\to{{\mathcal{X}}}$ and $\mathbf{\phi} = (\phi_1, \cdots, \phi_{{{\color{black}\ell}}}):{{\mathcal{Y}}}\to{\ensuremath{\llbracket{1},{M}\rrbracket}}^{{{\color{black}\ell}}}$ satisfying $$\begin{aligned} P_e(W_{Y|X}, f, \phi){\ensuremath{\triangleq}}\frac{1}{M}\sum_{w=1}^M \sum_{y\in{{\mathcal{Y}}}}W_{Y|X}(y|f(w)){\ensuremath{\mathds{1}\!\left\{\text{for all } i\in{\ensuremath{\llbracket{1},{{{\color{black}\ell}}}\rrbracket}}~\phi_i(y) \neq w\right\}}} {\leqslant}\epsilon.\end{aligned}$$ Furthermore, for a given function $\nu:{{\mathcal{X}}}\times{{\mathcal{Y}}}\to \mathbb{R}$, an encoder $f:{\ensuremath{\llbracket{1},{M}\rrbracket}}\to{{\mathcal{X}}}$, and ${{\color{black}\ell}}\in{\ensuremath{\llbracket{1},{M}\rrbracket}}$, we define a universal list decoder $\phi[\nu, f, {{\color{black}\ell}}]: {{\mathcal{Y}}}\to {\ensuremath{\llbracket{1},{M}\rrbracket}}^{{\color{black}\ell}}$ as $$\begin{aligned} \phi[\nu, f, {{\color{black}\ell}}](y) {\ensuremath{\triangleq}}\mathop{\text {argmax}}_{(w_1, \cdots, w_{{{\color{black}\ell}}}):w_1<\cdots<w_{{{\color{black}\ell}}}} \sum_{i=1}^{{{\color{black}\ell}}} \nu(f(w_i), y).\end{aligned}$$ In other words, $\phi[\nu, f, {{\color{black}\ell}}](y)$ outputs the ${{\color{black}\ell}}$ distinct indices in ${\ensuremath{\llbracket{1},{M}\rrbracket}}$ with the largest $\nu(f(\cdot), y)$. Let $({{\mathcal{X}}}, W_{Y|X}, {{\mathcal{Y}}})$ be a channel, and $P_X$ be a distribution over ${{\mathcal{X}}}$. If $F:{\ensuremath{\llbracket{1},{M}\rrbracket}}\to{{\mathcal{X}}}$ is a random encoder such that $F(1), \cdots, F(M)$ are according to $P_X$, the following lemma upper-bounds the expected value of the probability of error for the random list code. \[lm:one\_shot\_list\] For any function $\nu:{{\mathcal{X}}}\times{{\mathcal{Y}}}\to\mathbb{R}$, $$\begin{aligned} {{\mathbb{E}_{F}}{\left(P_e(W_{Y|X}, F, \phi[\nu, F, {{\color{black}\ell}}])\right)}} {\leqslant}\sum_{x, y}P_X(x)W_{Y|X}(y|x) \min{{\left(1, {{\left(\frac{eMq(x, y)}{{{\color{black}\ell}}}\right)}}^{{{\color{black}\ell}}} \right)}}\label{eq:one-shot-list}\end{aligned}$$ where $q(x, y) {\ensuremath{\triangleq}}\sum_{\tilde{x}}P_X(\tilde{x}) {\ensuremath{\mathds{1}\!\left\{\nu(x, y) {\leqslant}\nu(\tilde{x}, y)\right\}}}={{\mathbb{P}_{P_X}}{\left(\nu(X, y) {\geqslant}\nu(x, y)\right)}}$. Using the definition of probability of error and the symmetry of the messages, we have $$\begin{aligned} &{{\mathbb{E}_{F}}{\left(P_e(W_{Y|X}, F, \phi[\nu, F, {{\color{black}\ell}}])\right)}} \nonumber\\ &= \sum_{x_1, \cdots, x_M, y } \prod_{w=1}^M P_X(x_w) \frac{1}{M}\sum_{w'=1}^M W_{Y|X}(y|x_{w'}){\ensuremath{\mathds{1}\!\left\{\exists w''_1 < \cdots <w''_{{{\color{black}\ell}}}:\forall i\in{\ensuremath{\llbracket{1},{{{\color{black}\ell}}}\rrbracket}}~ \nu(x_{w''_i}, y) {\geqslant}\nu(x_{w'}, y), w_i'' \neq w'\right\}}}\displaybreak[0]\nonumber\\ &= \sum_{x, y}W_{Y|X}(y|x) P_X(x) \sum_{x_2, \cdots, x_M } \prod_{w=2}^M P_X(x_w) {\ensuremath{\mathds{1}\!\left\{\exists 1 < w''_1 < \cdots < w''_{{{\color{black}\ell}}}:\forall i\in{\ensuremath{\llbracket{1},{{{\color{black}\ell}}}\rrbracket}}~ \nu(x_{w''_i}, y) {\geqslant}\nu(x, y)\right\}}}\displaybreak[0]\nonumber\\ &{\leqslant}\sum_{x, y}W_{Y|X}(y|x) P_X(x) \sum_{x_1, \cdots, x_M } \prod_{w=1}^M P_X(x_w) {\ensuremath{\mathds{1}\!\left\{\exists w''_1 < \cdots <w''_{{{\color{black}\ell}}}:\forall i\in{\ensuremath{\llbracket{1},{{{\color{black}\ell}}}\rrbracket}}~ \nu(x_{w''_i}, y) {\geqslant}\nu(x, y)\right\}}}.\label{eq:list_error}\end{aligned}$$ Hence, if we define $B(x, y)$ as a random variable distributed according to a Binomial$(M, q(x, y))$ distribution, we have $$\begin{aligned} \sum_{x_1, \cdots, x_M } \prod_{w=1}^M P_X(x_w) {\ensuremath{\mathds{1}\!\left\{\exists w''_1 < \cdots <w''_{{{\color{black}\ell}}}:\forall i\in{\ensuremath{\llbracket{1},{{{\color{black}\ell}}}\rrbracket}}~ \nu(x_{w''_i}, y) {\geqslant}\nu(x, y)\right\}}} = {{\mathbb{P}_{}}{\left(B(x, y) {\geqslant}{{\color{black}\ell}}\right)}}.\end{aligned}$$ Following the same steps as in [@merhav2014list], we obtain$$\begin{aligned} {{\mathbb{P}_{}}{\left(B(x, y) {\geqslant}{{\color{black}\ell}}\right)}} &{\leqslant}\min{{\left(1, e^{-{{\color{black}\ell}}{{\left(\ln \frac{{{\color{black}\ell}}}{Mq(x,y)} - 1\right)}}}\right)}}\\ &= \min{{\left(1, {{\left(\frac{eM q(x,y)}{{{\color{black}\ell}}}\right)}}^{{{\color{black}\ell}}}\right)}}, \label{eq:binom}\end{aligned}$$ and the result follows by substituting into . We are now ready to use the one-shot result Lemma \[lm:one\_shot\_list\] to prove that a random code has good performance for all states given that the list size is properly chosen. Fixing the function $\nu$, the decoder is universal and does not depend on the channel. In particular, we use the standard choice of $\nu(\mathbf{x}, \mathbf{y}) {\ensuremath{\triangleq}}I(\mathbf{x} \wedge \mathbf{y})$ and follow the analysis of [@merhav2014list] to upper-bound the right-hand side of . For $\nu(\mathbf{x}, \mathbf{y}) {\ensuremath{\triangleq}}I(\mathbf{x}\wedge\mathbf{y})$, Lemma \[lm:one\_shot\_list\] implies that $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbf{A}^b \notin {{\mathcal{L}}}^b\right)}} {\leqslant}\sum_{\mathbf{x}, \mathbf{y}}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}(\mathbf{x})W_{\mathbf{Y}|\mathbf{X}\mathbf{S}}(\mathbf{y}|\mathbf{x},\mathbf{s}) \min{{\left(1, {{\left(\frac{e2^kq(\mathbf{x}, \mathbf{y})}{{{\color{black}\ell}}^b}\right)}}^{{{\color{black}\ell^b}}}\right)}},\end{aligned}$$ for $$\begin{aligned} q(\mathbf{x}, \mathbf{y}) {\ensuremath{\triangleq}}\sum_{\tilde{\mathbf{x}}}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}(\tilde{\mathbf{x}}) {\ensuremath{\mathds{1}\!\left\{I(\mathbf{x}\wedge \mathbf{y}) {\leqslant}I(\tilde{\mathbf{x}}\wedge \mathbf{y})\right\}}}.\end{aligned}$$ To upper-bound $q(\mathbf{x}, \mathbf{y})$, let $\mathbf{y} \in {{\mathcal{T}}}_{Q_Y}$ and $\mathbf{x} \in {{\mathcal{T}}}_{Q_{X|Y}}(\mathbf{y})$ for some $Q_Y$ and $Q_{X|Y}$. Then, $$\begin{aligned} q(\mathbf{x}, \mathbf{y}) &= \sum_{\widetilde{\mathbf{x}}}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}(\widetilde{\mathbf{x}}) {\ensuremath{\mathds{1}\!\left\{I(\mathbf{x}\wedge \mathbf{y}) {\leqslant}I(\widetilde{\mathbf{x}}\wedge \mathbf{y})\right\}}}\displaybreak[0]\\ &= \sum_{\widetilde{Q}_{X|Y}\in{{\mathcal{P}}}_n({{\mathcal{X}}}|{{\mathcal{Y}}})}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}({{\mathcal{T}}}_{\widetilde{Q}_{X|Y}}(\mathbf{y})) {\ensuremath{\mathds{1}\!\left\{I(Q_Y, Q_{X|Y}) {\leqslant}I(Q_Y, \widetilde{Q}_{X|Y})\right\}}}\displaybreak[0]\\ &\stackrel{(a)}{{\leqslant}} \sum_{\widetilde{Q}_{X|Y}\in{{\mathcal{P}}}_n({{\mathcal{X}}}|{{\mathcal{Y}}})}2^{-n{{{\mathbb{D}}\!\left({\widetilde{Q}_{X|Y}\VertP_X|Q_Y}\right)}}} {\ensuremath{\mathds{1}\!\left\{I(Q_Y, Q_{X|Y}) {\leqslant}I(Q_Y, \widetilde{Q}_{X|Y})\right\}}}\displaybreak[0]\\ &\stackrel{(b)}{{\leqslant}} (n+1)^{{\ensuremath{\left|{{{\mathcal{X}}}}\right|}}{\ensuremath{\left|{{{\mathcal{Y}}}}\right|}}}2^{-n{{\left(\min_{\widetilde{Q}_{X|Y}:I(Q_Y, Q_{X|Y}) {\leqslant}I(Q_Y, \widetilde{Q}_{X|Y}) } {{{\mathbb{D}}\!\left({\widetilde{Q}_{X|Y}\VertP_X|Q_Y}\right)}}\right)}} }\displaybreak[0]\\ &= (n+1)^{{\ensuremath{\left|{{{\mathcal{X}}}}\right|}}{\ensuremath{\left|{{{\mathcal{Y}}}}\right|}}}2^{-n{{\left(\min_{\widetilde{Q}_{X|Y}:I(Q_Y, Q_{X|Y}) {\leqslant}I(Q_Y, \widetilde{Q}_{X|Y}) } I(Q_Y, \widetilde{Q}_{X|Y}) + {{{\mathbb{D}}\!\left({\widetilde{Q}_{X|Y} \circ Q_Y\VertP_X}\right)}} \right)}}}\displaybreak[0]\\ &{\leqslant}(n+1)^{{\ensuremath{\left|{{{\mathcal{X}}}}\right|}}{\ensuremath{\left|{{{\mathcal{Y}}}}\right|}}} 2^{-nI(Q_Y, Q_{X|Y})}\displaybreak[0]\\ &= 2^{-n{{\left(I(\mathbf{x}\wedge \mathbf{y}) + O{{\left(\frac{\log n}{n}\right)}}\right)}}}\end{aligned}$$ where $(a)$ follows from [@csiszar2011information Equation 2.8], and $(b)$ follows from ${\ensuremath{\left|{{{\mathcal{P}}}_n({{\mathcal{X}}}|{{\mathcal{Y}}})}\right|}} {\leqslant}(n+1)^{{\ensuremath{\left|{{{\mathcal{X}}}}\right|}}{\ensuremath{\left|{{{\mathcal{Y}}}}\right|}}}$. Hence, we obtain $$\begin{gathered} \sum_{\mathbf{x}, \mathbf{y}}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}(\mathbf{x})W_{\mathbf{Y}|\mathbf{X}\mathbf{S}}(\mathbf{y}|\mathbf{x},\mathbf{s})\min{{\left(1, {{\left(\frac{e2^kq(\mathbf{x}, \mathbf{y})}{{{\color{black}\ell}}^b}\right)}}^{{{\color{black}\ell}}^b}\right)}}\\ \begin{split} &{\leqslant}\sum_{\mathbf{x}, \mathbf{y}}P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}(\mathbf{x})W_{\mathbf{Y}|\mathbf{X}\mathbf{S}}(\mathbf{y}|\mathbf{x},\mathbf{s}) 2^{-n {{\color{black}\ell}}^b\left[I(\mathbf{x}\wedge\mathbf{y}) - \frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+}\\ &= \sum_{V_{XY|S} \in {{\mathcal{P}}}_n({{\mathcal{X}}}\times {{\mathcal{Y}}}|{{\mathcal{S}}})}(P_X^{{{ \mathchoice{\raisebox{1pt}{$\displaystyle\otimes$}} {\raisebox{1pt}{$\otimes$}} {\raisebox{0.5pt}{\scalebox{0.7}{$\scriptstyle\otimes$}}} {\raisebox{0.4pt}{\scalebox{0.6}{$\scriptscriptstyle\otimes$}}}}n}}\times W_{\mathbf{Y}|\mathbf{X}\mathbf{S}})({{\mathcal{T}}}_{V_{XY|S}}(\mathbf{s}))2^{-n {{\color{black}\ell}}^b\left[I(V_{XY|S}\circ P_S) - \frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+}\\ &{\leqslant}\sum_{V_{XY|S} \in {{\mathcal{P}}}_n({{\mathcal{X}}}\times {{\mathcal{Y}}}|{{\mathcal{S}}})} 2^{-n{{{\mathbb{D}}\!\left({V_{XY|S}\VertP_X\times W_{Y|XS}|P_S}\right)}}} 2^{-nL\left[I(V_{XY|S}\circ P_S) -\frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+}\\ &\stackrel{(a)}{{\leqslant}} 2^{-n{{\left(\min_{V_{XY|S}}{{{\mathbb{D}}\!\left({V_{XY|S}\VertP_X\times W_{Y|XS}|P_S}\right)}} +{{\color{black}\ell}}^b\left[I(V_{XY|S}\circ P_S) -\frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+\right)}}}, \end{split}\end{gathered}$$ where $(a)$ follows since $\log |{{\mathcal{P}}}_n({{\mathcal{X}}}\times {{\mathcal{Y}}}|{{\mathcal{S}}})| = O(\log n)$. This completes the proof of . We now turn to the proof of . To use , we consider two cases for $V_{XY|S}$. If $V_{XY|S}$ is such that $I(V_{XY|S}\circ P_S) - \frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} > \frac{1}{2}\zeta$, since ${{{\mathbb{D}}\!\left({V_{XY|S}\VertW\times P_X|P_S}\right)}} {\geqslant}0$, we obtain for any $0<\xi < \frac{{{\color{black}\ell}}^b}{2}\zeta $, $$\begin{aligned} 2^{-n \left[{{{\mathbb{D}}\!\left({V_{XY|S}\VertW\times P_X|P_S}\right)}} + {{\color{black}\ell}}^b\left[I(V_{XY|S}\circ P_S) - \frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+\right]} {\leqslant}2^{-\xi n},\end{aligned}$$ Otherwise, we have $I(V_{XY|S}\circ P_S) -\frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} {\leqslant}\frac{1}{2}\zeta$. Recall that $\frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} = {{{\mathbb{I}}\!\left(X;Y\right)}} - \zeta$ by , and as a result, $$\begin{aligned} I(V_{XY|S}\circ P_S) - \frac{\log e}{n} - {{{\mathbb{I}}\!\left(X;Y\right)}} + \zeta{\leqslant}\frac{1}{2}\zeta. \end{aligned}$$ Note that ${{{\mathbb{I}}\!\left(X;Y\right)}} = I((P_X \times W_{Y|XS}) \circ P_S)$. Thus, for large enough $n$, $$\begin{aligned} I(V_{XY|S}\circ P_S) - I((W_{Y|XS}\times P_X)\circ P_S) {\leqslant}-\frac{1}{3} \zeta.\end{aligned}$$ By the continuity of mutual information, we know that there exists $\epsilon > 0$ such that $$\begin{aligned} {{{\mathbb{V}}\!\left(V_{XY|S} \circ P_S, (W_{Y|XS}\times P_X) \circ P_S\right)}}{\geqslant}~\epsilon.\end{aligned}$$ Moreover, this $\epsilon$ can be chosen independent of $P_S$. Applying Pinsker’s inequality, we obtain $$\begin{aligned} {{{\mathbb{D}}\!\left({V_{XY|S} \circ P_S\Vert (W_{Y|XS}\times P_X) \circ P_S}\right)}} {\geqslant}\epsilon^2.\end{aligned}$$ Finally, by the convexity of KL-divergence, we conclude that $$\begin{aligned} {{{\mathbb{D}}\!\left({V_{XY|S}\Vert W_{Y|XS}\times P_X|P_S}\right)}} {\geqslant}{{{\mathbb{D}}\!\left({V_{XY|S}\circ P_S\Vert (W\times P_X) \circ P_S}\right)}} {\geqslant}\epsilon^2.\end{aligned}$$ which implies that $$\begin{aligned} 2^{-n \left[{{{\mathbb{D}}\!\left({V_{XY|S}\VertW_{Y|XS}\times P_X|P_S}\right)}} + {{\color{black}\ell}}^b\left[I(V\circ P_S) -\frac{1}{n}{{\left(k - \log \frac{e}{{{\color{black}\ell}}^b}\right)}} + O{{\left(\frac{\log n}{n}\right)}}\right]^+\right]} {\leqslant}2^{-\epsilon^2 n}.\end{aligned}$$ Thus, equation holds for any $0<\xi < \min{{\left(\epsilon^2, \frac{\zeta}{2}\right)}}$. Proof of Lemma \[lm:avc-code\] {#sec:layer-sec} ============================== We first derive a super-exponential bound for the probability that a randomly chosen code is not secure for the adversary’s channel; such bounds have already been used to prove the achievability for the wiretap channel Type [II]{} [@Goldfeld2016a; @nafea2016new], and we provide here an alternative proof that relies on upper-bounding the mutual information by an average of KL-divergence terms. This approach simplifies the argument and may be of independent interest. We first establish our results in a one-shot setup and then extend them to $n$ channel uses. Consider $K$ independent bits $\mathbf{W} = (W_1, \cdots, W_K)$ that are encoded through an encoder $f: \{0,1\}^K \to {{\mathcal{X}}}$ and assume that the codeword $f(\mathbf{W})$ is transmitted over the channel $({{\mathcal{X}}}, W_{Z|X}, {{\mathcal{Z}}})$. For a fixed $f$ and $m$, ${{{\mathbb{I}}\!\left(W_1, \cdots, W_m; Z\right)}}$ is the information leaked about the first $m$ bits of the message using the encoder $f$. The following lemma establishes bounds for the leakage obtained with random codes. \[lem:one\_shot\_secrecy\] Suppose $F:\{0,1\}^K\to {{\mathcal{X}}}$ is a random encoder such that $ \{F(\mathbf{w}):\mathbf{w} \in \{0,1\}^K\}$ are according to $P_X$. For $P_Z {\ensuremath{\triangleq}}W_{Z|X} \circ P_X$, $\mu_Z {\ensuremath{\triangleq}}\min_{z:P_Z(z) > 0} P_Z(z)$, and any $\gamma$, we have $$\begin{aligned} \label{eq:one_e_i} {{\mathbb{E}_{F}}{\left({{{\mathbb{I}}\!\left(W_1, \cdots, W_m; Z\right)}}\right)}} {\leqslant}\log \left(\frac{1}{\mu_Z} +1\right){{\mathbb{P}_{P_X \times W_{Z|X}}}{\left(\log \frac{W_{Z|X}(Z|X)}{P_Z(Z)} {\geqslant}\gamma \right)}}+ 2^{-K + m + \gamma+1}.\end{aligned}$$ Furthermore, for all $\eta > {{\mathbb{E}_{F}}{\left({{{\mathbb{I}}\!\left(W_1, \cdots, W_m; Z\right)}}\right)}}$, we have $$\begin{aligned} {{\mathbb{P}_{F}}{\left({{{\mathbb{I}}\!\left(W_1, \cdots, W_m; Z\right)}} {\geqslant}(1+\epsilon) \eta'\right)}} {\leqslant}2^{-\frac{2^m\epsilon^2\eta}{2\ln 2 \log \frac{1}{\mu_Z}}}. \label{eq:one_p_i}\end{aligned}$$ The first inequality in follows from standard channel resolvability results [@hayashi2006general; @bloch2013strong], but for completeness, we provide the proof here using our notation. Let $\widehat{P}_{W_1\cdots W_m Z}$ denote the of $(W_1, \cdots, W_m, Z)$ for a particular realization of the encoder $F$. Notice that $$\begin{aligned} {{{\mathbb{I}}\!\left(W_{1}, \cdots, W_{m};Z\right)}} &= {{{\mathbb{D}}\!\left({\widehat{P}_{W_{1} \cdots W_{m}Z}\Vert\widehat{P}_{W_1 \cdots W_m}\times \widehat{P}_Z}\right)}}\\ &{\leqslant}{{{\mathbb{D}}\!\left({\widehat{P}_{W_{1} \cdots W_{m}Z}\Vert\widehat{P}_{W_1 \cdots W_m}\times P_Z}\right)}}\\ &= \sum_{w_1, \cdots, w_m}\frac{1}{2^m} {{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1 \cdots W_m=w_m}\VertP_Z}\right)}}\label{eq:i_sum_d} .\end{aligned}$$ Additionally, for a fixed $w_1, \cdots, w_m$, $$\begin{aligned} \widehat{P}_{Z|W_1 \cdots W_m}(z|w_1, \cdots, w_m) = \frac{1}{2^{K-m}}\sum_{w_{m+1},\cdots, w_K} W_{Z|X}(z|f(w_1, \cdots, w_K)),\end{aligned}$$ so that applying the one-shot channel resolvability upper-bound [@tahmasbi2017second Lemma 3], we obtain $$\begin{aligned} {{\mathbb{E}_{F}}{\left({{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1 \cdots W_m=w_m}\VertP_Z}\right)}}\right)}} &{\leqslant}\log \left(\frac{1}{\mu_Z} +1\right){{\mathbb{P}_{P_X \times W_{Z|X}}}{\left(\log \frac{W_{Z|X}(Z|X)}{P_Z(Z)} {\geqslant}\gamma \right)}}+2^{-K + m + \gamma+1}.\end{aligned}$$ To obtain , we show that the expression in is sum of independent random variables so that we can use Chernoff bound. To this end, notice that ${{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1, \cdots, W_m=w_m}\VertP_Z}\right)}}$ depends on the codewords corresponding to indices $\{(w_1, \cdots, w_m, w_{m+1}, \cdots, w_K):(w_{m+1}, \cdots, w_K)\in\{0, 1\}^{K-m}\}$, and for distinct $w_1, \cdots, w_m$, these sets are disjoint. Since the codewords are generated independently, $$\begin{aligned} \left\{{{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1, \cdots, W_m=w_m}\VertP_Z}\right)}}: (w_1, \cdots, w_m)\in\{0,1\}^m\right\}\end{aligned}$$ are also independent. Furthermore, to find a uniform upper-bound for ${{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1, \cdots, W_m=w_m}\VertP_Z}\right)}}$, note that $$\begin{aligned} {{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1, \cdots, W_m=w_m}\VertP_Z}\right)}} &= \sum_{z} \widehat{P}_{Z|W_1, \cdots, W_m}(z|w_1, \cdots, w_m) \log \frac{\widehat{P}_{Z|W_1, \cdots, W_m}(z|w_1, \cdots, w_m)}{P_Z(z)} \\ &{\leqslant}\sum_{z} \widehat{P}_{Z|W_1, \cdots, W_m}(z|w_1, \cdots, w_m) \log \frac{1}{P_Z(z)} \\ &{\leqslant}\log \frac{1}{\mu_Z}.\end{aligned}$$ Thus, [@ahlswede1989identification Lemma LD] implies that $$\begin{aligned} {{\mathbb{P}_{F}}{\left({{{\mathbb{I}}\!\left(W_1, \cdots, W_m; Z\right)}} {\geqslant}(1+\epsilon)\eta\right)}} &{\leqslant}{{\mathbb{P}_{F}}{\left(\frac{1}{2^m}\sum_{w_1, \cdots, w_m} {{{\mathbb{D}}\!\left({\widehat{P}_{Z|W_1=w_1 \cdots W_m=w_m}\VertP_Z}\right)}} {\geqslant}(1+\epsilon)\eta\right)}}\\ &{\leqslant}\exp_2{{\left(-\frac{2^m\epsilon^2\eta}{2\ln 2 \log \frac{1}{\mu_Z}}\right)}}.\end{aligned}$$ We now apply Lemma \[lem:one\_shot\_secrecy\] to an $W_{Z|XS}$ to prove Lemma \[lm:s\_z\_bound\]. First note that for memory-less channels, the information density $\log \frac{W_{Z|X}(Z|X)}{P_Z(Z)}$ concentrates around its expectation, which is mutual information. Furthermore, since the number of codewords is increasing exponentially with the block-length, by Lemma \[lem:one\_shot\_secrecy\], we obtain a doubly-exponential upper-bound on the probability that a random code is not secure for a specific state, and therefore, we can use union bound to a bound for the probability that a code is secure for all states whose number increases exponentially. For a fixed $\mathbf{s}^b \in {{\mathcal{S}}}^n$, to apply Lemma \[lem:one\_shot\_secrecy\], let $(\mathbf{X}, \mathbf{Z})$ be distributed according to $\prod_{i=1}^n {{\left(P_X \times W_{Z|XS=s_i^b}\right)}}$, $P_{Z|S=s} {\ensuremath{\triangleq}}W_{Z|XS=s} \circ P_X$, $\mu_Z {\ensuremath{\triangleq}}\min_{z\in{{\mathcal{Z}}}, s\in{{\mathcal{S}}}P_{Z|S}(z|s) > 0 } P_{Z|S}(z|s)$, and $ \gamma {\ensuremath{\triangleq}}{{\left(\mathbb{I}_Z^\alpha + \frac{1}{2}\zeta\right)}} n$. Define $$\begin{aligned} \Lambda {\ensuremath{\triangleq}}\max_{s, x, z: W_{Z|XS}(z|xs)P_X(x) > 0} \left|\log \frac{W_{Z|XS}(z|xs)}{P_{Z|S}(z|s)} - I(P_X, W_{Z|XS=s})\right|,\end{aligned}$$ and note that $\Lambda < \infty$. Therefore, by Hoeffding’s inequality, $$\begin{aligned} {{\mathbb{P}_{}}{\left(\log \prod_{i=1}^n\frac{W_{Z|XS}(Z_i|X_i,s_i)}{P_{Z|S}(Z_i|s_i)} {\geqslant}\gamma \right)}} &= {{\mathbb{P}_{}}{\left(\sum_{i=1}^n \log\frac{W_{Z|XS=s_i}(Z_i|X_i)}{P_{Z|S}(Z_i|s_i)} {\geqslant}\gamma\right)}}\displaybreak[0]\\ &{\leqslant}\exp_2{{\left(-\frac{(\gamma - n\mathbb{I}_Z^\alpha)^2}{2n\Lambda^2}\right)}}\displaybreak[0]\\ &=\exp_2{{\left(-\frac{n\zeta^2}{8\Lambda^2}\right)}}.\end{aligned}$$ Further, by definition of $m^{b+1}$ in Eq.  and $\gamma$, $$\begin{aligned} 2^{- k + m(\mathbf{s}) + \gamma+1 } &= 2^{-k + k - \left \lceil (\mathbb{I}_Z^\alpha + \zeta)n\right\rceil + (\mathbb{I}_Z^\alpha + \frac{1}{2}\zeta)n + 1}\\ &{\leqslant}2^{-\frac{1}{2}\zeta n + 1}.\end{aligned}$$ Accordingly, ${{\mathbb{E}_{F^b}}{\left({{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} \right)}}$ is upper-bounded by $$\begin{aligned} \ n\log{{\left(\frac{1}{\mu_Z} + 1\right)}}\exp_2{{\left(-\frac{n\zeta^2}{8\Lambda^2}\right)}} + 2^{-\frac{1}{2}\zeta n + 1}.\end{aligned}$$ By choosing $0<\xi < \min\left(\frac{1}{2}\zeta,{\frac{\zeta^2}{8\Lambda^2}}\right)$, for large enough $n$, ${{\mathbb{E}_{F^b}}{\left({{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} \right)}}{\leqslant}\frac{2}{3}2^{-\xi n}{\ensuremath{\triangleq}}\eta $. Thus, with $\epsilon = \frac{1}{2}$, yields that $$\begin{aligned} {{\mathbb{P}_{F^b}}{\left({{{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} }{\geqslant}{{\left(1+\frac{1}{2}\right)}}\eta\right)}} &={{\mathbb{P}_{F^b}}{\left( {{{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} }){\geqslant}2^{-\xi n} \right)}} \\ &{\leqslant}\exp_2{{\left(-\frac{2^{m^{b+1}}2^{-\xi n}}{12\ln( 2) n\log\frac{1}{\mu_Z}}\right)}}.\end{aligned}$$ As a result, by the union bound, $$\begin{aligned} \label{eq:p_bound_i_multi_lemma} {{\mathbb{P}_{F^b}}{\left(\text{there exits } \mathbf{s}^b: {{{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} }{\geqslant}2^{-\xi n} \right)}} & {\leqslant}\sum_{\mathbf{s}} {{\mathbb{P}_{F^b}}{\left( {{{{\mathbb{I}}\!\left(\overline{\mathbf{A}}^b; \mathbf{Z}^b | \mathbf{s}^b\right)}} }{\geqslant}2^{-\xi n}\right)}} \\ & {\leqslant}|{{\mathcal{S}}}|^n \exp_2{{\left(-\frac{2^{m^{b+1}}2^{-\xi n}}{12\ln( 2) n\log\frac{1}{\mu_Z}}\right)}}.\end{aligned}$$ Notice that $m^{b+1} {\geqslant}\zeta n$ since $k {\geqslant}{{\left(\mathbb{I}_Z^{\max} + 2 \zeta\right)}}n+1$. Thus, for $0<\xi < \frac{1}{3} \zeta $, we obtain that for large $n$, $$\begin{aligned} |{{\mathcal{S}}}|^n \exp_2{{\left(-\frac{2^{m^{b+1}}2^{-\xi n}}{12\ln (2) n\log\frac{1}{\mu_Z}}\right)}} {\leqslant}2^{-2^{\xi n}}.\end{aligned}$$ Proof of Lemma \[lm:estimation\_performance\] {#sec:estimation-lem} ============================================= We first show that for all $\epsilon>0$, $$\begin{aligned} \label{eq:estimation_error} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon\right)}} {\leqslant}2\exp{{\left(-\frac{t(p_1-p_0)^2\epsilon^2}{2}\right)}} + 2 \exp{{\left(-\frac{\epsilon^2t}{2}\right)}}.\end{aligned}$$ To prove this, notice that $$\begin{aligned} \label{eq:mb_estimation_bayes} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon\right)}} &= \sum_{\mathbf{j}}{{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}=\mathbf{j}\right)}} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon|\mathbf{J}^{b-1}=\mathbf{j}\right)}}.\end{aligned}$$ To upper-bound the above summation, we split it into two terms. First, if we define $$\begin{aligned} {{\mathcal{E}}}{\ensuremath{\triangleq}}\left\{\mathbf{j} = (j_1, \cdots, j_t):~j_1< \cdots < j_t,~\left| \frac{1}{t}\sum_{i=1}^t s_{j_i}^{b-1}-\overline{\alpha}^{b-1}\right| > \frac{\epsilon}{2}\right\},\end{aligned}$$ we have $$\begin{aligned} \label{eq:mb_first_trem} \sum_{\mathbf{j}\in{{\mathcal{E}}}}{{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}=\mathbf{j}\right)}} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon|\mathbf{J}^{b-1}=\mathbf{j}\right)}}&{\leqslant}\sum_{\mathbf{j}\in{{\mathcal{E}}}}{{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}=\mathbf{j}\right)}}\\ &= {{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}\in {{\mathcal{E}}}\right)}}.\end{aligned}$$ To upper-bound ${{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}\in {{\mathcal{E}}}\right)}}$, we express this probability in terms of the of a random variable with hypergeometric distribution. Let $H$ shows the number of successes in $t$ draws *without replacement* from a population of size $n'$ with exactly ${\ensuremath{\textnormal{wt}(\mathbf{s}^{b-1})}}$ successes in the population. Then, we have $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbf{J}^{b-1}\in {{\mathcal{E}}}\right)}} &= {{\mathbb{P}_{}}{\left(\left|H - t\frac{{\ensuremath{\textnormal{wt}(\mathbf{s}^{b-1})}}}{n'} \right| {\geqslant}t \frac{\epsilon}{2}\right)}}\\ &\stackrel{(a)}{{\leqslant}} 2\exp{{\left(-\frac{\epsilon^2 t}{2}\right)}}\end{aligned}$$ where $(a)$ follows from standard tail bounds for hypergeometric distribution (e.g., see [@hoeffding1963probability]). We now fix a $\mathbf{j} = (j_1, \cdots, j_t) \notin {{\mathcal{E}}}$ such that $j_1 < \cdots < j_t$ and define $$\begin{aligned} T_i &{\ensuremath{\triangleq}}\frac{{\ensuremath{\mathds{1}\!\left\{\overline{X}_{j_i}^{ b}=\overline{x}_0\right\}}} - p_0 }{p_1 - p_0 }.\end{aligned}$$ We know that $$\begin{aligned} {{\mathbb{E}_{}}{\left(T_i\right)}} &= s_{j_i}^{b-1}\\ \frac{ - p_0}{p_1-p_0 } &{\leqslant}T_i {\leqslant}\frac{1 - p_0 }{p_1 -p_0 }.\end{aligned}$$ Hence, by Hoeffding’s inequality, we obtain $$\begin{aligned} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon |\mathbf{J}^{b-1}=\mathbf{j}\right)}} &= {{\mathbb{P}_{}}{\left(\left|\frac{1}{t}\sum_{i=1}^t T_i - \overline{\alpha}^{b-1}\right| {\geqslant}\epsilon \right)}}\displaybreak[0]\\ &= {{\mathbb{P}_{}}{\left(\left|\frac{1}{t}\sum_{i=1}^t T_i -\frac{\sum_{i=1}^t s_{j_i}^{b-1}}{t} - \overline{\alpha}^{b-1} + \frac{\sum_{i=1}^t s_{j_i}^{b-1}}{t} \right| {\geqslant}\epsilon \right)}}\displaybreak[0]\\ &{\leqslant}{{\mathbb{P}_{}}{\left(\left|\frac{1}{t}\sum_{i=1}^t T_i -\frac{\sum_{i=1}^t s_{j_i}^{b-1}}{t}\right| {\geqslant}\epsilon - \left| - \overline{\alpha}^{b-1} + \frac{\sum_{i=1}^t s_{j_i}^{b-1}}{t} \right| \right)}}\displaybreak[0]\\ &{\leqslant}2\exp{{\left(-2t(p_1-p_0)^2{{\left(\epsilon - \left|\overline{\alpha}^{b-1} - \frac{\sum_{i=1}^t s_{j_i}^{b-1}}{t} \right|\right)}}^2\right)}}\displaybreak[0]\\ \label{eq:mb_second_trem} &\stackrel{(a)}{{\leqslant}} 2\exp{{\left(-\frac{t(p_1-p_0)^2\epsilon^2}{2}\right)}},\end{aligned}$$ where $(a)$ follows since $\mathbf{j}\notin {{\mathcal{E}}}$. Therefore, combining , , and , we obtain $$\begin{aligned} {{\mathbb{P}_{}}{\left(|\widehat{\alpha}^{b-1} - \overline{\alpha}^{b-1}| > \epsilon\right)}} &{\leqslant}2\exp{{\left(-\frac{t(p_1-p_0)^2\epsilon^2}{2}\right)}} + 2 \exp{{\left(-\frac{\epsilon^2t}{2}\right)}}.\end{aligned}$$ We now turn to the proof of a lower bound for ${{\mathbb{P}_{}}{\left(\widehat{m}^b {\leqslant}m^b\right)}}$. Notice that for $b=1$, by definition $\widehat{m}^b = m^b = 0$, and we have ${{\mathbb{P}_{}}{\left(\widehat{m}^b {\leqslant}m^b\right)}} = 1$. Therefore, we assume $b > 1$. Then, $$\begin{aligned} {{\mathbb{P}_{}}{\left(\widehat{m}^b {\leqslant}m^b\right)}} &= {{\mathbb{P}_{}}{\left( k - \left\lceil{{\left(\mathbb{I}_Z^{\widehat{\alpha}^{b-1}}+2\zeta\right)}}n\right\rceil {\leqslant}k - \left\lceil{{\left(\mathbb{I}_Z^{{\alpha}^{b-1}}+\zeta\right)}}n\right\rceil\right)}}\\ &{\geqslant}{{\mathbb{P}_{}}{\left({{\left(\mathbb{I}_Z^{{\alpha}^{b-1}}+\zeta\right)}}n+1 {\leqslant}{{\left(\mathbb{I}_Z^{\widehat{\alpha}^{b-1}}+2\zeta\right)}}n \right)}}\\ &={{\mathbb{P}_{}}{\left(\mathbb{I}_Z^{{\alpha}^{b-1}}-\mathbb{I}_Z^{\widehat{\alpha}^{b-1}} {\leqslant}\zeta-\frac{1}{n} \right)}}.\end{aligned}$$ Since $\mathbb{I}^\alpha_Z$ is uniformly continuous in $\alpha$, there exists $\zeta_2$, independent of $n$, such that if $|\alpha_1 - \alpha_2| {\leqslant}\zeta_2$, then $\mathbb{I}_Z^{\alpha_1}-\mathbb{I}_Z^{\alpha_2} {\leqslant}\zeta-\frac{1}{n}$. Hence, we have $$\begin{aligned} {{\mathbb{P}_{}}{\left(\mathbb{I}_Z^{{\alpha}^{b-1}}-\mathbb{I}_Z^{\widehat{\alpha}^{b-1}} {\leqslant}\zeta-\frac{1}{n} \right)}} &{\geqslant}{{\mathbb{P}_{}}{\left(|{{\alpha}^{b-1}}-{\widehat{\alpha}^{b-1}}| {\leqslant}\zeta_2 \right)}}\\\displaybreak[0] &= {{\mathbb{P}_{}}{\left(|{{\alpha}^{b-1}} - \overline{\alpha}^{b-1}-{\widehat{\alpha}^{b-1}} + \overline{\alpha}^{b-1}| {\leqslant}\zeta_2 \right)}}\\\displaybreak[0] &{\geqslant}{{\mathbb{P}_{}}{\left(|{{\alpha}^{b-1}} - \overline{\alpha}^{b-1}| + |-{\widehat{\alpha}^{b-1}} + \overline{\alpha}^{b-1}| {\leqslant}\zeta_2 \right)}}\\\displaybreak[0] &= {{\mathbb{P}_{}}{\left(\left | \frac{{\ensuremath{\textnormal{wt}(\mathbf{s}^{b-1})}} - \sum_{i=1}^t s_{J^{b-1}_i}^{b-1}}{n} - \frac{{\ensuremath{\textnormal{wt}(\mathbf{s}^{b-1})}}}{n'} \right| + |-{\widehat{\alpha}^{b-1}} + \overline{\alpha}^{b-1}| {\leqslant}\zeta_2 \right)}}\\ & = {{\mathbb{P}_{}}{\left(\left | \frac{{\ensuremath{\textnormal{wt}(\mathbf{s}^{b-1})}}t}{nn'} - \frac{ \sum_{i=1}^t s_{J^{b-1}_i}^{b-1}}{n} \right| + |-{\widehat{\alpha}^{{b-1}}} + \overline{\alpha}^{b-1}| {\leqslant}\zeta_2 \right)}}\\ &{\geqslant}{{\mathbb{P}_{}}{\left( |-{\widehat{\alpha}^{b-1}} + \overline{\alpha}^{b-1}| {\leqslant}\zeta_2 - \frac{2t}{n} \right)}}\\ &{\geqslant}1- 2\exp{{\left(-\frac{t(p_1-p_0)^2{{\left(\zeta_2 - \frac{2t}{n}\right)}}^2}{2}\right)}} + 2 \exp{{\left(-\frac{{{\left(\zeta_2 - \frac{2t}{n}\right)}}^2t}{2}\right)}}\end{aligned}$$ Therefore, for $\frac{t}{n}$ small enough, we can find $\xi > 0$ such that ${{\mathbb{P}_{}}{\left(\widehat{m}^b {\leqslant}m^b\right)}}{\geqslant}1 - 2^{-\xi t}$. By same argument, we can show that ${{\mathbb{P}_{}}{\left(\widehat{u}^b {\geqslant}u^b\right)}} {\geqslant}1 - 2^{-\xi t}$ and ${{\mathbb{P}_{}}{\left(\widehat{\ell}^b {\geqslant}\ell^b\right)}} {\geqslant}1 - 2^{-\xi t}$, which completes the proof of lemma. [^1]: This work was presented in part at the 2017 IEEE International Symposium on Information Theory [@tahmasbi2017learning]. This work was supported in part by NSF award CCF 1527074. [^2]: [[To be precise, ${{\mathcal{T}}}_{P_X}$ depends on $n$, but we drop $n$ from our notation for simplicity.]{}]{} [^3]: Despite conceptual similarities with layered secrecy coding [@zou2013layered], our problem formulation is different for technical reasons.

--- abstract: 'We introduce the concept of *continuous transportation* task to the context of multi-agent systems. A continuous transportation task is one in which a multi-agent team visits a number of fixed locations, picks up objects, and delivers them to a transportation hub. The goal is to maximize the rate of transportation while the objects are replenished over time . In this extended abstract, we present a hybrid of centralized and distributed approaches that minimize communications in the multi-agent team. We contribute a novel online partitioning-transportation algorithm with information gathering in the multi-agent team.' author: - | Chao Wang^^, Somchaya Liemhetcharat^§^, Kian Hsiang Low^^\ \ \ bibliography: - 'sigproc.bib' title: 'Multi-Agent Continuous Transportation with Online Balanced Partitioning' --- Introduction ============ We are interested in multi-agent coordination and continuous transportation, where agents visit locations in the environment to transport objects (passengers or items) and deliver them to a transportation hub. The objects replenish over time, and we consider Poisson model of object replenishment. Poisson model is suitable for scenarios with independently-occurring objects, such as passengers appear at the transit stops. Thus, the continuous transportation task is general and applicable to many real-life scenarios, e.g., first/last mile problem, package or mail collection and delivery. The most similar work has been done is in multi-robot foraging and delivery problem [@liemhetcharat2015continuous; @liemhetcharat2015multi]. A multi-robot team forages resources from environment to a home location or delivers items to locations on request while the resources replenish over time. The goal of continuous foraging is to maximize the rate of resource foraging. Thus, continuous foraging has similarities to continuous transportation, and thus we compare our algorithms with that of [@liemhetcharat2015continuous; @liemhetcharat2015multi]. In addition, continuous area sweeping (e.g., [@ahmadi2006multi]) problem has similarities to continuous transportation, with the main difference being that the agents have a carrying capacity and must periodically return to the transportation hub. Hence, we compare to the benchmark of [@ahmadi2006multi]. Besides, [@ahmadi2006multi] uses area partitioning algorithm that relies on communications and negotiations among robots. Thus, it is prone to message loss and inconsistency. In contrast, our approach does not require communications among agents, and the partitioning is conducted separately. Hence, we present a hybrid of centralized and distributed approaches, and the efficacy is demonstrated in the evaluation section. Problem And Approach ==================== In this section, we formally define the multi-agent continuous transportation problem in detail, and give an overview of our approach. In a continuous transportation task, a multi-agent team visits a number of fixed locations and transport objects to a transportation hub. Formal Problem Definition ------------------------- The multi-agent continuous transportation problem can be defined as follows: - $\mathcal{A}=\{a_1,...,a_{k}\}$ is the set of transportation agents, e.g., the autonomous vehicles. - $s_i$, $c_i$, and $y_i\ (\leq c_i)$ denote agent $a_i$’s speed, maximum capacity, and current load, i.e., the number of objects carried. - $\mathcal{L}=\{l_1,...,l_{n}\}$ is the set of locations, where $l_0$ is the transportation hub. - $v_{j,t}$ denotes the number of objects available at location $l_j$ at time step $t$, e.g., the number of passengers appear at the transit stop $l_j$. - $\hat{v}_{j,t}^{(i)}$ denotes $a_i$’s estimate of $v_{j,t}$ at time step $t$. - $o_j$ denotes the observation made at location $l_j$. When a transportation agent $a_i$ arrives at a location $l_j\ (j > 0)$, min($v_{j,t}$, $c_i-y_i$) objects at $l_j$ are picked up by $a_i$, and $a_i$ makes an observation $o_j$ of the number of objects remaining. When $a_i$ arrives at $l_0$, all $y_i$ objects carried by $a_i$ are transferred to $l_0$. - $D:\mathcal{L}\times\mathcal{L}\rightarrow\mathbb{R}^+$ is the distance function of the locations. - $t(a_i,l_j,l_k)=\ceil*{\frac{D(l_j,l_k)}{s_i}}$ is the time steps taken for an agent $a_i$ to move from location $l_j$ to $l_k$. The goal is to maximize the rate of objects delivered to the transportation hub $l_0$ within $T$ time steps, i.e., maximize $\frac{v_{0,T}}{T}$. Our Approach ------------ Our approach for solving the continuous transportation problem is: - We assume that $v_{j,t}$ follows a known model — in this paper, we assume that $v_{j,t}$ follows the Poisson model, where the number of objects replenished every time step follows a Poisson distribution with mean $\lambda_j$. However, the parameters of the models (i.e., $\lambda_j$) are not known in advance; - The estimates $\hat{v}^{(i)}_{j,t}$ are updated using the model and observations from transportation agents $a_i$ visiting location $l_j$; - Following [@liemhetcharat2015continuous], the transportation agents do not share their models $\hat{v}^{(i)}_{j,t}$ since it could be expensive subject to the communication bandwidth and team size. Different from [@liemhetcharat2015continuous], in our approach, the sharing of destination and load among agents are not required. We contribute an online algorithm that partition the locations based on their 2D-position and estimated replenishment rate, and controls the transportation agents $a_i$, that use $\hat{v}^{(i)}_{j,t}$ to plan their next destination within the cluster of locations assigned. The algorithm dynamically repartitions locations based on information gathered by the multi-agent team. Algorithm and Evaluation ======================== Our Online Balanced Partitioning (OBP) algorithm begins with statically partition the locations into clusters based on 2D-position of locations with $k$-means algorithm. The algorithm is inspired by algorithms proposed for continuous foraging [@liemhetcharat2015continuous]. The main difference is that the agents replan destinations in the cluster of locations assigned, instead of all the locations. The replanning of destination within each cluster is based on Greedy Rate [@liemhetcharat2015continuous], i.e., maximizing the transportation rate. Further, due to partitioning, communications of destinations and loads are not required. By only considering the 2D-position of locations, the workload of each agent might not be balanced, i.e., the total replenishment rate of some clusters could be so high that the agents cannot afford it, while some other agents are idle. On the other hand, the result of standard $k$-means algorithm depends on the choice of initial centroid which is randomly generated. In this case, we introduce online balanced partitioning, i.e., balance estimated total replenishment rate of each cluster with information gathering. The replenishment rate is not known in advance. The agents use preset estimated replenishment rate to update its estimated number of objects. The estimate can only be corrected when they visit the location and make an observation. Since we focus on Poisson replenishment model, where number of objects replenished per time step follows a mean value, we believe that estimated replenishment rate can be corrected with continuous observations, i.e., the total number of objects replenished divided by time steps elapsed. Our online partitioning algorithm dynamically corrects the estimated replenishment rate by visiting a location and making observation. Once the deviation of total replenishment rate of clusters are greater than a preset threshold, repartitioning will be done immediately. ![Comparison of our Online Balanced Partitioning (OBP) algorithm against the benchmark of Greedy Rate with Expected Observation (GR+EO) and Continuous Area Sweeping (CAS) algorithm.[]{data-label="fig:OBP"}](plot_poisson_cap10.png){width="50.00000%"} Figure \[fig:OBP\] shows the performance of our Online Balanced Partitioning (OBP) algorithm when the capacities of the agents are 10 and the number of locations are 20. In order to demonstrate the effectiveness of OBP, we compared OBP with GR+EO [@liemhetcharat2015continuous] which requires the presence of a reconnaissance agent. The solid red, blue, and gray lines show Online Balanced Partitioning (OBP), Greedy Rate with Expected Observation (GR+EO), and Random Transportation (R), and the shaded areas show the standard deviations of these algorithms. As the number of agents increase, OBP outperforms all other algorithms, even GR+EO ($p=3\times 10^{-23}$), primarily because of the dynamic partitioning with information gathering by the multi-agent team. It clearly illustrates the efficacy of our OBP algorithms over GR. Without the use of reconnaissance agent and communications among agents, our OBP algorithm can still outperform GR+EO algorithm. Acknowledgments {#acknowledgments .unnumbered} =============== This work was supported by NUS Computer Center through the use of its high performance computing facilities.

--- abstract: 'A review is given of the different ways to describe $\bar{p}p$ scattering. Next the Nijmegen partial-wave analyses of the $\bar{p}p$ data as well as the corresponding Nijmegen $\overline{N}\!N$ database are discussed. These partial-wave analyses are finally used as a tool to construct a better $\overline{N}\!N$ potential model and also to clarify questions raised in the literature.' author: - | [*J.J. de Swart and Th.A. Rijken*]{}\ Institute for Theoretical Physics, University of Nijmegen\ Nijmegen, The Netherlands\ and\ [*R. Timmermans*]{}\ Institute for Theoretical Physics, University of Nijmegen\ Nijmegen, The Netherlands\ and\ Theoretical Division, Los Alamos National Laboratory\ Los Alamos, NM 87545, USA title: | [$\bar{p}p$ partial-wave analysis and\ $\overline{N}\!N$ potentials]{} --- At the end of this tex-file we put the 4 figures (that belong to this manuscript), compressed with “uufiles”. Read there how to unpack this file. You can print the tex-file and the 4 ps-files (figures) separately, but using “dvips” you get the figures printed at the right place in the text. —————————————————————- Invited talk given by J.J. de Swart at the Second Biennial Workshop on\ Nucleon-Antinucleon Physics, NAN ’93, Moscow, Russia, September 1993. Introduction {#I} ============ Starting with an antiproton beam directed on a proton target many reactions are possible. First of all is the elastic scattering $\bar{p}p \rightarrow \bar{p}p$. For this reaction differential cross sections $\sigma_{el}(\theta)$, analyzing-power data $A_{el}(\theta)$ [@Kun88; @Ber89], and even some depolarization data $D_{yy}(\theta)$ [@Kun91] have been measured. We will discuss these extensively in this talk. The annihilation channel, $\bar{p}p \rightarrow$ mesons, is studied very intensively by theorists as well as by experimentalists. Many different reactions can be distinguished. For our purposes, however, only a global description will turn out to be sufficient. The charge-exchange reaction $\bar{p}p \rightarrow \bar{n}n$ has its threshold at $p_L = 99$ MeV/c. Important is that in the one-boson-exchange (OBE) picture only charged mesons can be exchanged. The most important of these are the $\pi^{\pm}$ and $\rho^{\pm}$ mesons. The study of this reaction allowed us to determine the coupling constant of the charged pion to the nucleons [@Tim91b]. Recently, excellent data for this charge-exchange reaction has been obtained at LEAR for the differential cross section $\sigma_{ce}(\theta)$ and for the analyzing power $A_{ce}(\theta)$ [@Bir90]. Very recently, even charge-exchange depolarization data have become available [@Bir93]. Excellent data are also available for the strangeness-exchange reaction $\bar{p}p \rightarrow \bar{\Lambda}\Lambda$ with threshold $p_L = 1.435$ GeV/c [@Bar87]. More data for this reaction are forthcoming as well as data for the other strangeness-exchange reactions $\bar{p}p \rightarrow \bar{\Lambda}\Sigma, \bar{\Sigma}\Lambda$ [@Bar90] and $\bar{\Sigma}\Sigma$. These reactions are very important for the precise determination of the $\Lambda N\!K$ and $\Sigma N\!K$ coupling constants and combining these with the $N\!N\pi$ coupling constant gives us information about flavor SU(3) [@Tim91a]. Antinucleon-nucleon potentials {#II} ============================== It is customary to start with some meson-theoretic $N\!N$ potential and then apply the $G$-parity transformation [@Pai52] to get the corresponding $\overline{N}\!N$ potential. This is a straightforward, but rather cumbersome procedure. When you ask people about details, then most people must confess that they do not know. We would like to point out that just charge conjugation, together with charge independence, without actually combining them to $G$, is sufficient for our purposes. To understand this, let us look at the $ppm^{0}$ vertex describing the coupling of a neutral meson $m^{0}$ to the proton $p$ with a coupling constant $g$. When we apply charge conjugation $C$ we have $$\bar{p} = Cp \rule{1cm}{0mm} {\rm and} \rule{1cm}{0mm} \overline{m^{0}} = Cm^{0}\ ,$$ and we describe now the $\bar{p}\bar{p}\overline{m^{0}}$ vertex. For nonstrange, neutral mesons $m^{0}$ one can define the charge parity $\eta_{c}$ by $\overline{m^{0}}=\eta_{c} m^{0}$. Charge-conjugation invariance of the interaction Lagrangian describing this $ppm^{0}$ vertex requires that the coupling constant $g$ of the meson $m^{0}$ to the proton $p$ is equal to the coupling constant of the antimeson $\overline{m^{0}}$ to the antiproton $\bar{p}$. The coupling constant $\bar{g}$ of the meson $m^{0}$ to the antiproton is then given by $$\bar{g} = \eta_{c} g\ .$$ For mesons of the type $Q\overline{Q}$, with relative orbital angular momentum $L$ and total spin $S$ the charge parity is $$\eta_{c} = (-)^{L+S}\ .$$ Therefore the pseudoscalar ($^{1}S_{0}$) mesons have $J^{PC} = 0^{-+}$, the vector ($^{3}S_{1}$) mesons have $J^{PC} = 1^{--}$, the scalar ($^{3}P_{0}$) mesons have $J^{PC} = 0^{++}$, etc. We see that from the important mesons only the vector mesons have negative charge parity and therefore the coupling constants of the vector mesons change sign when going from the nucleons to the antinucleons. In the OBE picture the $pp$ potential $V(pp)$ is the sum of the exchanges of the pseudoscalar meson $\pi$, the vector mesons $\rho$ and $\omega$, the scalar meson $\varepsilon(760)$, etc. That is $$V(pp) = V_{\pi} + V_{\rho} + V_{\omega} + V_{\varepsilon} + \ldots$$ The potential $V(\bar{p}p)$ described in the same OBE picture is then given by $$V(\bar{p}p) = V_{\pi} - V_{\rho} - V_{\omega} + V_{\varepsilon} + \ldots$$ In these reactions only neutral mesons are exchanged. When we want to describe the charge-exchange reaction $\bar{p}p \rightarrow \bar{n}n$, then it is easiest to recall charge independence. Charge independence requires that the coupling constant $g_{c}$ of the charged meson to the nucleons is given by $g_{c} = g\sqrt{2}$, and to the antinucleons by $\bar{g}_{c} = \bar{g}\sqrt{2}$. The charge-exchange potential is therefore given by $$V_{ce} = 2(V_{\pi} - V_{\rho} + \ldots)\ .$$ The diagonal potential in the $\bar{n}n$ channel is, using charge independence, given by $V(\bar{n}n) = V(\bar{p}p)$. [*What can we learn from the $N\!N$ potentials about the $\overline{N}\!N$ potentials*]{}?\ The $pp$ central force is relatively weak due to the cancellations between the repulsive contribution of the vector mesons $\omega$ and $\rho$ and the attractive contribution of the scalar mesons $\varepsilon(760)$, etc. The $\bar{p}p$ central force is strongly attractive, because the vectors mesons have now an attractive contribution which adds coherently to the attractive contribution of the scalar mesons, giving a very strong overall central force. Also the tensor force in $N\!N$ is relatively weak, because $\pi$ and the important $\rho$ contributions have opposite sign. In the $\overline{N}\!N$ case these mesons add coherently again to give a very strong tensor force. This strong tensor force is responsible for the importance of the transitions $$\ ^3S_1 \leftrightarrow\ ^3D_1 \rule{1cm}{0mm} , \rule{1cm}{0mm} \ ^3P_2 \leftrightarrow\ ^3F_2 \rule{1cm}{0mm} , \rule{1cm}{0mm} \ ^3D_3 \leftrightarrow\ ^3G_3 \rule{1cm}{0mm} {\rm etc.}$$ Various models {#III} ============== Black-disk model {#III.1} ---------------- One of the simplest models for the description of the elastic and inelastic cross section is the black-disk model. This model gives $$\sigma_{el} = \sigma_{ann} = \frac{1}{2} \sigma_{T} = \pi R^{2}\ ,$$ where $R$ is the radius of the black disk. This relation is satisfied very approximately. It shows that the annihilation cross section predicts radii for the black disk which are energy dependent and pretty large. In the momentum interval 200 MeV/c $< p_L <$ 1 GeV/c this radius $R$ varies from more than 2 fm to about 1.4 fm. Boundary-condition model {#III.2} ------------------------ The boundary-condition model in $\overline{N}\!N$ was first introduced by M. Spergel in 1967 [@Spe67]. Later many more people used this now more than a quarter century old model (see  Refs. [@Dal77; @Del78]). The model is based on the observation that the interaction for large values of the radius is often well-known, while the interaction for small radii is very hard to describe. This problem is then solved by just specifying a boundary condition at $r=b$. For this boundary condition one takes the logarithmic derivative of the radial wave function at the boundary radius $b$ $$P=b \left( \frac{d\psi}{dr}/\psi \right)_{r=b}\ .$$ Outside this radius one assumes that the interaction can be described by a known potential $V_{L}$. This long-range interaction is made of meson exchanges as described in section \[II\] and it contains of course also the electromagnetic interaction. A nice, instructive example is the [*modified black disk*]{}, where $V_{L} = 0$ and $P = -ipb$. The $P$ matrix contains a negative imaginary part, implying absorption of flux at the boundary. The boundary $b$ is a measure for the annihilation radius. This modified black disk is specified by only one parameter: the radius $b$. When one looks how these boundary-condition models have been used, then one sees that in these extremely simple models every time only very few parameters have been introduced. The conclusion is that such a few-parameter model can fit possibly some data, but it will never be able to fit all the available $\overline{N}\!N$ data, with the same set of only a few parameters. Optical-potential model {#III.3} ----------------------- The optical-potential models have become quite an industry in the $\overline{N}\!N$ community. The first such model was from R. Bryan and R. Phillips in 1968 [@Bry68]. In the optical-potential model the interaction between the antinucleon and the nucleon is described by a complex potential from $r=0$ to infinity. For the basic potential one takes a meson-theoretic potential, obtained from some known $N\!N$ potential by using the charge-conjugation operation. Then, in order to get annihilation, to this potential is added another complex potential, $$V(r) = (U-iW) f(r)\ .$$ Here $U$ and $W$ are constants and $f(r)$ is some radial function. This radial function can be the Woods-Saxon form, a Gaussian form, or even a square well. Let us give you a DO-IT-YOURSELF-KIT called: [*How to make your own optical potential?*]{} Instructions:\ 1. Look through the literature and decide which $N\!N$ potential your want to use.\ 2. Apply to this potential charge conjugation, so that you obtain the corresponding $\overline{N}\!N$ potential.\ 3. Pick your favored functional form for $f(r)$. This will contain a range parameter $b$. After you have made this choice, find some arguments, which sound like QCD, to justify this chosen form.\ 4. Pick one of the beautiful differential cross sections as measured by Eisenhandler  [@Eis76] and adjust $U$, $W$, and $b$ such that a reasonable fit (at least at sight) is obtained for this particular cross section.\ Your model is now a three-parameter model, which fits [*some*]{} of the data (at least the Eisenhandler data at one energy) reasonably well (at sight), but it cannot possibly fit [*all*]{} the $\overline{N}\!N$ data, because the model does not have enough freedom. After Bryan and Phillips many people have constructed similar models (see  Refs. [@Dov80; @Koh86; @Hip89; @Car91]). Also in Nijmegen we made such an optical-potential model, which we optimized by making a least-squares fit to our database, which contained at that time $N_{d} = 3309$ data. Because we actually performed a fit to all the $\overline{N}\!N$ data we think that we will have about the lowest $\chi^2$ of all the available two- or three-parameter optical-potential models. For our model $\chi_{\rm min}^{2} = 6\ 10^{9}$. This enormously large number is NOT a printing error, but just an expression of the total failure of such simple models. It is, therefore, astonishing to see that regularly new measurements from LEAR are compared to one or more of these few-parameter optical-potential models (see  Ref. [@Kun91]), as if something can be learned from such a comparison! There is only one group, the theory group of R. Vinh Mau in Paris, that has seriously tried to fit all available $\overline{N}\!N$ data with an optical-potential model. In 1982 they got a fit with $\chi^{2}/N_{d} = 2.8$, where they compared with the then available pre-LEAR data [@Cot82]. In 1991 they published an update [@Pig91], where they fitted now also the LEAR data. For the real part of the potential they took the $G$-conjugated Paris $N\!N$ potential [@Lac80]. Because the inner region of this $N\!N$ potential is treated totally phenomenologically it is impossible to take that over to $\overline{N}\!N$, so something has been done there and probably some extra parameters have been introduced. The imaginary part of the potential they write as $$\begin{aligned} W(r) & = & \left\{ \rule{0mm}{6mm} g_{c} (1+f_{c}T_{L})+g_{ss}(1+f_{ss}T_{L})\, \mbox{\boldmath $\sigma$}_{1} \cdot \mbox{\boldmath $\sigma$}_{2} \right. \\ && \left. +\, g_{T} S_{12} + g_{LS}\ {\bf L} \cdot {\bf S}\ \frac{1}{4m^{2}r} \frac{d}{dr} \right\} \frac{K_{0}(2mr)}{r} \ ,\end{aligned}$$ where $T_L$ is the lab kinetic energy and the parameters are the $g$’s and the $f$’s. For each isospin a set of 6 parameters is fitted, so that the imaginary part is described by about 12 real parameters. In total the Paris $\overline{N}\!N$ potential uses at least 12, possibly about 22, parameters. The correct number used is not so important, what is important, is that the number is much larger than 3. The Paris group do fit then to 2714 data and get $\chi^{2}/N_{d} = 6.7$. The quality of this fit is very hard to assess, because the Paris group did not try to make their own selection of the data, but tried to fit all the available data, many of which are contradictory. It would be interesting to see their fit to the Nijmegen $\bar{p}p$ database [@Tim93] (see section \[VIII\]), where all the contradictory sets have been removed. An important lesson could have been learned already in 1982 from this Paris work. An optical-potential model needs at least about 15 parameters to be able to give a reasonable fit to the $\overline{N}\!N$ data. This means that practically all few-parameter optical-potential models published after 1982 should have been rejected by the journals. Coupled-channels model {#III.4} ---------------------- Another way to introduce inelasticity in our formalism is to introduce explicitly couplings from the $\overline{N}\!N$ channels to annihilation channels. This was done in 1984 by P. Timmers  in the Nijmegen coupled-channels model: CC84 [@Tim84]. Fitting to the then available pre-LEAR data resulted in a quite satisfactory fit with $\chi^{2}/N_{d} = 1.39$. Several people have have later tried similar models [@Liu90; @Dal90]. An update of the old model CC84 was made in 1991 in Nijmegen in the thesis of R. Timmermans [@Tim91c]. This new coupled-channels model, which we would like to call the Nijmegen model CC93, gives $\chi^{2}/N_{d} = 1.58$, when fitted to $N_{d} = 3646$ data. We will come back to this model somewhat later. Antiproton-proton partial-wave analysis {#VII} ======================================= In Nijmegen we have for almost 15 years been busy with partial-wave analyses of the $N\!N$ data. We have now developed rather sophisticated and accurate methods to do these PWA’s [@Ber88; @Ber90; @Klo93]. A few years ago we realized that it was possible to do a PWA of all the available $\bar{p}p$ data in [*exactly the same way*]{} as our $N\!N$ PWA. Before this realization we always thought that such a PWA would be almost impossible in $\overline{N}\!N$. Luckily, it is not impossible. We will try to give a short description of our PWA [@Tim93]. In an energy-dependent partial-wave analysis one needs a model to describe the energy dependence of the various partial-wave amplitudes. Our model is a mixture of the boundary-condition model and the optical-potential model. We choose the boundary at $b=1.3$ fm. This value is determined by the width of the diffraction peak and cannot be chosen differently, without deteriorating the fit to the data. The long-range potential $V_{L}$ for $r>b$ is $$V_L = V_{\overline{N}\!N} + V_{C} + V_{M\!M}\ .$$ Here $V_{C}$ is the relativistic Coulomb potential, $V_{M\!M}$ the magnetic-moment interaction, and $V_{\overline{N}\!N}$ is the charge-conjugated Nijmegen $N\!N$ potential, Nijm78 [@Nag78]. We solve the relativistic Schrödinger equation [@Swa78] for each energy and for each partial wave, subject to the boundary condition $$P = b \left( \frac{d\psi}{dr}/\psi \right)_{r=b} \ ,$$ at $r=b$. This boundary condition may be energy dependent. To get the value of $P$ as a function of the energy we use for the spin-uncoupled waves (like $^{1}S_{0}$, $^{1}P_{1}$, $^{1}D_{2}$, $\ldots$ and $^{3}P_{0}$, $^{3}P_{1}$, $^{3}D_{2}$, $\ldots$) the optical-potential picture. We take a square-well optical potential for $r \leq b$. This short-ranged potential $V_{S}$ we write as $$V_{S} = U_{S} - iW_{S}\ .$$ In this way we get in each partial wave and for each isospin the parameters $U_{S}$ and $W_{S}$. Using these potentials we can calculate easily the boundary condition $P$ and the scattering amplitudes. For example in all singlet waves $^{1}S_{0}$, $^{1}P_{1}$, $^{1}D_{2}$, $\ldots$, we get $U=0$ and $W \approx 100$ MeV. For the triplet waves we take $W$ independent of the isospin. The parameters for the $^{3}P_{0}$ wave are  $W=159 \pm 9$ MeV and independent of the isospin, and $U (I=0) = -132 \pm 9$ MeV and $U (I=1) = 178 \pm 19$ MeV. To describe all relevant partial waves we need in our $\overline{N}\!N$ PWA 30 parameters. In our fit to the data we use all available data in the momentum interval 119 MeV/c $< p_{L} <$ 923 MeV/c. The lowest momentum is determined by the fact that for lower momenta no data are available. The highest momentum is determined by several considerations. In $N\!N$ we use all data up to $T_{L} = 350$ MeV, which corresponds to $p_{L} = 810$ MeV/c. Because we wanted to include all the elastic backward cross sections of Alston-Garnjost  [@Als79], we need to go to $p_{L} = 923$ MeV/c which corresponds to $T_{L} = 454$ MeV. At this energy the potential description in $N\!N$ is still valid, and therefore we feel that also here our description must work at least up to this momentum. Our final dataset contains $N_{d}=3646$ experimental data. In our analyses we need to determine $N_{n}=113$ normalizations and $N_{p}=30$ parameters. This leads to the number $N_{df}$ of degrees of freedom $N_{df} = N_{d} - N_{n} - N_{p} = 3503$. When the dataset is a perfect statistical ensemble and when the model to describe the data is totally correct, then one expects for $\chi^2_{\rm min}$: $$\langle \chi^{2}_{\rm min} \rangle = N_{df} \pm \sqrt{2N_{df}}\ .$$ Thus expected is $\langle \chi^{2}_{\rm min}(\bar{p}p) \rangle/N_{df} = 1.000 \pm .024$.\ In our PWA we obtain $\chi^{2}_{\rm min}(\bar{p}p)/N_{df} = 1.085$. We see that we are about 3.5 standard deviations away from the expectation value. To get a feeling for these numbers let us compare with the $pp$ data and the $pp$ PWA. The number of data is now $N_{d} = 1787$ and we expect $$\langle \chi^{2}_{\rm min}(pp) \rangle /N_{df} = 1.000 \pm 0.035\ .$$ In the latest Nijmegen analysis, NijmPWA93 [@Klo93], we get $$\chi^{2}_{\rm min}(pp)/N_{df} = 1.108,$$ which is 3 standard deviations too high. We see that our $\bar{p}p$ analysis compares favorable with a similar analysis for the $pp$ data. This means therefore that we have a statistically rather good solution and also that this solution will be essentially correct. The Nijmegen ${\bf \overline{N}\! N}$ database {#VIII} ============================================== An essential ingredient in our successfully completed PWA, as well as an important product of this PWA, is the Nijmegen $\overline{N}\!N$ database [@Tim93]. As pointed out before, we use all data with $p_{L} < 925$ MeV/c or $T_{L} < 454$ MeV. This means that our momentum range is similar to the momentum range used in the Nijmegen $N\!N$ PWA’s. We will compare regularly with the $N\!N$ case to show that the same methods, which work well in $N\!N$, work also well in $\overline{N}\!N$ and that the results are also similar. The number of data $N_{d}$ in the various final datasets are $N_{d}(\bar{p}p) = 3543$, $N_{d}(pp) = 1787$, and $N_{d}(np) = 2514$. In the processes to come to these final datasets we had to reject data. We do not want to go into details [@Ber88] about what are the various criteria to remove data from the dataset. We would like to point out, however, that in $pp$ scattering there is a long history about which datasets are reliable, and which not. We did not invent the method of discarding incorrect data, we just followed common practice and used common sense. In the $\bar{p}p$ case we needed to reject 744 data, which is 17% of our final dataset. In the $pp$ case we discarded 292 data or 14% of the final dataset, and in the $np$ case we rejected 932 data, which amounts to 27% of the final dataset. It is clear that the $\bar{p}p$ case does not seem to be out of bounds. Of course, it is unfortunate that so many data have to be rejected, because these data represented many man-years of work and a lot of money and effort. However, when one wants to treat the data in a statistically correct manner, then often one cannot handle all datasets, but one must reject certain datasets. This does not mean that all these rejected datasets are “bad” data, it only means, that if we want to apply statistical methods, then, unfortunately, certain datasets cannot be used. ------------------------- ------ ------ ------ ------ LEAR rest LEAR rest $\sigma_{T},\sigma_{A}$ 124 - - 63 $\sigma(\theta)$ 281 2507 91 154 $A(\theta)$ 200 29 89 - $D$ 5 - 9 - 610 2536 189 217 ------------------------- ------ ------ ------ ------ : Number of elastic and charge-exchange data divided over various categories.[]{data-label="tab.I"} In Table \[tab.I\] we give the number of data points divided into elastic versus charge exchange, LEAR versus the rest, and total cross sections $\sigma_{T}$, annihilation cross sections $\sigma_{A}$, differential cross sections $\sigma(\theta)$, analyzing-power data $A(\theta)$, and depolarization data $D(\theta)$. This table give some interesting information.\ The most striking fact is that:\ [*Of the final dataset only $22$% of the data comes from LEAR.*]{}\ This is after 10 years operation of LEAR. Remember the promises (or were it boasts) from CERN, made before LEAR was built. They were something like: “Only one day running of LEAR will produce more scattering data then all other methods together.” Unfortunately, this promise of CERN did not work out. Also it is clear that LEAR has not given much valuable information about $\sigma(\theta)$ for the elastic reaction. A lot of the elastic $\sigma(\theta)$ data from LEAR unfortunately needed to be rejected [@Tim93]! This does not mean that LEAR did not produce beautiful data. Some of the charge-exchange data and the strangeness-exchange data are really of high quality. Another striking fact is the virtual absence of spin-transfer and spin-correlation data. For the elastic reaction below 925 MeV/c there are only 5 depolarizations measured with enormous errors [@Kun91]. Very recently, some depolarization data of good quality have become available for the charge-exchange reaction [@Bir93]. A valid question is therefore:\ [*Can one do a PWA of the $\bar{p}p$ data, when there are essentially no “spin data”?*]{}\ The answer is yes! The proof that it can be done lies in the fact that we actually produced a $\bar{p}p$ PWA with a very good $\chi^{2}/N_{d}$. We have also checked this at length in our $pp$ PWA’s. We convinced ourselves that a $pp$ PWA using only $\sigma(\theta)$ and $A(\theta)$ data gives a pretty good solution. Of course, adding spin-transfer and spin-correlation data was helpful and tightened the error bands. However, most spin-transfer and spin-correlation data in the $pp$ dataset actually did not give any additional information. Fits to the data {#IX} ================ It is of course impossible to show here how well the various experimental data are fitted. From our final $\chi^{2}/N_{d} = 1.085$ one can draw the conclusion, that almost every dataset will have a contribution to $\chi^{2}$ which is roughly equal to the number of data points as is required by statistics. Let us look at some of the experimental data. In Fig. \[fig.1\] we present total cross sections from PS172 [@Clo84]. The fit gives for these 75 data points $\chi^{2} = 88.4$. In the same Fig. \[fig.1\] one can also find 52 annihilation cross sections from PS173 [@Bru87]. These points contributed $\chi^{2} = 65.3$ to the total $\chi^{2}$. In Fig. \[fig.2\] we plot the elastic differential cross section $\sigma(\theta)$ at $p_L = 790$ MeV/c as measured in 1976 by Eisenhandler  [@Eis76] The 95 data points contribute $\chi^{2} = 101.5$. The vertical scale is logarithmic. The nice fit reflects the high quality of these pre-LEAR data. The differential cross section of the charge-exchange reaction $\bar{p}p \rightarrow \bar{n}n$ at $p_{L}=693$ MeV/c as measured by PS199 is given in Fig. \[fig.3\]. The 33 data contribute $\chi^{2} = 39.3$. This dataset can be considered important, because it is very constraining. One needs all partial waves up to $L=10$ to get a satisfactory fit to these data. Coupled-channels potential model {#X} ================================ Having finished our discussion of the Nijmegen $\bar{p}p$ partial-wave analysis we can look at the $\overline{N}\!N$ potentials. We decided to update the old coupled-channels model Nijmegen CC84 of Timmers  [@Tim84]. Because of our experience with the various datasets this was not very difficult, just very computer-time consuming. The result was the new Nijmegen CC93 model [@Tim91c]. In this model we treat the $\overline{N}\!N$ coupled channels on the particle basis. We therefore have a $\bar{p}p$ channel as well as a $\bar{n}n$ channel. This allows us to introduce the charge-independence breaking effects of the Coulomb interaction in the $\bar{p}p$ channel and of the mass differences between the proton and neutron as well as between the exchanged $\pi^{0}$ and $\pi^{\pm}$. These $\overline{N}\!N$ channels are coupled to annihilation channels. We assume here that annihilation can happen only into two fictitious mesons; into one pair of mesons with total mass 1700 MeV/c$^{2}$, and into another pair with total mass 700 MeV/c$^{2}$. Moreover, we assume that these annihilation channels appear in both isospins $I=0$ as well as $I=1$. We end up with 6 coupled channels for each of the $\bar{p}p$ channels: $^{1}S_{0}$, $^{1}P_{1}$, $^{1}D_{2}$, $^{1}F_{3}$, etc.and $^{3}P_{0}$, $^{3}P_{1}$, $^{3}D_{2}$, $^{3}F_{3}$, etc. Due to the tensor force we end up with 12 coupled channels for each of the $\bar{p}p$ coupled channels: $^3S_1+\,^3D_1$, $^3P_2+\,^3F_2$, $^3D_3+\,^3G_3$, etc. We use the relativistic Schrödinger equation in coordinate space. The interaction is then described by either a $6\times 6$ or a $12\times12$ potential matrix $$V = \left( \begin{array}{cc} V_{\overline{N}\!N} & V_{\!A} \\[2mm] \widetilde{V}_{\!A} & 0 \end{array} \right) \ .$$ The $2\times 2$ (or $4\times 4$) submatrix $V_{\overline{N}\!N}$ we write as $$V_{\overline{N}\!N} = V_{C} + V_{M\!M} + V_{O\!B\!E}\ ,$$ where for $V_{C}$ we use the relativistic Coulomb potential, $V_{M\!M}$ describes the magnetic-moment interaction, and for $V_{O\!B\!E}$ we use the charge-conjugated Nijmegen $N\!N$ potential Nijm78 [@Nag78]. We have assumed that we may neglect the diagonal interaction in the annihilation channels. The annihilation potential $V_{\!A}$ connects the $\overline{N}\!N$ channels to the two-meson annihilation channels. It is either a $2\times 4$ matrix or a $4\times 8$ matrix. This potential we write as $$V_{\!A}(r) = \left( V_{C} + V_{SS} \mbox{\boldmath $\sigma$}_{1} \cdot \mbox{\boldmath $\sigma$}_{2} + V_{T} S_{12} m_{a}r + V_{SO} {\bf L}\cdot {\bf S} \frac{1}{m_{a}^{2}r} \frac{d}{dr} \right) \frac{1}{1+e^{m_ar}}\ .$$ The factor $m_{a}r$ is introduced in the tensor force to make this potential identically zero at the origin. Here $m_{a}$ is the mass of the meson (either 850 MeV/c$^{2}$ or 350 MeV/c$^{2}$). This annihilation potential depends on the spin structure of the initial state. For each isospin and for each meson channel five parameters are introduced: $V_{C}$, $V_{SS}$, $V_{T}$, $V_{SO}$, and $m_a$. This gives a model with in total $4\times 5=20$ parameters. These parameters can then be fitted to the $\overline{N}\!N$ data. Doing this we obtained $\chi^{2}/N_{d}=3.5$. It is clear, of course, that although the old Nijmegen soft-core potential Nijm78 is a pretty good $N\!N$ potential, it is definitely not the ultimate potential. We decided therefore to introduce now as extra parameters the coupling constants of the $\rho$, $\omega$, $\varepsilon(760)$, pomeron, and $a_{0}(980)$. Adding these parameters allowed us quite a drop in $\chi^{2}$. Now we reached $$\chi^{2}/N_{d} = 1.58 \ ,$$ with a total of 26 parameters. The reaction $\bar{p}p \rightarrow \bar{\Lambda}\Lambda$ {#XI} ======================================================== It is perhaps not superfluous to point out here, that we also made a PWA of the strangeness-exchange reaction $\bar{p}p \rightarrow \bar{\Lambda} \Lambda$ [@Tim91a; @Tim92]. Fitting the $N_{d}=142$ data, we get $\chi^{2}_{\rm min}/N_{d}=1.027$. The first theoretical treatments of this reaction were by F. Tabakin and R.A. Eisenstein [@Tab85] and independently by P. Timmers in his thesis [@Tim85]. Many other treatments of this reaction can be found (see [*e.g.*]{} Refs. [@Nis85; @Koh86a; @Fur87; @Kro87; @Alb88; @LaF88; @Hai93]). In the meson-exchange models it is clear that next to $K(494)$ exchange, there is also the exchange of the vector meson $K^{*}(892)$. In Nijmegen we have been able to determine the $\Lambda N\!K$ coupling constant at the pole [@Tim91a]. We found $f^{2}_{\Lambda N\!K} = 0.071 \pm 0.007$. This value is in agreement with the value $f^{2}_{\Lambda N\!K}=0.0734$ used in the recent soft-core Nijmegen hyperon-nucleon potential [@Mae89]. When we determine also the mass of the exchanged pseudoscalar meson we find $m(K) = 480 \pm 60$ MeV in good agreement with the experimental value $m(K)=493.646(9)$ MeV. This shows that we are actually looking at the one-kaon-exchange mechanism in the reaction $\bar{p}p \rightarrow \bar{\Lambda}\Lambda$. When the data for the reactions $\bar{p}p \rightarrow \bar{\Lambda}\Sigma$ and $\bar{\Sigma}\Lambda$ are available, then also the $\Sigma N\!K$ coupling constant can be determined. When this can be done with sufficient accuracy, then information about the SU(3) ratio $\alpha = F/(F+D)$ can be obtained, and SU(3) for these coupling constants can then actually be studied. PWA as a TOOL {#concl} ============= We have presented here some of the results of the first, energy-dependent partial-wave analysis of the elastic and charge-exchange $\bar{p}p$ scattering data [@Tim93]. We also discussed the Nijmegen $\overline{N}\!N$ dataset, where we removed the contradictory or otherwise not so good data from the world $\bar{p}p$ dataset. The main reason that we have been able to perform a PWA of the $\bar{p}p$ scattering data is that practically all partial-wave amplitudes are dominated by the potential outside $r=1.3$ fm. This long range potential consists of the electromagnetic potential, the OPE potential, and the exchange potentials of the mesons like $\rho,\ \omega,\ \varepsilon$, etc. This long-range potential is therefore well known. In our PWA of the $N\!N$ scattering data [@Klo93] it was noticed by us that the long-range potential in the $N\!N$ case dominated the $N\!N$ partial-wave scattering amplitudes. In the $\bar{p}p$ case the long-range potential is much stronger (see section \[II\]), and the dominance in the $\bar{p}p$ case is therefore more marked. One could formulate this the following way. The important $\bar{p}p$ partial-wave amplitudes are “$\pi,\ \rho,\ \varepsilon$, and $\omega$ dominated.” This gives the most important energy dependence of these amplitudes. The slower energy dependence due to the short-range interaction can easily be parametrized. A second reason for the successful PWA is the availability and easy access to computers, because the methods used are very computer intensive. We want to stress the fact that our multienergy PWA can now be used as a [**tool**]{}. This tool allows us first of all to judge the quality of a particular dataset. This enabled to us to set up the Nijmegen $\overline{N}\!N$ database. Secondly, it can be used in the study of the $\overline{N}\!N$ interaction. To demonstrate these things let us look at the Meeting Report of the Archamps meeting from October 1991 [@Bra93]. Beforehand the participants were asked to discuss at the meeting such questions as:\ [*What is the evidence for one-pion exchange in the $\overline{N}\!N$ interaction?*]{}\ In Nijmegen we determined [@Tim91b; @Tim93], using the PWA as a tool, the $N\!N\pi$ coupling constant for charged pions from the data of the charge-exchange reaction $\bar{p}p \rightarrow \bar{n}n$. We found [@Tim93] $$f_{c}^{2} = 0.0732 \pm 0.0011\ .$$ This is only 64 standard deviations away from zero!! Using analogous techniques we could also determine this coupling constant for charged pions in our analyses of the $np$ scattering data. We found there [@Sto93] $$f_{c}^{2} = 0.0748 \pm 0.0003\ .$$ The same coupling constant can also be seen in analyses of the $\pi^{\pm}p$ scattering data. There the VPI&SU group finds $f_{c}^{2} = 0.0735\pm 0.0015$ [@Arn90]. In $pp$ scattering we have determined the $pp\pi^{0}$ coupling constant. Our latest determination gives [@Sto93] $$f_{p}^{2} = 0.0745 \pm 0.0006\ .$$ The nice agreement between these different values shows\ (1) the charge independence for these coupling constants and its shows that\ (2) the presence of OPE in the $\overline{N}\!N$ interaction is a 64 s.d. effect.\ [*What more evidence does one wants?*]{}\ We also played around with the pion masses. In $N\!N$ scattering we were able to determine the masses of the $\pi^{0}$ and $\pi^{\pm}$. We found there $m_{\pi^0}=135.6(1.3)$ and $m_{\pi^\pm}=139.4(1.0)$ MeV/c$^2$, to be compared to the particle-data values\ $m_{\pi^0}=134.9739$ and $m_{\pi^\pm}=139.56755$ MeV/c$^2$. We did not try to determine these masses again in $\overline{N}\!N$ scattering. However, we think we could have. We checked that changing the correct pion masses to an averaged $\pi$-mass raised our $\chi^{2}_{\rm min}(\bar{p}p)$ with 9. Another question posed before that meeting was [*“What is the evidence for the $G$-parity rule?”*]{} In our determination of the $N\!N\pi$ coupling constant in the charge-exchange reaction this G-parity rule was of course implicitly assumed. Our determination of $f_{c}^{2}$ and its agreement with the expected value can therefore be seen as a proof of this rule for pion exchange. When one looks through the literature one finds several, what we think, artificially created problems. Why is this done? Only to get beamtime? One of such problems is the statement: “The OBE model does not work.” We would like to point out that the OBE model works excellently [@Tim91c]. Other examples can be found in the already mentioned Archamps Meeting Report [@Bra93]. The authors of this report claim that the charge-exchange differential cross sections at low energy pose a [*challenge for every model*]{}. Let us look at those data. Contrary to what is stated in the Meeting Report these data are a part of our dataset, so we have sufficient knowledge to discuss them. The discussion concerns data of PS173 [@Bru86b]. At four momenta the differential cross section for $\bar{p}p \rightarrow \bar{n}n$ was measured. The results of our PWA for these measurements are:\ At $p_{L} = 183$ MeV/c there are 13 $d\sigma_{ce}/d\Omega$ data. 4 of these data are rejected because each of them contributes more than 9 to our $\chi^{2}$. This is the three-standard-deviation rule. The remaining 9 data contribute $\chi^{2}=8.3$.\ At $p_{L}=287$ MeV/c there are 14 $d\sigma_{ce}/d\Omega$ data, where 1 of these data points is discarded because it contributes more than 9 to our $\chi^{2}$. The remaining 13 data contribute $\chi^{2}=24.0$.\ At $p_{L}=505$ MeV/c there are 14 $d\sigma_{ce}/d\Omega$ data. One of them is discarded because of its too large $\chi^{2}$ contribution. The remaining 13 data contribute $\chi^{2}=30.1$.\ At $p_{L}=590$ MeV/c there are 15 $d\sigma_{ce}/d\Omega$ data, where 2 of them are discarded. The remaining 13 data points contribute $\chi^{2}=32.8$. What can we conclude? At the lowest momentum we rejected 30% of the data and the remaining dataset is then OK. However, at the other three momenta we find rather large contributions to $\chi^{2}$. A dataset of 13 data is, according to the three-standard-deviation rule, not allowed to contribute more than $\chi^{2}_{\rm max}=31.7$ to the $\chi^{2}_{\rm min}$ of our database. This means that we really should reject the data at $p_{L} = 590$ MeV/c. When we combine the 4 datasets to one dataset with 47 data points, we see that these data points contribute $\chi^{2}=95.2$ to $\chi^{2}_{\rm min}$ of our database. The rule says that a set of 47 data may not have a $\chi^{2}$-contribution larger than 78.5. This means that this whole dataset should be rejected. The only reason, that these dubious data are still contained in the Nijmegen $\overline{N}\!N$ database and not discarded, is that there are no other charge-exchange data at such low momenta. Our philosophy here was that these imperfect data are perhaps better than no data at all. The authors of the Archamps Meeting Report [@Bra93], two experimentalists and a phenomenologist, are obviously incorrect. Our PWA shows clearly that these data cannot pose “a challenge for every model,” because these data should really be discarded! Another [*challenge for models*]{} seems to be that [*“the strangeness-exchange reaction*]{} $\bar{p}p \rightarrow \bar{\Lambda}\Lambda$ [*takes place in almost pure triplet states.”*]{} Let us look for a moment in more detail at the beautiful data of PS185 [@Bar87]. These data have been studied by many people. In Nijmegen we performed also a PWA of these data [@Tim91a; @Tim92]. It is very clear from our PWA that in this reaction the tensor force plays a dominant role. The tensor force acts only in triplet waves. These triplet waves make up the bulk of the cross section. This result has been confirmed by several groups and clearly this is not a challenge, but only a case of strong tensor forces. In section \[II\] we already explained the reason for such strong tensor forces. A big deal is often made of the $\rho$ parameter, the real-to-imaginary ratio of the forward scattering amplitude. The extraction of this parameter from the available experimental data is based on a rather shaky theory and on not much better data, polluted by Molière scattering and, in our opinion, the underestimation of systematic errors. When we look for example at the seven $\rho$ determinations by PS173 [@Bru85] then we note that this group has published only at four of these energies the corresponding $d\sigma/d\Omega$ data [@Bru86a]! In our PWA we discard these data at three of the energies. We feel therefore strongly, that the $\rho$ determinations by PS173 should clearly be discarded and very probably the errors on the determinations by PS172 [@Lin87] should be enlarged considerably. This leads to the simple picture as shown in Fig. \[fig.4\]. Another curious trend is the direct comparison between predictions of meson-exchange models and of simple quark-gluon models for the strangeness-exchange reaction. There are even serious proposals [@Hai92; @Alb93] for experiments to distinguish between these models: it is proposed to measure the spin transfer in $\bar{p}p \rightarrow \bar{\Lambda}\Lambda$. Let us make it clear from the outset, that we believe that [**all**]{} data must eventually be explained in terms of quark-gluon exchanges, because this is the underlying theory. However, for the analogous $N\!N$ interaction one has unfortunately not yet succeeded to give a proper explanation of the meson-exchange mechanism in terms of quarks and gluons exchanges. The theory is not so advanced yet. In the $\overline{N}\!N$ reactions we are of course in exactly the same situation. Using our PWA as a tool we determined the $\Lambda N\!K$ coupling constant [**and**]{} the mass of the exchanged kaon. This way we established beyond [**any**]{} doubt that the one-kaon-exchange potential is present in the transition potential and that again the tensor potential dominates. It is absolutely not necessary to measure the spin transfer to distinguish between the $K(494)$- and $K^{*}(892)$-exchange picture and a simple quark-gluon-exchange picture. This distinction has already been made using our PWA as a tool and using just the differential cross sections and polarizations. It has become a fad to promote the measurements of spin-transfer and spin-correlation data, as if these data will solve all our troubles. For example in the already often mentioned Archamps Meeting Report [@Bra93] one can read that the longitudinal spin transfer is obviously a favorite of one of its authors. Somewhere else in the same report one finds the question: “Which new spin measurement would be crucial to confirm or rule out present models?” The answer to this last question is of course: None! When one measures differential cross sections, polarizations, spin transfers, or spin correlations, carefully enough, none of the present models will fit these new data, but adjustments will be made in the models in such a way that they do fit the data again. Physics is hard work from experimentalists as well as from theorists. One needs many and varied data and one single experiment has only a marginal influence. [**Acknowledgments**]{}\ Part of this work was included in the research program of the Stichting voor Fundamenteel Onderzoek der Materie (FOM) with financial support from the Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO). [99]{} R.A. Kunne [*et al*]{}., Phys. Lett. B [**206**]{}, 557 (1988); Nucl. Phys. [**B323**]{}, 1 (1989). R. Bertini [*et al*]{}., Phys. Lett. B [**228**]{}, 531 (1989); F. Perrot [*et al*]{}., [*ibid*]{}. [**261**]{}, 188 (1991). R.A. Kunne [*et al*]{}., Phys. Lett. B [**261**]{}, 191 (1991). R. Timmermans, Th.A. Rijken, and J.J. de Swart, Phys. Rev. Lett. [**67**]{}, 1074 (1991). R. Birsa [*et al*]{}., Phys. Lett. B [**246**]{}, 267 (1990); [*ibid*]{}. [**273**]{}, 533 (1991). R. Birsa [*et al*]{}., Phys. Lett. B [**302**]{}, 517 (1993). P.D. Barnes [*et al*]{}., Phys. Lett. B [**189**]{}, 249 (1987); [*ibid*]{}. [**199**]{}, 147 (1987); [**229**]{}, 432 (1989); Nucl. Phys. [**A526**]{}, 575 (1991). P.D. Barnes [*et al*]{}., Phys. Lett. B [**246**]{}, 273 (1990). R. Timmermans, Th.A. Rijken, and J.J. de Swart, Phys. Lett. B [**257**]{}, 227 (1991). A. Pais and R. Jost, Phys. Rev. [**87**]{}, 871 (1952); C. Goebel, Phys. Rev. [**103**]{}, 258 (1956); T.D. Lee and C.N. Yang, Il Nuovo Cimento [**3**]{}, 749 (1956). M.S. Spergel, Il Nuovo Cimento [**47A**]{}, 410 (1967). O.D. Dal’karov and F. Myhrer, Il Nuovo Cimento [**40A**]{}, 152 (1977). A. Delville, P. Jasselette, and J. Vandermeulen, Am. J. Phys. [**46**]{}, 907 (1978). R.A. Bryan and R.J.N. Phillips, Nucl. Phys. [**B5**]{}, 201 (1968); [*ibid*]{}. [**B7**]{}, 481(E) (1968). E. Eisenhandler [*et al*]{}., Nucl. Phys. [**B113**]{}, 1 (1976). C.B. Dover and J.-M. Richard, Phys. Rev. C [**21**]{}, 1466 (1980); [*ibid*]{}. [**25**]{}, 1952 (1982). M. Kohno and W. Weise, Nucl. Phys. [**A454**]{}, 429 (1986). T. Hippchen, K. Holinde, and W. Plessas, Phys. Rev. C [**39**]{}, 761 (1989). J. Carbonell, O.D. Dal’karov, K.V. Protasov, and I.S. Shapiro, Nucl. Phys. [**A535**]{}, 651 (1991). J. Côté, M. Lacombe, B. Loiseau, B. Moussallam, and R. Vinh Mau, Phys. Rev. Lett. [**48**]{}, 1319 (1982). M. Pignone, M. Lacombe, B. Loiseau, and R. Vinh Mau, Phys. Rev. Lett. [**67**]{}, 2423 (1991). M. Lacombe, B. Loiseau, J.-M. Richard, R. Vinh Mau, J. Côté, P. Pirès, and R. de Tourreil, Phys. Rev. C [**21**]{}, 861 (1980). P.H. Timmers, W.A. van der Sanden, and J.J. de Swart, Phys. Rev. D [**29**]{}, 1928 (1984). G.Q. Liu and F. Tabakin, Phys. Rev. C [**41**]{}, 665 (1990). O.D. Dal’karov, J. Carbonell, and K.V. Protasov, Sov. J. Nucl. Phys. [**52**]{}, 1052 (1990), (Yad. Fiz. [**52**]{}, 1670 (1990)). R. Timmermans, Ph.D. thesis, University of Nijmegen (1991); R. Timmermans, Th. A. Rijken, and J.J. de Swart, in preparation. J.R. Bergervoet, P.C. van Campen, W.A. van der Sanden, and J.J. de Swart, Phys. Rev. C [**38**]{}, 15 (1988). J.R. Bergervoet, P.C. van Campen, R.A.M. Klomp, J.-L. de Kok, T.A. Rijken, V.G.J. Stoks, and J.J. de Swart, Phys. Rev. C [**41**]{}, 1435 (1990). V.G.J. Stoks, R.A.M. Klomp, M.C.M. Rentmeester, and J.J. de Swart, Phys. Rev. C [**48**]{}, 792 (1993). R. Timmermans, Th.A. Rijken, and J.J. de Swart, submitted for publication. M.M. Nagels, T.A. Rijken, and J.J. de Swart, Phys. Rev. D [**17**]{}, 768 (1978). J.J. de Swart and M.M. Nagels, Fortschr. Phys. [**26**]{}, 215 (1978). M. Alston-Garnjost [*et al*]{}., Phys. Rev. Lett. [**43**]{}, 1901 (1979). A.S. Clough [*et al*]{}., Phys. Lett. B [**146**]{}, 299 (1984); D.V. Bugg [*et al*]{}., Phys. Lett. B [**194**]{}, 563 (1987). W. Brückner [*et al*]{}., Phys. Lett. B [**197**]{}, 463 (1987); [*ibid*]{}. [**199**]{}, 596(E) (1987); Zeit. Phys. [**A335**]{}, 217 (1990). R. Timmermans, Th.A. Rijken, and J.J. de Swart, Nucl. Phys. [**A479**]{}, 383c (1988); Phys. Rev. D [**45**]{}, 2288 (1992). F. Tabakin and R.A. Eisenstein, Phys. Rev. C [**31**]{}, 1857 (1985). P.H. Timmers, Ph.D. thesis, University of Nijmegen (1985). J.A. Niskanen, Helsinki preprint HU-TFT-85-28. M. Kohno and W. Weise, Phys. Lett. B [**179**]{}, 15 (1986); Phys. Lett. B [**206**]{}, 584 (1988); Nucl. Phys. [**A479**]{}, 433c (1988). S. Furui and A. Faessler, Nucl. Phys. [**A468**]{}, 669 (1987). P. Kroll and W. Schweiger, Nucl. Phys. [**A474**]{}, 608 (1987). M.A. Alberg, E.M. Henley, and L. Wilets, Phys. Rev. C [**38**]{}, 1506 (1988); Z. Phys. [**A331**]{}, 207 (1988). P. LaFrance, B. Loiseau, and R. Vinh Mau, Phys. Lett. B [**214**]{}, 317 (1988); P. LaFrance and B. Loiseau, Nucl. Phys. [**A528**]{}, 557 (1991). J. Haidenbauer, K. Holinde, and J. Speth, Nucl. Phys. [**A562**]{}, 317 (1993). P.M.M. Maessen, Th.A. Rijken, and J.J. de Swart, Phys. Rev. C [**40**]{}, 2226 (1989). F. Bradamante, R. Hess, and J.-M. Richard, Few-Body Systems [**14**]{}, 37 (1993). V. Stoks, R. Timmermans, and J.J. de Swart, Phys. Rev. C [**47**]{}, 512 (1993). R.A. Arndt, Z. Li, L.D. Roper, and R.L. Workman, Phys. Rev. Lett. [**65**]{}, 157 (1990). W. Brückner [*et al*]{}., Phys. Lett. B [**169**]{}, 302 (1986). W. Brückner [*et al*]{}., Phys. Lett. B [**158**]{}, 180 (1985). W. Brückner [*et al*]{}., Phys. Lett. B [**166**]{}, 113 (1986); Zeit. Phys. [**A335**]{}, 217 (1990). L. Linssen [*et al*]{}., Nucl. Phys. [**A469**]{}, 726 (1987). J. Haidenbauer, K. Holinde, V. Mull, and J. Speth, Phys. Lett. B [**291**]{}, 223 (1992). M.A. Alberg, E.M. Henley, L. Wilets, and P.D. Kunz, Seattle preprint DOE/ER/40427-31-N93.

--- abstract: '[**Amorphous systems have rapidly gained promise as novel platforms for topological matter. In this work we establish a scaling theory of amorphous topological phase transitions driven by the density of lattice points in two dimensions. By carrying out a finite-size scaling analysis of topological invariants averaged over discrete and continuum random geometries, we discover unique critical properties of Chern and $\mathbb{Z}_2$ glass transitions. Even for short-range hopping models the Chern glass phase may persist down to the fundamental lower bound given by the classical percolation threshold. While the topological indices accurately satisfy the postulated one-parameter scaling, they do not generally flow to the closest integer value in the thermodynamic limit. Furthermore, the value of the critical exponent describing the diverging localization length varies continuously along the phase boundary and is not fixed by the symmetry class of the Hamiltonian. We conclude that the critical behaviour of amorphous topological systems exhibit characteristic features not observed in disordered systems, motivating a wealth of new research directions.** ]{}' author: - 'Isac Sahlberg$^{1}$' - 'Alex Westström$^{2,3}$' - 'Kim Pöyhönen$^{2,4}$' - 'Teemu Ojanen$^{1,2}$' title: Topological phase transitions in glassy quantum matter --- [**Introduction–**]{} While topological classification of matter is in principle completely independent on the symmetry-breaking classification, a vast recent literature is devoted to the increasingly subtle interplay of topology and spatial order [@chiu]. However, by adopting a completely complementary starting point, a number of recent studies have identified amorphous systems *without* reference to a band structure as fruitful platforms for topological states [@shenoy; @mitchell; @xiao; @mansha; @bourne; @chern; @agarwala]. This crucial property sets amorphous topological systems apart from disordered and Anderson topological insulators [@li; @groth] where non-trivial topology relies on residual spatial order. Amorphous topological states are extremely appealing for two reasons. In contrast to crystalline topological states, they do not rely on the specific spatial distribution of their microscopic constituents, thus being exceptionally robust. Furthermore, the possibility to fabricate topological states through randomly located dopants could allow accessing a whole new class of designer topological systems. This aspect was recently highlighted in concrete proposals for amorphous topological superconductors [@poyh2] and insulators [@minarelli]. The existence of amorphous topological states has become a well-established fact with a rapidly growing number of novel proposals. ![ **a**-**c** illustrate the proliferation of topologically non-trivial wave functions through amorphous systems as the density of lattice sites is increased. The dots represent randomly distributed dopants that support electronic orbitals hybridizing with nearby sites. The red and orange regions represent the magnitude of wave functions with the largest spatial extent. Configuration **c** supports an extended state, thus defining a topologically non-trivial system. **d**: The course-grained geometry of large clusters of discrete percolation lattices is largely insensitive to the lattice-scale details shown in the inset. The square and triangular site percolation configurations shown in **e** give rise to statistically similar long-distance random geometry. **f**: The colored area represents the modulus of topological wavefunction $|\psi|$ on a critical percolating cluster on a square lattice. []{data-label="fig:fsurface"}](1schematic){width="0.99\linewidth"} Previously it has been shown that various hopping models with randomly distributed dopants will undergo a topological phase transition at sufficiently high density [@shenoy; @poyh2]. The purpose of this work is to establish a quantitative description of this phase transition in representative topological models with randomly-generated geometries in two dimensions. Our work highlights the interplay of physics of topological matter [@volovik; @bernevig1], the Anderson localization [@evers; @huck] and classical percolation theory [@stauffer]. In contrast to semiclassical or effective network studies [@kramer], our theory is based on fully quantum-mechanical evaluation of topological invariants of representative microscopic models averaged over random geometries and studying their finite-size scaling properties in the vicinity of the critical density. A topological phase transition in an amorphous system can be thought of as a proliferation of wave functions carrying topological indices through the random lattice as depicted in Fig. 1 a-c. We consider both continuum and discrete random geometries that are commonly studied in classical percolation theory. Discrete lattices with randomly occupied sites represent systems where random dopants have preferred absorption sites. Furthermore, when the correlation length of the underlying classical percolation problem is comparable or larger than the system size, discrete and continuum random lattices describe a similar large-scale geometry insensitive to short-range details, as illustrated in Fig. 1 d-f. We establish that the divergence of the localization length of the wave functions carrying topological quantum numbers satisfy the scaling behaviour $\xi=(p-p_c^Q)^{-\bar{\nu}}$, where $p$ denotes the probability of a lattice site to be occupied, $p_c^Q$ is the critical probability and $\bar{\nu}>0$ is a critical exponent. Continuum models are shown to obey a similar relation where the site occupation probability is replaced by density of particles $\rho$. Unexpectedly, the non-trivial phase for paradigmatic short range hopping models may survive down to the fundamental lower limit given by the classical site percolation threshold. Equally unexpectedly, the critical exponent $\bar{\nu}$ is non-universal and varies continuously within a symmetry class of the Hamiltonian. This behaviour is in sharp contrast to excessively studied topological transitions in disordered systems. Our findings stimulate a large number of research directions in the rapidly rising field of amorphous topological matter. [**Theoretical approach–**]{} The purpose of our work is to formulate a quantitative theory of topological phase transitions in amorphous 2d systems. We consider systems with broken time-reversal symmetry characterized by non-zero Chern numbers as well as time-reversal invariant systems classified by a $\mathbb{Z}_2$ invariant [@schnyder]. The minimal representative lattice models for these cases have two and four orbitals per lattice site. The minimal models are expected to capture a generic behaviour of the topological phase transition because amorphous systems do not exhibit spatial symmetries that would enforce additional band degeneracies. An effective tight-binding Hamiltonian for the Chern insulator can be expressed as $$H_1=\begin{pmatrix}\label{h1} (2-M)\delta_{ij}+T_{ij} & iT_{ij}e^{-i{\varphi}_{ij}} \\ iT_{ij}e^{i{\varphi}_{ij}} & -(2-M)\delta_{ij}-T_{ij} \end{pmatrix},$$ where $M$ is the onsite energy difference of the two orbitals in the units of characteristic hopping amplitude and $T_{ij}=-\frac12 e^{-r_{ij}/\eta}\theta(R-r_{ij})$ describes the spatial decay of the hopping amplitudes. Here $r_{ij}=|\mathbf{r}_i-\mathbf{r}_j|$ is the distance between sites $i,j$, the parameters $\eta, R$ describe the decay of hopping, and the phase factor is given by $e^{i{\varphi}_{ij}}=\frac{r_{ij}^x+ir_{ij}^y}{r_{ij}}$, where $r_{ij}^x = x_i-x_j$. Introducing two additional orbitals, the studied time-reversal invariant $\mathbb{Z}_2$ model can expressed as $$\label{h2} H_2=\begin{pmatrix} H_1 & \Gamma \\ \Gamma^\dagger & H_1^* \end{pmatrix},$$ and describes two time-reversed copies of $H_1$ coupled by the off-diagonal block $\Gamma=i\alpha\,\mathrm{Diag}\left[iT_{ij}e^{-i{\varphi}_{ij}}, iT_{ij}e^{i{\varphi}_{ij}}\right]$. This model can be thought of as an amorphous Bernevig-Hughes-Zhang (BHZ) model [@bernevig2] with a Rashba-type inversion-breaking term parametrized by $\alpha$. If the models are discretized on a square lattice with $\eta=\infty$ and $R$ just above the lattice constant, these models reduce to a paradigmatic two-band Chern insulator $H_1(k)=\sin k_x \sigma_x + \sin k_y \sigma_y+\left[2-M-\cos k_x -\cos k_y \right]\sigma_z$ and an inversion-broken BHZ model $H_2(k)=\mathrm{Diag}\left[ H_1(k),H_1(-k)^*\right]+\alpha \left[-\sin k_x \tau_y +\sin k_y \tau_x\sigma_z \right]$, where $\sigma_i$ and $\tau_i$ are two sets of Pauli matrices. {width="0.99\linewidth"} We will study the behaviour of the topological invariants of models and averaged over different random geometries as a function of the density of lattice sites, assuming that half of the orbitals are populated. To gain insight on the thermodynamic limit, we will study the finite-size scaling behaviour of the topological invariants. The invariants are calculated by the real-space Chern number and Bott index algorithms that require a single diagonalization of the Hamiltonian per configuration [@zhang2; @loring]. We will consider continuum problems with randomly distributed lattice sites characterized by a density per unit area $\rho$, as well as discrete random geometries familiar from the percolation theory shown in Fig. 1 d-f. Percolation theory studies the cluster properties on different lattices where the sites are independently occupied with probability $p$. A cluster in the discrete problem is defined as a collection of lattice points that are connected by a nearest-neighbour connecting path and in continuum problems as a set of points that belong to a union of site-centred discs of fixed radius. Above one dimension there generically exists a critical probability $0<p_c^{cl}<1$ or critical density $0<\rho_c^{cl}$ above which infinite systems contain an infinite cluster, defining a second-order phase transition. A course-grained long-distance random geometry generated by different percolation lattices is similar to continuum percolation problems. Indeed, as confirmed here explicitly, the critical properties for discrete and continuum random geometries are essentially similar. Our scaling hypothesis for the amorphous topological phase transition is formulated in terms of topological invariants averaged over random geometries. We propose that the averaged Chern number obeys the scaling $$\label{scale} \bar{\mathcal{C}}=f[(p-p_c^Q)L^{1/\bar{\nu}}]$$ as a function of the occupation probability and linear system size $L$. Here $f(x)$ is an a priori unknown scaling function which approaches to 0 (1) for $x\ll0$ ($x\gg0$). For continuum problems we postulate a similar expression with $p$ and $p_c^Q$ substituted by particles per unit area $\rho$ (in appropriate units) and its critical value $\rho_c^Q$. The scaling form has important ramifications. First, it implies that the phase transition is sharp in the thermodynamic limit, taking place at the critical value $p_c^Q$ which we call the topological quantum percolation threshold. Second, $\bar{\nu}$ is a critical exponent describing the diverging localization length of wave functions carrying the Chern number through the relation $\xi=(p-p_c^Q)^{-\bar{\nu}}$ near the transition. In addition, due to the connection between the Chern number and the Hall conductance, Eq. (\[scale\]) also describes the scaling of the Hall response (in the units of $e^2/h$) in amorphous systems. This provides a concrete connection between our theory and observables. The $\mathbb{Z}_2$ invariant can be evaluated through a configuration-averaged spin Bott index $\bar{\mathcal{C}}_s$ [@huang], for which we assume a similar scaling form. It should be stressed that while the scaling hypotheses for topological invariants are superficially similar to the scaling of percolation probability [@stauffer], they characterize the topology of quantum ground states of models , , which is conceptually completely independent on the classical percolation problem. [**Chern glass transition–**]{} We first consider the discrete random geometries. The topological phase diagrams as a function of density for square and triangular random lattices are illustrated in Fig. 2 **a**-**b**. The non-trivial phases in amorphous systems are protected by a mobility gap rather than a true energy gap. This is reflected in the fact that the gapped spectrum of a pristine system is rendered gapless for $p<1$ by randomly missing sites, as illustrated by the density of states (DOS) in Figs. 1 **c**-**f**. The location of the phase boundary between trivial and non-trivial phases and the corresponding critical filling $p_c^Q$ is seen to depend strongly on the value of the mass parameter $M$. Strikingly, for the optimal values of $M$ the topologically non-trivial state is seen to extend all the way down to the classical percolation threshold $p_c^{cl}$. While there is recent convincing evidence of a localization transition for nontopological quantum percolation models for $p<1$, it is thought to be possible only in the close vicinity of $p=1$ [@schubert1; @schubert2; @dillon]. Therefore it is remarkable that the topological phase may persist all the way down to the classical percolation threshold. At the site percolation threshold, at which point there is no longer a meaningful separation between bulk and edge, there appears a critical cluster characterized by a fractal dimension $d=1.9$. For the optimal values of $M$, the Chern glass model avoids Anderson localization even at $p_c^{cl}$. As illustrated in App. B of the Supplemental Information (SI), topological wave functions efficiently penetrate the bottlenecks of the critical cluster. It should be noted that $p_c^{cl}$ gives the lower bound for the existence of the topological phase in the thermodynamic limit since below that the system is disconnected and cannot support extended states for nearest-neighbour models. The signatures of a topological phase below $p_c^{cl}$ reflect finite-size effects and numerical accuracy. {width="0.99\linewidth"} The onset of the topological phase transition and its scaling behaviour is illustrated in Fig. 2 **g**-**j**. The method of calculating the configuration-averaged transition curves is explained in App. A in the SI. For a fixed $M$, the topological index exhibits very accurate finite-size scaling of the postulated form , implying that the localization length of the topologically non-trivial states diverges as $\xi_Q=(p-p_c^Q)^{-\bar{\nu}}$ near the critical filling $p_c^Q$. From the scaling collapse we can estimate the values of the critical filling $p_c^Q$ and critical exponent $\bar{\nu}$. While the phase diagram shows that $p_c^Q$ depends sensitively on $M$, the properties of Quantum Hall transitions in disordered systems [@huck; @kramer] might suggest that the critical exponent $\bar{\nu}$ could be universal in a symmetry class of Hamiltonians [@altland]. *However, this is not the case.* We observe that the value of $\bar{\nu}$ varies even by a factor of 1.6 as a function of $M$. As discussed in App. A, the relative accuracy of extracted $p_c^Q$ and $\bar{\nu}$ values is expected to be around $5\%-10\%$. This estimate follows from benchmarking our method of evaluating the transition curves for classical percolation problems with comparable system sizes and statistical sampling and comparing to values in the literature. In topological and classical problems the main source of error comes from the limited configuration sampling and is of the same order of magnitude. Therefore the numerical uncertainty is very small compared to the variation of $\bar{\nu}$ along the phase boundary and can be ruled out as a source for observed non-universality of $\bar{\nu}$. For the optimal $M$ values for which the topological threshold and the classical threshold agrees $p_c^Q\approx p_c^{cl}$, the extended non-trivial states are restricted to the critical cluster. Thus the linear extent of these states should be of the order of the diameter of the classical critical cluster and the value of $\bar{\nu}$ should agree with the correlation length exponent of classical percolation $\nu=4/3$. Indeed, our results reproduce this fact within numerical accuracy. The values of critical data for selected points are given in the table in Fig. \[fig:continuum\]. It is a remarkable feature of the Chern glass transition that the value of *a configuration-averaged Chern number is not universal at $p_c^Q$*, in particular it does not need to be a half-integer. This is in a striking contrast to the disordered quantum Hall transitions where the critical value is a half-integer and the Hall conductance flows to the closest integer multiple of $e^2/h$ in the thermodynamic limit [@pruisken]. An amorphous system can flow to $e^2/h$ even if the the initial value is, say, $0.3e^2/h$ as implied by Fig. 2 **g**. The topological properties of the continuum Chern glass are in qualitative agreement with the discrete models. First, to make a precise comparison with the classical disc percolation, we consider a continuum model without exponential decay ($\eta=\infty$) but with a constant hopping amplitude within the radius $R$. In Fig. 3 **a**-**c** we have plotted the topological phase diagram as a function of density $\rho$ and studied the scaling properties of the transition. We discover again that the non-trivial phase may reach the vicinity of the classical disc percolation threshold $\rho_c^{cl}$. The critical exponent $\bar{\nu}$ describing the diverging localization length through $\xi_Q=(\rho-\rho_c^Q)^{-\bar{\nu}}$, is also found to vary strongly along the phase boundary. However, at the optimal point where $\rho_c^{Q}\approx\rho_c^{cl}$, we again expect it to agree with the classical percolation result $\nu=4/3$ for the same reasons as in the discrete models. This expectation is confirmed within numerical accuracy. In Fig. 3 **d**-**f** we have plotted results for a physically more realistic exponentially attenuated hopping where the hard cutoff is taken to infinity $R=\infty$. In practice, the cutoff at $R$ is basically irrelevant when $R>4\eta$. The scaling behaviour in the exponential Chern glass model is the same as for the disc model. While the critical exponent does not exhibit universality, the discrete and continuum Chern glass models share qualitatively similar critical properties. [**$\boldsymbol {\mathbb{Z}_2}$ glass transition–**]{} Now we complement our results regarding the Chern glass transition by considering a $\mathbb{Z}_2$ model on a random square lattice with a nearest-neighbour hopping. The important technical difference to the Chern glass case is that now we have to work with four orbitals per lattice site and the evaluation of $\mathbb{Z}_2$ invariant the help of the spin Bott index [@huang] requires two diagonalizations per configuration. These facts make the study computationally substantially more expensive compared to the Chern glass case, thus we are restricted to smaller systems. Here we concentrate on the $\alpha\neq 0$ case since for $\alpha=0$ the model trivially reduces to the above studied Chern glass problem (mod 2). The phase diagram of the system is shown in Fig. 3 **g**, illustrating how the topological threshold of the $\mathbb{Z}_2$ transition is increased by the coupling between the time-reversed Chern blocks. Nevertheless, as shown in Figs. 3 **g**-**h**, the $\mathbb{Z}_2$ glass transition exhibits the same form of scaling near the critical density with a non-universal exponent $\bar{\nu}$. The uncertainty for the extracted values of $\bar{\nu}$ is much larger than for the Chern glass models due to the system size limitations. For example, we estimate that for the parameters in Fig. 3 **g** the uncertainty is of the order $\bar{\nu}=3.5\pm 1$. The studied $\mathbb{Z}_2$ glass phase diagram also highlights an essential qualitative difference from a phase diagram of a disordered BHZ model. A BHZ model with finite coupling between the Chern blocks $\alpha\neq 0$ and quenched disorder exhibits a well-documented metallic phase between topological and trivial insulating phases [@yamakage]. However, in the density-driven $\mathbb{Z}_2$ glass transition we do not observe signatures of such a metallic intermediate state. A metallic state would give rise to extended states on a fixed interval of $p$ and comprise a departure from the observed scaling form which implies that the strip of extended states (transition region) shrinks as the system size is increased. This fact further supports our findings that the topological properties of crystalline systems with quenched disorder are qualitatively different from those of amorphous systems. [**Discussion and outlook–**]{} Motivated by the rising interest in amorphous topological states, we introduced a scaling theory of density-driven topological phase transitions in glassy systems. Surprisingly, glassy topological systems support unique critical properties sharply different from extensively studied disordered systems. While we showed that glassy systems accurately obey the postulated finite-size scaling relation, the critical exponent describing the delocalization transition is not universal. Furthermore, the Hall conductance does not necessarily flow to the closest integer multiple of $e^2/h$ in the thermodynamic limit. Remarkably, the non-trivial Chern glass phase even for short-range hopping models may persist down to the ultimate lower bound, the classical percolation threshold, where the geometry of the random lattice becomes fractal. For the parameter values at which the topological and classical percolation thresholds meet, the critical exponent describing the delocalization transition is consistent with the correlation length exponent of 2d percolation. This seems to be the only universal aspect of the phase transition. While a numerical approach cannot provide a definite answer to what happens in the thermodynamic limit, we observe single parameter scaling in the accessible system sizes. There exists substantial evidence [@schubert1; @schubert2; @dillon] that localization in non-topological quantum percolation models in 2d is significantly distinct from the conventional one-parameter scaling theory of disordered systems [@abrahams]. Therefore it is to be expected that also the topological transition and critical properties of amorphous topological models have unique properties not encountered in disordered systems. Our findings open a whole new line of research in amorphous topological systems. To list a few topics, we mention a generalization of our results to higher dimensions, characterizing the fractal nature of topological states and transport properties near the critical density. Our work also provides a basis in generalizing the theory of topological semimetals to amorphous systems [@yang] . [16]{} C.-K. Chiu, J. C. Y. Teo, A. P. Schnyder, and S. Ryu, Rev. Mod. Phys. **88**, 035005 (2016). A. Agarwala and V.B. Shenoy, Phys. Rev. Lett. 118, 236402 (2017). N. P. Mitchell, L. M. Nash, D. Hexner, A. Turner, and W. T. M. Irvine, Nat. Phys. **14**, 380 (2018). M. Xiao and S. Fan, Phys. Rev. B **96**, 100202(R) (2017). S. Mansha and Y. D. Chong, Phys. Rev. B **96**, 121405(R) (2017). C. Bourne, E. Prodan, J. Phys. A: Math. Theor. **51** 235202 (2018). G.-W. Chern, arXiv:1809.10575. A. Agarwala, V. Juricic, B. Roy, arXiv:1902.00507. J. Li, R.-L. Chu, J. K. Jain, and S.-Q. Shen, Phys. Rev. Lett. **102**, 136806 (2009). C. W. Groth, M. Wimmer, A. R. Akhmerov, J. Tworzydlo, and C. W. J. Beenakker, Phys. Rev. Lett. **103**, 196805 (2009). K. Pöyhönen, I. Sahlberg, A. Westström and Teemu Ojanen, Nat. Comm. **9**, 2103 (2018). E. L. Minarelli, K. Pöyhönen, G. A. R. van Dalum, T. Ojanen, and Lars Fritz, arXix:1809.09578. G. E. Volovik, *The Universe in a Helium Droplet*, (Oxford University Press, 2003). B. A. Bernevig and T. L. Hughes, *Topological Insulators and Superconductors*, (Princeton University Press, 2013). F. Evers and A. D. Mirlin, Rev. Mod. Phys. **80**, 1355 (2008). B. Huckenstein, Rev. Mod. Phys. **67**, 357 (1995). D. Stauffer, A. Aharony, *Introduction to Percolation Theory, 2nd edition*, (Taylor & Francis, 1994). B. Kramer, T. Ohtsuki, S. Kettemann, Phys. Rep. **417**, 211 (2005). A. P. Schnyder, S. Ryu, A. Furusaki, and A. W. W. Ludwig, Phys. Rev. B **78**, 195125 (2008); B. A. Bernevig, T. L. Hughes, and S.-C. Zhang, Science **314**, 1757 (2006). Y. F. Zhang, Y. Y. Yang, Y. Ju, L. Sheng, D. N. Sheng, R. Shen, and D. Y. Xing, Chinese Phys. B **22**, 117312 (2013). T. A. Loring and M. B. Hastings, EPL **92** 67004 (2010). H. Huang, F. Liu, Phys. Rev. B **98**, 125130 (2018). A. Weiße, G. Wellein, A. Alvermann, and H. Fehske, Rev. Mod. Phys. **78**, 275 (2006). G. Schubert and H. Fehske, *Lecture Notes in Physics* **762**, pp. 135, Springer-Verlag (2008). G. Schubert and H. Fehske, Phys. Rev. B **77**, 245130 (2008). B. S. Dillon, H. Nakanishi, Eur. Phys. J. B **87**, 286 (2014). A. Altland and M. R. Zirnbauer, Phys. Rev. B **55**, 1142 (1997); A.M. M. Pruisken, Phys. Rev. Lett., **61**, 1297 (1988). J. Quintanilla, S. Torquato, and R. M. Ziff, J. Phys. A **33**, L399 (2000). A. Yamakage, K. Nomura, K.-I. Imura, and Y. Kuramoto, J. Phys. Soc. Jpn. **80**, 053703 (2011). E. Abrahams, P.W. Anderson, D.C. Licciardello and T.V. Ramakrishnan, Phys. Rev. Lett., **42**, 673 (1979). Y.-B. Yang, T. Qin, D.-L. Deng, L.-M. Duan, Y. Xu, arXix:1810.07710. Calculation of configuration-averaged topological transition curves {#a:transition} =================================================================== Discrete random geometries -------------------------- In this appendix we detail our method of calculating configuration-averaged Chern numbers $\bar{\mathcal{C}}(p)$ as a function of the occupation probability $p$ of a single lattice site. This procedure applies to $\mathbb{Z}_2$ transitions and classical percolation problems in essentially the same form. The key point is to first calculate the Chern data $\bar{\mathcal{C}}(n)$ as a function of the number of occupied sites $n$ randomly distributed in the lattice, and afterwards analytically deduce $\bar{\mathcal{C}}(p)$ for all $p$. This procedure has crucial advantages over directly sampling configurations with sites occupied by probability $p$. First of all, $p$, in contrast to $n$, is a continuous variable and fixing appropriate $p$ resolution for scaling studies is not a priori clear. Secondly, as made explicit below, $\bar{\mathcal{C}}(p)$ for a single $p$ value is obtained as a weighted average over $\bar{\mathcal{C}}(n)$ with several different $n$. This turns out to reduce statistical fluctuations in the transition curves and to allow more accurate evaluation of $p_c^Q$ and $\bar{\nu}$. This is especially important in the studied problem where evaluation of topological indices, unlike the percolation probability, requires diagonalization of the Hamiltonian, imposing limitations for practical system sizes. We benchmark our method by applying it to finite-size scaling of classical percolation problems and show that percolation thresholds $p_c$ and correlation length exponents $\nu$ can be extracted in good accuracy within the system sizes employed in our study. Assume we have acquired knowledge of the quantity $\bar{\mathcal {C}}(n)$ as a function of occupied sites $n$ on a finite lattice of $N=L^2$ sites. In practice, this is done by averaging the topological index over different configurations with the same $n$. In total, there are $2^N$ possible configurations with varying amounts of occupied sites $n$, and there are ${N\choose n}$ possible configurations for a fixed $n$. We would now like to know $\bar{\mathcal {C}}(p)$ for all $p \in [0,1]$. Given a probability $p$ for each lattice site to be occupied, the probability to get exactly $n$ occupied sites out of $N$ possible sites is given by the binomial distribution $P(n,p)$. Thus, the desired quantity $\bar{\mathcal {C}}(p)$ can be calculated using the following prescription: $$\label{C(p)toC(n)} \bar{\mathcal {C}}(p) = \sum_n P(n,p) \bar{\mathcal {C}}(n) = \sum_n {N\choose n} p^n (1-p)^{N-n} \bar{\mathcal {C}}(n) .$$ The validity of the procedure can be straightforwardly tested. The comparison of $\bar{\mathcal {C}}(p)$ calculated by applying shows that the method indeed reproduces the correct $\bar{\mathcal {C}}(p)$ curves, as well as drastically reduces the statistical fluctuations compared to sampling configurations with probability $p$ directly. Since the conversion from $\bar{\mathcal {C}}(n)$ is performed analytically, one can choose a much higher density for the probabilities without the need to perform additional diagonalization. In Fig. 4, we have compared the transition curves produced by either fixing $p$ directly, and beginning with even very course sampling of $\bar{\mathcal {C}}(n)$ and converting to $\bar{\mathcal {C}}(p)$ using . Note that all errors come from sampling from the finite set of configurations for fixed $n$—perfect knowledge of $\bar{\mathcal {C}}(n)$ would yield the *exact* transition curve $\bar{\mathcal {C}}(p)$. ![The calculation of the transition curves $\bar{\mathcal {C}}(p)$ is substantially more efficient by first calculating the data $\bar{\mathcal {C}}(n)$, and then transforming using Eq. , illustrated here. **a**: A rough sample average of $\bar{\mathcal {C}}(n)$ for $M=1.14$ on a $25\times 25$ square lattice, with a sample size of 50. **b**: The red dots represent a configuration average of $\bar{\mathcal {C}}(p)$ with a sample size of 1000, while the black graph is the transition curve obtained through the use of on the data in **a**. In **c** and **d**, the same method is illustrated for lattice percolation $\bar{\Pi}(p)$ on a $30\times 30$ square lattice, with the blue graph $\bar{\Pi}(n)$ and the red dots $\Pi(p)$ having a sample size of 100 and 500, respectively. **e**: The critical exponent calculated by collapsing two $\bar{\mathcal {C}}(p)$ curves for each $M$, as a function of the sample size. Each new set of configurations cumulatively increases the accuracy of the obtained curves, and hence the obtained critical values $p_c^Q, \bar{\nu}$. []{data-label="raw_vs_analytical_lattice"}](4lattice_perc){width="0.99\linewidth"} The method explained above for calculating the transition curves can be applied to $\mathbb{Z}_2$ systems and classical percolation by replacing $\bar{\mathcal {C}}(n)$ by a spin-Bott index $\bar{\mathcal {C}_s}(n)$, or $\bar{\Pi}(n)$, giving the probability of the existence of a percolating path connecting left and right edges of the sample. Applying gives access to $\bar{\mathcal {C}_s}(p)$ and $\bar{\Pi}(p)$. To benchmark our method of calculating transition curves, we apply the method to classical percolation problems. In Fig. 3 **j** in the main text we have collected our critical data. Comparison to the accurate reference values from the literature shows that the relative error in $p_c^{cl}$ and $\nu$ for the comparable system sizes available for topological studies is of the order of 5-10%. The results obtained for the critical exponents become more accurate as the sample size is increased, as the effects of the statistical sampling decrease, see Fig.  \[raw\_vs\_analytical\_lattice\] **e**. Since the main source of error for classical percolation and topological problems originates from the limited statistical sampling and both exhibit comparable fluctuations, we expect that the error in $p_c^Q, \bar{\nu}$ is of the same order. For classical problems we could go to much larger systems and obtain much more accurate data, but evaluation of topological indices is computationally much more expensive and we are limited to modest system sizes. Nevertheless, the above considerations suggest that we can reach good accuracy also in the topological data. Continuum random geometries --------------------------- For continuum models we need to carry out a configuration average of the topological index $\bar{\mathcal {C}}(\rho)$ as a function of density. In contrast to the above, we do not have a fixed underlying lattice where the lattice points are either occupied or unoccupied. First of all, we have two separate cases that we will consider: constant hopping in a finite radius $r \le R$, and hopping which decays exponentially $\sim e^{-r/\eta}$. We can denote the cutoff by $\theta (r-R) e^{-r/\eta}$, the former case corresponding to $\eta=\infty$, and the latter to $R=\infty$. We will first look at the former case; the latter will be handled analogously. Denote by $\rho$ the density $\rho = \frac{\mathrm{Sum\ of\ area\ of\ disks}}{\mathrm{Total\ area}} = \frac{n \pi R^2}{L^2} = D(n) \pi R^2$, where $n$ is the number of lattice points in our region of size $L^2$, and $D(n)=n/L^2$ is the particle number per system size. For our purposes, in order for $\rho$ to describe a meaningful density, it needs to depend on $R$. Given a large system with overall particle number density $D(n)$, if we choose a subsystem of size $L^2$, we will, on average, find $n = D(n) L^2$ lattice sites in it. Every lattice point is located either within the confines of the subsystem, or not, completely independently of all other points. What we have described is a Poisson distribution with the intensity $\lambda = D(n) L^2$. The probability to find $k$ lattice points in our system is $$P_\lambda(k) = \frac{\lambda^k e^{-\lambda}}{k!}, \quad k = 0,1,2,...$$ Thus, the quantity $\bar{\mathcal{C}}_R(D)$ for fixed particle number density $D$ and radius $R$ can be calculated using $$\label{C(rho)toC(n)} \bar{\mathcal{C}}_R(D) = \sum_n P_{\lambda(D)}(n) \bar{\mathcal{C}}_R(n) ,$$ where $\lambda(D) = D L^2$. Finally, in terms of the density $\rho$, we can give the desired quantity as $\bar{\mathcal{C}}(\rho) = \bar{\mathcal{C}}(D \pi R^2) = \bar{\mathcal{C}}_R(D)$. For the case of $R=\infty$, where the exponential decay is controlled by a finite $\eta$, the quantity of interest is instead $\bar{\mathcal{C}}(\rho) = \bar{\mathcal{C}}(D \pi \eta^2) = \bar{\mathcal{C}}_\eta(D)$. ![Same as in Fig. \[raw\_vs\_analytical\_lattice\], except for continuous systems with no underlying lattice structure, and where the presription of calculating transition curves is now . **a** and **b** are for the continuum model Chern number $C$ for a 25$\times$25 system with $R=\infty$, and the parameter $M=1$. **c** and **d** are for the inclusive disc percolation for a 30$\times$30 system, where $R$ stands for the maximal hopping radius between lattice points. $N_0 = L^2$ is the number of lattice sites analogous to the maximum $N$ for finite lattices, with the difference that $n$ can here exceed $N_0$.[]{data-label="raw_vs_analytical_continuum"}](5continuum_perc){width="0.99\linewidth"} Again, the amount of densities $\rho$ to evaluate $\bar{\mathcal{C}}(\rho)$ can be chosen as high as desired as the conversion from $\bar{\mathcal{C}}(n)$ to $\bar{\mathcal{C}}(\rho)$ is calculated analytically. Note that, in contrast to the binomial distribution for the lattice problem above, the sum $\sum_n$ here goes to infinity, so one must include the terms until $P_{\lambda(D)}(n)$ becomes negligible. Again, the strongest confirmation of the validity of this method comes from benchmarking it for disc percolation. We can calculate the disc percolation probability $\bar{\Pi}(\rho)$ from Eq.  by replacing the topological index $\bar{\mathcal{C}}(n)$ by configuration-averaged percolation probability as function of lattice points $\bar{\Pi}(n)$. This enables us to extract the critical density and the critical exponent and to compare them to their known values $\rho_c=4.52$ [@quitanilla] and $\nu=4/3$. For system sizes up to $L=30$ with $N=L^2$ particles averaged over 1000 configurations we obtain the values $\rho_c=4.55$, $\nu=1.35$. Since the main source of error for the topological problem and disc percolation is the limited statistical sampling, we expect similar relative error for $\rho_c^{Q}$ and $\bar{\nu}$. Critical wave functions at the percolation threshold ==================================================== In the main text we pointed out that for optimal values of the mass parameter $M$, the topologically non-trivial phase may persist down to the site (disc) percolation threshold for short-range hopping models with discrete (continuum) random geometries. The percolation threshold provides a theoretical lower bound for the existence of a non-trivial topological phase for these models since no extended states can exist below the threshold when the underlying system consists of a collection of finite clusters. At the threshold, the underlying critical cluster is a random fractal with dimension d=1.9 in two dimensions [@stauffer]. Remarkably, as required by the presence of a non-trivial phase, extended states may survive even in this extreme limit. In the Fig. \[fig:wavefunctions\] below we have plotted the modulus of the wave functions living on critical clusters of square and triangular lattices. These extended states nucleate at $E=0$ at the critical density $p_c$ and the linear extent of these states is limited by the critical cluster. This provides a natural explanation for why the localization length exponent at the percolation threshold coincides with the correlation length exponent of the percolation cluster. ![ Modulus of critical wave functions $|\Psi|$ on percolating clusters for the nearest-neigbor Chern glass model on $100\times100$ lattices at the percolation threshold. All states have energies $E\lesssim10^{-4}M$. **a**-**b**: Square lattice percolations geometries with $M=1.14$ and $p=p_c=0.59$. **c**-**d**: Triangular lattice percolation geometries with $M=1$ and $p=0.52$, almost at the threshold $p_c=0.5$ []{data-label="fig:wavefunctions"}](6wavefunctions){width="0.99\linewidth"}

--- abstract: 'We show that the assumption of non-zero topological susceptibility of the vacuum in a fermion-free version of a theory, such as gravity or QCD, suffices to conclude the following: Once $N_f$ massless fermion flavors are added to the theory, they break the chiral flavor symmetry dynamically, down to a subgroup that would be anomaly-free under gauging; In both theories, the pseudo-Goldstone corresponding to axial $U(1)$-symmetry becomes massive; In QCD as well as in gravity the massless fermions are eliminated from the low energy spectrum of the theory. All the above conclusions are reached without making an assumption about confinement. Some key methods of our approach are: Reformulation of topological susceptibility in the language of a $3$-form gauge theory; Utilization of gravity in the role of a spectator interaction for the chiral anomaly-matching in QCD; Gauging chiral symmetries and matching their anomalies using the spectator Green-Schwarz type axions. Our observations suggest that breaking of chiral symmetries in QCD and in gravity can be described in unified topological language, and seemingly-disconnected phenomena, such as, the generation of $\eta''$-mass in QCD and breaking of global chiral symmetry by gravity may share a secret analogy. The described phenomenon may shed a new light - via contribution of micro black holes into the gravitational topological susceptibility of the vacuum - on incompatibility between black holes and global symmetries. It appears that explicit breaking is not the sole possibility, and like QCD, gravity may break global symmetries dynamically. As an useful byproduct, matching of gravitational anomalies provides a selection tool for compositeness, eliminating possibility of massless composite fermions where standard gauge anomaly matching would allow for their existence.' author: - Gia Dvali title: Topological Origin of Chiral Symmetry Breaking in QCD and in Gravity --- Introduction ============ An important role of topological susceptibility of the vacuum in defining the infrared properties of QCD has been long appreciated. A well-known example is the generation of the mass of the $\eta'$-meson via Witten-Veneziano mechanism [@WV], in the framework of QCD with ’t Hooft’s large number of colors [@planar]. This approach, in difference with the original solution by ’t Hooft [@tHooft1976; @determinant], does not explicitly rely on instantons, but on the topological susceptibility of the vacuum. Also, a direct connection between this entity and appearance of Goldstones with simultaneous solution to $U(1)$ problem in QCD, has been suggested by Veneziano in [@veneziano]. Yet, the power of vacuum topological susceptibility has not been fully exploited for understanding the vacuum structure and the dynamics of chiral symmetry breaking. First, to our knowledge, the question whether the assumption about vacuum topological susceptibility suffice for proving the further breaking of $SU(N_f)_L\times SU(N_f)_R$ chiral symmetry in QCD with $N_f$ massless quark flavor, is still open. Moreover, the role of a gravitational version of the topological susceptibility of the vacuum in generation of the mass gap and chiral symmetry breaking by fermion condensate was explored only relatively recently [@gia3form; @DJP; @DFF; @DFneu]. In this note we would like to expand on previous results and show that topological susceptibility of the vacuum gives a powerful tool both for establishing the fact and predicting the pattern of dynamical chiral symmetry breaking in generic class of theories. Some steps were already taken in this direction. It was shown [@gia3form] that in QCD the vacuum topological susceptibility allows to understand and re-formulate the mass generation - both for $\eta'$-meson as well as for axion - [*entirely*]{} in the language of a Higgs phase of a $3$-form gauge theory. Moreover, it was also shown there that the same connection persists in gravity: The existence of vacuum topological susceptibility in pure gravity, would imply the generation of mass gap with the participation of all the sectors (such as, chiral fermions or elementary axions) of the theory exhibiting a chiral gravitational anomaly. That is, a collective pseudo-scalar degree of freedom - transforming under the axial symmetry anomalous with respect to gravity - would become massive. It was further shown [@DJP] that the connection between the existence of topological Chern-Pontryagin density and generation of mass gap due to anomalous current is generic and can be formulated entirely in topological terms: In any theory with vacuum topological susceptibility and an anomalous current, one can predict a generation of mass of a pseudo-scalar meson corresponding to a would-be Goldstone of the anomalous current. The sectors most sensitive to the gravitational topological susceptibility would be axion and neutrinos. As shown in [@gia3form; @DFF], the existence of such topological susceptibility in the absence of massless fermions, would directly contribute into the mass of the axion, and thus, would jeopardize the Peccei-Quinn solution [@PQ] to the strong CP problem. As a way out [@gia3form; @DFF], in the presence of fermions with zero or small bare masses (e.g., a single neutrino species would suffice) the axion mass becomes protected: Such fermions contribute into the chiral gravitational anomaly and nullify the gravitational topological susceptibility of the vacuum. Simultaneously, the fermion sector must deliver a pseudo-scalar would-be Goldstone boson corresponding to the anomalous current. The mass of this boson is generated in a way very similar to the generation of $\eta'$ mass in QCD. The connection between the vacuum topological susceptibility in gravity and generation of mass gap for light fermions may find an interesting phenomenological application for neutrino masses [@DFneu] and neutrino-axion physics[@DFF]. In this light, settling the question of gravitational vacuum topological susceptibility and its role in chiral symmetry breaking is of direct important for phenomenology. Summarizing where we stand so far: The above studies showed that if the fermion-free version of a theory (e.g., QCD or gravity ) admits a non-zero value of topological susceptibility, then after introduction of fermions, they condense and break the anomalous $U(1)$ chiral symmetry spontaneously. In the same time a would-be Goldstone boson acquires a mass due to a $3$-form Higgs effect. On the other hand, to the best of our knowledge, the necessity of breaking - solely due to vacuum topological susceptibility - of chiral symmetry beyond the anomalous $U(1)_A$ has not been established firmly, neither in gauge theories nor in gravity. In this note we shall attempt to fill up this gap and show that from the assumption of non-zero topological susceptibility of the vacuum in a fermion-free version of a theory, such as QCD or gravity, follows the following: - Once $N_f$ massless fermion flavors are added to the theory, they form a condensate breaking the $U(N_f)$ chiral flavor symmetry dynamically, down to one of its maximal anomaly-free subgroups, i.e., a maximal subgroup that, if gauged, would be anomaly-free. - In both theories, the pseudo-Goldstone corresponding to axial $U(1)$-symmetry becomes massive. This generation of mass can be understood as a $3$-form Higgs effect. - Both in gravity and in QCD the massless fermions are eliminated from the low energy spectrum. - All the above phenomena are independent of assumption about the confinement. We shall speculate that the dynamical breaking of chiral symmetry due to topological susceptibility of the vacuum may shed a new light on incompatibility between black holes and global symmetries. The link could be provided by a contribution into the gravitational topological susceptibility of the vacuum from the micro black holes. That is, black holes could break global chiral symmetries dynamically via such a contribution. This adds a new flavor to the relation between black holes and global symmetries. The standard assumption is an explicit breaking of such symmetries by some high-dimensional operators. What we observe is that, similarly to QCD, the breaking of global chiral symmetry in gravity may be [*dynamical*]{}. Of course, the latter does not exclude the possibility of explicit breaking. In fact, as we shall see, for the axial $U(1)$ both types of breaking (explicit and spontaneous-dynamical) originate from the gravitational topological susceptibility of the vacuum. The $3$-form language nicely clarifies the coexistence of these two types of breaking: The generation of mass for a would-be Goldstone boson from explicit breaking by gravitational anomaly, can be viewed - in alternative language - as the mass coming from the $3$-form Higgs effect.\ Let us briefly mention some of the new approaches that we use in our treatment. In our analysis, we shall use the formulation [@gia3form], which describes phenomena of strong CP-violation and generation of the mass of $\eta'$-meson (or axion) in the language of a $3$-form gauge field. The convenience of this language is that it allows to understand the vacuum structure of QCD with and without massless quarks as the two phases - Higgs and Coulomb - of a $3$-form gauge theory. We must note that the possibility of modeling the $\eta'$ mass in QCD by a massive $3$-form, has been recognized for some time, in particular, by Aurilia, Takahashi and Townsend [@aurilia]. Similar effects have been noticed by several authors in different contexts. For example, Gabadadze and Shifman discussed the screening of $3$-form by axion, while analyzing the structure of axion domain walls [@walls]. The analysis in [@gia3form] is important for us in two respects. First, it shows that $3$-form Higgs effect accounts for the entire QCD contribution to the $\eta'$ and axion masses, including the one coming from ’t Hooft’s instantons, as well as, for an additional contribution in case of a non-zero gravitational topological susceptibility of the vacuum. Secondly, it fully accounts for the CP-invariance of QCD (or gravitational) vacuum. So it appears that the $3$-form Higgs language is not just an useful analogy for modeling the $\eta'$ (or axion) mass generation in QCD (or a similar effect in gravity), but rather, it represents a remarkably precise effective description of this phenomenon. We shall elaborate on this issue below. The presentation shall be self-sufficient, but for a more detailed discussion of this formulation the reader is referred to [@gia3form].\ The other two important ingredient of our analysis are the following.\ First, we shall heavily rely on ’t Hooft’s anomaly matching [@anomalymatching], but shall use as spectators not only fermions, but also Green-Schwarz type [@GS] axions. This allows us to weakly gauge and monitor the anomaly matching even with respect to the symmetries that are anomalous and are explicitly broken by instantons [@determinant; @tHooft1976] or other effects. This is important, since for example, in his original breakthrough paper on anomaly matching [@anomalymatching], ’t Hooft did not impose anomaly matching for the axial $U(1)_A$-symmetry, because it is explicitly broken by instantons. However, as we show below, for gauged $U(1)_A$ for which anomalies are cancelled by a spectator Green-Schwarz axion, the ’t Hooft determinant term generated by the instantons [@determinant] is manifestly gauge invariant under $U(1)_A$. Therefore, anomalies under this symmetry must be monitored. This monitoring provides us with some crucial information, for controlling the breaking of the non-abelian $SU(N_f)$ part of the chiral symmetry.\ Another new tool that we employ in QCD, is that we use [*gravity*]{} as a spectator interaction and impose matching of the gravitational chiral anomalies. The extra power provided by this method relies on the fact that, although composite fermions can be color-singlets, they cannot be “gravity-singlets". By general covariance, there exist no particles decoupled from gravity. This gives an additional consistency relation and eventually enables us to deduce the absence of massless fermions, both elementary and composite, and a complete breaking of any part of chiral symmetry that would exhibit anomalies under gauging.\ We must note that the employment of gravity in the role of a spectator interaction for the anomaly-matching, is a byproduct that can be used in composite model building irrespectively of the issue of vacuum topological susceptibility. Even if one ignores the latter and uses the confinement or a compositeness as a starting assumption, the gravitational anomaly can severely restrict the composite spectrum, e.g., by eliminating massless composites in cases where standard gauge anomaly-matching would allow their existence. Assumption and reasoning ======================== Let us first formulate our assumptions. We consider a generic theory (gauge or gravity) which satisfies the following conditions. In the version of the theory that includes no massless fermions, there exists a non-vanishing topological susceptibility of the vacuum with respect to some $3$-form $C$. As usual, the gauge-invariant zero-form field strength is $E \equiv ^*dC$, where the star stands for Hodge-dual. That is, we assume that there exist the following correlator, $$\label{Etop} \langle E, E\rangle_{p\to 0}\, = \, const \neq 0 \, .$$ where $p$ is momentum. Next, assume that after introduction of $N_f$-flavors of some light fermions there exists a chiral $U(1)_A$-current with the anomalous divergence $$\partial^{\mu} J_{\mu}^{(A)} \, = \, E \, , \label{divJ}$$ where for compactness we absorbed the unknown anomaly coefficient in the normalization. Our conclusions are independent of its value as long as it is non-zero.\ We then wish to prove that in such a theory fermions condense and spontaneously break chiral symmetry. The would-be Goldstone of the anomalous current (call it $\eta'$) is becoming massive. It is eaten-up by the three-form $C$ and becomes massive via an effect closely analogous to a three-form version of the Higgs (or Stückelberg) phenomenon [@gia3form]. Moreover, the fermions are eliminated from the low energy theory. The above picture of chiral symmetry breaking is independent from underlying structure of the theory. Therefore, it is equally applicable to gauge theories as well as to gravity. In addition we wish to show that in QCD, the $U(N_f)_L \times U(N_f)_R$ chiral symmetry is broken down to its subgroup that under gauging is anomaly-free. In gravity, the same holds. With $N$ massless fermions, the flavor symmetry is $U(N)$, which is spontaneously broken down to an anomaly-free subgroup. The natural candidate for such a subgroup is $SO(N)$. The symmetry breaking $U(N) \rightarrow SO(N)$ results in $N(N+1)/2$ would-be Goldstones. Out of these, one is getting its mass directly from nullifying the topological susceptibility. The remaining $N(N+1)/2 -1$ Goldstones could potentially get masses in case there exist operators that break chiral flavor symmetry explicitly. Such explicit-breaking sources are fully compatible with out treatment, but are not required for the consistency of reasoning. Therefore, $N(N+1)/2 - 1$ Goldstones could potentially stay massless.\ Although, we shall try to formulate our arguments in maximally generic way, at first, we shall keep making contact with QCD and later generalize to gravity.\ It shall become clear that the arguments leading to the chiral symmetry breaking and the elimination of massless fermions should equally apply to any theory that satisfies the conditions (\[Etop\]) and (\[divJ\]), regardless of the assumption about the confinement. However, the patters of symmetry breaking can vary from theory to theory.\ We start by more detailed formulation of the initial assumption. Consider a theory that contains a gauge three-form $C$, either elementary or composite, and no massless fermions present. A good example is QCD of pure glue, i.e., no quarks included. We shall not specify the number of colors, and allow ourselves to take it large, when needed. We assume that the topological susceptibility of the vacuum is non-zero, (\[Etop\]). For QCD the condition (\[Etop\]) takes the following form, $$\begin{aligned} \label{FFcorr} \langle F\tilde{F},F\tilde{F}\rangle_{p\to 0} \equiv \nonumber\\ \equiv \lim\limits_{p \to 0}\int d^4 x e^{ipx} \langle T [F\tilde{F}(x)F\tilde{F}(0)]\rangle =\mathrm{const}\neq 0\, , \end{aligned}$$ where $F$ is the usual QCD field strength two-form, $\tilde{F} \equiv ^*F$.\ One important thing that a non-zero topological susceptibility of the vacuum (\[FFcorr\]) tells us, is that the theory contains vacua that break CP-symmetry. These vacua are usually labeled by a CP-violating parameter ${\vartheta}$. The physical observability of this parameter is equivalent to (\[FFcorr\]). Correspondingly, when the correlator in (\[FFcorr\]) vanishes, CP-violation becomes unobservable and one can say that ${\vartheta}$-parameter becomes unphysical. Throughout our discussion we shall heavily rely on matching the descriptions - of certain physical phenomena - in the languages of high-energy and low-energy theories. In particular, these two descriptions must agree on the fact that in the absence of massless quarks the CP-violation is [*physical*]{} and that it becomes [*unphysical*]{} once the massless quarks are introduced. This matching of the two descriptions will lead us to the conclusion of necessity of chiral symmetry breaking by fermion condensate and of elimination of massless fermions from the low energy theory. We shall now discuss this in more details. Alice and Bob ------------- Let us introduce the two observers: A high energy observer Alice and a low energy observer Bob. These two physicists must both describe the same physics, but from two different perspectives. In particular, they both agree that in QCD without quarks there exist CP-violating vacua. For Alice, this is accounted by the introduction of the ${\vartheta}$-term in her Lagrangian, $$L_{Alice} \, = \, {\vartheta}F\tilde{F} \, - \, FF \,, \label{AliceT}$$ where the second term is the usual QCD field strength, with irrelevant constants absorbed in normalization. In the absence of quarks, the ${\vartheta}$-term is physical and can be measured. [^1]\ Let us explain how the same CP-violating vacua are described by Bob. For this, let us rewrite the theory in terms of the Chern-Simons three-form $C$ and its field strength $E \equiv ^*dC$, $$C\equiv A\wedge dA + {2 \over 3} A \wedge A \wedge A,~~~ E \equiv F\tilde{F}=^*\mathrm{d}C.\label{eqE}$$ Here, $A$ is the gluon gauge field one-form. Under the QCD gauge transformation $C$ shifts by an exterior derivative of a two-form, $$C \rightarrow C + d\omega \,, \label{gauge}$$ where the explicit form of $\omega$ is irrelevant and shall not be displayed here. When expressed in terms of the field $C$, the non-vanishing correlator (\[FFcorr\]) implies that the propagator of $3$-form $C$ has a pole at $p^2\, =\, 0$, $$\label{CCcorr} \langle C, C\rangle_{p\to 0}\, \propto \, \frac{1}{p^2} \, .$$ As noticed by Lüscher [@massless3form], this fact tells us that $C$ is a fully legitimate [*massless*]{} $3$-form gauge field, with all the usual properties. If we couple $C$ to some external probe sources, then at large distances $C$ will mediate interactions between them exactly the way an elementary $3$-form does. For example, if we source $C$ by a $3$-form current of some probe membrane, it will induce a distance-independent static force between the two parallel membranes (see, [@gia3form]). This force is very similar to static Coulomb force between the two point-like massive charges in $1+1$-dimensional Schwinger model. The similarity is not accidental, since the massless one-form field in $1+1$ does not propagate any degree of freedom. The latter property is fully shared by the massless $3$-form in $3+1$-dimensions. Because of this, the existence of a massless $3$-form in effective theory is in no conflict with usual intuition that a theory of a pure glue has a mass gap. Although a massless $3$-form does not propagate any physical degree of freedom, nevertheless, it can create a static long-range electric-type field, as discussed above. As we shall see, this static field is the key for describing CP-violation in low energy theory. The above knowledge combined with the invariance of the theory under the gauge shift (\[gauge\]) leads to the effective Lagrangian describing the topological structure of the QCD vacuum in form of a gauge theory containing solely the massless three-form $C$ [@gia3form]. In other words, at zero momentum all the massive states decouple and the vacuum structure of the theory is fully determined by its dependence on $C$. Since we are interested in the vacuum structure, i.e., the zero momentum limit of the theory, we shall work in effective low energy theory below the QCD scale, with all the massive excitations being integrated out. We shall relax the assumption about the mass-gap shortly, but it is useful to first go thought the argument under it. Since we are working in theory with a mass-gap, there are no propagating massless degrees of freedom in this theory. Then, by gauge invariance, the effective Lagrangian includes the following terms: 1) A function algebraic in $E$; and 2) the terms that depend on derivatives of $E$. Since no massless excitations have been integrated out, all the higher derivative terms must vanish in zero momentum limit. Let the algebraic function be ${\mathcal K}(E)$. Then, the effective Lagrangian has the form, $$\label{effQCD} L_{Bob} \, = \, {\mathcal K}(E) + \, ({\rm terms~depending~on~} \partial_{\mu} E ) \,.$$ We shall now look for the vacuum solutions of the equations of motion obtained by taking variation with respect to $C$. These are solutions with constant $E$. Because of this, the contribution from the derivative terms is automatically zero and the equation of motion reduces to, $$\label{eqEeff} \partial_{\mu} \left ( {\partial {\mathcal K}(E) \over \partial E} \right ) \, = \, 0,$$ This equation is satisfied for an arbitrary constant value of $E$ (except the values for which derivative of ${\mathcal K}$-function may be singular). From now on we shall work in units of the scale $\Lambda$ that sets the relation $F\tilde{F} = E\Lambda^2$, where $E = ^*dC$ is written for $C$ with canonical normalization. The value of this scale is determined by the vacuum topological susceptibility and its precise value is unimportant for us. In units of this scale, we can parameterize the constant vacuum expectation value (VEV) of the electric field that solves (\[eqEeff\]) as $$E = {\vartheta}\,. \label{Tvac}$$ Thus, the vacua of the theory are labeled by an integration constant ${\vartheta}$ and correspond to vacua with different values of the static “electric field" $E$ [^2] Notice that the vacua with different values of ${\vartheta}$ are physically distinct, since they differ by the amount of CP violation: The electric field $E$ is CP-odd. We may try to impose the condition that the initial function ${\mathcal K}(E)$ is CP-invariant, meaning that it only contains even powers of $E$, ${\mathcal K}(E) = {1 \over 2} E^2 + ...$. Nevertheless, re-expanding the theory around the vacuum $E={\vartheta}$, we end up with a linear CP-violating term in the Lagrangian, $$\label{EXP} L_{Bob} \, = \, {\vartheta}E + {1 \over 2} E^2 + ... \,.$$ For the purpose of our discussion we can limit ourselves with infinitesimal values of ${\vartheta}$ (and correspondingly $E$).\ The direct connection between the existence of non-zero topological susceptibility (\[FFcorr\]) and a static electric field $E \neq 0$, removes any need for the assumption of the mass gap, which we shall now relax. Indeed, the existence of a constant electric field $E$ means that in the theory (in this case pure glue) there exist no [*mobile charges*]{} that source the $3$-form $C$. If such charges were to exist, they would give the mass to $C$ and would correspondingly screen the electric field $E$, in a way very similar to what happens to an ordinary Maxwell electric field in a superconductor and/or in the Higgs phase. This can be understood directly at the level of the effective Lagrangian for $C$ and is very important. Naively, one could say that if we do not demand a mass gap, then the effective Lagrangian for $C$ could contain “non-local" terms resulting from integrating-out some massless degrees of freedom, which could potentially ruin the static electric field solutions. However, the gauge-invariant “non-local" terms that could do such a job (e.g., the terms of the type $C \Pi C$, where $\Pi$ is a transverse projector) can be explicitly shown to come from integrating-out the mobile sources of $C$, such as, massless axions or massless membranes (see [@gia3form]). But, their existence would be equivalent of having zero topological susceptibility of the vacuum in theory of pure glue, which would contradict to our starting assumption. In other words, the screening of $E$ would imply that the pole at $p^2=0$ is removed, or equivalently, the vacuum topological susceptibility (\[FFcorr\]) would vanish. This argument shows that the existence of the constant vacuum electric field $E={\vartheta}$ is a direct consequence of non-zero vacuum topological susceptibility (\[FFcorr\]) and does not require any extra assumption about a mass gap in the theory. Conversely, introduction of mobile charges that screen $E$, would render $C$ massive, and thus, would make the topological susceptibility of vacuum zero. As we shall see later, this is exactly what happens once we introduce massless fermions in the theory. However, for the time being, we restrict ourselves by pure glue. Since in the theory there exist no dynamical sources for $C$, there are no transitions between the vacua with different ${\vartheta}$. Hence, the theory splits into infinity of super-selection sectors labeled by the angle ${\vartheta}$. These vacua are nothing but the famous ${\vartheta}$-vacua of QCD [@tHooft1976; @thetavacua], which Bob has “rediscovered" in the language of a $3$-form gauge theory. These are exactly the same vacua that Alice labels by the ${\vartheta}$-term in (\[AliceT\]). We thus arrive to the following matching of the descriptions of the two physicists: Alice’s ${\vartheta}$-vacua are understood by Bob as the vacua with different values of the static CP-violating electric field $E$, $$\left ({\vartheta}-{\rm vacua} \right )_{{\rm Alice}}\, = \, \left (E-{\rm vacua} \right )_{{\rm Bob}} \, . \label{AB}$$ That is, what for Alice serves as a parameter of the theory, for Bob is a label of the vacuum. When Alice changes the theory by choosing a different value of ${\vartheta}$, Bob is making a corresponding choice by selecting the vacuum with a different value of the electric field $E$.\ \[Here one could be tempted to enter into a semantics discussion and ask: Which term is more appropriate for the use with respect to ${\vartheta}$, a “parameter" or an “integration constant"? Usually, the difference between the two notions is that the former labels a theory, whereas the latter labels a particular solution (in this case a static electric field) in a given theory. However, when the choices obey a super-selection rule, the two notions become equivalent. For example, since there exist no charges that could source $C$ in theory of pure glue, no transitions occur between the sectors with different values of $E$. The Hilbert spaces constructed on top of these vacua are orthogonal. So Bob could equally well call $E$ a parameter, rather than a “solution". The two languages describe the same physics, as they should, and matching the two descriptions gives us an useful tool.\]\ Notice, the vacua have different energies, but due to absence of sources there are no transitions among them. The lowest energy vacuum is the CP conserving one. Alice explains this using Vafa-Witten theorem [@VafaWitten]. For Bob, the same conclusion is reached because the lowest energy state is the one with the vanishing electric field, $E = 0$. Notice, as explained in [@gia3form], the requirement that ${\vartheta}$ be defined modulo $2\pi$ translates as a constraint on the ${\mathcal K}$-function that its derivative must be inverse of a periodic function. Since we are working with small values of $E$, this constraint is not important for our discussion.\ Introducing Massless Quarks --------------------------- Let us now introduce in the theory some massless quark flavors. It is useful to introduce them in form of left-handed Weyl fermions in fundamental $\psi_{i}$ and anti-fundamental $\bar{\psi}^{\bar{j}}$ representations of the color group. Here the bar does not stand for Dirac conjugation, but for distinguishing the fundamental and anti-fundamental representations of the color group. $i,j = 1,2,...N_f$ and $\bar{i},\bar{j} = 1,2, ...N_f$ are the flavor indexes. The color index is not written explicitly. In the absence of the fermion masses the theory exhibits a $U(N_f)_L\times U(N_f)_R$ flavor symmetry, where $U(N_f)_L$ and $U(N_f)_R$ act on $\psi_{i}$ and $\bar{\psi}^{\bar{j}}$ independently. The fermions thus form a bi-fundamental representation $(\bar{N}_f, N_f)$ of the flavor group. It is useful to write the flavor group as $SU(N_f)_L\times SU(N_f)_R \times U(1)_V\times U(1)_A$. Where, $U(1)_V$ and $U(1)_{A}$ are flavor-blind vector and axial symmetries that act on the fermions as $$\psi_{i} \rightarrow {\rm e}^{i\alpha} \psi_{i}, ~~~ \bar{\psi}^{\bar{j}} \rightarrow {\rm e}^{-i\alpha} \bar{\psi}^{\bar{j}} \label{vector}$$ and $$\psi_{i} \rightarrow {\rm e}^{i\alpha} \psi_{i}, ~~~ \bar{\psi}^{\bar{j}} \rightarrow {\rm e}^{ i\alpha} \bar{\psi}^{\bar{j}} \label{axial}$$ respectively. Here $\alpha$, in each case, is a transformation parameter. We now wish to show that the fermions form a condensate that breaks $U(N_f)_L\times U(N_f)_R$ chiral symmetry spontaneously at least down to its anomaly-free subgroup, i.e., a subgroup that would be anomaly free if the entire $U(N_f)_L\times U(N_f)_R$-symmetry were to be gauged.\ We shall organize the argument in four steps:\ - First, we shall make sure that the introduction of quarks makes the vacuum topological susceptibility zero. Both for Alice and Bob this implies that vacuum becomes CP-conserving. - Next, following [@gia3form] we show that in Bob’s description this means that $3$-form gauge theory enters the Higgs phase, i.e., $C$ acquires a non-zero mass. By gauge-invariance (\[gauge\]) this is only possible if $C$ “eats" a pseudo-scalar degree of freedom, which we denote by $\eta'$. We determine that the only degree of freedom that matches the required properties is the phase of the quark-condensate. Hence, quark condensate must break $U(1)_A$-symmetry spontaneously. This part of the argument reinforces the results already presented in [@gia3form; @DJP; @DFF; @DFneu]. It is also in agreement with the conclusion of [@veneziano], reached in a different approach. - Next, we do two things. First, we introduce gravity as a spectator interaction. Secondly, we gauge the axial symmetry weakly and cancel the gauge-gauge and mixed gauge-gravity anomalies by a spectator Green-Schwarz axions. Then, by matching anomalies in theories of Alice and Bob we conclude that there cannot exist any massless fermions at low energies. - Finally, by weakly gauging the different subgroups of the flavor group and applying ’t Hooft’s anomaly matching conditions with spectator fermions, we deduce that the quark condensate must break the $U(N_f)_L \otimes U(N_f)_R$-symmetry down to its anomaly-free subgroup. Necessity of fermion condensate and mass-generation for $\eta'$ --------------------------------------------------------------- Both Alice and Bob must agree on the fact that after introduction of the massless quarks the topological susceptibility must vanish, or equivalently, the parameter ${\vartheta}$ must become unphysical. This is an effect that can be established experimentally: $E$ is a CP-odd quantity, which can be measured. Thus, unambiguously, after introduction of massless fermions both Alice and Bob must measure zero CP-violation in their theories. However, they explain this fact differently.\ Alice knows that ${\vartheta}$ is unphysical because it can be rotated away by the chiral transformation (\[axial\]). Due to Adler-Bell-Jackiw chiral anomaly [@ABJ] the Lagrangian shifts under this transformation as $$\delta L_{Alice} \, = \, \alpha F\tilde{F} \, = \alpha E \, , \label{shiftC}$$ and any initial ${\vartheta}$-term can be eliminated by the suitable choice of $\alpha$. Here and below, we absorb the anomaly coefficient in definition of $F\tilde{F}$. One of the consequences of this freedom is that CP is unbroken. Thus, the topological susceptibility must vanish. This means that the pole in $3$-form correlator must be moved away from $p^2=0$, $$\label{eq:CGap} \langle C, C\rangle_{p\to 0}=\frac{1}{p^2 - m^2}\;.$$\ The equation (\[eq:CGap\]) explains how the same phenomenon is accounted by Bob. For Bob the physicality of ${\vartheta}$ was equivalent to the fact that vacuum could support a constant electric field $E$, which breaks CP-symmetry. This is only possible if the low energy theory contains a massless three-form $C$. The elimination of ${\vartheta}$ - that Alice understands as a freedom of changing the fermion phases by a chiral rotation (\[axial\]) - for Bob means that the $3$-form $C$ became [*massive*]{}. And indeed, this is fully confirmed by the equation (\[eq:CGap\]). Thus, in order to match Alice’s story, Bob needs to explain why $C$ has acquired a non-zero mass after quarks were introduced. How can a $3$-form gauge field $C$ acquire a mass? The massive $3$-form, unlike the massless one, does propagate one (pseudo-scalar) degree of freedom. The key point is that the generation of the mass gap must happen with the full respect of the gauge redundancy (\[gauge\]), i.e., the $3$-form must acquire the longitudinal pseudo-scalar degree of freedom as a result of a Higgs-like effect. Let us denote this - yet to be identified - pseudo-scalar degree of freedom by $\eta'$. Then, up to irrelevant higher order terms, the gauge invariant Lagrangian describing the mass-generation of the three-form is uniquely fixed as [@gia3form], $$\label{effectiveQCDeta} L \, = \, {1 \over 2} E^2 \, - \, {\eta' \over f_{\eta}} \, E \, + \, {1 \over 2} \partial_{\mu}{\eta'} \partial^{\mu} {\eta'},$$ where as before we work in units of the scale $\Lambda$. $f_{\eta}$ is a parameter that encodes information about the origin of $\eta'$, which we still need to discover. Also notice: the Higher order derivative terms cannot modify the value of the mass gap, since they all vanish at zero momentum. The equations of motions for $C$ and $\eta'$ are $$\label{eqEeff} \partial_{\mu} \left( E - {\eta' \over f_{\eta}} \right)=0,$$ and $$\label{boxeta1} \Box \, \eta' = \, - \, {1 \over f_{\eta} } E \,,$$ respectively. Solving the equation of motion for $C$, we obtain the following expression for the electric field $$\label{solEeff} E \, = \, \left({\eta' \over f_{\eta}}-{\vartheta}\right),$$ where, just as in the fermion-free case (\[Tvac\]), the parameter ${\vartheta}$ is an integration constant. The crucial novelty that coupling with $\eta'$ introduces is that ${\vartheta}$ is [*unphysical*]{}: For arbitrary choice of ${\vartheta}$, the vacuum expectation value of the electric field $E$ vanishes and simultaneously the mass gap is generated. Indeed, inserting (\[solEeff\]) in (\[boxeta1\]) we get the following effective equation of motion for $\eta'$, $$\label{boxeta} \Box\, \eta' + {1 \over f_{\eta} }\left({\eta' \over f_{\eta}} -{\vartheta}\right) = 0.$$ Here we immediately observe: 1\) The fields $\eta'$ and $C$ combine and form a single propagating massive pseudo-scalar field of mass $m_{\eta'}^2 = 1/ f_{\eta}^2$; and 2\) The vacuum expectation value of this field, ${\eta' \over f_{\eta}} ={\vartheta}$, is exactly such that it forces $E = 0$, as it is clear from (\[solEeff\]). Thus, irrespective of the choice of the integration constant ${\vartheta}$, Bob detects zero CP-violation, since the VEV of the CP-odd electric field $E$ is zero. Notice, this is a manifestation in Bob’s language of Vafa-Witten theorem [@VafaWitten], which implies that once ${\vartheta}$ is promoted into a dynamical field, it relaxes to CP-conserving minimum.\ What we are witnessing is the analog of the Higgs phase for the $3$-form $C$. As it is clear from the second term in (\[effectiveQCDeta\]) the $3$-form $C$ is sourced by the topological $3$-form current $d\eta'$. Thus, the pseudo-scalar $\eta'$ plays the role of the mobile charge that screens the electric field and gives mass to $C$.\ [*To summarize: So far - using the gauge-invariance (\[gauge\]) and the matching of the descriptions of Alice and Bob - we have established that the introduction of fermions is accompanied by the appearance of a pseudo-scalar $\eta'$, which renders the $3$-form $C$ massive.* ]{}\ We now need to identify the origin of $\eta'$. The first step is to understand that $\eta'$ is a phase excitation of some order parameter formed out of fermions, which transforms non-trivially under $U(1)_A$. That is, $\eta'$ is a would-be Goldstone boson of spontaneously broken chiral $U(1)_A$-symmetry. The above understanding can again be achieved by matching the descriptions of Alice and Bob. Alice, being an UV-observer, has no information about the existence of $\eta'$. But, she knows that ${\vartheta}$ is unphysical, because she can eliminate it by arranging a shift of the Lagrangian (\[shiftC\]) by performing a chiral $U(1)_A$-transformation of fermions (\[axial\]). So, by this transformation Alice can shift ${\vartheta}$ as $${\vartheta}\rightarrow {\vartheta}\, + \, \alpha \,. \label{shiftTA}$$ Since the CP-invariance is universal for both observers, the shift of ${\vartheta}$ by the chiral rotation performed by Alice, in theory of Bob must be exactly matched by the shift of $\eta'$. This is also clear from the equation (\[solEeff\]) which shows that in Bob’s theory the vacuum is always at ${\eta' \over f_{\eta}}-{\vartheta}= 0$, where $E$ vanishes. Therefore, the shift of ${\vartheta}$ implies the corresponding shift of $\eta'$, such that the equality $E=0$ is kept intact. In the normalization in which ${\vartheta}$ denotes a coefficient of $F\tilde{F}$ and the anomalous coupling of $\eta'$ is $${\eta' \over f_{\eta}} F\tilde{F} \, , \label{cefficient}$$ the shift of $\eta'$ induced by the Alice’s chiral transformation is $${\eta' \over f_{\eta}} \rightarrow {\eta' \over f_{\eta}} \, + \, \alpha \,. \label{shiftEta}$$ The above is the key point and let us reiterate it. For Alice the non-existence of physical CP-violation is in arbitrariness of ${\vartheta}$, which she can change at will by the chiral rotation (\[axial\]). This shift in the theory of Bob corresponds to a shift of the integration constant ${\vartheta}$ in (\[solEeff\]). The non-existence of CP-violation for Bob means that the zero value of the electric field $E$ is insensitive to this shift. This in only possible if the shift of ${\vartheta}$ (\[shiftTA\]) is exactly compensated by the shift of $\eta'$ in (\[shiftEta\]). This shows that $\eta'$ is a phase degree of freedom of some fermion composite order parameter that transforms under $U(1)_A$. The matching between the descriptions of Alice and Bob is summarized in table \[BHtable\].\ **Alice** **Bob** -------------------------------------------------------------- -------------------------------------------------------------- No fermions (physical ${\vartheta}$) Coulomb phase ($E \neq 0$) Chiral Fermions (unphysical ${\vartheta}$) Higgs phase ($\eta'$ screens $E$) $U(1)_A$-shift: ${\vartheta}\rightarrow {\vartheta}+ \alpha$ $U(1)_A$-shift: $\eta' \rightarrow \eta' + f_{\eta} \alpha $ : The table summarizing correspondence between the descriptions of Alice and Bob.[]{data-label="BHtable"} Notice, the $3$-form language[@gia3form] that Bob is using to conclude the existence of massive pseudo-scalar $\eta'$, fully agrees with the standard language of ’t Hooft about generation of this mass from instantons. In ’t Hooft’s language, the explicit breaking is due to the instanton-generated term in the effective Lagrangian [@determinant], $$L_{'t Hooft} \, = \, e^{-i{\vartheta}} det(\bar{\psi}\psi)\, \rightarrow \, e^{- i({\vartheta}- {\eta' \over f_{\eta}})} |\langle det(\bar{\psi}\psi) \rangle |\,. \label{detFermion}$$ After assuming that the VEV of the above order parameter is non-zero, this term generates the mass to exactly the same degree of freedom $\eta'$ as Bob sees. The advantage of Bob’s language is that the spontaneous breaking of $U(1)_A$ is not an input assumption, but is a necessity, dictated by the $3$-form Higgs effect. In addition, this language captures the entire QCD contribution to the mass of $\eta'$ that comes through axial anomaly, since all such sources must contribute through (\[FFcorr\]) and thus must be accounted by Bob’s effective Lagrangian (\[effectiveQCDeta\]). Indeed, imagine that there exists an additional contribution, $\mu_{\eta}$, to the mass of $\eta'$, which is not accounted by the $3$-form Higgs effect. Then, such contribution would be added to (\[effectiveQCDeta\]) in form of an additional mass term ${1 \over 2} \mu_{\eta}^2 \eta'^2$ and correspondingly would modify the equation (\[boxeta\]) as $$\label{boxeta1} \Box\, \eta' + {1 \over f_{\eta} }\left({\eta' \over f_{\eta}} -{\vartheta}\right) \, + \, \mu_{\eta}^2 \eta' \, = \, 0.$$ The CP-odd electric field then would have a non-zero VEV: $$\label{Eneq0} E \, = \, - \, {\vartheta}{(\mu_{\eta}f_{\eta})^2 \over 1 + (\mu_{\eta} f_{\eta})^2 } \, .$$ This would make no sense, since it is incompatible with Alice’s description, which says that massless fermions remove CP violation. Thus, the contribution from ’t Hooft determinant to the mass of $\eta'$ is already fully included in (\[effectiveQCDeta\]) and should [*not*]{} be counted as extra. The above shows that one should be very careful not to superimpose the two languages. This will lead to double-counting!\ Coming back to were we are, we have concluded that the pseudo-scalar $\eta'$ that Higgses the $3$-form $C$, must come from the phase of an order parameter composed out of fermions. However, at this point we do not know which order parameter is the right one. For example, if the only non-zero order parameter were the entire ’t Hooft fermion determinant of $N_f$-flavors, the $Z_{2N_f}$ subgroup of $U(1)_A$-symmetry together with $SU(N_f)_L\times SU(N_f)_R\times U(1)_V$-group would survive unbroken. In such a case $\eta'$ would represent the over-all phase of the entire fermion determinant. If instead, for example, the order parameter were an universal fermion bilinear $\langle \bar{\psi}\psi \rangle$, the flavor symmetry $U(1)_A$ would be broken all the way down to $SU(N_f)_{L+R} \otimes U(1)_V\times Z_2$.\ \[As a consistency check, matching equation (\[boxeta1\]) with the anomalous divergence of the axial current (\[divJ\]) we get $$\partial_{\mu} \eta' \, \propto \, J_{\mu}^{(A)}\, , \label{etacurrent}$$ which reconfirms that $\eta'$ is a Goldstone boson of broken $U(1)_A$ axial symmetry.\]\ Thus, the quark sector of the theory must deliver an order parameter that transforms non-trivially under $U(1)_A$.\ Breaking of $SU(N_f)_L \otimes SU(N_f)_R$ flavor symmetry. ---------------------------------------------------------- What we need to prove now is that this order parameter must break not only $U(1)_A$, but also the chiral flavor symmetry $SU(N_f)_L \otimes SU(N_f)_R$ at least down to its anomaly-free subgroup. Introducing a notation $\Phi_{i}^{\bar{j}} \equiv \bar{\psi}^{\bar{j}} \psi_{i}$, we can construct a Lorentz-invariant order parameter that is invariant under $SU(N_f)_L \otimes SU(N_f)_R \otimes U(1)_V$ and transforms nontrivially under $U(1)_A$. This order parameter has the form of ’t Hooft determinant: $det \Phi \equiv \epsilon^{i_1,...i_{N_f}} \epsilon_{\bar{j}_1,...\bar{j}_{N_f}} \Phi_{i_1}^{\bar{j}_1}...\Phi_{i_{N_f}}^{\bar{j}_{N_f}}$. Notice that it is therefore invariant under an anomaly-free discrete $Z_{2N_f}$-subgroup of $U(1)_A$. For spontaneous breaking of $U(1)_A$ and delivering a Goldstone boson $\eta'$ with anomalous coupling the necessary condition is that the VEV $\langle det \Phi \rangle$ is non-zero. [^3] If this were to be the only available non-zero order parameter, $\langle det \Phi \rangle \neq 0$, the flavor symmetry would be broken down to $SU(N_f)_L \otimes SU(N_f)_R \otimes U(1)_V$. We need to understand what happens to the rest of the flavor group. Let us now show that the flavor symmetry $SU(N_f)_L \otimes SU(N_f)_R$ must be spontaneously broken down to its anomaly-free subgroup. ### Absence of massless fermions at low energies: Gravity as a spectator As a first step towards this proof, let us convince ourselves that in the low energy theory of Bob there is no place for any massless fermions. We shall try to prove this by contradiction. Let us assume that such massless fermions do exist and denote them by $\lambda_{\alpha}, ~ \alpha = 1,2,...n$, where $n$ is unspecified. For definiteness, assume all of them to be written in left-handed Weyl basis. Since we are making no assumption about the confinement, $\lambda$-fermions can be either a sub-set of Alice’s $\psi$-fermions or their massless composites. We wish to convince ourselves that this set is empty. As the first step, notice that if $\lambda$-s exist they must form an anomaly-free set with respect $U(1)_A$. In order to see this we shall weakly gauge the $U(1)_A$-symmetry by introducing a gauge field $X_{\mu}$, which under the local version of the axial transformation (\[axial\]) shifts as $$X_{\mu} \rightarrow X_{\mu}\, + \, { 1 \over g_X} \partial_{\mu} \alpha(x)\,, \label{gaugedX}$$ where $g_X$ is the gauge coupling. The gauged version of $U(1)_A$-symmetry we shall denote by $U(1)_X$, in order to distinguish it from the global axial symmetry. This gauging produces gauge chiral anomaly. In order to cancel it we shall introduce a Green-Schwarz axion, $a$, that under the gauge transformation (\[gaugedX\]) shifts as $${a \over f_a} \rightarrow {a \over f_a} \, - \, \alpha(x) \,. \label{AxionX}$$ Here $f_a$ is a decay constant, which we shall later take to infinity.\ The axion $a$ should not be confused with Weinberg-Wilczek axion [@axion], which is usually invoked for solving the strong-CP problem via Peccei-Quinn mechanism [@PQ]. Rather, the axion $a$ is the axion of Green-Schwarz type [@GS], the role of which is the cancellation of gauge $U(1)_X$ anomalies. The CP-invariance in our case is maintained by $\eta'$, as it was discussed in details above. In this sense, in our setup the role of Weinberg-Wilczek axion [@axion] is taken up by $\eta'$. We shall explicitly show below, how $a$ and $\eta'$ share their jobs of canceling anomalies and maintaining CP-invariance.\ The resulting gauge invariant Lagrangian of Alice has the following form $$\begin{aligned} \label{AliceLX} L_{Alice} \, &=& \, \sum_{\psi} i \overline{\psi} \gamma^{\mu} D_{\mu} \psi \, -\nonumber\\ && - \, F_X^2 \, - \, F^2 + \, {\vartheta}F\tilde{F} \nonumber\\ && + \, (f_a)^2\left (g_XX_{\mu} + \partial_{\mu}{a \over f_a} \right )^2 \, + \, \nonumber\\ && + \, {a \over f_a} \left (F_X\tilde{F}_X \, + \, F\tilde{F} \right ) \,. \end{aligned}$$ The first row describes the gauge invariant coupling of $\psi$-fermions to gluons and to $X_{\mu}$, with $D_{\mu}$ denoting a standard covariant derivative with respect to both interactions. The contraction of color and flavor indexes is obvious and is not written explicitly. The second row describes the usual gauge field strengths, with irrelevant constants absorbed, plus the ${\vartheta}$-term for QCD. The ${\vartheta}$-term for the abelian group is irrelevant. The third row describes the gauge invariant mass of $X_{\mu}$, in which $a$ plays the role of the Stückelberg field. Finally, the fourth row describes the coupling of the axion to the dual field strengths. The normalization is such that under $U(1)_X$ the fermions contribute an anomalous shift, $$\delta L_{fermion} = \alpha \left (F_X\tilde{F}_X \, + F\tilde{F} \right )\,, \label{fermionSX}$$ which is exactly compensated by the shift of axion (\[AxionX\]).\ Notice, if we switch for a moment to the conventional language of explicit symmetry-breaking by instantons, we will need to add the ’t Hooft determinant term to Alice’s Lagrangian. This term has a form $$L_{'t Hooft} \, = \, e^{i \left ({a \over f_a} - {\vartheta}\right)} det(\bar{\psi}\psi)\, + \,... \, , \label{detFermion1}$$ and is manifestly gauge invariant under $U(1)_X$. This shows clearly that - as was stressed in the introduction - the instantons [*cannot*]{} avoid requirement of anomaly matching for $U(1)_X$-symmetry. This constraint was not taken into the account in ’t Hoofts original paper [@anomalymatching].\ Notice, that on top of the anomaly-free local symmetry, there continues to exist the “old" anomalous global symmetry $U(1)_A$, which only acts on fermions and not on either $X_{\mu}$ or $a$. Because of this, the existence of new gauge symmetry $U(1)_X$ is not affecting the solution of the strong-CP problem: Alice can still shift away the ${\vartheta}$-term by the old anomalous global chiral rotation. This is also clear from the fact that the new degrees of freedom that we have added to the theory can be made arbitrarily massive, by taking the quantity $m_a \equiv g_Xf_a$ large. Below the scale $m_a$ the modes $X_{\mu}$ and $a$ decouple and can be integrated out. Alice’s theory then is left with Anomalous global symmetry $U(1)_A$ as before. The same continues to be true for arbitrarily small, but finite, value of $m_a$. Correspondingly, in the low energy theory of Bob the story with $\eta'$ persists as before, modulo a small mixing with $a$. By taking the couplings $g_X$ and $1/f_a$ small, we can make $X_{\mu}$ and $a$ to interact with the QCD sector arbitrarily weakly. In this way they only play the role of spectators for monitoring the anomaly matching.\ Now, first we need to show that in Bob’s theory, in the absence of $\lambda$-fermions, the gauge anomaly is fully matched by $\eta'$, and that the decoupling limit, $$g_X \rightarrow 0,~~f_a \rightarrow \infty, ~~m_a=finite, \label{decouplingX}$$ is smooth. The relevant part of Bob’s theory without $\lambda$-fermions has the following form, $$\begin{aligned} \label{BobLX} L_{Bob} (bosons) \, &&= \, {1 \over 2} E^2 \, - \,\left ( {\eta' \over f_{\eta}} + {a \over f_a} \right) E \, - \nonumber\\ && - \, F_X^2 \, + \, {1 \over 2} (f_a)^2\left (g_XX_{\mu} + \partial_{\mu}{a \over f_a} \right )^2 \, + \, \nonumber\\ && + \, {1 \over 2} (f_{\eta})^2\left (g_XX_{\mu} - \partial_{\mu}{\eta' \over f_{\eta}} \right )^2 \,, \end{aligned}$$ (where we have restored some numerical factors for convenience). The relative signs inside the brackets are important, since they ensure the gauge invariance under the above-discussed normalization. The gauge-invariance of the theory under $U(1)_X$ is ensured by the corresponding shift of $\eta'$: $${\eta' \over f_{\eta} } \rightarrow {\eta' \over f_{\eta} } \, + \, \alpha(x) \,. \label{ETAX-shift}$$ This shift is fixed from the fact that $\eta'$ transforms as the phase of the fermion ’t Hooft determinant. It is clear that although spectators ensure the invariance of the theory under the local $U(1)_X$-transformation, they do not affect Bob’s solution of the strong-CP problem, and the decoupling limit (\[decouplingX\]) is smooth. Indeed, solving the equation of motion for $E$ we obtain, $$\label{EX} E \, = \, \left( {\eta' \over f_{\eta}} + {a \over f_a} -{\vartheta}\right),$$ where ${\vartheta}$ as before is an integration constant that is matched with Alice’s ${\vartheta}$. We see that the electric field $E$ continues to be nullified by the following degree of freedom, $$\label{Etilde} \tilde{\eta'} \, \equiv \, { f_a\eta' + f_{\eta} a \over \sqrt{ f_a^2 + f_{\eta}^2}}\,.$$ Diagonalizing the mass-matrix, we can see easily that it is exactly this degree of freedom that is eaten-up by the QCD $3$-form $C$ and is getting a mass, $$m_{\tilde{\eta'}}^2 \, = \, {1 \over f_a^2 } + {1 \over f_{\eta}^2} \, . \label{massEX}$$ The orthogonal combination, $$\label{Atilde} \tilde{a} \, \equiv \, { f_{\eta'} \eta' - f_{a} a \over \sqrt{ f_a^2 + f_{\eta}^2}} ,$$ is eaten up by the $X_{\mu}$ gauge field and is getting the following mass, $$m_{\tilde{a}}^2 \, = \, g_X^2 \, \left (f_a^2 + f_{\eta}^2 \right ) \,. \label{massEX}$$ That is, $\tilde{a}$ becomes a longitudinal polarization of $X_{\mu}$ and the two form a Proca field of mass $m_{\tilde{a}}$.\ \[As a consistency check, notice that $\tilde{\eta'}$ is exactly the combination that would get mass from ’t Hooft determinant (\[detFermion1\]), if we were to use the conventional language of instantons. This is obvious from plugging the last expression in (\[detFermion\]) into (\[detFermion1\]). This clearly shows that the two languages are alternatives to each other and must not be used simultaneously, in order to avoid double-counting.\]\ In the decoupling limit (\[decouplingX\]) we have $$\tilde{\eta'} \rightarrow \eta',~{\rm and}~ \tilde {a} \rightarrow a \,, \label{limitX}$$ and the two sectors decouple. Thus, the limit is smooth and the strong-CP problem is solved for arbitrary values of $g_X$ and $f_a$. The anomalous shifts generated by $\eta'$ and $a$ exactly cancel each other for arbitrary choice of these parameters. From here it is clear that $\lambda$-fermions must be anomaly-free with respect to $U(1)_X$. Correspondingly, they must be anomaly-free under the global $U(1)_A$-symmetry as well, since the two act on fermions in the same way.\ Let us now show that from the last statement it follows that $\lambda$-s cannot exist, because them being anomaly-free with respect to $U(1)_A$, contradicts to the anomaly matching with respect to another spectator interaction that we shall now introduce - [*gravity*]{}.\ We thus, take into the account coupling to gravity. We shall treat gravity as a spectator interaction, keeping in mind a limit of infinitesimally small Newtonian coupling. Notice, since $\lambda$-s are massless, there always exists a chiral symmetry $$\lambda \rightarrow e^{i\beta} \lambda \,, \label{YLambda}$$ which is anomalous with respect to gravity. We shall gauge this symmetry and call it $U(1)_Y$. The important thing is that $U(1)_Y$ must be different from $U(1)_X$, since, as we just concluded, under the latter $\lambda$-s are anomaly-free. The gauge field of $U(1)_Y$ we shall denote by $Y_{\mu}$. This chiral symmetry exhibits both $(U(1)_Y)^3$ as well as a mixed $U(1)_Y$-gravity-gravity anomaly [@gravityanomaly], due to which the Lagrangian shifts as, $$\label{RRgravY} \delta L_{Bob} \, = \, \beta(x) (R\tilde{R} + F_Y\tilde{F}_Y) \, .$$ Once again, we absorb the anomaly coefficients into the normalizations of $R\tilde{R}$ and $F_Y\tilde{F}_Y$. In order to cancel the above anomalies we introduce a new spectator Green-Schwarz axion $b$, with a decay constant $f_b$. The $U(1)_Y$-gauge transformation acts on $Y_{\mu}$ and $b$ as $$Y_{\mu} \rightarrow Y_{\mu}\, + \, { 1 \over g_Y} \partial_{\mu} \beta(x)\,,~~ {b \over f_b} \rightarrow {b \over f_b} \, - \, \beta(x) \,. \label{AxionY}$$ where $g_Y$ is the gauge coupling. The gauge invariant version of fermionic part of Bob’s Lagrangian becomes, $$\begin{aligned} \label{BobFermion} && L_{Bob}(fermi) \, = \, \sum_{\lambda} i \overline{\lambda} \gamma^{\mu} D_{\mu} \lambda \, - \, F_Y^2 \, + \nonumber\\ && + \, (f_b)^2 \left (g_YY_{\mu} + \partial_{\mu}{b \over f_b} \right )^2 \, + \, {b \over f_b} \left (F_Y\tilde{F}_Y \, + \, R\tilde{R} \right ) \, + \nonumber\\ && + \, ({\rm pure ~ gravitational~action}) \, , \end{aligned}$$ where, of course derivatives are replaced by their covariant versions with respect to all gauge interactions including gravity. Obviously, the gravitational anomalies of $U(1)_X$ are cancelled among $\eta'$ and $a$ and that sector is not displayed explicitly. Notice, in case if Bob’s action contains some higher order interactions among $\lambda$-s that break the global $U(1)_Y$-symmetry explicitly down to some discrete $Z_n$-subgroup, with $n$ some integer, this is no obstacle for gauging it. We just need to promote the coefficient of every such operator into a [*spurion*]{} that transforms under gauge $U(1)_Y$ in the appropriate way. For this the simple prescription is: Each fermionic field $\lambda$ or $\bar{\lambda}$ entering such a vertex, must be accompanied by a factor $e^{i{b \over f_b}}$ or $e^{-i{b \over f_b}}$ respectively. For example, an operator $\lambda^n e^{i{b \over f_b}n}$ is manifestly gauge invariant. Now let us move into the theory of Alice. All the spectator interactions including gravity, $Y_{\mu}$ and the axion $b$, continue to be part of Alice’s theory, but - by definition - not the $\lambda$-fermions. Under the gauge $U(1)_Y$-transformation $b$ continues to generate the anomalous shifts of the Lagrangian, which would be matched by (\[RRgravY\]) if $\lambda$-fermions were around. Thus, we are ready to ask: What matches the $U(1)_Y$-anomalies generated by the axion $b$ in theory of Alice? Alice only has $\psi$-fermions at her disposal. But, the only symmetry anomalous with respect to gravity that acts on $\psi$-s is axial $U(1)_X$. This creates an inconsistency: First, we know that $U(1)_X$ and $U(1)_Y$ must be different; Secondly, all the $U(1)_X$-anomalies, including the mixed gravitational one, are fully taken care by the axion $a$. So in Alice’s theory there are no fermions that could compesate the gravitational anomaly of $U(1)_Y$.\ Thus, the assumption that in Bob-s theory there can exists some massless $\lambda$-fermions lead us to a contradiction. Thus, there is no room for any massless fermions (composite or elementary) in the theory of Bob.\ Notice, for reaching this conclusion is was important to use gravity as a spectator interaction. Without gravity, we could only prove that the low energy massless fermions must be color singlets, but would not be able to exclude their existence. The power of gravity is that there are no “gravity-singlets" in nature and existence gravitational anomaly is universal for any set of massless fermions. This gave us an additional matching condition, which is absent in pure QCD. ### Breaking of $SU(N_f)_L \otimes SU(N_f)_R$ Chiral Symmetry As a result of the above analysis, we are left with the following two logical possibilities:\ [*(1)*]{} The fermions stay elementary, but acquire effective mass-terms from symmetry breaking;\ or\ [*(2)*]{} The massless fermions disappear at low energies, because they all form bound-states. Since we have already excluded the possibility of massless fermionic bound-states, any bound-state formed by the fermions must be either massive or bosonic. The formation of the bound-states may or may not be due to confinement, about which we make no assumption.\ The above two logical possibilities are not inter-exclusive. We need to prove that each of of them implies that the chiral symmetry is completely broken.\ [*Option (1)*]{}. From the option [*(1)*]{} the complete breaking of chiral symmetry follows in an obvious way. It is enough to notice that the fermion mass matrix $M_{\bar{j}}^{i}\, \bar{\psi}^{\bar{j}} \psi_{i}$ transforms as a bi-fundamental representation $(\bar{N}_f, N_f)$ of $U(N_f)_L\times U(N_f)_R$ flavor symmetry. For an arbitrary bi-fundamental matrix $M_{\bar{j}}^{i}$ with all the eigenvalues non-zero the maximal little group is $SU(N_f)_{R+L} \otimes U(1)_V\otimes Z_{2}$. Hence, this is the largest sub-group that can survive unbroken, under the condition that all fermions are massive.\ [*Option (2)*]{}. The option [*(2)*]{} also implies breaking of $SU(N_f)_L \otimes SU(N_f)_R$, at least down to its maximal anomaly-free subgroup. The proof, follows from applying ’t Hooft’s anomaly matching condition with spectator fermions [@anomalymatching]. The reason in our case it allows us to make a stronger statement is because we have already gained a small advantage with respect to ’t Hooft’s starting position: Thanks to the matching anomalies of weakly-gauged axial symmetry, while using gravity and Green-Schwarz axions as spectators, we have already eliminated the possibility of any massless fermions altogether. This simplifies our task significantly. The argument can be presented as a particular case of a more general situation.\ Consider a theory based on a global chiral symmetry group $G_{chiral}$ and with the set of fermions $\psi$ transforming under it. We shall allow $\psi$-fermions to also interact via some other gauge interaction $H$. Now, assume that, due to unspecified dynamics of $H$, below certain scale $\Lambda$ all the $\psi$-fermions form bound-states, so that there are no fermions transforming under $G_{chiral}$ in the low energy theory. Then, we can prove that in deep infrared the chiral group $G_{chiral}$ must be spontaneously broken at least down to its maximal anomaly-free subgroup.\ Following ’t Hooft, let us assume that $G_{chiral}$ is gauged weakly. (In fact, would be enough to gauge different sets of chiral abelian subgroups of $G_{chiral}$.) We do this by introducing the set of gauge fields $B_{\mu}^{(m)}$ with $m = 1,2, ...$. This gauging creates the chiral anomalies of $(G_{chiral})^3$-type. In order to cancel them, we introduce the set of massless colorless chiral fermions $\chi$ only charged under $G_{chiral}$. Now, above the scale $\Lambda$ the $(G_{chiral})^3$ anomalies are cancelled among $\chi$-s and $\psi$-s. But, below the scale $\Lambda$, by assumption there are no massless fermions available from the $\psi$-sector. So, if the gauge fields $B_{\mu}^{(m)}$ of $G_{chiral}$ survive massless at low energies, it would be impossible to match the anomalies created by $\chi$-s. Thus, the only consistent possibility is that below the scale $\Lambda$ the group $G_{chiral}$ is Higgsed down to its anomaly-free subgroup, $G_{free} \subset G_{chiral}$, by some order parameters formed out of $\psi$-s. Since the argument goes through for arbitrarily-weak gauge coupling $g_B$, it also should persists for the case when $G_{chiral}$ is global. Thus, we conclude that even if all the $\psi$-fermions at low energies are in either massive or bosonic bounds-states, nevertheless the anomalous part of the chiral flavor group $G_{chiral}$ must be fully broken.\ It is trivial to apply the above general argument to QCD with $N_f$ flavors. We just have to take $G_{chiral} = SU(N_f)_L\times SU(N_f)_R$ and assume that the role of $H$ is played by the gauge color group. We then conclude that the flavor group must be spontaneously broken down to an anomaly-free subgroup $G_{free}$.\ Let us reiterate that this is a stronger conclusion than what one would reach by applying ’t Hooft’s anomaly matching without knowing that massless fermions are forbidden at low energies. In such a case, the surviving low-energy group would not need to be anomaly-free. It would only need to deliver the same anomaly as the high-energy one, in order to match the anomalies of the spectator $\chi$-fermions. This would give a milder constraint.\ The physical meaning of the spontaneous breaking of chiral flavor symmetry by the condensate of $\psi$-fermions is very transparent. This dynamical symmetry-breaking is necessary, in order for $\psi$-fermions to deliver the Nambu-Goldstone bosons, the pions $\pi^{(m)}$, which effectively assume the role of Stückelberg-Wess-Zumino-Green-Schwarz type axions and cancel the anomalies of $\chi$-fermions in Bob’s theory. For example, consider a pion $\pi^{(m)}$ resulting from the spontaneous breaking of a given anomalous abelian subgroup $U(1)_{chiral}^{(m)}$. Integrating out $\psi$-fermions, this pion acquires an anomalous couplings, $$\label{BobBb} L_{Bob} \, = \, {\pi^{(m)} \over f_{\pi^{(m)}}} F_{(m)}\tilde{F}_{(m)} \, + \, ... \, ,$$ where $F_{(m)}$ is the field strength of $B_{\mu}^{(m)}$ and $\tilde{F}_{(m)}$ its dual (again, the coefficient has been rescaled in normalization). Under, $U(1)_{chiral}^{(m)}$ the pion shifts as $${\pi^{(m)} \over f_{\pi^{(m)}}} \, \rightarrow {\pi^{(m)} \over f_{\pi^{(m)}}} \, + \, \beta^{(m)}(x) \,. \label{shiftPionQCD}$$ and cancels $U(1)_{chiral}^{(m)}$-anomaly generated by the $\chi$-fermions in Bob’s theory. This completes the argument. Generalization to gravity ========================= Although in our discussion we used the QCD-terminology, we tried to keep it maximally general and independent from the specifics of the QCD dynamics. Correspondingly, we can attempt to apply our arguments to any theory which has a non-zero topological susceptibility in the absence of fermions and a chiral anomaly in their presence. Every step of our discussion can be repeated to conclude that in any such theory the fermions must condense and break Chiral symmetry. Moreover, there must exist a pseudo-Goldstone analog of $\eta'$ that gets a mass in this process, by being eaten-up by a $3$-form. Beyond this point the specifics of the theory must be taken into the account and discussion becomes more subtle. We shall now apply our reasoning to gravity. It has already been noticed [@gia3form] that gravity contains both of the required ingredients. First, it contains $C$-field in form of a gravitational Chern-Simons $3$-form, $$C_g\equiv \Gamma \wedge d \Gamma + {2 \over 3} \Gamma\wedge \Gamma\wedge \Gamma,~~~ E_g \equiv R\tilde{R}= ^*\mathrm{d}C_g\,, \label{eqEG}$$ where $\Gamma$ is the connection, $R$ is the Riemann tensor and $\tilde{R}$ its Hodge-dual. Secondly, the chiral current exhibits the well-known chiral gravitational anomaly [@gravityanomaly], $$\partial^{\mu} J_{\mu}^{(A)} \, = \, R\tilde{R} \, , \label{divRR}$$ where, the current includes all the fermions of the theory $\psi_{(j)}$, which can be written in the same handiness basis (e.g., all left-handed). This introduces a difference that in gravity, with the same number of fermion degrees of freedom, the flavor group is larger than in QCD. Ignoring other interactions, the flavor group of gravity is $U(N)$, where $N$ counts number of all the massless Weyl fermions. At present we have no knowledge about the existence of non-zero topological susceptibility of the vacuum in pure gravity (i.e., gravity without massless fermions and/or axions). Therefore, we take this as our starting [*assumption*]{}. We assume that due to some unspecified physics the following correlator is non-zero in a fermion-free version of quantum gravity, $$\label{ETGrav} \langle E_g, E_g\rangle_{p\to 0}\, = \, const \neq 0 \, .$$ Einstein gravity exhibits no mass gap (i.e., graviton is massless). However, as we have discussed, the existence of the constant electric field $E$ in the vacuum of pure glue, directly follows from the assumption of non-zero vacuum topological susceptibility and is independent from the assumptions of mass gap and/or confinement. The same argument goes through in gravity: From the assumption of non-zero vacuum topological susceptibility (\[ETGrav\]), it directly follows that $C_g$ is a massless $3$-form field, with its propagator having a pole at $p^2=0$. Correspondingly, the vacuum houses a constant CP-odd electric field $$E_g = {\vartheta}_g \,. \label{Egrav}$$ Just like in case of pure glue, this electric field is the way Bob understands the physical CP violation, which in Alice’s description is due to physical observability of the gravitational ${\vartheta}_g$-term, $$\label{RRgrav} L_{Alice} \, = \, {\vartheta}_g R\tilde{R} \, + \, ...\, .$$ When Alice changes the value of ${\vartheta}_g$, Bob changes the value of $E_g$. [^4]\ The next step is to introduce some massless fermions. In order to make comparison with the case of QCD, let us introduce the same number of fermionic degrees of freedom. What in QCD would be $N_f$ flavors, by gravity are seen as $N = 2N_f$ massless Weyl fermions, all of which can be written in left-handed basis, $\psi_{i}$, where $i =1,2,...N$. These fermions form a fundamental representation of $U(N)$ flavor group. The $U(1)_{A}$ subgroup of this symmetry, which acts on the fermions as, $$\psi_{i} \rightarrow {\rm e}^{i\alpha} \psi_{i} \,, \label{axialG}$$ is anomalous with respect to gravity. The corresponding current $J_{\mu}^{(A)}$ exhibits an anomalous divergence [@gravityanomaly] given by (\[divRR\]). Therefore, under the chiral transformation (\[axialG\]) the Lagrangian shifts as $$\delta L_{Alice} \, = \, \alpha R\tilde{R} \,. \label{varRR}$$ Once again, all additional coefficients are absorbed in normalization of $R\tilde{R}$. Due to this, just as in QCD, the introduction of massless fermions makes the gravitational analog of ${\vartheta}$ unphysical and correspondingly makes the vacuum topological susceptibility zero. Correspondingly, CP is unbroken. The restoration of CP-invariance is [*universal*]{} and Alice and Bob must agree on this fact.\ In Alice’s language this is understood as a freedom to arbitrarily shift ${\vartheta}_g$ by $U(1)_A$ chiral transformation (\[axialG\]).\ In Bob’s theory, this must be matched by the appearance of a pseudo-scalar $\eta_g'$ that Higgses the gravitational Chern-Simons $3$-form $C_g$ exactly in the same way as $\eta'$-meson Higgses the analogous Chern-Simons $3$-form in QCD. The effective Lagrangian describing this phenomenon is fixed by the gauge invariance and the anomalous coupling, and therefore, is very similar to (\[effectiveQCDeta\]) $$\label{effectiveG} L_{Bob} \, = \, {1 \over 2} E_g^2 \, - \, {\eta_g' \over f_{g}} \, E_g \, + \, {1 \over 2} \partial_{\mu}{\eta_g'} \partial^{\mu} {\eta_g'},$$ Of course, here and everywhere the generally-covariant contraction of indexes is assumed. Solving the equation of motion for $E_g$, $$\label{solEG} E_g \, = \, \left({\eta_g' \over f_{g}}-{\vartheta}_g\right),$$ and plugging it into the equation for $\eta'_g$, we get, $$\label{boxetaG} \Box\, \eta'_g + {1 \over f_{g} }\left({\eta'_g \over f_{g}} -{\vartheta}_g \right) = 0 \,.$$ Thus, exactly as it happened in QCD, we conclude that the vacuum is at ${\eta_g' \over f_{g}}={\vartheta}_g$ so that the VEV of the CP-odd electric field vanishes, $E_g=0$. This is the way Bob understands CP-conservation in his theory. CP-invariance of the vacuum implies that under the chiral transformation (\[axialG\]) performed by Alice, $\eta_g'$ must shift as $${\eta'_g \over f_{g}} \rightarrow {\eta'_g \over f_{g}} \, + \, \alpha \,, \label{shiftEtaG}$$ in order to compensate the shift of ${\vartheta}_g$, $${\vartheta}_g \rightarrow {\vartheta}_g \, + \, \alpha \,, \label{shiftThetaG}$$ which follows from (\[varRR\]). Thus, making the exact same matching - as was done in the case of QCD - between the descriptions of Alice and Bob, we arrive to the conclusion that $\eta'_g$ is a pseudo-Goldstone boson, which originates from a spontaneous breaking of $U(1)_A$ symmetry (\[axialG\]) triggered by a fermion condensate. Introducing a notation $\Phi_{ij} \equiv \psi_i \psi_j$, we construct the gravitational version of the order parameter for breaking the $U(1)_A$-symmetry: $det \Phi \equiv \epsilon^{i_1,...i_N} \epsilon^{j_1,...j_N} \Phi_{i_1j_1}...\Phi_{i_Nj_N}$, which is invariant under the $SU(N)$ subgroup of the original $U(N)$ flavor group. Notice, $\Phi_{ij}$ being a Lorentz-invariant of Majorana-type is symmetric in $i,j$-indexes and transforms as $N(N+1)/2$-dimensional representation of $SU(N)$ flavor group. For spontaneous breaking of $U(1)_A$ the necessary condition is that the vacuum expectation value $\langle det \Phi \rangle$ is non-zero. Existence of $\eta_g'$ indicates that this is the case.\ We now wish to understand what is the fate of $SU(N)$-part of the flavor symmetry. For this we first need to check whether a conclusion that we reached in QCD - namely, the absence of massless fermions in Bob’s theory - holds in gravity. In QCD we reached this conclusion by gauging the axial symmetry and matching anomalies. We shall do exactly the same in gravity. We promote global $U(1)_A$ into a gauge $U(1)_X$ symmetry and cancel anomaly by a Green-Schwarz axion $a$. Then, Alice’s anomaly-free Lagrangian is, $$\begin{aligned} \label{AliceXG} L_{Alice} \, &=& \, \sum_{\psi} i \overline{\psi} \gamma^{\mu} D_{\mu} \psi \, -\nonumber\\ && - \, F_X^2 \, + \, {\vartheta}_g R\tilde{R} \nonumber\\ && + \, (f_a)^2\left (g_XX_{\mu} + \partial_{\mu}{a \over f_a} \right )^2 \, + \, \nonumber\\ && + \, {a \over f_a} \left (F_X\tilde{F}_X \, + \, R\tilde{R} \right ) \, + \nonumber\\ && + \, M_P^2 {\mathcal R} \,. \end{aligned}$$ where the last row stands for pure-gravitational part with $M_P$ the Planck mass and ${\mathcal R}$ the Ricci scalar. Cancellation of gravitational anomalies works exactly as in the case of QCD: The would-be anomalous shift of the Lagrangian (\[varRR\]) is compensated by the respective gauge shift of the axion (\[AxionX\]).\ [**\[**]{} At this point we have to open a parenthesis, because here comes the followings [**caveat**]{}: In QCD, both interactions - $U(1)_X$ and gravity - were introduced as spectators and both could be chosen to be [*arbitrarily weak*]{}. In other words, there was no problem in taking the decoupling limit (\[decouplingX\]) smoothly, since simultaneously we could take $M_P \rightarrow \infty$. But, in gravity the story is different, since gravity can not play the role of its own spectator. Correspondingly, if we want to make sure that the decoupling limit is smooth we should be able to take (\[decouplingX\]) while keeping $M_P$ [*finite*]{}. One may argue that this is not legitimate. There are the following two (inter-related) reasons for that. The first one follows from a wide-spread view of incompatibility between gravity and global symmetries. Subscribing to this view, we have to be worried about taking the decoupling limit (\[decouplingX\]) with finite $M_P$, since in this limit we are left with a global $U(1)_A$-symmetry coupled to gravity. The second reason is based on a stronger constraint that goes under the name of “Weak Gravity Conjecture" [@weakgravity]. According to this conjecture, not only global, but also gauged symmetries are incompatible with gravity, if their strength is weaker than the strength of gravity. Again, one could say that this argument precludes us from taking the limit (\[decouplingX\]), while keeping a finite $M_P$. However, given what we are after, the above would result into a circular argument that would get us to nowhere. We are not satisfied with a limited knowledge that gravity is inconsistent with global symmetries and we want to understand why and how she gets rid of them. In other words, our goal is precisely to understands how gravity responds to global symmetries dynamically. For this, we must allow ourselves to take the global limit and see what happens. As we shall see, just as in case of QCD, gravity responds by eliminating massless fermions and breaking all the chiral symmetries dynamically. In this respect, the results of the present paper can be viewed as a dynamical justification for incompatibility between global (or weakly gauged) symmetries with gravity: If we try to insist on such global (or weakly-gauged) symmetry, the fermions condense and break it [*dynamically*]{}! [**\]**]{}\ With above in mind we allow ourselves to take $g_X$ and $f_a^{-1}$ arbitrarily small, while keeping finite $M_P$. The rest of the arguments closely follow the QCD reasoning. First, in theory of Bob the axion shift is matched by $\eta_g'$. The corresponding gauge invariant Lagrangian is just a gravitational version of (\[BobLX\]), $$\begin{aligned} \label{BobLXG} L_{Bob} (bosons) \, &&= \, {1 \over 2} E_g^2 \, - \,\left ( {\eta_g' \over f_g} + {a \over f_a} \right) E_g \, - \nonumber\\ && - \, F_X^2 \, + \, {1 \over 2} (f_a)^2\left (g_XX_{\mu} + \partial_{\mu}{a \over f_a} \right )^2 \, + \, \nonumber\\ && + \, {1 \over 2} (f_g)^2\left (g_XX_{\mu} - \partial_{\mu}{\eta_g' \over f_g} \right )^2 + \nonumber\\ && + \, M_P^2 {\mathcal R} \,, \end{aligned}$$ where the gauge shifts of $\eta'_g$ and $a$ compensate each other. From here, performing exactly the same steps as in QCD, we can conclude that in Bob’s theory there cannot exist any massless fermions, since chiral anomaly is fully booked by $\eta'_g$.\ Indeed, if either a subset of Alice’s $\psi$-fermions or their composites could survive massless in Bob’s theory, this surviving massless set $\lambda$ must be anomaly-free with respect to $U(1)_X$, since this anomaly cancels among the axion $a$ and $\eta'_g$-meson. But, since $\lambda$-s are massless, there inevitably exists a chiral symmetry $U(1)_Y$ acting on them as (\[YLambda\]). Just as in the case of QCD, we gauge this symmetry introducing the spectator gauge field $Y_{\mu}$ and a Green-Schwarz axion $b$, which transform under $U(1)_Y$ as (\[AxionY\]). The axion $b$ cancels $U(1)_Y$ anomalies of $\lambda$-fermions. This sector of Bob’s Lagrangian has the form (\[BobFermion\]). So far so good. But now, going to Alice’s theory, there are no fermions there able to cancel the anomalous shift of $b$-axion. Obviously, $\psi$-s cannot do this job, since the only symmetry anomalous with respect to gravity under which they transform is $U(1)_X$, which - by construction - is different from $U(1)_Y$. Thus, we see that Bob’s theory is not hospitable to any massless fermion species. This leaves us with the same two options as in QCD: The fermions must either get masses from the chiral symmetry breaking or form some massive composites (or both). Again, gauging various chiral subgroups of the $SU(N)$ group and performing ’t Hooft’s anomaly matching, we can conclude that chiral symmetry must be broken down to its anomaly-free subgroup. The only technical novelty in the matching argument is that, unlike QCD, we cannot introduce the [*massless*]{} spectator $\chi$-fermions, since all fermions interact gravitationally and this will simply be equivalent to the increase of $N$. Thus, we have to use Wess-Zumino-Green-Schwarz-type axions for such anomaly matching.\ The spectator axions can be introduced in a straightforward way for the entire $SU(N)$-symmetry or for its subgroups. As an useful guide-line, one can think of these axions as originating from integrating-out the set of left-handed $\chi$-fermions transforming as anti-fundamental representation of $SU(N)$-group. Then, the $\psi_i$- and $\chi^i$-fermions together form an anomaly-free set of the $SU(N)$-group. Now assume that $\chi$-fermions get large masses from the Yukawa couplings to the VEVs of some elementary scalars that Higgs the entire $SU(N)$-group. The Higgs scalars do not have any Yukawa couplings to the $\psi$-fermions, but only to $\chi_i$-s and some additional $SU(N)$-singlet fermions. As a result $\chi$-fermions mix with the additional gauge singlet fermions and form heavy Dirac particles. Whereas, the $\psi$-fermions remain massless. The Higgsing of $SU(N)$-group results into $N^2-1$ Goldstone bosons $b^{(m)}~~(m=1,2,...N^2-1)$, which are eaten up by the corresponding $SU(N)$ gauge bosons, $B_{\mu} ^{(m)}$, and become massive. By a suitable choice of the VEV of the Higgs fields and the value of $SU(N)$ gauge coupling, we can make the $\chi$-fermions [*arbitrarily heavy*]{} and the gauge bosons [*arbitrarily light*]{}. After integrating out the heavy $\chi$-fermions, the anomalies of $\psi$ fermions are cancelled by the Goldstone bosons $b^{(m)}$ via effective Wess-Zumino-Witten terms [@WZW]. In other words, the Goldstone bosons $b^{(m)}$ are the low energy representatives of the heavy $\chi$-fermions “delegated" for the anomaly matching. We must stress, however, that since in our case we are finally taking the decoupling limit, we can introduce the compensating Goldstone-bosons $b^{(m)}$ directly, without any explicit reference to the heavy fermions. Now, we know that below some scale $\Lambda$ the $\psi$-fermions must decouple, since the low energy theory of Bob cannot contain any massless fermions. Because the scale $\Lambda$ is totally independent of $SU(N)$-gauge dynamics, we can make the masses of $SU(N)$-gauge bosons arbitrarily smaller than $\Lambda$. By consistency, the anomalies generated by the spectator Goldstone bosons $b^{(m)}$ must continue to cancel in effective theory below the scale $\Lambda$. This is only possible if $\psi$-fermions delegate their own representative Nambu-Goldstone bosons, $\pi^{(m)}$, for canceling the anomalies of spectator Wess-Zumino-Green-Schwarz axions, $b^{(m)}$. Thus, the $SU(N)$-symmetry must be spontaneously broken by $\psi$-condensate down to an anomaly-free subgroup, resulting into a set of Nambu-Goldstone bosons, $\pi^{(m)}$. We shall refer to them as gravi-pions.\ For completeness of presentation, and in order to see how $\pi^{(m)}$-s and $b^{(m)}$-s share their jobs, let us explicitly go through the story for some chiral abelian sub-group of $SU(N)$-symmetry. We denote it by $U(1)_{(m)}$ and the corresponding generator by $Q_{(m)}$. This symmetry acts on fermion species as $$\psi \rightarrow e^{i Q_{(m)} \beta^{(m)}} \psi \, , \label{Qgravity}$$ where $\beta^{(m)}$ is a transformation parameter. Let us denote the corresponding gauge field by $B_{\mu}^{(m)}$. Since, the generator $Q_{(m)}$ has a zero trace, the mixed anomalies of the type $U(1)_{(m)}-U(1)_X-U(1)_X$ and $U(1)_{(m)}$-gravity-gravity are absent. The anomaly $U(1)_{(m)}-U(1)_{(m)}-U(1)_X$ will be taken care by the coupling of axion $a$. So we have to take care of $(U(1)_{(m)})^3$ anomaly. For this we need to introduce the corresponding Green-Schwarz axion, $b^{(m)}$. We normalize its shifts under (\[Qgravity\]) as $$\begin{aligned} B^{(m)}_{\mu} &&\rightarrow \, B^{(m)}_{\mu} + \, { 1 \over g_m} \partial_{\mu} \beta^{(m)} (x)\,, \nonumber\\ {b^{(m)} \over f_m} \, && \rightarrow {b^{(m)} \over f_m} \, - \, \beta^{(m)}(x) \, , \label{gaugeBb} \end{aligned}$$ where $g_m$ is the gauge coupling and $f_m$ is the axion decay constant. The resulting gauge invariant Lagrangian of Alice has the following form $$\begin{aligned} \label{AliceBb} L_{Alice} \, &=& \, \sum_{\psi} i \overline{\psi} \gamma^{\mu} D_{\mu} \psi \, - F_{(m)}^2 + \, \nonumber\\ && + \, (f_m)^2\left (g_mB_{\mu}^{(m)} + \partial_{\mu}{b^{(m)} \over f_m} \right )^2 \, + \, \nonumber\\ && + \, \left( c_m {a \over f_a}\, + \, {b^{(m)} \over f_m} \right) F_{(m)}\tilde{F}_{(m)} \, +...\, .\, \end{aligned}$$ where $D_{\mu}$ is covariant also with respect to (\[gaugeBb\]). The rest of the terms denoted by “..." are the same as in (\[AliceXG\]). The constant $c_m$ is a relative anomaly coefficient that remains after we absorb all other constants in normalizations of the dual field strengths. Then, the anomaly generated by the fermion transformation (\[Qgravity\]) is cancelled by the respective shift of axion $b^{(m)}$ given by (\[gaugeBb\]).\ We now go to theory of Bob, where, as we know, there are no massless fermions. However, the spectator Green-Schwarz axions continue to generate anomalies by their gauge shifts (\[gaugeBb\]). What degrees of freedom cancel these anomalies? The answer is that the $\psi$-fermions that cannot penetrate into Bob’s theory in form of fermions, must condense and deliver the Nambu-Goldstone degrees of freedom, which have anomalous couplings and shift appropriately under the gauge symmetry. That is, the only possibility to cancel the anomalous shift generated by (\[gaugeBb\]) is by the shift of a Goldstone boson which comes from the phase of a fermion condensate that breaks $U(1)_{(m)}$ spontaneously. This Nambu-Goldstone boson plays the role analogous to pion in QCD. The relevant part of Bob’s Lagrangian is, $$\begin{aligned} \label{BobBb} L_{Bob} (\pi_{(m)}) \, &=& - F_{(m)}^2 + \, \left({\pi^{(m)} \over f_{\pi^{(m)}}} + {b^{(m)} \over f_m} \right) F_{(m)}\tilde{F}_{(m)} \, + \nonumber\\ && + \, (f_m)^2\left (g_mB_{\mu}^{(m)} + \partial_{\mu}{b^{(m)} \over f_m} \right )^2 \, + \nonumber\\ && + \, (f_{\pi^{(m)}} )^2\left (g_mB_{\mu}^{(m)} - \partial_{\mu}{\pi^{(m)} \over f_{\pi^{(m)}}} \right )^2 \, + \nonumber\\ && +\, \ c_m {a \over f_a} F_{(m)}\tilde{F}_{(m)} \, + ...\, . \end{aligned}$$ where “ ..." includes (\[BobLXG\]) and $f_{\pi^{(m)}}$ is the gravi-pion decay constant. The gauge shift (\[gaugeBb\]) is compensated by the corresponding shift of pion, $${\pi^{(m)} \over f_{\pi^{(m)}}} \, \rightarrow {\pi^{(m)} \over f_{\pi^{(m)}}} \, + \, \beta^{(m)}(x) \,. \label{shiftPion}$$ We see that the degree of freedom, $$\label{Btilde} \tilde{b}^{(m)} \, \equiv \, { f_{\pi^{(m)}} \pi^{(m)} - f_{m} b^{(m)} \over \sqrt{ f_m^2 + f_{\pi^{(m)}}^2}} ,$$ is eaten up by the $B_{\mu}^{(m)}$ gauge field and is getting the following mass, $$m_{\tilde{b}^{(m)}}^2 \, = \, g_m^2 \, \left ( f_m^2 + f_{\pi^{(m)}}^2 \right ) \,. \label{massBtilde}$$ The orthogonal combination, $$\label{Etilde} \tilde{\pi}^{(m)} \, \equiv \, { f_{\pi^{(m)}} b^{(m)} + f_{m} \pi^{(m)} \over \sqrt{ f_m^2 + f_{\pi^{(m)}}^2}} \,,$$ remains massless. In the decoupling limit $g_m \rightarrow 0,~f_{m} \rightarrow \infty$ we have $$\tilde{\pi}^{(m)} \rightarrow \pi^{(m)} ,~{\rm and}~ \tilde{b}^{(m)} \rightarrow b^{(m)} \,, \label{limitB}$$ and the two sectors decouple. Thus, we see that the anomaly matching by the spectator Wess-Zumino-Green-Schwarz axions shows that $SU(N)$ flavor group must be spontaneously broken by the fermion condensate down to its anomaly-free subgroup. Since in principle $SU(N)$ allows for different inequivalent embeddings of anomaly-free subgroups, we cannot say with certainty down to which one the symmetry must be broken. The $SO(N)$ subgroup could be one natural outcome resulting in total $N^2/2 + N/2 -1$ Goldstones.\ For phenomenological considerations the $SU(48)$ flavor group would be of particular interest, as this is the flavor group of Standard Model fermions in the limit in which all interactions except gravity are switched off. This flavor group naturally admits the gauging of a “diagonal" anomaly-free $SO(10)_{G}$ subgroup defined by the following sequence of embeddings, $$\begin{aligned} SO(10)_{G} &&\subset SO(10)\times SO(10)\times SO(10) \subset \nonumber\\ \subset && SU(16)\times SU(16)\times SU(16) \subset SU(48) \, . \label{SO10} \end{aligned}$$ The $48$-plet of fermions then reduces to three copies of $16$-dimensional spinor representation of $SO(10)_{G}$, which can serve as a symmetry group of grand unification. ### Persistence of dynamical symmetry breaking in presence of high-dimensional operators So far we were ignoring other sources of symmetry breaking. In gravity we cannot exclude existence of high-dimensional operators that could break $U(N)$-flavor symmetry explicitly. Can a presence of such operators change our conclusion about the dynamical symmetry breaking? The answer is negative, since such operators do not disturb anomaly matching by axions. This is because they are no obstacle for gauging the relevant flavor symmetries. Using the corresponding Green-Schwarz axions, such operators can be completed into their gauge-invariant versions. After performing the anomaly matching, and establishing the fact of dynamical spontaneous symmetry breaking, we can decouple axions and recover back the operator that adds an explicit breaking of a [*spontaneously-broken*]{} global symmetry. This contribution generates non-zero masses to the corresponding pions, promoting them into pseudo-Goldstone bosons, but is not effecting their existence. It is enough to demonstrate this for the chiral abelian subgroups. The prescription is the following. Consider an arbitrary high dimensional operator ${\mathcal O}( \bar{\psi},\psi)$ that depends on fermions and their conjugates and breaks the flavor symmetry explicitly. Let the charges of this operator under the set of chiral abelian symmetries $U(1)_{(m)} \times U(1)_X$ be $(q_m, q_X)$, in the units of charges of corresponding Green-Schwarz axions. Then, this operator can be easily promoted into its gauge invariant version under these symmetries by means of multiplication by an appropriate exponential factor, $$e^{- i \left (\sum_m q_m {b^{(m)} \over f_m } + q_X{a \over f_a} \right )} {\mathcal O}( \bar{\psi},\psi) \, . \label{Ogauge}$$ With this arrangement, we make sure that the gauge anomaly matching condition holds for arbitrarily-small values of the couplings between the spectator axions and fermions. Thus, the previous conclusion - that the anomaly-matching demands the existence of pions in theory of Bob - is unaffected by high-dimensional operators. In the decoupling limit, $f_m,f_a \rightarrow \infty$, we are left with global flavor symmetries that are, on one hand, spontaneously broken by the fermion condensate and, on the other hand, are explicitly broken by the high-dimensional operators (\[Ogauge\]). Correspondingly, the phenomenon of dynamical symmetry breaking by fermion condensate persists also in the presence of high-dimensional operators. The only important effect of such operators in that they contribute to the masses of those gvari-pions on which the depend explicitly. In other words, masses of gravi-pions become non-zero if there exist operators that break corresponding generators of the global flavor group explicitly. This masses however will be suppressed by the ratio of scales of the fermion condensate and the scale of the high-dimensional operator, which is unknown, but can be as large as the Planck mass.\ In summary, we showed that under the assumption of non-zero topological susceptibility in fermion-free version of gravity, after introduction of fermions with zero bare masses, the following happens: - Fermions condense and form a non-zero order parameter that breaks the chiral $SU(N)\times U(1)_A$-symmetry spontaneously down to its anomaly-free subgroup. - The would-be Goldstone boson of $U(1)_A$, which we call $\eta_g'$, becomes massive by being eaten by a three-form $C_g$, via the effect closely analogous to Higgs phenomenon [@gia3form]. This ensures the CP-invariance of gravitational vacuum. - No massless fermions exist in the low energy theory of gravity. - The spontaneous symmetry breaking $SU(N)$ down to its anomaly-free subgroup results into Goldstones, which we called gravi-pions. The breaking to the largest anomaly-free subgroup $SO(N)$ results into $N^2/2 +N/2 -1$ Goldstones. - There may exist some high-dimensional operators that explicitly break $SU(N)$ flavor symmetry. These can generate masses for gravi-pions. However, they abolish neither the phenomenon of dynamical symmetry breaking by fermion condensate, nor the generation of the fermion mass gap. Possibility of dynamical symmetry breaking by gravity is striking. In the next section we shall try to understand the fundamental meaning of this phenomenon. The role of micro black holes? ============================== We have seen that the three phenomena - breaking of chiral symmetry, generation of the $\eta'$-meson mass and decoupling of fermions from the low energy theory - follow from a single assumption of non-zero topological susceptibility of the vacuum in fermion-free version of the theory, without any explicit assumption about the confinement. Applying this reasoning to gravity, we are lead to a rather striking conclusion that gravity responds to the existence of massless fermions by dynamical breaking of chiral symmetry and generation of mass gap in fermion spectrum.\ This raises the following question:\ What underlying physics gives a non-zero correlator (\[ETGrav\]) in pure gravity, and why is it incompatible with the existence of massless fermion and chiral symmetry?\ We think that the answer may be provided by black holes and this connection can shed light on another known puzzle in gravity. It is well-known that black holes are incompatible with continuous global symmetries. This knowledge often goes under the name of “folk theorem", which says that continuous global symmetries must either be gauged or broken. The argument, that will not be repeated here, relies on no-hair properties of classical black holes. But the question is: How the black holes manage to explicitly break the symmetry at the microscopic level? If we look closely at our previous analysis, we may have a partial answer to this question. At least the part that concerns the chiral symmetry. Indeed, existence of massless fermions coupled to gravity, would imply existence of $U(N)$-flavor symmetry. This would be incompatible with black holes. But, we have shown that microscopic gravity could take care of this potential inconsistency via topological susceptibility: It generates a fermionic mass gap and simultaneously breaks the chiral symmetry dynamically. The “bootstrap" can be provided by the black holes themselves. Namely, by their contribution into the topological susceptibility of the vacuum. The small black holes are natural candidates for contributing into the correlator [@DFneu]. Once we accept that black holes states are inevitable part of the gravity spectrum, then, everything shall fall into places, provided we assume that the micro-black hole states $|BH_k\rangle$ contribute into the vacuum correlator $$\label{BH} \langle R\tilde{R}, R\tilde{R} \rangle_{p\to 0}\, = \sum_k \langle R\tilde{R} |BH_k\rangle \langle BH_k| R\tilde{R} \rangle \, + \, ... \neq 0 \, .$$ This contribution provides the link between the black holes and the breaking of global chiral symmetry.\ Of course, without more detailed control of quantum gravity, we cannot compute the value of the correlator (\[BH\]), however, some reasonable guesses can be made. First, it is reasonable to expect that the dominant contribution in (\[BH\]) comes from [*quantum*]{} black holes, i.e., the microscopic black holes of the smallest size and mass. Their existence follows from the existence of macroscopic (classical) black holes. The simple way to see this is to wait until the black hole becomes so small that its evaporation rate becomes highly non-thermal. Deviations from thermality are set by inverse entropy [@giaThermal], and therefore we can expect that the contribution into (\[BH\]) from the black holes with smallest entropy costs less suppression. Simultaneously, the corresponding size of a smallest entropy black hole marks the scale below which the quantum-gravity becomes strong. Next, we can give arguments why the scale of the correlator can naturally be smaller than the Planck scale, $M_P$. In theory with $N$ particle species there is an upper bound on the size of semi-classical black holes, which is given by the length $ = \sqrt{N} M_P^{-1}$ [@giaN]. So, in such theories the quantum gravity scale is set by $M_P/\sqrt{N}$ rather than by $M_P$. Secondly, we should take into account the non-perturbative suppression, which could be as strong as $e^{-N}$. What is interesting, is the universal nature of fermion condensate. Since the gravitational chiral anomaly involves all the existing fermion species, the generation of fermion condensate and chiral symmetry breaking must be universal throughout all the fermions of the theory. Of course, the most sensitive to the effect are the light fermions, and in the case of the Standard Model, the neutrinos. So it makes sense to ask whether such a condensate can be a significant contributor into neutrino masses [@DFneu; @Lena]. On the phenomenological side, the universal nature of the effect suggests that the experimental lower bound on the masses of the lightest fermions give us an experimental knowledge of the topological structure of the gravitational vacuum. In particular, they give upper bound on the value of gravitational vacuum topological susceptibility somewhere not much above the neutrino mass scale. Relation with confinement? ========================== The natural question to ask is whether there is any relation between our observations and phenomenon of the confinement, about which we have made no assumption. [^5] Of course, this assumption cannot apply to gravity. Naively, it seems that our arguments diminish the role of confinement in the story of chiral symmetry breaking, as they show that solely the assumption of vacuum topological susceptibility suffices to do the job. However, this is a superficial view. We never proved that topological susceptibility of the vacuum can be non-zero in a non-confining theory. The only example of $3+1$-dimensional Poincare invariant theory for which we are more or less confident that this quantity is non-zero, is massless-quark-free version of QCD, which is believed to be confining. So, it could very well be that there is an intrinsic connection between this vacuum correlator and confinement. Of course, having confinement as a necessary condition for its existence, would eliminate the possibility of non-zero gravitational topological susceptibility of the vacuum. Obviously, in such a case our arguments for chiral symmetry breaking would not apply to gravity. What speaks against such a possibility is a seemingly-independent “folk theorem" argument, showing that black holes cannot tolerate chiral symmetry. As discussed above, this cries for a “bootstrap" scenario in which the black holes make this intolerance self-consistent via breaking the chiral symmetry precisely by contributing into the gravitational topological susceptibility of the vacuum. In this light, it would be extremely important to better understand the contribution from black hole states in (\[BH\]). The positive result would create a precedent of breaking chiral symmetry in non-confining theory. This would also give reason for rethinking the role of confinement in chiral symmetry breaking also in confining theories, such as QCD. Acknowledgements {#acknowledgements .unnumbered} ================ We thank Cesar Gomez, for many lively and stimulating discussions, Gerard ’t Hooft, for discussions on anomaly matching using spectator axions and gravity, and Nico Wintergerst for useful discussions. We thank Gabriele Veneziano, for pointing out ref. [@veneziano] and for discussions. We thank Lena Funcke for discussions and collaboration on related phenomenological topics. This work was supported in part by the Humboldt Foundation under Humboldt Professorship, ERC Advanced Grant 339169 “Selfcompletion”, by TR 33 “The Dark Universe”, and by the DFG cluster of excellence “Origin and Structure of the Universe”. [10]{} E. Witten, “Current Algebra Theorems for the U(1) Goldstone Boson”, Nucl. Phys. **B156**, 269 (1979); G. Veneziano, “U(1) Without Instantons”, Nucl. Phys. **B159**, 213 (1979). G. ’t Hooft, “A Planar Diagram Theory for Strong Interactions", Nucl.  Phys.  [**B72**]{}, 461, (1974). G. ’t Hooft, “Symmetry Breaking through Bell-Jackiw Anomalies”. Phys. Rev. Lett. **37**, 8 (1976). G. ’t Hooft, “Computation of the Quantum Effects Due to a Four-Dimensional Pseudoparticle", Phys. Rev.  [**D14**]{}, 3432-3450 (1976), Erratum: Phys. Rev.  [**D18**]{}, 2199 (1978); Phys.  Rep.  [**142**]{}, 357 (1986). G. Veneziano, “Goldstone Mechanism From Gluon Dynamics", Published in Phys. Lett.  [**B 95**]{}, 90-92, (1980) G. Dvali, “Three-form gauging of axion symmetries and gravity,” arXiv:hep-th/0507215; “A Vacuum accumulation solution to the strong CP problem," Phys. Rev. [**D74**]{} 025019, (2006), hep-th/0510053. G. Dvali, R. Jackiw and S.-Y. Pi, “Topological mass generation in four dimensions,” Phys. Rev. Lett.  [**96**]{}, 081602 (2006), arXiv:hep-th/0511175. G. Dvali, S. Folkerts, and A. Franca, “How neutrino protects the axion”, Phys. Rev. D **89**, 105025 (2014), arXiv:1312.7273 \[hep-th\]. G. Dvali and L. Funcke, “Small neutrino masses from gravitational ${\vartheta}$-term”, Phys. Rev. D **93**, 113002 (2016), arXiv:1602.03191 \[hep-ph\]; “Domestic Axion", arXiv:1608.08969 \[hep-ph\] R. D. Peccei and H. R. Quinn, “$CP$ Conservation in the Presence of Instantons,” Phys. Rev. Lett.  [**38**]{}, 1440 (1977). The U(1) Problem and the Higgs Mechanism in Two-dimensions and Four-dimensions A.  Aurilia, Y. Takahashi, and P.K. Townsend, Phys.  Lett. [**95B**]{}, 265 (1980). G. Gabadadze, M. A.  Shifman, “D-Walls and Junctions in Supersymmetric Gluodynamics in the Large N Limit Suggest the Existence of Heavy Hadrons", Phys.  Rev. [**D61**]{}, 075014 (2000), hep-th/9910050 G. Ôt Hooft, “Recent Developments in Gauge Theories" (Plenum Press, 1980) 135; reprinted in Unity of Forces in the Universe Vol. II, A. Zee ed. (World Scientific, 1982) 1004. M. B. Green, J. H.  Schwarz, “Anomaly cancellations in supersymmetric D = 10 gauge theory and superstring theory”. Phys. Lett.   [**B 149**]{}, 117 (1984). M. Lüscher,“ The secret long range force in quantum field theories with instantons", Phys.  Lett. [**B78**]{}, 465 (1978). C.G. Callan, Jr., R.F.  Dashen, D.J.  Gross, “The Structure of the Gauge Theory Vacuum", Phys. Lett.  [**63B**]{}, 334-340 (1976); R. Jackiw, C. Rebbi, “Vacuum Periodicity in a Yang-Mills Quantum Theory”, Phys.  Rev.  Lett.   [**37**]{}, 172-175 (1976). C. Vafa and E. Witten, “Parity Conservation in Quantum Chromodynamics", Phys. Rev. Lett. [**53**]{}, 535 (1984) S. L. Adler, “Axial-Vector Vertex in Spinor Electrodynamics”, Phys. Rev. **177**, 2426 (1969); J. S. Bell and R. Jackiw, “A PCAC puzzle: $\pi_0\rightarrow\gamma\gamma$ in the $\sigma$-model”, Nuovo Cimento **A60**, 47 (1969). S. Weinberg, “A New Light Boson?,” Phys. Rev. Lett.  [**40**]{}, 223 (1978); F. Wilczek, “Problem of Strong $P$ and $T$ Invariance in the Presence of Instantons,” Phys. Rev. Lett.  [**40**]{}, 279 (1978). R. Delbourgo and A. Salam, “The gravitational correction to PCAC”, Phys. Lett. **B40**, 381 (1972); T. Eguchi and P. Freund, “Quantum Gravity and World Topology”, Phys. Rev. Lett. **37**, 1251 (1976); S. Deser, M. J. Duff, C. J. Isham, “Gravitationally induced $CP$ effects”, Phys. Lett. **B93**, 419 (1980); L. Alvarez-Gaume and E. Witten, “Gravitational anomalies”, Nucl. Phys. **B234**, 269 (1983). S. Deser, R. Jackiw and S. Templeton, “Three-Dimensional Massive Gauge Theories”, Phys. Rev.  Lett.  [**48**]{}, 975Ð978 (1982); R. Jackiw, S.-Y  Pi, . “ChernÐSimons modification of general relativity”. Phys.  Rev.  [**D68**]{}, 104012, (2003). N. Arkani-Hamed, L.  Motl, A.  Nicolis, C.  Vafa, “The String landscape, black holes and gravity as the weakest force", JHEP 0706 (2007) 060; hep-th/0601001. There have been several applications of this conjecture to the strength of axion coupling. For most recent discussion, see, A.  Hebecker, P.  Soler, “The Weak Gravity Conjecture and the Axionic Black Hole Paradox", arXiv:1702.06130 \[hep-th\], and references therein. J. Wess and B. Zumino, “Consequences of anomalous ward identities",  Phys.  Lett. [**B 37**]{}, 95 (1971); E. Witten, “Global Aspects of Current Algebra", Nucl.  Phys. [**B 223**]{}, 422 (1983). For a nice overview discussion on effective field theory treatment of anomalies, including generation of Wess-Zumino term from heavy fermions, see, J. Preskill, “Gauge anomalies in an effective field theory", Annals  Phys. [**210**]{}, 323 (1991). G. Dvali, “Non-Thermal Corrections to Hawking Radiation Versus the Information Paradox", Fortsch. Phys. [**64**]{}, 106-108 (2016), arXiv:1509.04645 \[hep-th\]. G. Dvali, *Black Holes and Large N Species Solution to the Hierarchy Problem*, Fortschr. Phys. [**58**]{} (2010), 528, [arXiv:0706.2050 \[hep-th\]]{}.\ G. Dvali and M. Redi, *Black Hole Bound on the Number of Species and Quantum Gravity at LHC*, Phys. Rev. [**D77**]{} (2008), 045027, [arXiv:0710.4344 \[hep-th\]]{}.\ G. Dvali and C. Gomez, *Quantum Information and Gravity Cutoff in Theories with Species*, Phys. Lett. [**B674**]{} (2009), 303, [arXiv:0812.1940 \[hep-th\]]{}. Lena Funcke, private communications. [^1]: For example, this can be accomplished by introduction of some heavy external sources, (e.g., a heavy pseudo-scalar) coupled to $F\tilde{F}$, and by the measurement of CP-violating effects in their interactions. The strength of these effects will be controlled by ${\vartheta}$. The precise methodology is unimportant for our discussion. [^2]: The term “electric field" is used in the analogy with electrodynamics, with $C$ playing the role of the electromagnetic vector potential. For example, the constant value of $E \equiv \epsilon^{\alpha\beta\mu\nu} \partial_{\alpha } C_{\beta\mu\nu} $ corresponds to a choice $C_{\mu\nu\alpha} = \epsilon_{\mu\nu\alpha\beta} x^{\beta}$. [^3]: This can be seen by the following simple argument. Imagine that $\Phi_{i}^{\bar{j}}$ has a VEV, but such that $\langle det \Phi \rangle = 0$. Then, without loss of generality we can always make the VEV real by an anomaly-free $SU(N_f)_L \otimes SU(N_f)_R$ transformation. Thus, we can eliminate all the phases without $U(1)_A$-rotation and hence no $\eta'$ will emerge with anomalous coupling. [^4]: Some implications of the gravitational analog of ${\vartheta}$-term, in theories with lower dimensions or on backgrounds with broken Poincare invariance, where previously discussed in [@CSgrav]. Our focus here is very different. Since we assume non-zero topological susceptibility, the term (\[RRgrav\]) is physical without the need of violation of Poincare symmetry. [^5]: In fact, if this assumption is added, the proof of chiral symmetry breaking in QCD becomes simpler.

--- abstract: 'The behavior of several nuclear properties with temperature is analyzed within the framework of the Finite Temperature Hartree-Fock-Bogoliubov (FTHFB) theory with the Gogny force and large configuration spaces. Thermal shape fluctuations in the quadrupole degree of freedom, around the mean field solution, are taken into account with the Landau prescription. As representative examples the nuclei $^{164}$Er, $^{152}$Dy and $^{192}$Hg are studied. Numerical results for the superfluid to normal and deformed to spherical shape transitions are presented. We found a substantial effect of the fluctuations on the average value of several observables. In particular, we get a decrease in the critical temperature ($T_c$) for the shape transition as compared with the plain FTHFB prediction as well as a washing out of the shape transition signatures. The new values of $T_c$ are closer to the ones found in Strutinsky calculations and with the Pairing Plus Quadrupole model Hamiltonian.' author: - 'V. Martin' - 'J.L. Egido' - 'L.M. Robledo' bibliography: - 'fluc.bib' title: Thermal shape fluctuation effects in the description of hot nuclei --- Introduction {#introduction .unnumbered} ============ Since the advent of the new generation of $ 4\pi $ gamma ray detectors and the improved accuracy in the channel selection new possibilities have opened up in the study of nuclear structure. Besides this, the availability of faster computers has made possible to perform realistic theoretical investigations with large configuration spaces. The high excitation energy is specially interesting since new features may take place. For example, in the quasicontinuum, the high level density gives rise to the unexpected phenomenon of the damping of the rotational motion. In the limit of high excitation energies (or temperature $T$) quantum effects become less relevant or may even disappear. Thus one expects that in a heated nucleus physical effects like superfluidity or shape deformations are washed out when $T$ increases. This expectation can be easily understood in terms of the shell model since, by increasing $T$, one promotes particles from levels below the Fermi surface to levels above it. In the case of pairing correlations, blocking levels amounts to destroying Cooper pairs. In the case of shape deformation, by depopulating the deformation driving levels (intruders) one gets on the average less deformation. Experimental information about nuclear shape changes can be obtained by means of the Giant Dipole Resonance (GDR) built on excited states. Exclusive experiments studying the GDR strength at a given excitation energy (or $T$) of the nucleus have been carried out in refs. [@AGH.90; @NBH.92; @NBH.99; @HBB.03]. The understanding of these phenomena is relevant because it affects important features like the fission barriers and the stability of the nucleus itself. For a recent review on hot nuclei see ref. [@ER.93]. The shape transitions have been object of many studies, most of them with [*schematic models*]{}, separable forces, and [*small configuration spaces*]{} [@Mo.73; @Goo.81; @ERM.86; @Goo.86; @RR.94]. The theoretical approaches used in the calculations are based on the mean field approximation, mainly the Finite Temperature Hartree-Fock-Bogoliubov theory (FTHFB). The mean field approximations predict sharp shape transitions, whereas for finite systems, however, one expects washed out transitions instead. The fact that the predicted critical temperatures are rather high (around 2-3 MeV) indicates that not only the most probable deformation is relevant but that there is a finite (in some cases very large) probability for the system to have other shapes which should be taken into account. Calculations beyond mean field including thermal fluctuations have confirmed the expectation of washed out transitions [@EDR.86; @8]. Theoretical studies with [*effective forces*]{} and [*large configuration spaces*]{} have been performed at the FTHFB level with density dependent forces, Skyrme [@BQ.74a; @BQ.74b; @QF.78] and recently with the Gogny [@EMR.00] force. Additional calculations have been done in the relativistic mean field (RMF) approximation [@AT.00; @GML.00]. Calculations including thermal fluctuations in conjunction with large configurations spaces and effective forces have been performed only very recently [@AT.00; @MER.03]. From these studies a discrepancy has emerged since, while the [*mean field approaches*]{} (FTHFB) with [*effective forces*]{} (like Skyrme, the Gogny force or the relativistic approaches), provide the view of a sharp shape transition at a relatively high critical temperature ($T_c \approx 2.7$ MeV for $^{164}$Er), schematic models (like the Pairing plus Quadrupole) and Strutinsky calculations provide also a sharp transition though at a much lower critical temperature ($T_c \approx 1.7$ MeV for $^{164}$Er). Furthermore, the discordant point of a “sharp” transition for a small system, like the nucleus, predicted by both approaches requires further investigation. Earlier calculations with the Pairing plus Quadrupole Hamiltonian [@ERI.85; @EDR.86] have pointed out the relevance of including fluctuations in mean field approaches at finite temperature as a step forward to clarify some aspects of these problems. It is the aim of this paper to investigate the problems just mentioned as well as other related high excitation energy topics, level densities, etc., within a beyond mean field theory. Towards this end, the FTHFB calculations of ref. [@EMR.00] with the Gogny force and large configuration spaces will be generalized to include fluctuations in the quadrupole moment degree of freedom. The finite range density dependent Gogny force has the advantage of providing the particle-hole (Hartree-Fock) and the particle-particle (pairing) matrix elements from the same interaction, at variance with relativistic theories and most Skyrme calculations. In the fitting of the D1S [@1] parametrization no excited states or spin dependent data was used, however, it has produced good results in the description of nuclear properties not only at zero [@2; @3] but also at large spin [@254NoSpin] and, more recently [@EMR.00], in calculations at high excitation energy. Since our purpose is to study the behavior of shell effects and fluctuations with temperature, we have selected both theoretically and experimentally well known nuclei that display a variety of shapes in the ground state: strongly deformed ($ ^{164} $Er), oblate ($ ^{192} $Hg ) and rather soft ($ ^{152} $Dy). Theory {#theory .unnumbered} ====== To study the behavior of nuclei with increasing temperature we use the D1S [@1] parametrization of the finite range density dependent Gogny force [@2; @3] in the FTHFB framework [@12; @Goo.81]. The Gogny force, at variance with most of the Skyrme parametrizations and the relativistic models, allows full selfconsistent calculations since it provides the particle-hole and pairing fields from the same force. At finite temperature, as at temperature zero, the basic approximation is the mean field theory. Its most sophisticated version, the FTHFB, has been developed in refs. [@6a; @Goo.81; @ERM.86]. For convenience we will give here a short outline. For a system at constant temperature $T$ and with chemical potential $\mu$, the equilibrium state can be obtained from the variational principle over the grand canonical potential $$\Omega =E-TS-\mu N. \label{gcp}$$ The energy, $E$, entropy, $S$, and particle number, $N$, are thermal averages defined by $$\begin{array}{lllll} E & \equiv & \left\langle \hat{H}\right\rangle_T & = & Tr(\hat{D}\hat{H}),\\ S & \equiv & \left\langle -k\ln \hat{D}\right\rangle_T & = & -k \,Tr(\hat{D}\ln \hat{D}),\\ N & \equiv & \left\langle \hat{N}\right\rangle_T & = & Tr(\hat{D}\hat{N}), \end{array}$$ with $ Z $ the grand partition function and $ \hat{D} $ the density operator given by $$\begin{array}{ccc} Z & = & Tr[\exp (-\beta (\hat{H}-\mu \hat{N}))],\\ \hat{D} & = & Z^{-1}\exp (-\beta (\hat{H}-\mu \hat{N})), \end{array}$$ with $ \beta =1/kT $. In the FTHFB approach the density operator is approximated by $$\hat D_0~=~{ {e^{\hat{\cal H}/T}} \over Z_0}, \label{E2.8}$$ where $\hat{\cal H}$ is the most general Hermitian single particle operator, to be determined by the variational principle and $Z_0$ is the partition function. It can be shown [@ER.93] that ${\cal H}$ is given by $${\cal H}~=~\left(\begin{array}{cc}h&\Delta\\- \Delta^*&-h^*\end{array}\right),$$ with $ \Delta $ the pair potential and $ h $ the HF hamiltonian. $ h $ is given in terms of the kinetic energy $ t $, the HF field, $ \Gamma $, and the chemical potential, $ \mu $, $$\begin{array}{ccc} h & = & t +\Gamma -\mu, \\ \Gamma _{ij} & = & \sum _{kl}v_{ikjl}\rho _{lk},\\ \Delta _{ij} & = & \frac{1}{2}\sum _{kl}v_{ijkl}\kappa _{kl}. \end{array}$$ The density matrix, $ \rho $, and the pairing tensor, $ \kappa $, are given by $$\begin{array}{ccc} \rho & = & UfU^{+}+V^{*}(1-f)V^{t},\\ \kappa & = & UfV^{+}+V^{*}(1-f)U^{t}, \end{array}$$ and $$f_{i}=\frac{1}{1+e^{\beta E_{i}}}.$$ The matrices $(U,V)$ provide the relation between the quasiparticle and the single particle basis : $$\alpha^+_m~=~\sum_k U^{}_{km} c^+_k + V_{km} c^{}_k.$$ They are determined, together with the quasiparticle energies, $ E_{i} $, by the FTHFB equation $$\label{hfbeq} \begin{array}{lllll} \left( \begin{array}{cc} h & \Delta \\ -\Delta ^{*} & -h^{*} \end{array}\right) & \left( \begin{array}{c} U_{k}\\ V_{k} \end{array}\right) & = & \left( \begin{array}{c} U_{k}\\ V_{k} \end{array}\right) & E_{k}. \end{array}$$ The solution of this equations provides us with the configuration that minimizes the grand canonical potential. With $U, V$ and $f$ known one can determine the density operator $\hat{D}_0$ and calculate any expectation value. For density dependent forces like Skyrme or Gogny, the formalism remains unchanged except in the evaluation of the one body Hamiltonian $h$. Due to the dependence on the density of the interaction, $h$ gets [@MER.03] an extra term, $\partial \Gamma$, which is usually referred to as the ”rearrangement potential” and is given by $$\partial \Gamma_{m m'} = \left< \frac{\partial H}{\partial \rho_{m'm}} \right>_T.$$ The FTHFB solution gives us the most probable shapes, quadrupole, hexadecupole, etc, as well as the most probable gap parameters and so on. At finite temperatures, however, we have statistical (or thermal) fluctuations around this solution. In principle one could consider fluctuations in the most relevant degrees of freedom. For nuclei, at high excitation energy, the most important one is the quadrupole deformation, and we therefore shall consider only the fluctuations in the quadrupole moment $\langle {\hat Q}_{20} \rangle $ in this paper. To generate the solutions with different shapes we solve the grand canonical potential, Eq. (\[gcp\]), with an additional constraint on the quadrupole moment, i.e., we minimize $ \Omega = E - T S - \mu N - \lambda_{Q_{20}} q$. The Lagrange multiplier $\lambda_{Q_{20}}$ is adjusted in such a way that the thermal expectation value $\langle \hat{Q}_{20} \rangle = Tr(\hat{D}\hat{Q}_{20})$, has the required value $q$. According to Landau [@LL.59] the probability $ P(q)$ to obtain a certain value $q$ of the deformation is characterized by the free energy $F(q)= E(q) - T S(q)$ of the system with deformation $q$ $$P(q)~\propto~e^{-F(q)/T}. \label{probq}$$ Using classical statistics, therefore, for the ensemble average of an observable $\hat{\cal O}$ one obtains the expression $$\label{averages} \overline{\cal O} =\frac{\int {\cal O}(q)\exp (-F(q)/T)dq} {\int \exp (-F(q)/T)dq},$$ where ${\cal O}(q)$ is the thermal expectation value of the operator $ \hat{\cal O}$ calculated for the system with the deformation $q$, and $dq$ is the volume element in deformation space. In our case the set $ q $ corresponds to the quadrupole deformation $q_{20}$, thus $dq = dq_{20}$, with metric equal to one. The limits in the thermal average integrals, see Eq. (\[averages\]), are chosen to span the full $\beta_2$ region in which the probability of having one of these values, given by Eq. (\[probq\]), is not negligible. This covers both prolate and oblate regions. High temperature calculations require large configuration spaces. In order to maintain the computational burden within reasonable limits we restrict ourselves to axial symmetry. We are aware that for soft nuclei and/or high temperature, triaxiality may play an important role. In the calculations we use an axially deformed harmonic oscillator (HO) basis with a size defined by the condition $$2b_{\rho} n_{\rho} + b_z n_z < N_0,$$ where $ n_{\rho} $ and $ n_{z} $ are the HO axial quantum numbers, $b_{\rho}=q^{1/3} $ and $ b_{z}=q^{-2/3} $ with $ q=R_{z}/R_{\rho}, $ the nuclear axis ratio. In our case we have used $ q=1.5 $ and $ N_{0}=15 $ which allows for deformations big enough to reach the fission barrier and provides room enough for the temperature induced excitations. However, as an additional check, we have also used $ N_{0}=17 $ for some selected calculations. Reflection asymmetry is allowed in the calculations, i.e. the nuclei may develop octupole deformations. In order to compare our Gogny force results with the ones, more conventional and popular, of the schematic Pairing plus Quadrupole model (PPQ), we have also performed calculations with this force. The configuration space (the spherical oscillator shells $N = 5, 6$ for neutrons and $N = 4, 5$ for protons) and the force parameters used are the one of Baranger-Kumar [@ppq]. The calculations have been performed in exactly the same way as in ref. [@8] Results {#results .unnumbered} ======= We have performed FTHFB calculations with the D1S parameter set of the Gogny force in several nuclei to study the evolution of shell effects with temperature. Nuclei with different ground state deformations have been selected to illustrate their different behavior. As an example of a nucleus with a strongly prolate deformed ground state we used the thoroughly studied $^{164}$Er. The soft $^{152}$Dy$_{86}$ is a transitional nucleus between the clearly spherical Dysprosium isotopes with $N \leq 84 $ and the well deformed ones with $N \geq 88$. This nucleus was selected for its rich shell structure and shape coexistence. The heavier $ ^{192} $Hg has been chosen due to its oblate ground state shape. The maximum temperature studied has been kept below 3 MeV, such that continuum contributions can be safely disregarded [@18; @18b]. The thermal fluctuations are represented through averages calculated according to Eq. (\[averages\]). The deviations around the mean values can be studied by the standard deviation value $$\sigma ({\cal O}) = \sqrt{\; \overline{ {\hat {\cal O}}^2} - \left[ \; {\overline{\hat {\cal O}}} \; \right]^2}. \label{sig}$$ This quantity is presented in some cases for further clarification of the results obtained. Before entering in the discussion of the shape and pairing phase transitions we will start by presenting first a general view. In Fig. \[Fig:free\] we present the free energy, $F(\beta_2)$, and the quantity $ P(\beta_2)~\propto~\exp (-F(\beta_2)/T) $ versus the quadrupole deformation, $\beta_2$, at different temperatures. $ P(\beta_2)$ provides the weight of a given shape $\beta_2$ in the evaluation of thermal average values. The results for $^{164}$Er are displayed in the left column, for $^{152}$Dy in the middle one and for $^{192}$Hg in the right one. For the rare earth nuclei $^{164}$Er and $^{152}$Dy, where calculations within the two shells configuration space mentioned above are feasible, we also present results with the PPQ model. Results with the Gogny force are displayed by continuous lines and those with the PPQ force by dashed ones (thick lines represent $F(\beta_2)$ and thin ones $ P(\beta_2)$). The well depths are measured from the point with $\beta_2=0$ and $ P(\beta_2)$ has been normalized is such a way that the most probable deformation takes the value of unity. Let us discuss first the low temperature calculations ($T = 0.3 $ MeV ) where we can observe the intrinsic shapes of the ground states. In the Gogny calculations for $^{164}$Er there is a deep prolate minimum at $\beta_2 \approx 0.3$ and about 4.5 MeV higher an oblate one. With the PPQ model the same gross features are observed, though the minima are not so deep. The probability distribution $ P(\beta_2)$, however, is similar in both calculations. For the nucleus $^{152}$Dy, in the Gogny case, the prolate minimum is at $\beta_2 \approx 0.15$ and a bit higher in energy the oblate one. The PPQ model, for this nucleus, provides a broad minimum around the spherical shape. $ P(\beta_2)$, in contrast with $^{164}$Er, looks quite different in the Gogny case than in the PPQ one. In both nuclei the free energy surfaces are broader with the Gogny force than with the PPQ one. For $^{192}$Hg, we find the minimum at an oblate deformation of $\beta_2 \approx -0.15$ and about 1.7 MeV higher a small prolate minimum. At higher temperatures the expected disappearance of shell effects becomes clear, in particular the vanishing of the barriers when several minima are available and the development of only one spherical minimum. Further finite temperature effects like the widening of the free energy curve and the more important role of fluctuations with increasing temperature appear in the Gogny calculations but not in the PPQ ones. In the PPQ case the free energy surfaces, with increasing temperatures, become flatter but not broader. They even become narrower! This unphysical effect has to do, obviously, with the size of the configuration space (two shells). As we can see already at ($T = 0.3 $ MeV ), at large deformations $F(\beta_2)$ increases very steeply because there are no orbitals with high-$j$ (deformation driving ) coming down. The mechanism to soften the free energy surface at high temperature by enhancing the probability to occupy high-lying orbitals (among them the high-$j$ ones) works only with large configuration spaces. The anomalous behavior of $ P(\beta_2)$ in $^{152}$Dy, in the PPQ approach, at $T = 0.6 $ MeV as compared with $T = 0.3 $ MeV is due to the fact that at $T = 0.6 $ the neutron pairing gap vanishes and the ground state becomes slightly prolate. Figures \[Fig:164ErAll\], \[Fig:152DyAll\] and \[Fig:192HgAll\] show the detailed calculations for all three nuclei. These figures include both, the results at the FTHFB level and with shape fluctuations calculated as described above. Dashed lines and open symbols indicate the FTHFB results. Solid lines and filled symbols are used for averaged, fluctuations including, calculations. Figs. \[Fig:152DyAll\] and \[Fig:192HgAll\] show only Gogny results. The nucleus $^{164}Er$ ---------------------- In Fig. \[Fig:164ErAll\] we display the results of the calculations for the nucleus $ ^{164} $Er with the Gogny force and with the PPQ model Hamiltonian. In panel (a) we show the selfconsistent FTHFB (i.e., calculated with the solution of Eq. \[hfbeq\]) and the averaged results (i.e., calculated according to Eq. \[averages\]) for the $\beta_2, \beta_4$ and $\beta_6$ deformation parameters as a function of the temperature with the Gogny interaction. Let us first discuss the deformation parameter $\beta_2$. For temperatures $0<T<1.0$ MeV, both predictions behave similarly, as one would expect for a nucleus with a well pronounced minimum. For temperatures $1.0<T<2.0$ MeV, the FTHFB $\beta_2$-values decrease rather smoothly while the averaged ones undergo a strong reduction. For $T>2.0$ MeV the selfconsistent values decrease very steeply and collapse, finally, to zero deformation at $T=2.7$ MeV. The averaged values, on the contrary, change tendency decreasing very smoothly in such a way that an almost constant value of $\beta_2$ is eventually obtained. The behavior of $\beta_4$ and $\beta_6$ is similar to the one of $\beta_2$ though not that spectacular. The same plot for $\beta_2$ but with the PPQ interaction is represented in panel (b). Quantitatively the main differences with the Gogny results are the faster collapse of the selfconsistent value, at $T\approx 1.8$ MeV, and the reduction of the temperature interval where the averaged values are smaller than the selfconsistent ones. Looking at the probability distribution in Fig. \[Fig:free\] one can easily understand these differences. The temperature value at which the [*mean field*]{} (mf) deformation parameter collapses, which we will denote $T^{mf}_{c}$, has often been used in earlier mean field studies to signal a shape phase transition. It is obvious from panels (a) and (b) that in theories [*beyond mean field*]{} things look quite different and that definition of the critical temperature must be carefully considered. The big difference in $T^{mf}_{c}$ as predicted by effective forces, like the Gogny force, and the PPQ is also known from calculations with Skyrme forces [@BQ.74a; @BQ.74b] and the relativistic mean field approximation [@AT.00; @10b]. The standard deviation in the deformation parameter $\beta_2$, $\sigma(\beta_2)$, calculated according to Eq. \[sig\], is presented in panel (c). One can distinguish three well defined zones : in the first one at low temperature, when pairing is still strong, the deformation is kept almost constant and fluctuations raise slowly. The Gogny and PPQ calculations predict about the same equilibrium shape in this zone. For temperatures higher than the one corresponding to the pairing collapse, $\sigma(\beta_2)$ increases rapidly up to a maximum value, remaining more or less at this value at higher temperatures. This step behavior is characteristic of a shape phase transition region. In fact, one could define the shape transition temperature as the one at which $\sigma(\beta_2)$ has a maximum. With this criterion one obtains $T=1.4$ MeV for the PPQ result and $T=1.7$ MeV in the Gogny case. Note the large difference in $\sigma(\beta_2)$ between the Gogny and PPQ results at high temperature. In this comparison, however, one must keep in mind that the PPQ model hamiltonian is restricted to a configuration space of two oscillator shells which strongly constraints the ability to produce fluctuations. This is clearly seen in Fig. \[Fig:free\] : the PPQ results rapidly develop a narrow parabolic shape with increasing temperature. This lack of fluctuations was already identified as partially responsible for the low multiplicity seen in the collective E2 quasicontinuum spectra in gamma decay calculations when compared to experiment [@16; @16b]. An additional confirmation of the importance of the fluctuations in calculations with the Gogny interaction as compared to the PPQ case is provided by the different behavior of the averaged deformation parameters with respect to the mean field within the same model, see panels a) and b). The value of the average $\beta_2$ parameter [*at high temperatures*]{} illustrates the deviation of the free energy surface from a parabolic behavior. A value close to zero is expected for a parabola, e.g., in the PPQ case, while a larger one, as in the Gogny case, indicates the softness of the prolate side as compared with the oblate one. One should nevertheless keep in mind that only axially symmetric deformed shapes are allowed in the calculations. In panel d) of Fig. \[Fig:164ErAll\], the proton and neutron pairing energies are displayed for the Gogny force. Up to $T \approx 0.2 $ MeV the pairing energies are rather constant but for higher $T$ values they decrease in absolute value very fast up to $T = 0.7$ MeV where they vanish. Thermal shape fluctuations, as expected in the low temperature regime, have little effect on the pairing correlations. Pairing fluctuations which would be more relevant [@ERI.85] are not considered in this work. It is interesting to take a look at the internal excitation energy, $E^*$, and also analyze it through the behavior of its derivative, the specific heat $C_V(T)=\partial E^*/\partial T$, since the appearance of peaks in this quantity is customarily interpreted as a signature in the search for phase transitions. The evolution of $E^*$ with temperature for both the FTHFB, $E^*$, and the average, $\overline{E^*}$, calculations, is presented in panels e) (Gogny force) and f) (PPQ force) of Fig. \[Fig:164ErAll\]. With the Gogny force and in the low temperature regime, we can see the pairing collapse which is visible as a change in the slope of $E^*$. At higher temperatures a fairly quadratic behavior is observed in the excitation energy, which is slightly modified when the transition to the spherical phase takes place at high temperature around $T \sim 2.7$ MeV. There, a weaker change in the slope, hardly seen in the scale of the plot, is found. The change is more abrupt in $E^*$ than in $\overline{E^*}$, again as expected in the picture of a thermally faded transition. The same facts are observed in the PPQ plot, where one can additionally observe that at high temperature the energy behaves more linear than quadratic as a function of T. The different behavior of $E^*$ and $\overline{E^*}$ is also interesting : a) At temperatures below 0.8 MeV, both energy values coincide, b) between 0.8 MeV and the corresponding $T^{mf}_{c}$ (around 2.7 MeV for Gogny and 1.8 MeV for PPQ), $E^*$ is always below $\overline{E^*}$, c) at the critical temperature both energies do coincide and d) at higher temperatures they are rather similar. This behavior has a simple explanation if one considers the entropy as a function of the deformation at fixed temperature, see Fig. \[Fig:Entro\], and the fact that in general $F \leq \overline{F} $ [^1], with $F$ the selfconsistent FTHFB free energy at the given $T$. At low temperatures ($\leq 0.8$ MeV) and at temperatures above $T^{mf}_{c}$, the entropy is almost shape independent, i.e., $\overline{S} \approx S $ (with $S$ the selfconsistent FTHFB entropy at the given $T$), and the free energy behaves like a parabola, $ \overline{F}\approx F $, consequently $ \overline{E^*} \approx E^*$. At temperatures $0.8 \leq T \leq T^{mf}_{c} $, $F < \overline{F}$ and $S < \overline{S}$, see Figs. \[Fig:free\] and \[Fig:Entro\], and consequently, since $\overline{F}= \overline{E} - T \overline{S}$, $E^* < \overline{E^*}$. The change in slope in $\overline{E^*}$ as compared with $E^*$ indicates that the corresponding specific heats, $ \partial \overline{E^*} / \partial T$ and $\partial E^* / \partial T$, will be rather different. This can be seen in Fig. \[Fig:Cv\] where we observe two peaks in the selfconsistent results, both in the PPQ and in the Gogny calculations [^2]. In between we find the typical linear behavior for a Fermi gas ($ C_V = 2aT $). The low temperature peak is associated with the superfluid to normal transition and the high temperature one with the deformed to spherical shape transition. The low temperature peak remains nearly unaffected by the inclusion of fluctuations. At the mean field level this transition takes place at a temperature low enough such that shape fluctuations are irrelevant. Comparing the FTHFB results with the PPQ force with those obtained with the Gogny interaction for the shape transition, we find again the different temperature predictions. Using as critical temperature, $T_{c}$, the temperature where $C_V$ changes curvature, the same values are obtained as when the $\beta_2=0$ rule was used (1.8 MeV for PPQ [*vs.*]{} 2.7 MeV for Gogny). However, when fluctuations are taken into account, the difference in the predicted $T_{c}$ by Gogny and PPQ gets smaller and the sharpness of the peaks reduced, indicating a less abrupt transition as is expected in a mesoscopic system. This is more evident in the Gogny results, where the peak becomes a broad bump, providing another clue of the greater importance of fluctuations in the Gogny case. By contrast, the PPQ peak, although broader than in the mean field case, is still sharp. Furthermore the PPQ specific heat levels off, showing how the limited configuration space available is a clear disadvantage of this model. If we now look for the changes in curvature we find $T_{c}=1.4$ MeV in the PPQ and $T_{c}=1.7$ MeV with the Gogny interaction. It is interesting to notice that these values agree very well with the ones obtained looking at $\sigma(\beta_2)$. The nucleus $^{152}Dy$ ---------------------- As we have seen in Fig. \[Fig:free\], $^{152}Dy$ displays a potential energy surface with energetically close prolate and oblate minima. It could illustrate a nucleus with shape coexistence, that means, already at temperatures near to zero there is a finite probability of populating more than one minimum. For this nucleus we will not perform a discussion as exhaustive as for $^{164}Er$ but we will consider the most relevant facts. In Fig. \[Fig:152DyAll\] the results of the calculations for the Gogny force are displayed. In panel a) the mean field and the averaged values of the pairing energies are plotted. The pairing energy of the neutron (proton) system collapses at $T \approx 0.5$ ($T \approx 1.0$) MeV. The averaged values, as expected, almost coincide with the mean field ones. In panel b) the $\beta$-deformation parameters are shown. As we can observe in the behavior of the $\beta_2$ parameter the effect of the shape fluctuations in this case is already noticeable at very small temperatures. This is due to the fact that, in the ground state $^{152}Dy$ is a much less deformed nucleus than $ ^{164}$Er and that the energy difference between the oblate and prolate minimum is small, amounting to only $\sim 0.55$ MeV for $T<0.5$ MeV. Hence the averaging formula assigns finite weights to the oblate side already at small $T$’s causing the observed steep decrease in the average $\beta_2$ value. Although the FTHFB, searching for the strict minimum, provides a deformed ground state for $^{152}$Dy, we see how even at the lowest temperatures the average deformation is very small, in agreement with the experimental data. At the high temperature limit we observe that above 1.4 MeV, the average deformation stays rather constant, or slightly increases, up to 0.04. This anomalous behavior is due to the fact that the superdeformation driving orbitals are being occupied at this temperature range. The $ \beta _{4} $ and $ \beta _{6} $ deformation parameters follow closely the $\beta_2$ behavior. In particular, in the FTHFB description they become zero at the same temperature as $\beta_2$. In panel c) the excitation energy is depicted. In the FTHFB approach the changes in slope at temperatures of 0.5, 1.0 and 1.4 MeV are due to the neutron and proton pairing collapse and to the shape transition. The behavior of $\overline{E^*}$ can be understood in terms of the entropy plots as explained for the $^{164}$Er nucleus. The changes in slope in the energy plots are magnified in the specific heat versus temperature plot depicted in panel d). In the mean field approach we find a broad composite peak, corresponding to the proton and neutron pairing collapse, showing substructures around $T\approx 0.5$ and 1.0 MeV. Furthermore, one sees a second peak at 1.4 MeV corresponding to the FTHFB shape transition. If shape fluctuations are included in the calculations we obtain only one broad peak. The small peak at 1.4 MeV, however, is not there anymore indicating that it has been shifted to the pairing transition bump or simply washed out. In fact, the almost identical broad energy mean field peak and the single one obtained with fluctuations could be seen as a clue that the shape transition in this soft nucleus is inexistent, since there is no higher temperature peak but only a small modification of the “pairing” one. To check this hypothesis we have also performed calculations with the particle-particle channel of the Gogny force set to zero, i.e. plain FTHF. In this way we obtain $C_V$ curves without the pairing transition peaks. The results are plotted, superimposed to the standard FTHFB calculations, using square symbols in panel d) of Fig. \[Fig:152DyAll\]. Again, open squares are for the FTHF calculations and full ones for averaged ones. As it was expected, the FTHF curves for $C_V$ show no peak for the pairing transition and the only peak present is the one corresponding to the shape transition at 1.4 MeV, which coincides with the one obtained with the full FTHFB calculations, since at these temperatures pairing is already zero. The thermal averaging results show a broad shoulder approximately in the same temperature region in which the pairing transition was located. This rather soft bump is a clear indication of the above mentioned situation, i.e., at small temperatures no clearly predominant minimum exists and at high temperatures the nucleus does not become exactly spherical. The nucleus $^{192}$Hg ---------------------- As we have seen in Fig. \[Fig:free\] at very small temperatures this nucleus presents an oblate deformed ground state and about 1.7 MeV above a prolate minimum. The results for the nucleus $^{192}$Hg are shown in Fig. \[Fig:192HgAll\]. The pairing energies are displayed in panel a). The proximity of the $Z=82$ shell closure causes the vanishing of the proton pairing energies for all temperatures. The neutron system, on the contrary, has a large pairing energy at $T=0$ MeV, which vanishes at $T=0.8$ MeV. As before, the shape fluctuations have almost no effect on the pairing energies. In panel b) the behavior of the deformation parameters with increasing temperature is plotted. For $^{192}$Hg, in the mean field approximation, we obtain an oblate ground state deformation of $\beta_2 = -0.135 $ which gets more oblate for increasing temperatures as the pairing energies go to zero. For larger $T$ values the deformation decreases and around $T \approx 1.4$ MeV the nucleus becomes spherical. As before, the effect of the shape fluctuations is mainly characterized by the the prolate-oblate ground state energy difference, see Fig. \[Fig:free\], which in this case amounts to 1.7 MeV at zero temperature and only above $T=0.5$ MeV start to diminish. Around this temperature the deformation gets smaller and around 1.4 MeV the average deformation is zero. Interestingly, the average $\beta_2$ and $\beta_4$ deformation parameters become positive in the limit of high temperatures due to the fact that at these temperatures the $F(T)$ curves are softer in the prolate than in the oblate side. In panel c) the excitation energies in both approximations, in the mean field and with shape fluctuations, are plotted versus the temperature. In the mean field calculations we find slope changes at $T=0.8$ MeV and $T=1.4$ MeV associated with the pairing collapse and the oblate-spherical shape transition. The inclusion of shape fluctuations affects mainly the region between $0.8 \leq T \leq 1.4$ MeV. The general behavior of both curves can be easily understood in the same terms as for $^{164}$Er. In panel d) the corresponding specific heats are represented. In the mean field approximation the expected peaks are clearly visible. The inclusion of the shape fluctuations produces a single, broader bump extending above the critical temperature for the pairing collapse. As in the $^{152}$Dy case this peak might be a superposition of the pairing and the shape transition peaks. To isolate the shape transition peak we have performed again calculations with the particle-particle channel of the Gogny force set to zero. The results of the calculations, without (empty squares) and with (filled squares) shape fluctuations, are represented in the same panel. In both approximations the results above $T=0.9$ MeV are obviously the same as before. Below this temperature and in the mean field approximation, as expected, the pairing peak is gone. However, in the calculations with the shape fluctuations we find a broad peak extending from $T=0.4$ MeV up to $T=1.0$ MeV with a change in curvature around $T=0.9$ MeV. Looking at the standard deviation $\sigma(\beta_2)$ of this nucleus (not shown here) we find a maximum at $T=0.9$ MeV, an additional indication of the shape transition. Level densities and nuclear radii --------------------------------- Level densities, $\rho(E^*)$, can be microscopically evaluated in the saddle point approximation, see for example eq. (2B-14) of ref. [@BM.75]. In Fig. \[Fig:lden\] the total level densities for the three nuclei under study are displayed against the excitation energy in the mean field approximation and on average, i.e., with the inclusion of shape fluctuations. In both cases we observe the overall expected exponential dependence and the well known abnormal behavior at very small excitation energies. For $^{164}$Er and up to 10 MeV excitation energy ($T \approx 0.8$ MeV) we find a good agreement between both predictions. Then, up to 70 MeV ($T \approx 2.3$ MeV), we observe an increase in the level density in the average description as compared with the mean field one. In particular, around 20 to 30 MeV excitation energy, the average prescription provides almost two orders of magnitude larger densities than the mean field one. This behavior can be easily understood looking at Figs. \[Fig:free\] and \[Fig:Entro\] and taking into account that the level density is proportional to the exponential of the entropy. At low ($T < 0.7 $ MeV) and high excitation energies ($T > 2.5 $ MeV) the entropy is rather shape independent, that means, the average value of the level density is very close to the one in the selfconsistent minimum. Consider now $T=1.4$ MeV. In this case the selfconsistent minimum is prolate ($\beta_2 \approx 0.3$), and the entropy at this shape and $T$ is smaller than for all the other shapes at this temperature. That means, since $\rho(\beta_2) \propto e^{S(\beta_2)}$, the average level density will always be larger than the selfconsistent one. Similar arguments apply to understand the behavior of the level densities of $^{152}$Dy and $^{192}$Hg. The fact that in these nuclei we do not find a larger difference between both descriptions is obviously due to the smoother behavior of the entropy with the deformation at the relevant temperatures. In Fig. \[Fig:msr\] the root mean squared (rms) radii of the three nuclei are plotted versus the temperature. In general we find that the rms radii are rather constant up to a given temperature, 2 MeV for $^{164}$Er and 1 MeV for $^{152}$Dy and $^{192}$Hg, and that in this temperature range the average values are rather similar to the FTHFB ones. From this temperature on the average values are larger than the FTHFB ones due to the fact that, at these temperatures, the probability for a given shape peaks at the spherical shape and that for a given volume the spherical shape corresponds to the one with the smallest rms radii. That means, fluctuations around the spherical minimum provide always larger rms radii. We also observe, at the highest temperatures, the expected increase of the rms radii. Discussion and conclusions {#discussion-and-conclusions .unnumbered} ========================== We have seen in the previous section that, in calculating average properties, the behavior of the entropy with the deformation parameter $\beta_2$ plays a major role. This behavior is by itself, indeed, quite interesting. The general behavior, see Fig. \[Fig:Entro\], is the following : In the high $T$ limit where the temperature effects dominate, we find, as expected, small shape dependence. At very low $T$, where the temperature effects are very small, we observe that to increase the entropy by 5 units we have to increase $T$ by 0.5 MeV and that the entropy is rather independent of the shape of the nucleus. Of course, in this region where pairing correlations are present it is difficult to make more precise statements. However, at moderate temperatures, which are however high enough to allow for significant quasiparticle occupation numbers but not too high in order that shell effects are still present, one can find a large dependence of the entropy on the nuclear shape. In this region spherical shapes, as expected, have larger entropy than axially deformed ones. Since the maxima of the entropy are associated with the minima of the grand potential a correspondence between Fig. \[Fig:free\] and Fig. \[Fig:Entro\] does not necessarily exist. One of the main outcomes of our research is the finding that shape fluctuations have a large effect on the description of shape transitions. In fact, the [*mean field approach*]{} (FTHFB) with effective forces (like Skyrme, the Gogny force or the relativistic approaches), provides the view of a sharp shape transition at a relatively high critical temperature ($T_c \approx 2.7$ MeV for $^{164}$Er). On the other hand, Strutinsky calculations or schematic models (like the Pairing plus Quadrupole) provide also a sharp transition though at a much lower critical temperature ($T_c \approx 1.7$ MeV for $^{164}$Er). It has been argued [@GML.00] that the different predictions for the critical temperature are due to the small effective mass obtained in the mean field approach with effective forces ($m^*/m \approx 0.7$, with $m$ the nucleon bare mass ) as compared to the Strutinsky or the PPQ model ($m^*/m \approx 1.0$). This argument is obviously restricted to the mean field approach. In theories beyond mean field it does not apply anymore because with increasing correlations the effective mass eventually becomes the bare mass. In calculations at finite temperatures two kinds of correlations have to be considered, on one hand the quantum ones and on the other the statistical or classical ones. Their relevance depends obviously on the excitation energy (or temperature), at low $T$’s the former are very important and at high $T$’s, the latter ones. At the temperatures where the shape transition is predicted to take place in the mean field approach, the probability of having a shape different to the selfconsistent one is very large. Therefore, it is obvious that, first, shape fluctuations must be included and second that the characterization of the shape transition must be considered more carefully. In the mean field approach a criterion for shape transition is just to look at the temperature at which the nucleus becomes spherical or alternatively to look for a peak at the specific heat as a function of the temperature. In theories beyond mean field, usually the second one is used because the average deformation can become very small but not zero. As we have seen, the inclusion of shape fluctuations provides a specific heat rather different from the mean field one, because a) it is not a sharp peak what we obtain but a rather broad bump (this is consistent with the fact that it is a very small system) and b) it appears at temperatures much lower than the ones predicted by the mean field approximation in agreement with the Strutinsky calculations. It is also interesting to notice that the predictions based on the specific heat analysis coincide with the ones of the standard deviation $\sigma(\beta_2)$. Of course one could ask about the effect of considering quantum correlations in our predictions. Canosa, Rossignoli and Ring [@CRR.99] have shown in model calculations based on the static path plus random phase approximation that at finite temperature quantum effects are observable dependent. In particular, they find that the specific heat remains practically unaffected when quantum correlations are taken into account. One could conclude therefore that the prediction of the shape transition by the inclusion of thermal shape fluctuations with effective forces is reliable. In conclusion, we have performed extensive calculations with the Gogny force and a large configuration space for three representative nuclei in the FTHFB framework. We have further studied the effect of thermal shape fluctuations and found that they strongly affect, among others, the traditional shape transition “view” of the FTHFB approach. They do it in two aspects, first the critical temperature for the transition is very much lowered and second, the specific heat peaks are not sharp but rather broad. Besides this, the peaks showing up in FTHFB calculations of the specific heat in soft nuclei, like $^{152}$Dy, are (almost) completely washed out when thermal shape fluctuations are taken into account indicating the absence of any shape transition. However, in strongly deformed nuclei, like $^{164}$Er, the shape fluctuations reconfirm the presence of a shape transition though of a different character. We also find a strong enhancement in the level density in the presence of a shape transition. The superfluid to normal phase transition is not affected by the inclusion of shape fluctuations. This work has been supported in part by DGI, Ministerio de Ciencia y Tecnología, Spain, under Project BFM2001-0184. [^1]: The equal sign is valid only in the case that the free energy is a parabolic function of the deformation. [^2]: The results displayed in this figure have been calculated by evaluating numerically the derivative of the energy. In ref. [@EMR.00] the expression $ C_V = T \partial S/\partial T $ was used instead and $S $ was calculated analytically in the mean field approach.

--- author: - 'Guo-Hua Mu' - 'Wei-Xing Zhou' bibliography: - 'E:/Papers/Auxiliary/Bibliography.bib' title: Relaxation dynamics of aftershocks after large volatility shocks in the SSEC index --- Introduction ============ Financial markets are complex systems, from which numerous empirical regularities have been documented in the Econophysics community [@Mantegna-Stanley-2000; @Bouchaud-Potters-2000; @Sornette-2003; @Malevergne-Sornette-2006; @Zhou-2007]. A large proportion of these stylized facts deal with volatility, which is an important measure of risk of financial assets. The dynamics of asset volatility exhibit significant similarity in scaling compared to seismic activities such as the Gutenberg-Richter law [@Gutenberg-Richter-1956-BSSA] and Omori’s law [@Omori-1994-JCSIUT]. We note that the Gutenberg-Richter law in Finance [@Kapopoulos-Siokis-2005-EL] has deep connection with the inverse cubic law in the right tail distribution of volatility [@Liu-Gopikrishnan-Cizeau-Meyer-Peng-Stanley-1999-PRE]. Both scaling laws concern the dynamic behavior of volatility after stock market crashes. Indeed, there are quite a few studies on the dynamic behavior of volatility after stock market crashes. Sornette and coworkers found that the implied variance of the Standard and Poor’s 500 (S&P 500) index after the infamous Black Monday (10/19/1987) decays as a power law decorated with log-periodic oscillations [@Sornette-Johansen-Bouchaud-1996-JPIF]. Lillo and Mantegna investigated 1-minute logarithmic changes of the S&P 500 index during 100 trading days after the Black Monday and found that the occurrence of events larger than some threshold exhibits power-law relaxation for different thresholds [@Lillo-Mantegna-2004-PA]. This Omori law was also found to hold after two other crashes on 10/27/1997 and 08/31/1998 [@Lillo-Mantegna-2003-PRE]. There is more evidence from other stock indexes. Selçuk investigated daily index data from 10 emerging stock markets (Argentina, Brazil, Hong Kong, Indonesia, Korea, Mexico, Philippines, Singapore, Taiwan and Turkey) and observed Omori’s law after two largest crashes in each market [@Selcuk-2004-PA]. Selçuk and Gençay utilized the 5-minute Dow Jones Industrial Average 30 index (DJIA) and identified Omori’s law after 10/08/1998 and 01/03/2001 [@Selcuk-Gencay-2006-PA]. In a recent paper, Weber [*[et al.]{}*]{} extended the procedure adopted in the aforementioned studies on two aspects [@Weber-Wang-VodenskaChitkushev-Havlin-Stanley-2007-PRE]: (1) They have smoothed the volatility (absolute of return) a moving average over an aggregated time scale in order to remove insignificant fluctuations; and (2) They found that the Omori law holds not only after big crashes but also after intermediate shocks. These two issues are of essential importance, which enable us to discuss the Omori law in alternative manners. On one hand, the volatility can be defined by the average of absolute returns in a given time interval rather than the absolute of return over the same time scale. On the other hand, the main shocks in financial markets are not necessary to be defined by large crashes. Instead, we can investigate volatility shocks so that rallies are also included besides crashes. Based on these considerations, we shall study in this work the volatility dynamics of the Shanghai Stock Exchange Composite (SSEC) index after large volatility shocks. The rest of this paper is organized as follows. We first present preliminary information about the composition of data sets, the definition of volatility, and the mathematical description of Omori’s law. We then provide an objective procedure for the identification of volatility shocks and investigate the aftershock dynamics of daily and minutely volatility after large shocks for different thresholds. Preliminary information {#s1:PreInfo} ======================= The data sets ------------- We analyze two high-frequency data sets recording the SSEC index $I(t)$ at two different sampling frequencies. The first data set contains the close prices at the end of every minute during the continuous double auction so that its sampling interval is one minute. This data set covers five and half years from February 2001 to August 2006. The size of the data exceeds 318,602. The sampling frequency of the second data set fluctuates along time and is roughly 10 realizations per minute. This data set records every quotations during the continuous double auction (from 9:30 a.m. to 11:30 a.m. and from 13:00 p.m. to 15:00 p.m.) released by the Shanghai Stock Exchange and displayed on the terminals that investors can watch on. The data span from January 2004 to June 2006. The size of the data set is 1,253,440. It is worth noting that this data set is not ultra-high-frequency but high-frequency, since ultra-high-frequency data record all transactions [@Engle-2000-Em]. Definition of volatility ------------------------ The return $r(t)$ over time scale $\delta t$ is defined as follows $$r(t)=\ln[I(t)]-\ln[I(t-\delta t)]~, \label{Eq:return}$$ where $\delta t$ is time resolution for each data set (one minute for the first data set and one step in unit of event time for the second one). In the Econophysics literature including the aforementioned work on financial Omori law, the absolute of return $|r(t)|$ is frequently adopted as a measure of volatility on the time interval $(t-\Delta{t},t]$. Actually, there are various methods proposed for estimating daily volatility in financial markets utilizing intraday data [@Pasquini-Serva-1999-EL; @Pasquini-Serva-1999-PA; @Pasquini-Serva-2000-EPJB; @Bollen-Inder-2002-JEF]. We utilize the following definition for volatility [@Andersen-Bollerslev-Diebold-Labys-2001-JASA; @Andersen-Bollerslev-Diebold-Labys-2001-JFE; @Bollen-Inder-2002-JEF] $$V(t)=\left[\sum_{t-\Delta{t}<\tau\leqslant{t}}{r^2(\tau)}\right]^{1/2}~, \label{Eq:Vt}$$ which is the root of the sum of squared returns[^1]. In this work, we consider minutely and daily volatilities. For daily volatility ($\Delta{t}$ is one day), we adopt the first data set so that $\delta{t}$ is one minute. For minutely volatility ($\Delta{t}$ is one minute), we use the second data set so that $\delta{t}$ equals to one step of event time. Omori’s law vs. power-law relaxation ------------------------------------ It is well-known that there is long memory and clustering in the volatility. Large shocks in the volatility are often followed by a series of aftershocks and the occurrence number of events with the volatility exceeding a given threshold decreases with time. Recently, the so-called Omori’s law was borrowed from Geophysics to describe the dynamics of financial aftershocks. Omori’s law states that the number of aftershocks decays with some power law of the time after large shocks: $n(t)\propto t^{-p}$. In order to avoid divergence at $t=0$, Omori’s law is often rewritten as $$n(t)= K(t+\tau)^{-p}~, \label{Eq:nt}$$ where $K$ and $\tau$ are two positive constants, and $n(t)$ is the occurrence rate of aftershocks during $(0,t]$. Equivalently, the cumulative number of aftershocks after large volatility shocks can be expressed as follows $$N(t)= \left\{ \begin{array}{lll} K[(t+\tau)^{1-p}-\tau^{1-p}]/(1-p)~, &~~& p\neq1\\ K\ln(t/\tau+1)~,&~~&p=1 \end{array} \right.. \label{Eq:Nt}$$ In the empirical analysis, the parameters $K$, $p$ and $\tau$ are estimated using nonlinear least-squares regression. Empirical analysis {#s1:EmpAna} ================== Identifying large volatility shocks ----------------------------------- When crashes are concerned, one have to address the problem how to define a crash, on which a consensus is still lack. A quite feasible and unambiguous definition is based on large drawdowns [@Johansen-Sornette-1998-EPJB; @Johansen-Sornette-2001-JR; @Sornette-2003-PR]. An alternative option is to seek for large price drops within different time windows [@Mishkin-White-2002-NBER], which was essentially the same idea used by Selçuk [@Selcuk-2004-PA]. A third method is to investigate those large price drops identified as crashes by academics and professionals. These three methods identify partially overlapping examples of crashes, that were used in the previous studies of the dynamics of volatility after crashes [@Lillo-Mantegna-2004-PA; @Lillo-Mantegna-2003-PRE; @Selcuk-2004-PA; @Selcuk-Gencay-2006-PA; @Weber-Wang-VodenskaChitkushev-Havlin-Stanley-2007-PRE]. The situation is different in this study. We concern with large volatility shocks, which correspond not only to crashes but also to rallies. For the daily volatility, we first select the seven largest volatilities. If the time interval between two events are smaller than 30 trading days, the smaller one is excluded as a foreshock or aftershock. We then determine the duration of the impact by identifying the local minimum of volatility after the main shock within 60 trading days. This selection procedure is detailed in the following. Figure \[Fig:DayV\](a) illustrates the evolution of daily volatility for the SSEC index constructed from minutely returns. The initial seven days are 10/23/2001, 11/16/2001, 05/21/2002, 06/24/2002, 02/02/2004, 10/29/2004 and 07/05/2006, respectively. We observe that 11/16/2001 is very close to 10/23/2001. Since the volatility of 10/23/2001 is larger than that of 11/16/2001, we treat the former as a main shock and the latter its aftershock. Similarly, the dates 05/21/2002 and 06/24/2002 are also close to each other and the volatility of the latter is much larger than that of the former. The latter is thus regarded as the main shock and the former is considered as its foreshock. Since 06/05/2006 is near the end of the sample period, it is excluded from investigation. We are thus left with four events on 10/23/2001, 06/24/2002, 02/02/2004 and 10/29/2004 for analysis, as indicated in Fig. \[Fig:DayV\](a) with arrows. ![\[Fig:DayV\] Identification of four large volatility shocks. The four dates are indicated in the daily volatility series (a) and in the evolution of SSEC index (b).](DayV.eps "fig:"){width="6.5cm"} ![\[Fig:DayV\] Identification of four large volatility shocks. The four dates are indicated in the daily volatility series (a) and in the evolution of SSEC index (b).](price.eps "fig:"){width="6.5cm"} Figure \[Fig:DayV\](b) shows the locations of the four selected events in the price trajectory of the SSEC index. It is worthy of noting that four pieces of information took place on those four days. The Chinese stock market entered an antibubble regime since August 2001, which was triggered by the promulgation of the [*[Tentative Administrative Measures for Raising Social Security Funds through the Sale of State-Owned Shares]{}*]{} by the State Council of China on 06/24/2001 [@Zhou-Sornette-2004a-PA]. China’s Securities Regulatory Commission, however, suspended the fifth rule of the “Provisional Rules” in the evening of 10/22/2001 and the SSEC index rose up by 9.86%[^2]. On 2002/06/24, the State Council of China announced the removal with immediate effect of the provisions of the [*[Tentative Administrative Measures for Raising Social Security Funds through the Sale of State-Owned Shares]{}*]{}. This caused an increase of 9.25% in the SSEC index. In the weekend right before 02/01/2004 (Monday), the State Council of China released [*[Some Opinions of the State Council on Promoting the Reform, Opening and Steady Growth of Capital Markets]{}*]{} \[[*[Effective]{}*]{}\] and the market fluctuated remarkably with a 2.08% rise. In the same year, the decision of the People’s Bank of China to raise the benchmark lending and deposit interest rates, lift the ceiling of RMB loan interest rates, and allow a downward movement of RMB deposit interest rates took effect from October 29. On the same day, the market swang a lot and closed with a -1.58% drop. Volatility shocks with different magnitude may have distinct durations of impact. In order to investigate the relaxation behavior of volatility after a shock, it is crucial to determine this impact duration. The simplest but crude way would be to fix the impact duration for all shocks, say, 60 or 100 trading days. The shortcoming of this rule is obvious. In this work, we use a relatively objective way in which the local minimum of volatility after an identified shock is regarded as the end of its impact. Specifically, the local minimum is determined within a 60-trading-day window ensuing the shock. The main information of the four identified large volatility shocks are presented in Table \[TB:dailyV\]. As expected, the impact duration of the main shock increases approximately with the relative magnitude $V_{max}/\sigma$, where $\sigma$ is the sample average of the volatilities. In addition, it is interesting to note that the four large volatility shocks are more relevant to rallies rather than crashes. $t_0$ $t_1$ $T$ $V_{max}/\sigma$ $V_{max}$ $V_{min}$ $r(t_0)$ ------------ ------------ ----- ------------------ ----------- ----------- ----------- 10/23/2001 11/29/2001 28 12.9 $0.00543$ $0.00019$ $+9.86\%$ 06/24/2002 09/13/2002 60 13.7 $0.00578$ $0.00015$ $+9.25\%$ 01/02/2004 04/07/2004 48 5.5 $0.00230$ $0.00022$ $+2.08\%$ 10/29/2004 11/25/2004 20 4.4 $0.00186$ $0.00023$ $-1.58\%$ : Description of the four large volatility shocks for the SSEC index. The sample standard deviation of daily volatility $\sigma=0.000421$. $t_0$ is the date of volatility shock and $t_1$ is the last date of shock impact. $T$ is the duration of shock impact from $t_0$ to $t_1$ in unit of trading day. $V_{\max}$ and $V_{\min}$ are the volatilities at time $t_0$ and $t_1$, respectively. $r(t_0)$ is the daily return on day $t_0$ referenced to the previous trading day.[]{data-label="TB:dailyV"} In order to investigate the relaxation behavior of large volatility shocks at different time scales, we also calculated minutely volatility and used the same shocks identified in daily volatility for comparison. Since the minutely time series is shorter, we have two large shocks left on 01/02/2004 and 10/29/2004. Compared with the large fluctuations in the Chinese stock market, these two shocks correspond to neither crashes nor rallies. Aftershock dynamics in daily volatility --------------------------------------- We have calculated the cumulative number $N(t)$ of aftershocks larger than some fixed threshold after each main shock. The threshold $\theta$ is presented based on the sample standard deviation of daily volatility, which is $\sigma=4.21\times10^{-4}$ within the time period concerned. Four thresholds are selected for each main shocks: $\theta/\sigma=0.6$, $0.7$, $0.8$, $0.9$ for 10/23/2001, 06/24/2002, and 10/29/2004 and $\theta/\sigma=0.7$, $0.8$, $0.9$, $1.0$ for 02/02/2004. The resulting functions $N(t)$ are plotted in Fig. \[Fig:dailyN\]. Note that the selection of thresholds is not arbitrary. If $\theta\ll\sigma$, all trading days are identified as aftershocks such that $N(t)=t$, which is illustrated in each panel of Fig. \[Fig:dailyN\]. If $\theta$ is too large, $N(t)$ becomes constant shortly after the main shock. This phenomena can be observed in Fig. \[Fig:dailyN\]. ![\[Fig:dailyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 10/23/2001, (b) 06/24/2002, (c) 02/02/2004, and (d) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=4.21\times10^{-4}$. The thresholds are $0.6\sigma$, $0.7\sigma$, $0.8\sigma$, and $0.9\sigma$ for (a) 10/23/2001, (b) 06/24/2002, and (d) 10/29/2004 and $0.7\sigma$, $0.8\sigma$, $0.9\sigma$, and $1.0\sigma$ for (c) 02/02/2004. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). The dashed line is $N(t)=t$.](Fig_DailyN_2001_10_23.eps "fig:"){width="6cm"} ![\[Fig:dailyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 10/23/2001, (b) 06/24/2002, (c) 02/02/2004, and (d) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=4.21\times10^{-4}$. The thresholds are $0.6\sigma$, $0.7\sigma$, $0.8\sigma$, and $0.9\sigma$ for (a) 10/23/2001, (b) 06/24/2002, and (d) 10/29/2004 and $0.7\sigma$, $0.8\sigma$, $0.9\sigma$, and $1.0\sigma$ for (c) 02/02/2004. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). The dashed line is $N(t)=t$.](Fig_DailyN_2002_06_24.eps "fig:"){width="6cm"} ![\[Fig:dailyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 10/23/2001, (b) 06/24/2002, (c) 02/02/2004, and (d) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=4.21\times10^{-4}$. The thresholds are $0.6\sigma$, $0.7\sigma$, $0.8\sigma$, and $0.9\sigma$ for (a) 10/23/2001, (b) 06/24/2002, and (d) 10/29/2004 and $0.7\sigma$, $0.8\sigma$, $0.9\sigma$, and $1.0\sigma$ for (c) 02/02/2004. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). The dashed line is $N(t)=t$.](Fig_DailyN_2004_02_02.eps "fig:"){width="6cm"} ![\[Fig:dailyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 10/23/2001, (b) 06/24/2002, (c) 02/02/2004, and (d) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=4.21\times10^{-4}$. The thresholds are $0.6\sigma$, $0.7\sigma$, $0.8\sigma$, and $0.9\sigma$ for (a) 10/23/2001, (b) 06/24/2002, and (d) 10/29/2004 and $0.7\sigma$, $0.8\sigma$, $0.9\sigma$, and $1.0\sigma$ for (c) 02/02/2004. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). The dashed line is $N(t)=t$.](Fig_DailyN_2004_10_29.eps "fig:"){width="6cm"} The power-law relaxation model (\[Eq:Nt\]) is calibrated for each empirical $N(t)$ function shown in Fig. \[Fig:dailyN\]. The estimated relaxation exponents $p$ and characteristic time scales $\tau$ are digested in Table \[TB:daily\]. All the exponents $p$ are significantly larger than 1 except one case ($t_0=06/24/2002$, $\theta=0.8\sigma$, $p=0.99$). We notice that these exponents are much larger than the relaxation exponents of daily volatility for many other emerging markets [@Selcuk-2004-PA]. In other words, the daily volatility after a main shock relaxes much faster in the Chinese stock market. In addition, the relaxation exponent $p$ increases with the threshold $\theta$ implying that larger aftershocks decay faster. This behavior is analogous to that of the intraday volatility of the S&P 500 index [@Lillo-Mantegna-2003-PRE]. ------------ -- ------------- ------------- ------------- ------------- -- ------------- ------------- ------------- ------------- -- -- $t_0$ $0.6\sigma$ $0.7\sigma$ $0.8\sigma$ $0.9\sigma$ $0.6\sigma$ $0.7\sigma$ $0.8\sigma$ $0.9\sigma$ 10/23/2001 1.70 1.94 1.83 2.13 28.2 20.9 23.6 38.8 06/24/2002 1.86 1.86 0.99 2.14 28.1 14.8 2.1 7.9 02/02/2004 1.48 1.62 2.10 2.40 43.7 36.6 38.4 29.1 10/29/2004 1.79 1.80 1.91 2.31 23.2 23.7 18.4 12.7 ------------ -- ------------- ------------- ------------- ------------- -- ------------- ------------- ------------- ------------- -- -- : Exponents and characteristic time scales of daily volatility relaxation after four large volatility shocks in Chinese stock market. The sample mean of daily volatility is $\sigma=4.21\times10^{-4}$. The four thresholds for 02/02/2004 are $0.7\sigma$, $0.8\sigma$, $0.9\sigma$, and $1.0\sigma$, different from those shown in the table. The unit of $\tau$ is trading day.[]{data-label="TB:daily"} Aftershock dynamics in minutely volatility ------------------------------------------ We have also calculated the cumulative number $N(t)$ of aftershocks larger than some fixed threshold after each main shock in minutely SSEC volatility. In this case, the sample standard deviation of minutely volatility is $\sigma=1.49\times10^{-4}$ within the time period of the data. Three thresholds are selected for each main shocks: $\theta/\sigma=1.5$, $1.75$, and $2.0$ for both shocks. Again, the selection of thresholds is not arbitrary. Smaller thresholds give straight lines $N(t)\approx{t}$, while too large thresholds produce constant $N(t)$ shortly after the main shock. The resulting functions $N(t)$ are plotted in Fig. \[Fig:minutelyN\]. We observe that the three curves for 02/02/2004 on panel (a) of Fig. \[Fig:minutelyN\] exhibit nice power-law curvature. However, the three curves on panel (b) are not different remarkably from straight lines. Hence, the minutely SSEC volatility relaxation after big shocks behaves differently from that for the S&P 500 index, in which power-law behavior is unveiled after mediate shocks [@Weber-Wang-VodenskaChitkushev-Havlin-Stanley-2007-PRE]. We submit that the difference between the construction of volatility in the two studies can not account for this discrepancy. ![\[Fig:minutelyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 02/02/2004 and (b) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=1.49\times10^{-4}$. The thresholds are $1.5\sigma$, $1.75\sigma$, and $2.0\sigma$. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). ](Fig_MinutelyN_2004_02_02.eps "fig:"){width="6.5cm"} ![\[Fig:minutelyN\] Cumulative number $N(t)$ of aftershocks larger than some fixed threshold after the main shock occurred on (a) 02/02/2004 and (b) 10/29/2004. The sample standard deviation of daily volatility is $\sigma=1.49\times10^{-4}$. The thresholds are $1.5\sigma$, $1.75\sigma$, and $2.0\sigma$. The solid lines are best fits to the data with Eq. (\[Eq:Nt\]). ](Fig_MinutelyN_2004_10_29.eps "fig:"){width="6.5cm"} We fitted the empirical $N(t)$ functions to Eq. (\[Eq:Nt\]), as shown in Fig. \[Fig:minutelyN\] with smooth curves. The estimated relaxation exponents $p$ and characteristic time scales $\tau$ are digested in Table \[TB:minutely\]. All the exponents $p$ except for $p=0.97$ are significantly larger than 1. Similar to the daily volatility case, the relaxation exponent $p$ for minutely volatility increases with the threshold $\theta$. This observation is different from those in other stock markets [@Lillo-Mantegna-2004-PA; @Lillo-Mantegna-2003-PRE; @Selcuk-Gencay-2006-PA; @Weber-Wang-VodenskaChitkushev-Havlin-Stanley-2007-PRE]. It is interesting to note that the characteristic timescale $\tau$ for 10/29/2004 is quite large. Applying Taylor’s expansion, we have $(1+t/\tau)^{1-p}=1+(1-p)(t/\tau)+o(t/\tau)$ and $\ln(t/\tau+1) =t/\tau+o(t/\tau)$ when $t\ll\tau$. Instituting these two linear approximations into Eq. (\[Eq:Nt\]) gives $$N(t) = K\tau^{-p}t \label{Eq:Nt2}$$ This simple algebraic derivation explains the almost-linear behavior of $N(t)$ for 10/29/2004 shown in Fig. \[Fig:minutelyN\](b). In this case, $p$ is a parameter controlling the slope of the linear $N(t)$. Therefore, the observation that $p$ increases with $\theta$ and the fact that the slope of $N(t)$ decreases with $\theta$ by definition are consistent with each other. ------------ -- ------------- -------------- ------------- -- ------------- -------------- ------------- -- -- -- -- $t_0$ $1.5\sigma$ $1.75\sigma$ $2.0\sigma$ $1.5\sigma$ $1.75\sigma$ $2.0\sigma$ 02/02/2004 0.97 1.58 1.72 9.4 16.3 18.0 10/29/2004 1.54 1.74 1.76 72.7 104.0 43.7 ------------ -- ------------- -------------- ------------- -- ------------- -------------- ------------- -- -- -- -- : Exponents and characteristic time scales of minutely volatility relaxation after two large volatility shocks in the Chinese stock market. The sample mean of minutely volatility is $\sigma=1.49\times10^{-4}$. The unit of $\tau$ is trading day.[]{data-label="TB:minutely"} Concluding remarks {#s1:conc} ================== In summary, we have investigated the dynamic behavior of the Chinese SSEC volatility after large volatility shocks. Daily and minutely volatilities are considered, which are calculated based on high-frequency data at finer scales. We stress that large volatility shocks are adopted as financial earthquake, rather than the crashes. These large volatility shocks are selected objectively, together with the durations of shock impact. In this way, no main shock is qualified as a crash in our analysis. Instead, two rallies are selected, which have large volatility. The selection of the volatility threshold $\theta$ is no arbitrary. Too small or too large $\theta$ gives $N(t)\sim t$ or $N(t)=const.$, respectively. We have found that the cumulative number $N(t)$ of aftershocks with magnitude exceeding a given threshold $\theta$ increases as a power-law to the time distance $t$ to the main shock. The power-law exponent $p$ value is a increasing function of the volatility threshold $\theta$. The aftershock dynamics is very different for the Chinese SSEC index volatility in the sense that most of the values of $p$ are significantly larger than 1. Hence, the power-law relaxation behavior is different from the Omori law where $p$ is close to 1. This study thus adds a new ingredient to the effort in searching for idiosyncratic behaviors in the Chinese stock markets [@Zhou-Sornette-2004a-PA; @Zhou-Yuan-2005-PA; @Jiang-Guo-Zhou-2007-EPJB]. This work was partly supported by the National Natural Science Foundation of China (Grant No. 70501011), the Fok Ying Tong Education Foundation (Grant No. 101086), and the Shanghai Rising-Star Program (No. 06QA14015). [^1]: The definition we adopt is different slightly from that of Ref. [@Andersen-Bollerslev-Diebold-Labys-2001-JASA; @Andersen-Bollerslev-Diebold-Labys-2001-JFE; @Bollen-Inder-2002-JEF], in which the intraday returns are filtered with an MA(1) model and the variance $V^2(t)$ is considered instead of $V(t)$. [^2]: There is a price limit of $\pm10\%$ fluctuation compared with the closure price on the last trading day. A rise of 9.86% in the SSEC means that most of the stocks went up by 10%.

--- bibliography: - 'COSMO\_arxiv\_bibliography.bib' --- @space@setup[ @preskip=5mm plus 2mm minus 2mm @postskip=@preskip ]{}

--- abstract: 'We investigate the possibility of explaining the enhancement in semileptonic decays of $\bar B \to D^{(*)} \tau \bar \nu$, the anomalies induced by $b\to s\mu^+\mu^-$ in $\bar B\to (K, K^*, \phi)\mu^+\mu^-$ and violation of lepton universality in $R_K = Br(\bar B\to K \mu^+\mu^-)/Br(\bar B\to K e^+e^-)$ within the framework of R-parity violating (RPV) MSSM. Exchange of down type right-handed squark coupled to quarks and leptons yield interactions which are similar to leptoquark induced interactions that have been proposed to explain the $\bar B \to D^{(*)} \tau \bar \nu$ by tree level interactions and $b\to s \mu^+\mu^-$ anomalies by loop induced interactions, simultaneously. However, the Yukawa couplings in such theories have severe constraints from other rare processes in $B$ and $D$ decays. Although this interaction can provide a viable solution to $R(D^{(*)})$ anomaly, we show that with the severe constraint from $\bar B \to K \nu \bar \nu$, it is impossible to solve the anomalies in $b\to s \mu^+\mu^-$ process simultaneously.' author: - 'N.G. Deshpande$^1$[^1], Xiao-Gang He$^{2,3,4}$[^2]' title: 'Consequences of R-Parity violating interactions for anomalies in $\bar B\to D^{(*)} \tau \bar \nu$ and $b\to s \mu^+\mu^-$' --- Introduction ============ Recent experimental data have shown deviations from standard model (SM) predictions in the ratio of $R(D^{(*)}) = Br(\bar B \to D^{(*)} \tau \nu)/Br(\bar B \to D^{(*)} l \nu)$ with $l = e,\;\mu$ and also in $b\to s \mu^+ \mu^-$ induced $B$ decays. Experimental values for $R(D^{(*)})$[@RD-1; @RD-2; @RD-3] are larger than the SM predictions[@RD-sm]. This anomalous effect is significant, at about 4$\sigma$ level[@HFAG]. The anomalies due to $b\to s \mu^+\mu^-$ induced processes show up in[@b2s-1] $B \to (K,K^*,\phi)\mu^+\mu^-$ decays. The observed branching ratios in these decays are lower than SM predictions[@b2s-sm; @b2s-the-1]. Also a deficit is shown in the ratio $R_K = Br(B\to K\mu^+\mu^-)/Br(B\to K e^+ e^-)$[@RK]. The SM predicts $R_K$ to be close to one, but experimental data give[@RK] $0.745^{+0.090}_{-0.074}\pm 0.036$. These effects are at 2 to 3 $\sigma$. Needless to say that these anomalies need to be further confirmed experimentally and we also need to understand SM predictions better. The latter processes involved are rare processes and therefore are sensitive to new physics. These anomalies have attracted a lot of theoretical attentions trying to solve the problems using new physics beyond SM[@RD-sm; @b2s-sm; @b2s-the-1; @RD-the-1; @desh-menon; @leptoquark; @b2s-the-2; @RD-b2s; @bauer-neubert]. In this work we study the possibility of using R-parity violating interaction to solve these anomalies. Previously, R-parity violation was invoked to explain[@desh-menon] only $R(D^{(*)})$. Exchange of down type right-handed squark coupled to quarks and leptons yield interactions, which are similar to leptoquark induced interactions that have been proposed to explain the $\bar B \to D^{(*)}\to \tau \bar \nu$ and $b\to s \mu^+\mu^-$ induced anomalies simultaneously[@bauer-neubert]. However, the Yukawa couplings have severe constraints from other rare processes in $B$ and $D$ decays. This interaction can provide a viable solution to $R^{(*)}$ anomaly. But with severe constraint from $\bar B \to K \nu \bar \nu$, it proves to be impossible to solve the anomalies induced by $b\to s \mu^+\mu^-$process. The most general renomalizable R-parity violating terms in the superpotentials are[@R-parity] $$\begin{aligned} W_{RPV} = \mu_i L_i H_u + {1\over 2} \lambda_{ijk}L_i L_j E^c_k + \lambda^\prime_{ijk}L_i Q_j D^c_k + {1\over 2} \lambda{''}_{ijk}U^c_iD^c_jD^c_k\;,.\end{aligned}$$ We will assume that $\lambda''$ term is zero to ensure proton stability. Since the processes we discuss involve leptons and quarks, the $\lambda'$ term should remain. In fact the interactions induced by this term at the tree and one loop level can contribute to $\bar B \to D^{(*)} \tau \bar \nu$ and $b\to s \mu^+\mu^-$ induced processes. It is tempting to see if these interactions can solve the related anomalies already. Although a combination of $\lambda'$ and $\lambda$ terms can also contribute, the resulting operators are disfavored by $\bar B\to D^{(*)} \tau \bar \nu$ process. We shall limit ourselves to exchange of right-handed down type squark, $\tilde d^k_R$, which are expected to have the necessary ingredients to explain the anomalies in $B$ decays. This model is similar to the leptoquark exchange discussed by many authors[@b2s-the-2], except a general leptoquark also has a right-handed couplings to $SU(2)_L$ singlets, which is forbidden in SUSY. These additional right-handed couplings turn out to be important for explaining the $g-2$ anomaly of muon, but do not play an essential role in explaining the B anomalies discussing here. The object of our paper is a careful consideration of the constraints from various $B$ and $D$ decays and analysie structure of Yukawa couplings $\lambda^\prime_{ijk}$ to see if the B anomalies can be resolved simultaneously. The paper by Bauer and Neubert[@bauer-neubert] is closest in spirit to our paper, but we are able to bring out the tension between different experimental constraints, and find that it is impossible to solve the $R(D^{(*)})$ and $b \to s \mu^+\mu^-$ anomalies simultaneously. The $R(D^{(*)})$ and $b\to s\mu^+\mu^-$ anomalies occur at tree level and loop level in the SM, respectively. To simultaneously solve these anomalies using a simple set of beyond SM interactions faces more constraints[@RD-b2s; @bauer-neubert] than just solving one of them as has been done in most of the studies. We find that by exchanging right-handed down type of squark, it is possible to solve the $R(D^{(*)})$ anomaly with tree interaction provided $\lambda^\prime_{33k}$ is sizable, of order $\sim 3$. For anomalies induced by $b\to s \mu^+\mu^-$, to obtain the right chirality for operators $O_9$, one needs to go to one loop level. The allowed couplings $\lambda^\prime_{ijk}$ are constrained from various experimental data, such as $K \to \pi \nu \bar \nu$, $\bar B \to K(K^*) \nu \bar \nu$ and $D^0 \to \mu^+\mu^-$. The strongest constraint comes from $\bar B \to K(K^*) \nu \bar \nu$ making the model impossible to explain anomalies induced by $b\to s \mu^+\mu^-$. R-parity violating interactions and $\bar B\to D^{(*)} \tau \bar \nu$ ===================================================================== Expanding the $\lambda'$ term in terms of fermions and sfermions, we have $$\begin{aligned} L = \lambda^{\prime}_{ijk}\left [ \tilde \nu^i_L \bar d^k_R d^j_L + \tilde d^j_L \bar d^k_R\nu^i_L + \tilde d^{k*}_R \bar \nu^{ci}_L d^j_L- \tilde l^i_L \bar d^k_R u^j_L - \tilde u^j_L \bar d^k_R l^i_L - \tilde d^{k*}_R \bar l^{ci}_L u^j_L\right ]\;,\end{aligned}$$ where the “tilde” indicates the sparticles, and “c” indicates charge conjugated fields. Working in the basis where down quarks are in their mass eigenstates, $Q^T = (V^{KM\dagger}u_L, d_l)$, one replaces $u^j_L$ in the above by $(V^{KM\dagger}u_L)^j$. Here $V^{KM}$ is the Kobayashi-Maskawa (KM) mixing matrix for quarks. If experimentally, the mass eigenstate of neutrino are not identified, one does not need to insert the PMNS mixing matrix for lepton sector. The neutrinos in the above equation are thus in the weak eigenstates. For leptoquark interactions discussed in eq. (6) in Ref.[@bauer-neubert], the reference seems to indicate that new parameters are involved due to rotation matrix $U_e$ in the lepton sector. However, since neutrinos are not in the mass basis in our work, it seems that provided we are always in the weak basis, no matrix is required in the lepton sector. We will assume sfermions are in their mass eigenstate basis. For a discussion of the choice of basis see Ref.[@R-parity] Exchanging sparticles, one obtains the following four fermion operators at the tree level $$\begin{aligned} {\cal L}_{eff} &=& {\lambda'_{ijk}\lambda^{'*}_{i'j'k}\over 2 m^2_{\tilde d^k_R}} \left [ \bar \nu^{i'}_L \gamma^\mu \nu^i_L \bar d^{j'}_L \gamma_\mu d_L^j + \bar e^{i'}_L \gamma^\mu e^i_L (\bar u_L V^{KM})^{j'}\gamma_\mu (V^{KM\dagger}u_L)^j \right .\nonumber\\ &&\hspace{1.7cm} -\nu^{i'}_L \gamma^\mu e^i_L \bar d^{j'}_L \gamma_\mu (V^{KM\dagger}u_L)^j -\bar e^{i'}_L \gamma^\mu \nu^i_L (\bar u_LV^{KM})^{j'} \gamma_\mu d^j_L\left . \right ]\nonumber\\ &-&{\lambda'_{ijk}\lambda^{'*}_{i'jk'}\over 2m^2_{\tilde d^j_L} } \bar \nu^{i'}_L \gamma^\mu \nu^i_L \bar d^k_R \gamma_\mu d^{k'}_R -{\lambda'_{ijk}\lambda^{'*}_{i'jk'}\over 2m^2_{\tilde u^j_L}} \bar e^{i'}_L \gamma^\mu e^i_L \bar d^k_R \gamma_\mu d^{k'}_R \\ &-&{\lambda'_{ijk}\lambda^{'*}_{ij'k'}\over 2m^2_{\tilde e^i_L}} (\bar u_{L\beta} V^{KM})^{j'} \gamma^\mu (V^{KM\dagger}u_{L\alpha})^j \bar d^k_{R\alpha} \gamma_\mu d^{k'}_{R\beta} - {\lambda'_{ijk}\lambda^{'*}_{ij'k'}\over 2m^2_{\tilde \nu^i_L}} \bar d^{j'}_{L\beta} \gamma^\mu d^j_{L\alpha} \bar d^k_{R\alpha} \gamma_\mu d^{k'}_{R\beta}, \nonumber \label{intr}\end{aligned}$$ In the above $\alpha$ and $\beta$ are color indices. At the tree level, besides the SM contributions to $\bar B\to D^{(*)} l \bar \nu$, there are also R-parity violating contributions, they are given by the term proportional to $ - (\lambda'_{l3k}\lambda^{\prime*}_{l'mk}/2 m^2_{\tilde d^k_R})\bar l_L \gamma^\mu \nu^{l'}_L (\bar u_LV^{KM})^{m} \gamma_\mu b_L$ in the above equation. Including the SM contributions one obtains[@desh-menon] $$\begin{aligned} H_{eff} &=& -{4G_F\over \sqrt{2}}V_{m3} (\delta^{l'}_l + \Delta^{l'm}_l) \bar l \gamma^\mu P_L \nu_{l'} \bar u^m\gamma_\mu P_L b_L\;,\nonumber\\ \Delta^{l,m}_{l'} &=& {\sqrt{2}\over 4 G_F} {\lambda^\prime_{l3k} \lambda^{\prime *}_{l' j' k}\over 2 m^2_{\tilde d^k_R}}{V_{mj'}\over V_{m3}} \;.\label{Heffbc}\end{aligned}$$ where $V_{ij}$ are elements in $V^{KM}$. Identifying different charged leptons in the final states, we find the ratio $R^{SM}_l(c) = Br(\bar B\to D^{(*)} l \nu)/Br(\bar B \to D^{(*)} l \nu)_{SM}$ of branching ratios compared with SM predictions to be given by $$\begin{aligned} &&R^{SM}_\tau(c) = \vert \Delta_1^{3,2}\vert^2 + \vert \Delta_2^{3,2}\vert ^2 + \vert 1+\Delta_3^{3,2}\vert^2 \;,\nonumber\\ &&R^{SM}_\mu(c) = \vert \Delta_1^{2,2}\vert ^2 + \vert 1+\Delta_2^{2,2}\vert ^2 + \vert \Delta_3^{2,2}\vert ^2\;,\nonumber\\ &&R^{SM}_e(c) = \vert 1+\Delta_1^{1,2}\vert ^2+\vert \Delta_2^{1,2}\vert ^2 +\vert \Delta_3^{1,2}\vert ^2 \;. \label{RSM}\end{aligned}$$ One can define a similar quantity $R^{SM}_l(u)$ for $Br(\bar B\to (\rho, \pi) l \nu)/Br(\bar B \to (\rho, \pi) l \nu)_{SM}$ and $Br(\bar B\to l \nu)/Br(\bar B \to l \nu)_{SM}$, and have $$\begin{aligned} R^{SM}_l(u)={Br(\bar B\to (\rho, \pi) l \nu)\over Br(\bar B \to (\rho, \pi) l \nu)_{SM}}={Br(\bar B\to l \nu)\over Br(\bar B \to l \nu)_{SM}}\;.\end{aligned}$$ Experimentally, $R^{SM}_{e}$ deviations from SM predictions is small, that is $R^{SM}_{e} \approx 1$, therefore we require $\Delta_i^{1,2}$, to be close to zero, which can be achieved by setting $\lambda^\prime_{1jk} = 0$, so that no linear terms in $\Delta^{i,j}_k$ contribute to $\bar B\to D^{(*)} e \bar \nu_e$. No large deviation has been observed in $R^{SM}_\mu$. However in $b\to s\mu^+\mu^-$ induced anomalies involves $\mu$ couplings, we will bare in mind that effect may have some impact for $R^{SM}_\mu(c)$. One may even contemplate that a somewhat enhanced $\bar B\to D^{(*)} \mu \bar \nu_\mu$ must be there if one tries to solve the $b\to s \mu^+\mu^-$ anomalies simultaneously. Although such a large deviation has not been established, theoretical calculations for the absolute values for the SM predictions and the experimental measurements may have some errors, so a certain level of deviation can be tolerated. We will take a conservative attitude to only allow up to 10% deviation from SM value, in $R^{SM}_\mu(c)$. We find that even such modest requirement put stringent constraint and making the attempt of simultaneously solve the two types of anomalies difficult. Defining $r(\bar B \to D^{(*)} \tau \bar \nu) = R(\bar B \to D^{(*)} \tau \bar \nu)/R(\bar B \to D^{(*)} \tau \bar \nu)_{SM}$, we have $$\begin{aligned} &&r(\bar B \to D^{(*)} \tau \bar \nu) = {2R^{SM}_\tau(c)\over R^{SM}_\mu(c)+R^{SM}_e(c)}\;. \label{BD}\end{aligned}$$ Changing $c$ to $u$, one can obtain the R-parity vilating contributions to $R(\bar B \to (\rho, \pi) \tau \nu)$. With the same approximation as above, we have $$\begin{aligned} &&r(\bar B \to \tau \bar \nu)=r(\bar B \to (\rho, \pi) \tau \bar \nu)={2R^{SM}_\tau(u)\over R^{SM}_\mu(u)+R^{SM}_e(u)}\;.\label{Bp}\end{aligned}$$ The linear terms in $r(\bar B \to D^{(*)} \tau \bar \nu)$ and $r(\bar B \to \tau \bar \nu)$ are proportional to $$\begin{aligned} (2 \lambda^{\prime}_{33k}\lambda^{\prime *}_{31k} - \lambda^{\prime}_{23k}\lambda^{\prime *}_{21k}){V_{cd}\over V_{cb}} + (2\lambda^{\prime}_{33k}\lambda^{\prime *}_{32k}- \lambda^{\prime}_{23k}\lambda^{\prime *}_{22k}){V_{cs}\over V_{cb}} +(2 \lambda^{\prime}_{33k}\lambda^{\prime *}_{33k} - \lambda^{\prime}_{23k}\lambda^{\prime *}_{23k})\nonumber\end{aligned}$$ and $$\begin{aligned} (2 \lambda^{\prime}_{33k}\lambda^{\prime *}_{31k} - \lambda^{\prime}_{23k}\lambda^{\prime *}_{21k}){V_{ud}\over V_{ub}} + (2\lambda^{\prime}_{33k}\lambda^{\prime *}_{32k}- \lambda^{\prime}_{23k}\lambda^{\prime *}_{22k}){V_{us}\over V_{ub}} +(2 \lambda^{\prime}_{33k}\lambda^{\prime *}_{33k} - \lambda^{\prime}_{23k}\lambda^{\prime *}_{23k})\;,\nonumber\end{aligned}$$ respectively. Note that there is a large enhancement factor $(V_{ud}/V_{ub})/(V_{cb}/V_{cd})$ for the first term in the expression for $r(\bar B \to \tau \bar \nu)$ compared with $r(\bar B \to D^{(*)}\tau \bar \nu)$. This may cause potential problem for a small deviation from 1 in $r(\bar B \to D^{(*)}\tau \bar \nu)$ to a large deviation in $r(\bar B \to \tau \bar \nu)$. One can avoid such a large enhancement by setting $\lambda^{\prime}_{31k, 21k}$ to be much smaller than other terms. In our later discussions we will set $\lambda^{\prime}_{31k}$ to be zero. The $\lambda^{\prime}_{21k}$ is also constrained to be small from $D^0 \to \mu^+\mu^-$ decay to be discussed in the following. But may play some important role in $b\to s \mu^+\mu^-$ decay. We will keep it in our discussions. The SM predictions and experimental measurements for $R(D^{(*)})$ are[@HFAG] $$\begin{aligned} &&R(D)_{SM} = 0.300\pm 0.008\;, \;\;\;\;\;R(D) = 0.397 \pm 0.040 \pm 0.028\;,\nonumber\\ &&R(D^*)_{SM} = 0.252\pm 0.003\;, \;\;\;\;R(D^*) = 0.316 \pm 0.016 \pm 0.010\;.\end{aligned}$$ The R-parity violating contributions to both $R(D)$ and $R(D^*)$ occur in a similar way, we use the averaged $r(\bar B \to D^{(*)} \tau \bar \nu)_{ave} = 1.266\pm 0.070$ of $r(\bar B\to D \tau \bar \nu)$ and $r(\bar B \to D^* \tau \bar \nu)$ to represent the anomaly. In the SM, $r_{ave} = 1$. To obtain a $r_{ave}$ within the $1\sigma$ region, $\lambda^\prime_{33k}$ is typically of order $\sim 3$. This large coupling makes it worrisome for this scenario from unitarity consideration. In more general terms, the unitarity limits concern the upper bound constraints on the coupling constants imposed by the condition of a scale evolution between the electroweak and the unification scales, free of divergences or Landau poles for the entire set of coupling constants. If so, the R-parity couplings are constrained to be about one at TeV scale[@R-parity]. A value of 3 is not consistent. The requirement of no Landau pole up to unifications scale may be not necessary if some new physics appear. One cannot for sure rule out the possibility of reaching unitarity bound of $\sqrt{4\pi}$ at a lower energy. However when attempt to also solve $b\to s \mu^+\mu^-$ induced anomalies, the model become much more constrained. Constraints from other tree level processes =========================================== Several other rare processes may receive tree level R-parity violating contributions. The constraints from these processes should be taken into account. We now study a few of the relevant ones: $K \to \pi \nu\bar \nu$, $\bar B \to K (K^{*}) \nu \bar \nu$, and $D^0 \to \mu^+\mu^-$. The possible terms generating these decays are $$\begin{aligned} &&{\lambda'_{ijk}\lambda^{'*}_{i'j'k}\over 2 m^2_{\tilde d^k_R}} \bar \nu^{i'}_L \gamma^\mu \nu^i_L \bar d^{j'}_L \gamma_\mu d_L^j \;,\;\;{\lambda'_{ijk}\lambda^{'*}_{i'j'k}\over 2 m^2_{\tilde d^k_R}} \bar e^{i'}_L \gamma^\mu e^i_L (\bar u_L V^{KM})^{j'}\gamma_\mu (V^{KM\dagger}u_L)^j \;,\nonumber\\ &&{\lambda'_{ijk}\lambda^{'*}_{i'jk'}\over 2m^2_{\tilde d^j_L} } \bar \nu^{i'}_L \gamma^\mu \nu^i_L \bar d^k_R \gamma_\mu d^{k'}_R\;, \;\;{\lambda'_{ijk}\lambda^{'*}_{i'jk'}\over 2m^2_{\tilde u^j_L}} \bar e^{i'}_L \gamma^\mu e^i_L \bar d^k_R \gamma_\mu d^{k'}_R\;.\end{aligned}$$ If $\lambda^\prime_{ijk}$ is non-zero for $k$ restricted to only one value, the two terms on the second line in the above equation will not induce the decays in question. For simplicity, we will work with this assumption [^3]. $D^0\to \mu^+\mu^-$ decay in the SM is extremely small. In our case, there are tree contributions which are therefore constrained severly. We have $$\begin{aligned} &&H_{eff} = - {1\over 2 m^2_{\tilde d^k_R}} C^k_{D\mu\mu}\mu_L\gamma_\mu \mu_L \bar u_L \gamma^\mu c_L\;,\nonumber\\ &&C^k_{D\mu\mu} = \lambda^\prime_{2jk}\lambda^{\prime *}_{2j'k}V_{1j'}V^*_{2j}\nonumber\\ &&= ( \lambda^\prime_{21k}V^*_{21} + \lambda^\prime_{22k}V^*_{22} + \lambda^\prime_{23k}V^*_{23} )(\lambda^{\prime *}_{21k}V_{11} + \lambda^{\prime *}_{22k}V_{12} + \lambda^{\prime *}_{23k}V_{13})\;.\end{aligned}$$ The decay width is given by $$\begin{aligned} \Gamma(D^0\to \mu^+\mu^-) = {1 \over 128 \pi } \left \vert {C^k_{D\mu\mu}\over m^2_{\tilde d^3_R}} \right \vert ^2 f_D^2 m_D m^2_\mu \sqrt{1-{4 m^2_\mu\over m^2_D}}\;,\end{aligned}$$ where $f_D = 212(1)$ MeV[@fD] is the $D^0$ decay constant. Using experimental upper bound[@PDG] $6.2\times 10^{-9}$ at 90% C.L. for $D^0\to \mu^+\mu^-$, we have $\vert C^k_{D\mu\mu}{(1\mbox{TeV})^2 /m^2_{\tilde d^k_R}}\vert < 6.1\times 10^{-2}$. With $\lambda^\prime_{21k, 22k}$ set to zero, $C^k_{D\mu\nu}$ is give by $C^k_{D\mu\mu} = \lambda^{\prime}_{23k}\lambda^{\prime *}_{23k}V_{ub}V_{cb}^*$. We have $\lambda^\prime_{23k} \lambda^{\prime *}_{23k}(1\mbox{TeV})^2 / m^2_{\tilde d^k_R} < (20)^2$. $\lambda^\prime_{23k}$ is only very loosely constrained from $D^0\to \mu+\mu^-$. If just $\lambda^\prime_{21k}$ or $\lambda^\prime_{22k}$ is non-zero, they are constrained as $$\begin{aligned} \lambda^\prime_{21k} \lambda^{\prime *}_{21k}{(1\mbox{TeV})^2 \over m^2_{\tilde d^k_R}}, \lambda^\prime_{22k} \lambda^{\prime *}_{22k}{(1\mbox{TeV})^2 \over m^2_{\tilde d^k_R}} < 0.28\;.\label{dmumu} \end{aligned}$$ These constraints on $\lambda^\prime_{21k}$ and $\lambda^\prime_{22k}$, make their effects on $b\to s\mu^+\mu^-$ small. Later we will show that even a small $\lambda^\prime_{22k}$ may play some important role in having a better coherent explanation of $R(D^{(*)})$ and $b\to s \mu^+\mu^-$ anomalies. For $K \to \pi \nu\bar \nu$, the ratio of $R_{K\to \pi \nu\bar \nu} =\Gamma_{RPV}/\Gamma_{SM}$ is given by[@ddh] $$\begin{aligned} &&R_{K\to \pi \nu\bar \nu} = \sum_{i=,e,\mu,\tau}{1\over 3}\left \vert 1 +{\Delta^{RPV}_{\nu_i\bar \nu_i}\over X_0(x_t) V_{ts}V^*_{td}} \right \vert ^2 + {1\over 3} \sum_{i\neq i'} \left \vert {\Delta^{RPV}_{\nu_i\bar \nu_{i'} }\over X_0(x_t)V_{ts}V^*_{td} }\right \vert ^2\;,\nonumber\\ &&\Delta^{RPV}_{\nu_i \bar \nu_{i'}} = {\pi s^2_W \over \sqrt{2} G_F \alpha} \left \vert -{\lambda^{\prime}_{i2k}\lambda^{\prime *}_{i' 1 k} \over 2 m^2_{\tilde d^k_R}} \right \vert ^2\;,\;\;\;\;X_0(x) = {x(2+x)\over 8(x - 1)} + {3x(x-2)\over 8(x-1)^2}\ln x\;,\end{aligned}$$ where $x_t = m^2_t/m^2_W$. Combining the SM prediction[@kpiuu-sm] for the branching ratio and experimental information[@PDG] $Br = (1.7\pm1.1)\times 10^{-10}$, at $2\sigma$ level, $\lambda^{\prime}_{i2k}\lambda^{\prime *}_{i'1k}$ are constraint to be less than a few times of $10^{-3} (m^2_{d^k_R}/ (1\mbox{TeV})^2)$. Since we will set $\lambda^{\prime *}_{i1k}=0$, this process is not affected at tree level. The expressions for $R_{\bar B \to \pi \nu\bar \nu}$ and $R_{\bar B \to K(K^*)\nu\bar \nu}$ of $\bar B\to \pi \nu\bar \nu$ and $\bar B \to K(K^*)\nu\bar \nu$ can be obtained from Eq.(14) by replacing $V_{ts}V^*_{td}$ to $V_{tb}V^*_{td}$ and $V_{tb}V^*_{ts}$, respectively. The corresponding $\Delta^{RPV}_{\nu_i \bar \nu_{i'}}$ are $$\begin{aligned} &&\mbox{For}\; \bar B\to \pi \nu \bar \nu: \;\;\;\;\;\;\;\;\;\;\Delta^{RPV}_{\nu_i \bar \nu_{i'}} = {\pi s^2_W \over \sqrt{2} G_F \alpha} \left \vert -{\lambda^{\prime}_{i3k}\lambda^{\prime *}_{i' 1 k} \over 2 m^2_{\tilde d^k_R}} \right \vert ^2\;,\nonumber\\ &&\mbox{For}\; \bar B\to K(K^*) \nu \bar \nu: \;\;\Delta^{RPV}_{\nu_i \bar \nu_{i'}} = {\pi s^2_W \over \sqrt{2} G_F \alpha} \left \vert -{\lambda^{\prime}_{i3k}\lambda^{\prime *}_{i' 2 k} \over 2 m^2_{\tilde d^k_R}} \right \vert ^2\;.\end{aligned}$$ For $B\to \pi \nu\bar \nu$, since we have set $\lambda^\prime_{i1k} = 0$, it is again not affected by R-pairty violating interactions in this model. The process $\bar B\to K(K^{*}) \nu\bar \nu$ will be affected. We have the following non-zero $\Delta^{RPV}_{\nu\bar\nu}$ $$\begin{aligned} &&\Delta^{RPV}_{\nu_\mu\bar \nu_\mu} = -{\lambda^{\prime}_{23k}\lambda^{\prime *}_{22k} \over 2 m^2_{d^k_R}}{\pi s^2_W\over \sqrt{2} G_F \alpha}\;,\;\;\Delta^{RPV}_{\nu_\tau\bar \nu_\tau} = -{\lambda^{\prime}_{33k}\lambda^{\prime *}_{32k} \over 2 m^2_{d^k_R}}{\pi s^2_W\over \sqrt{2} G_F \alpha}\;,\nonumber\\ &&\Delta^{RPV}_{\nu_\tau\bar \nu_\mu} = -{\lambda^{\prime}_{33k}\lambda^{\prime *}_{22k} \over 2 m^2_{d^k_R}}{\pi s^2_W\over \sqrt{2} G_F \alpha}\;,\;\; \Delta^{RPV}_{\nu_\mu\bar \nu_\tau} = -{\lambda^{\prime}_{23k}\lambda^{\prime *}_{32k} \over 2 m^2_{d^k_R}}{\pi s^2_W\over \sqrt{2} G_F \alpha}\;. \label{bnunu}\end{aligned}$$ Experimental data from BaBar[@Babar] and Belle[@Belle] give, $R_{B\to K(K^*) \nu\bar \nu} <4.3 (4.4)$ implying $\lambda^{\prime}_{23k}\lambda^{\prime *}_{22k}$, $\lambda^{\prime}_{33k}\lambda^{\prime *}_{32k}$, $\lambda^{\prime}_{33k}\lambda^{\prime *}_{22k}$ and $\lambda^{\prime}_{23k}\lambda^{\prime *}_{32k}$ are constrained from $\bar B\to K(K^*)\nu\bar \nu$. We shall return to this process later. Loop contributions for $b\to s \mu^+\mu^-$ induced anomalies ============================================================ The anomalous effects in $b\to s\mu^+\mu^-$ induced processes are only 2 to 3 $\sigma$ effects and need to be confirmed further. They may be due to our poor understanding of hadronic matrix elements involved, and may also be caused by new physics beyond SM. We now discuss how R-parity violating interaction may help to solve the problems. New physics contributes to $b\to s l \bar l$ can be parametrized as $H^{NP}_{eff} = \sum C^{NP}_i O_i$. Some of the most studied operators $O_i$ are $$\begin{aligned} &&O_9 = {\alpha\over {4\pi} }\bar s \gamma^\mu P_L b \bar \mu\gamma_\mu \mu\;,\;\;\;\;\;\;\;\;\;O^{\prime}_9 = {\alpha\over {4\pi}} \bar s \gamma^\mu P_R b \bar \mu\gamma_\mu \mu\;,\nonumber\\ &&O_{10} = {\alpha \over {4\pi}} \bar s \gamma^\mu P_L b \bar \mu\gamma_\mu \gamma_5 \mu\;,\;\;\;\;O^{\prime}_{10} = {\alpha \over {4\pi}} \bar s \gamma^\mu P_R b \bar \mu\gamma_\mu \gamma_5 \mu\;,\end{aligned}$$ where $P_{L,R} = (1\mp \gamma_5)/2$. The SM predictions are $C^{SM}_9 \approx - C^{SM}_{10} = 4.1$. A global analysis shows that to solve the anomalies in decays induced by $b\to s \mu^+\mu^-$, there are few scenarios where the anomalies can be solved with high confidence level and all cases $C^{NP}_9$ need to be around $-1$[@b2s-sm]. For example with $C^{NP}_9 = -1.09$ and $C^{NP}_{10}$, $C^{\prime, NP}_{9,10} = 0$ with a 4.5 pull; the cases with $C^{NP}_9 = - C^{\prime, NP}_{9}$, the best fit values are: $C^{NP}_{9} = -C^{\prime, NP}_{9} = -1.06$ and others equal to zero with a 4.8 pull; And the case with $C^{NP}_9 = - C^{NP}_{10}$, the best fit values are: $C^{NP}_{9} = -C^{NP}_{10} = -0.68$ and others equal to zero with a 4.2 pull. Here the number of “pulls” indicates by how many sigmas the best fit point is preferred over the SM point for a given scenario. The higher the pull, the better fit between theory and experimental data is reached. In our case, the R-parity violating contribution to be discussed belongs to the last case. For this case, the $1\sigma$ allowed range is[@b2s-sm], $-0.85 \sim -0.5$. With negative value for $C^{NP}_9$, the new physics contribution reduces $b\to s \mu^+\mu^-$ and therefore helps to explain why $B \to (K,K^*,\phi)\mu^+\mu^-$ branching ratios and $R_K$ are smaller than those predicted by SM. There is a potential contribution to $b\to s \mu^+\mu^-$ at tree level due to a term proportional to $\lambda'_{ijk}\lambda^{'*}_{i'jk'}/2m^2_{\tilde u^j_L} \bar e^{i'}_L \gamma^\mu e^i_L \bar d^k_R \gamma_\mu d^{k'}_R$. However, since we assume that there is only one non vanishing value for $k$, $b\to s \mu^+\mu^-$ is not induced by this contribution. One needs to include one loop contributions. At one loop level, exchanging $\tilde d^k_R$ in the loop, contributions with $C^{NP}_{9} = -C^{NP}_{10}$ can be generated with$$\begin{aligned} C^{NP,l \bar l'}_{9} &\approx& {m^2_q\over 8\pi \alpha} {1\over m^2_{\tilde d^k_R}} \lambda^{\prime}_{ l b k}\lambda^{\prime *}_{\bar l' m k} {V_{qm}V^{ *}_{ts}\over V_{tb} V^{ *}_{ts}}\nonumber\\ &-&{\sqrt{2}\over 64 \pi \alpha G_F} {\ln (m^2_{\tilde d^k_R}/m^2_{\tilde d^{k'}_R}) \over m^2_{\tilde d^k_R} - m^2_{\tilde d^{k'}_R}} \lambda^{\prime}_{ibk}\lambda^{\prime *}_{is k'} \lambda^{\prime}_{l j k'}\lambda^{\prime *}_{\bar l' j k} {1 \over V_{tb} V^{*}_{ts}}\;,\end{aligned}$$ where $m_q$ is the up type quark mass. The first term is induced by exchanging a $W$ boson and a sparticle $\tilde d^k_R$, and the second term is by exchanging two sparticles $\tilde d^k_R$ in the loops. The term of interest corresponds to $l=2$, $\bar l'=2$, $s=2$ and $b=3$ for the process $b\to s \mu^+ \mu^-$. One can relabel them with different numbers for other process. The first term is dominated by $q=t$, its contribution to $C_9^{NP,\mu\bar \mu}$ is about $0.15 \lambda^\prime_{23k}\lambda^{\prime *}_{23k}(1\mbox{TeV}/m_{\tilde d^k_R})^2$. This is a “wrong sign” contribution to solve $b\to s\mu^+\mu^-$ induced anomalies[^4]. With $\lambda^\prime_{1jk} = 0$ and $\lambda^\prime_{i1k} =0$ from considerations of no processes with electron has shown anomalies and $K\to \pi \nu \bar \nu$ constraint, and restricting $k$ to have only one value, we have $$\begin{aligned} C^{NP, l \bar l'}_{9} &\approx& {m^2_t\over 8\pi\alpha} {1\over m^2_{\tilde d^k_R}} \lambda^{\prime}_{ l3 k}\lambda^{\prime *}_{\bar l' 3 k} \nonumber \\ &-&{\sqrt{2}\over 64 \pi \alpha G_F} {1\over m^2_{\tilde d^3_R}} (\lambda^{\prime}_{23k}\lambda^{\prime *}_{22 k}+ \lambda^{\prime}_{33k}\lambda^{\prime *}_{32 k})(\lambda^\prime_{l2k}\lambda^{\prime *}_{\bar l' 2k}+\lambda^\prime_{l3k}\lambda^{\prime *}_{\bar l' 3k} ) {1 \over V_{tb} V^{*}_{ts}}\\ &=& \left (0.157 \lambda^{\prime}_{ l3 k}\lambda^{\prime *}_{\bar l'3 k} +2.0 (\lambda^{\prime}_{23k}\lambda^{\prime *}_{2 2 k}+ \lambda^{\prime}_{33k}\lambda^{\prime *}_{32 k})(\lambda^\prime_{l2k}\lambda^{\prime *}_{\bar l'2k}+\lambda^\prime_{l 3k}\lambda^{\prime *}_{\bar l' 3k} )\right ){(1\mbox{TeV})^2\over m^2_{\tilde d^k_R}}\;.\nonumber \label{c9t}\end{aligned}$$ Numerical analysis ================== We are now in a position to put things together to see if R-parity violating interactions may be able to solve the $R(D^{(*)})$ and $b\to s \mu^+\mu^-$ anomalies simultaneously. For the KM parameters we use those given in Particle Data Group[@PDG]. The aim is to produce values for $r(\bar B\to D^{(*)}\tau \bar \nu)_{ave}$, $C_{9}^{NP}$ as close as possible to their central values, $1.266$ and $-0.68$. At the same time we have to restrict $R_{\bar B \to K(K^*)\nu \bar \nu}$ to be less than 4.3 to satisfy experimental bound. If one just needs to solve the $R(D^{(*)})$ anomaly, one just can easily obtain the central value of $r_{ave}-1 = 0.266$ by setting all other $\lambda^\prime_{ijk}$ to zero except $\lambda^\prime_{33k}$ with its vale given by $2.95(m_{\tilde d^k_B}/1\mbox{TeV})$. If $m_{\tilde d^k_R}$ is way above TeV, then the coupling will violate the unitarity bound of $\sqrt{4\pi}$. Therefore for the theory to work purterbatively , one expect the squark mass to be less than a TeV or so which can be looked for at the LHC. With this choice of $\lambda^\prime$ the SM predictions for $R_{\bar B\to K \nu\bar \nu,\;K\to \pi \nu \bar \nu}$ and $\Gamma(D^0\to \mu^+\mu^-)$ will not be affected, and $R^{SM}_{e,\mu}(c,u) = 1$. One also predicts $r(\bar B \to D^{(*)}\tau \bar \nu) = r(\bar B\to \tau \bar \nu)=1.26$. This can be tested by future experimental data. This is the scenario discussed in Ref.[@desh-menon]. One can try to ease the unitarity bound by including the $\lambda^\prime_{33k}\lambda^{\prime *}_{32k}$ term with positive sign so that a smaller $\lambda^\prime_{33k}$ value is now allowed. We now discuss the contributions to $C^{NP}_9$ from eq.(\[c9t\]). Note that the first term in that equation is positive definate, one needs a larger second term with negative sign to produced the required value. If one just needs to satisfy this equation, one can easily find solutions. For example, taking $\lambda^\prime_{23k} = 3.0$, one just needs to have $\lambda^\prime_{23k}\lambda^{\prime*}_{22k} + \lambda^\prime_{33k}\lambda^{\prime*}_{32k}$ to be about -0.046 to produce $C^{NP}_9\sim -0.68$. One, however, has to consider other strong constraints. A particularly important constraint is from eq.(\[bnunu\]), to satisfy $R_{\bar B\to K \nu\bar \nu} <4.3$. To produce a negative $C^{NP}_9$, $\lambda^\prime_{23k}\lambda^{\prime*}_{22k} + \lambda^\prime_{33k}\lambda^{\prime*}_{32k}$ needs to be negative. From $R_{\bar B \to K \nu \bar \nu}$ constraint, each of $r_{2322}=\lambda^\prime_{23k}\lambda^{\prime*}_{22k}$ and $r_{3332}=\lambda^\prime_{33k}\lambda^{\prime*}_{32k}$ is constrained by be larger than $-0.09$. But in general they appear together in order to produce the value required for $C^{NP}_9$. This also leads to non-zero values for $r_{2332}=\lambda^\prime_{23k}\lambda^{\prime*}_{32k}$ and $r_{3322}=\lambda^\prime_{33k}\lambda^{\prime*}_{22k}$ increasing the value for $R_{\bar B \to K \nu \bar \nu}$. We find that with all $r_{ijkl}=-0.0436$ having the same value maximizes the size of $C^{NP}_9$ while minimize $R_{\bar B\to K\nu \bar \nu}$. For this case, using $C^{NP}_9 = -0.68$ and $R_{\bar B \to K \nu \bar \nu} <4.3$, we find $\lambda^\prime_{33k,23k} = 6.3$ and a small value for $\lambda^\prime_{22,32} =-0.0068$. With the above values for $\lambda^\prime$, the constraints from $D^0\to \mu^+\mu^-$ can be satisfied. However, the predicted values for $r_{ave}$ becomes 1.48 and $R^{SM}_\mu(c)$ is about 2.9. These values are completely ruled out by existing data. Also the solution with $\lambda^\prime_{33k,23k} = 6.3$ is problematic because it violates the unitarity bound and therefore is not a viable solution neither. In figure 1, we show $C^{NP}_9$, $r_{ave}$ and $R^{SM}_{\mu}(c)$ as functions of $\lambda^\prime_{23k}$. We see a smaller $C^{NP}_9$ in size may relax the situation, but within 1$\sigma$ range for $C^{NP}_9$, The value for $R^{SM}_{\mu,\tau}(c)$ are too large to allow the model to be a viable one. ![$C^{NP}_9$, $r_{ave}$ and $R^{SM}_{\mu}(c)$ as functions of $\lambda^\prime_{23k}$ from left to right, respectively. To get $R^{SM}_{\mu}(c)-1$ down to 10%, one needs to go to the lower range the $3\sigma$ range for $C^{NP}_9$ to about -0.18[@b2s-sm]. However, in that case, $r_{ave}$ also comes down and cannot explain the observed $R(D^{(*)})$ anomaly.[]{data-label="fig"}](fig1_c9.eps "fig:") ![$C^{NP}_9$, $r_{ave}$ and $R^{SM}_{\mu}(c)$ as functions of $\lambda^\prime_{23k}$ from left to right, respectively. To get $R^{SM}_{\mu}(c)-1$ down to 10%, one needs to go to the lower range the $3\sigma$ range for $C^{NP}_9$ to about -0.18[@b2s-sm]. However, in that case, $r_{ave}$ also comes down and cannot explain the observed $R(D^{(*)})$ anomaly.[]{data-label="fig"}](fig2_rave.eps "fig:") ![$C^{NP}_9$, $r_{ave}$ and $R^{SM}_{\mu}(c)$ as functions of $\lambda^\prime_{23k}$ from left to right, respectively. To get $R^{SM}_{\mu}(c)-1$ down to 10%, one needs to go to the lower range the $3\sigma$ range for $C^{NP}_9$ to about -0.18[@b2s-sm]. However, in that case, $r_{ave}$ also comes down and cannot explain the observed $R(D^{(*)})$ anomaly.[]{data-label="fig"}](fig3_Rmu.eps "fig:") We have searched a wide range of parameter space for $\lambda^\prime$ including with complex numbers and found no solutions which can simultaneously satisfy bounds on $R_{\bar B \to K\nu\bar \nu}$ and $R^{SM}_\mu(c)$ and at the same time to solve anomalies in $R(D^{(*)})$ and $b\to s \mu^+\mu^-$. Conclusions =========== We have studied the possibility of explaining the enhancement in semileptonic decays of $\bar B \to D^{(*)} \tau \bar \nu$ and the anomalies induced by $b\to s\mu^+\mu^-$ within the framework of R-parity violating (RPV) MSSM. Exchange of down type right-handed squark coupled to quarks and leptons yield interactions which are similar to leptoquark induced interactions which have been proposed to explain the $\bar B \to D^{(*)}\to \tau \bar \nu$ by tree level interactions and $b\to s \mu^+\mu^-$ induced anomalies by loop interactions, simultaneously. However, we find that the Yukawa couplings have severe constraints from other rare processes in $B$ and $D$ decays. This interaction can provide a viable solution to $R^{D(*)}$ anomaly. But with the severe constraint from $\bar B \to K \nu \bar \nu$, it proves impossible to solve the anomalies induced by $b\to s \mu^+\mu^-$. This conclusion also applies equally to the leptoquark model proposed in Ref.[@bauer-neubert]. This work was supported by a University of Oregon Global Studies Institute grant awarded to NGD and XGH. XGH was supported in part by MOE Academic Excellent Program (Grant No. 102R891505), NCTS and MOST of ROC (Grant No. MOST104-2112-M-002-015-MY3), and in part by NSFC (Grant Nos. 11175115 and 11575111) and Shanghai Science and Technology Commission (Grant No. 11DZ2260700) of PRC. XGH thanks the Institute of Theoretical Science, Department of Physics, University of Oregon for hospitality where this work was done. We thank M. Schmidt for bringing Ref.[@new] to our attention.\ \ [**Note Added**]{} Please note that soon after our submission Becirevic et al.[@new] have submitted a paper to arXiv which reaches similar conclusion on the inadmissibility of single leptoquark explanation of anomalies. [999]{} BaBar Collaboration, J. P. Lees et al., Phys. Rev. Lett. 109 (2012) 101802, \[arXiv:1205.5442\]. BaBar Collaboration, J. P. Lees et al., Phys. Rev. D88 (2013), no. 7 072012, \[arXiv:1303.0571\]. Belle Collaboration, M. Huschle et al., Phys. Rev. D92 (2015), no. 7 072014, \[arXiv:1507.03233\]. Belle Collaboration, A. Abdesselam et al., arXiv:1603.06711. LHCb Collaboration, R. Aaij et al., Phys. Rev. Lett. 115 (2015), no. 11 111803, \[arXiv:1506.08614\]. \[Addendum: Phys. Rev. Lett. 115 (2015), no.15 159901\]. HPQCD Collaboration, H. Na, C. M. Bouchard, G. P. Lepage, C. Monahan, and J. Shigemitsu, Phys. Rev. D92 (2015), no. 5 054510, \[arXiv:1505.03925\]; S. Fajfer, J. F. Kamenik, and I. Nisandzic, Phys. Rev. D85 (2012) 094025, \[arXiv:1203.2654\]. Heavy Flavor Averaging Group (HFAG), www.slac.stanford.edu/XORG/hfag. LHCb Collaboration, PRL 111 (2013) 191801, arXiv:1308.1707 \[hep-ex\]; LHCb Collaboration, JHEP 1406 (2014) 133, arXiv:1403.8044 \[hep-ex\]; R. Aaij et al. \[LHCb Collaboration\], arXiv:1512.04442 \[hep-ex\]; LHCb Collaboration, JHEP 1307 (2013) 084, arXiv:1305.2168 \[hep-ex\]; LHCb Collaboration, JHEP 1504 (2015) 064, arXiv:1501.03038 \[hep-ex\]. S. Descotes-Genon, L. Hofer, J. Matias and J. Virto, JHEP [**1606**]{}, 092 (2016) A. Ali, arXiv:1607.04918 \[hep-ph\]. LHCb Collaboration, Phys. Rev. Lett. 113 (2014) 151601, arXiv:1406.6482 \[hep-ex\]. K. Kiers and A. Soni, Phys. Rev. D [**56**]{}, 5786 (1997); M. Tanaka and R. Watanabe, Phys. Rev. D 82, 034027 (2010); A. Datta, M. Duraisamy and D. Ghosh, Phys. Rev. D 86, 034027 (2012); D. Becirevic, N. Kosnik and A. Tayduganov, Phys. Lett. B 716, 208 (2012); X. G. He and G. Valencia, Phys. Rev. D [**87**]{}, no. 1, 014014 (2013); Y. Sakaki and H. Tanaka, Phys. Rev. D [**87**]{}, no. 5, 054002 (2013); A. Celis, M. Jung, X. Q. Li and A. Pich, JHEP [**1301**]{}, 054 (2013); \[ P. Ko, Y. Omura and C. Yu, JHEP [**1303**]{}, 151 (2013); A. Crivellin, A. Kokulu and C. Greub, Phys. Rev. D [**87**]{}, no. 9, 094031 (2013); R. Dutta, A. Bhol and A. K. Giri, Phys. Rev. D [**88**]{}, no. 11, 114023 (2013); Eur. Phys. J. C [**74**]{}, no. 5, 2861 (2014); A. Soffer, Mod. Phys. Lett. A [**29**]{}, no. 07, 1430007 (2014); J. Zhu, H. M. Gan, R. M. Wang, Y. Y. Fan, Q. Chang and Y. G. Xu, Phys. Rev. D [**93**]{}, no. 9, 094023 (2016); S. Nandi, S. K. Patra and A. Soni, arXiv:1605.07191 \[hep-ph\]. N. G. Deshpande and A. Menon, JHEP [**1301**]{}, 025 (2013) \[arXiv:1208.4134 \[hep-ph\]\]. S. Fajfer, J. F. Kamenik, I. Nisandzic and J. Zupan, Phys. Rev. Lett.  [**109**]{}, 161801 (2012) \[arXiv:1206.1872 \[hep-ph\]\]; Y. Sakaki, M. Tanaka, A. Tayduganov and R. Watanabe, Phys. Rev. D [**88**]{}, no. 9, 094012 (2013); M. Freytsis, Z. Ligeti and J. T. Ruderman, Phys. Rev. D [**92**]{}, no. 5, 054018 (2015); \[ S. Bhattacharya, S. Nandi and S. K. Patra, Phys. Rev. D [**93**]{}, no. 3, 034011 (2016); B. Dumont, K. Nishiwaki and R. Watanabe, Phys. Rev. D [**94**]{}, no. 3, 034001 (2016); X. Q. Li, Y. D. Yang and X. Zhang, arXiv:1605.09308 \[hep-ph\]; A. K. Alok, D. Kumar, S. Kumbhakar and S. U. Sankar, arXiv:1606.03164 \[hep-ph\]. J. Matias, F. Mescia, M. Ramon and J. Virto, JHEP 1204 (2012) 104; G. Hiller and M. Schmaltz, Phys. Rev. D 90 (2014) 054014; A. J. Buras, F. De Fazio and J. Girrbach, JHEP 1402 (2014) 112; S. Biswas, D. Chowdhury, S. Han and S. J. Lee, JHEP [**1502**]{}, 142 (2015); D. Aristizabal Sierra, F. Staub and A. Vicente, Phys. Rev. D 92, 015001 (2015); S. L. Glashow, D. Guadagnoli and K. Lane, Phys. Rev. Lett. 114 (2015) 091801; A. Crivellin, G. D’Ambrosio and J. Heeck, Phys. Rev. Lett. 114, 151801 (2015); C. J. Lee and J. Tandean, JHEP 1508, 123 (2015); C. W. Chiang, X. G. He and G. Valencia, Phys. Rev. D [**93**]{}, no. 7, 074003 (2016) D. Becirevic, O. Sumensari and R. Zukanovich Funchal, Eur. Phys. J. C [**76**]{}, no. 3, 134 (2016); T. Hurth, F. Mahmoudi and S. Neshatpour, Nucl. Phys. B [**909**]{}, 737 (2016); D. Guadagnoli, D. Melikhov and M. Reboud, Phys. Lett. B [**760**]{}, 442 (2016); P. Koppenburg, Z. Dolezal and M. Smizanska, Scholarpedia [**11**]{}, 32643 (2016); C. H. Chen, T. Nomura and H. Okada, arXiv:1607.04857 \[hep-ph\]; S. M. Boucenna, A. Celis, J. Fuentes-Martin, A. Vicente and J. Virto, arXiv:1608.01349 \[hep-ph\]. B. Bhattacharya, A. Datta, D. London and S. Shivashankara, Phys. Lett. B [**742**]{}, 370 (2015); D. Das, C. Hati, G. Kumar and N. Mahajan, arXiv:1605.06313 \[hep-ph\]; S. M. Boucenna, A. Celis, J. Fuentes-Martin, A. Vicente and J. Virto, arXiv:1608.01349 \[hep-ph\]. M. Bauer and M. Neubert, Phys. Rev. Lett.  [**116**]{}, no. 14, 141802 (2016). R. Barbier, C. Berat, M. Besancon, M. Chemtob, A. Deandrea, E. Dudas, P. Fayet and S. Lavignac et al., Phys. Rept. 420, 1 (2005) \[hep-ph/0406039\]. J. L. Rosner, S. Stone and R. S. Van de Water, \[arXiv:1509.02220 \[hep-ph\]\]. K.A. Olive et al. (Particle Data Group), Chin. Phys. C, 38, 090001 (2014). N. G. Deshpande, D. K. Ghosh and X. G. He, Phys. Rev. D [**70**]{}, 093003 (2004) \[hep-ph/0407021\]. A. J. Buras, F. Schwab and S. Uhlig, Rev. Mod. Phys.  [**80**]{}, 965 (2008) J. P. Lees [*et al.*]{} \[BaBar Collaboration\], Phys. Rev. D [**87**]{}, no. 11, 112005 (2013) \[arXiv:1303.7465 \[hep-ex\]\]. O. Lutz [*et al.*]{} \[Belle Collaboration\], Phys. Rev. D [**87**]{}, no. 11, 111103 (2013) doi:10.1103/PhysRevD.87.111103 \[arXiv:1303.3719 \[hep-ex\]\]. M. Aaboud [*et al.*]{} \[ATLAS Collaboration\], \[BaBar Collaboration\], arXiv:1605.09637 \[hep-ex\]. M. Bona, talk presented at ICHEP 2016, August 3 - 10, Chicago, IL, USA. D. Becirevic, N. Kosnik, O. Sumensari and R. Zukanovich Funchal, JHEP [**1611**]{}, 035 (2016) doi:10.1007/JHEP11(2016)035 \[arXiv:1608.07583 \[hep-ph\]\]. [^1]: Electronic address: desh@uoregon.edu [^2]: Electronic address: hexg@phys.ntu.edu.tw [^3]: If $k$ can take more than one values, to avoid potential problems from other terms in Eq.(10), one may resort to the scenario that $\tilde d_L$, $\tilde u_L$, $\tilde e_L$, and $\tilde \nu_L$ to be much heavier than $\tilde d_R$ so that their contributions are suppressed. [^4]: In our earlier version, we had neglected this contribution and obtained erroneous conclusions which we correct them here.

--- abstract: 'In models with an extra $\text{SU}(2)_R$ gauge group and an extended scalar sector, the cascade decays of the $W''$ boson can provide various multiboson signals. In particular, diboson decays $W'' \to WZ$ can be suppressed while $W'' \to WZX$, with $X$ one of the scalars present in the model, can reach branching ratios around 4%. We discuss these multiboson signals focusing on possible interpretations of the ATLAS excess in fat jet pair production.' --- [**Multiboson production in $W'$ decays**]{} J. A. Aguilar–Saavedra$^{a}$, F. R. Joaquim$^{b}$\ [$^a$ Departamento de Física Teórica y del Cosmos, Universidad de Granada,\ E-18071 Granada, Spain]{}\ [$^b$ Departamento de Física and CFTP, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal]{} Introduction ============ A $3.4\sigma$ local excess in boson-tagged jet pair ($JJ$) production reported by the ATLAS Collaboration [@Aad:2015owa], near an invariant mass $m_{JJ} = 2$ TeV, stands out as the most prominent anomaly that the first run of the Large Hadron Collider (LHC) has left. This excess appears in a dedicated search for heavy resonances decaying into two gauge bosons $WZ$ that subsequently decay hadronically, each boson resulting in one fat jet ($J$). The CMS analysis of the same $JJ$ final state [@Khachatryan:2014hpa] also shows some excess at roughly the same invariant mass. But, intriguingly, complementary searches in the $\ell \nu J$ channel, corresponding to the leptonic decay $W \to \ell \nu$ ($\ell = e,\mu$) and $Z$ hadronic decay, give null results [@Khachatryan:2014gha; @Aad:2015ufa], even if — as in the case of the ATLAS search — they are more sensitive to the presence of a resonance. Consequently, the limits from the non-observation of a signal in this decay mode are in tension with the cross section required to explain the excess in ref. [@Aad:2015owa]. The $\ell^+ \ell^- J$ channel with $Z \to \ell^+ \ell^-$ and $W$ decaying hadronically is less sensitive. In the case of the CMS Collaboration [@Khachatryan:2014gha] there is some $\sim 2\sigma$ excess at a smaller invariant mass $m_{\ell \ell J} \sim 1.8$ TeV but the ATLAS analysis [@Aad:2014xka] gives a SM-like result. In addition, heavy resonances decaying into two gauge bosons $VV$ ($V=W,Z$) are also expected to decay into $Vh^0$, with $h^0$ the Higgs boson. Searches for $Vh^0$ in the $JJ$ channel by the CMS Collaboration [@Khachatryan:2015bma] do not show any excess, while a preliminary $Wh^0$ resonance search in the $\ell \nu J$ final state [@CMS:2015gla], less sensitive than the former, yields a $2.2 \, \sigma$ excess at $m_{Wh^0} = 1.8$ TeV (see ref. [@Aguilar-Saavedra:2015rna] for a detailed discussion). In order to address the tension between the ATLAS diboson excess [@Aad:2015owa] in the $JJ$ channel and the limits on a possible signal from the other channels [@Khachatryan:2014hpa; @Khachatryan:2014gha; @Aad:2015ufa; @Aad:2014xka], the hypothesis that this excess is due to diboson production plus an extra particle $X$ was put forward by one of us [@Aguilar-Saavedra:2015rna]. Two production and decay topologies were identified, with a heavy resonance $R$ decaying into $VVX$ via an intermediate on-shell resonance $Y$, as depicted in figure \[fig:topVY\]. In both cases, the $VVX$ final state could give a diboson-like signal in the ATLAS analysis [@Aad:2015owa], while not showing up so conspicuously in the rest of diboson resonance searches. In this paper we present an explicit example of a model where such processes can occur, with $R$ a charged spin-$1$ boson ($W'$), $Y$ a charged ($H^\pm$) or neutral ($H_1^0$) scalar and $X$ a pseudo-scalar ($A^0$) or the Higgs boson ($h^0$). Key ingredients in the model are an additional $\text{SU}(2)'$ gauge group, whose charged member is the $W'$ boson, and an additional scalar doublet to provide the scalars $H^\pm$, $H^0$ and $A^0$. ----- -- ----- (a) (b) ----- -- ----- We note that many interpretations of the ATLAS excess in terms of a spin-$1$ resonance decaying into $WZ$, $WW$ or $ZZ$ have appeared in the literature [@spin1], several with an extended scalar sector that couples to $\text{SU}(2)_L$ as well as to a new $\text{SU}(2)'$ gauge group. This is the case, for example, of left-right (LR) models. However, only direct decays $R \to VV$ have been considered, overlooking the tension between the $JJ$ and $\ell \nu J$ analyses or atributing it to statistical fluctuations.[^1] Direct $R \to JJ$ decays have also been considered in interpretations in terms of a new spin-$0$ resonance [@spin0] and other related work [@other]. As we will show in this paper, if the extra scalars present in models with an extra $\text{SU}(2)'$ symmetry group are lighter than the $W'$ boson, their cascade decays can provide multiboson signals. An alternative explanation of the absence of signals in the $\ell \nu J$ final state is that the diboson excess is due to some new particle having a mass close to the $W$ and $Z$ masses, with hadronic decays, as proposed for example in ref. [@Allanach:2015blv]. Nevertheless, this hypothesis does not explain why a significant excess has not been seen by the CMS Collaboration in their $JJ$ resonance search. In the remainder of this paper, we will first present in section \[sec:2\] the models to be used as a framework. Multiboson $W'$ decays will be discussed in section \[sec:3\], focusing on the dependence of the different (diboson, triboson) signals on the mixing in the scalar sector of the model. The possible multiboson cross sections will be investigated in section \[sec:4\]. After this general analysis, we give in section \[sec:5\] a couple of benchmark examples where either the triboson signals dominate, or have similar size as diboson signals. We summarise our results in section \[sec:6\]. Framework {#sec:2} ========= When considering models that can give a $VVX$ signal corresponding to any of the two topologies in figure \[fig:topVY\], we restrict ourselves to particles with spin $0$, $1/2$ or $1$, as those already found in Nature. Furthermore, we consider that $VV=WZ$, since the local significance of the excess with this fat jet selection is larger ($3.4\sigma$) than for $ZZ$ ($2.9\sigma$) and $WW$ ($2.6\sigma$) selections. In order to reproduce the diboson kinematics, the extra particle $X$ should have a mass $m_X = 100-200$ GeV, and the secondary resonance $Y$ should have a mass below the TeV. We will assume that the resonance $R$ decaying into $WZX$ is a charge $\pm 1$ particle and $X$ is a neutral one, because a relatively light charged particle $X^\pm$ would be copiously produced in pairs through its gauge coupling to the photon, leading to a dijet pair signal, so far unobserved [@ATLAS:2012ds; @Chatrchyan:2013izb; @Aaltonen:2013hya; @Khachatryan:2014lpa]. If $R$ is a heavy $W'$ boson, it would also explain (see for example refs. [@Deppisch:2014qpa; @Heikinheimo:2014tba; @Aguilar-Saavedra:2014ola]) a $2.8\sigma$ excess in $e^+ e^- jj$ production found by the CMS Collaboration [@Khachatryan:2014dka], at an invariant mass $m_{eejj} \simeq 2$ TeV. On the other hand, for a charged scalar resonance it is harder to justify the required production cross section (see however ref. [@Chen:2015xql]). These arguments motivate us to extend the SM gauge symmetry with an additional $\text{SU}(2)'$. For the secondary resonance $Y$, the simplest possibility is to have a new scalar. An additional vector boson, perhaps appearing by enlarging the $\text{SU}(2)'$ group, could yield the production and decay topologies in figure \[fig:topVY\] too. However, a lighter gauge boson with a mass of few hundreds of GeV, otherwise undetected, should be (almost) fermiophobic, in contrast with the $W'$ boson resonance produced in the $s$ channel. It is unclear that such possibility is viable. Then, we are led to enlarge the scalar sector of the SM. The mixing of the SM scalar sector with additional $\text{SU}(2)_L$ singlets or triplets is very constrained by Higgs couplings measurements [@Aad:2015pla] and precision electroweak data [@deBlas:2014mba], therefore we extend the scalar sector with an additional doublet. The four complex scalar fields in the two $\text{SU}(2)_L$ doublets must transform non-trivially under $\text{SU}(2)'$, in order to couple to the $W'$ boson. It seems more natural to arrange them into two doublets. One possibility is to have a bidoublet, as in LR models; another possibility is that the two $\text{SU}(2)_L$ doublets are $\text{SU}(2)'$ doublets too. We will restrict ourselves to the first option. Also, some of the quark fields must transform non-trivially under $\text{SU}(2)'$, so as to have a $W'$ coupling to quarks. The requirement of gauge invariance of Yukawa terms implies that the $\text{SU}(2)'$ doublets must include right-handed quark fields. In the lepton sector, new neutral leptons $N_R$ can be introduced, embedding the right-handed lepton fields into $\text{SU}(2)'$ doublets. (Alternatively, the $W'$ boson can be leptophobic if the right-handed as well as the left-handed lepton fields are $\text{SU}(2)'$ singlets.) With these assignments, we can identify $\text{SU}(2)'$ with a $\text{SU}(2)_R$ gauge group. In this work we will discuss two models, which differ in the way the extended gauge group SU(2)$_L\times$SU(2)$_R\times$U(1)$_{B-L}$ is broken to the standard model (SM) one SU(2)$_L\times$U(1)$_{Y}$. We consider two distinct scenarios: the triplet [@Pati:1974yy] and the doublet [@Senjanovic:1978ev] left-right models (TLRM and DLRM, respectively). In the TLRM, the SU(2)$_R\times$U(1)$_{B-L}$ breaking occurs through the vacuum expectation value (VEV) of a $\text{SU}(2)_{R}$ triplet $$\Delta_R=\left(\! \begin{array}{cc} \Delta_R^+/\sqrt{2} &\Delta_R^{++}\\ \Delta_R^0 &-\Delta_R^+/\sqrt{2} \end{array} \right)\sim (1,3,2)\;,\; \langle \Delta_R \rangle= \frac{1}{\sqrt 2} \left(\! \begin{array}{cc} 0 &0\\ v_R &0 \end{array} \! \right)\,, \label{deltaRdef}$$ while in the DLRM, a $\text{SU}(2)_R$ doublet $\chi_R$ is added instead, $$\chi_R=\left(\! \begin{array}{c} \chi^+_R\\ \chi^0_R \end{array} \!\right)\sim (1,2,1)\;,\;\langle\chi_R \rangle=\frac{1}{\sqrt 2} \left(\! \begin{array}{c} 0\\ v_R \end{array} \!\right)\,. \label{def3}$$ Gauge interactions of $\Delta_R$ and $\chi_R$ are given by the covariant derivatives $$\begin{aligned} D^\mu \Delta_R &= \partial^\mu \Delta_R- i g_R \,\left[\frac{\vec{\tau}}{2} \cdot \vec {W}_{R}^\mu,\Delta_R\right]- i {g'}B^\mu \Delta_R\,, \notag \\ D^\mu \chi_R &= \partial^\mu \chi_R - i g_R \,\frac{\vec{\tau}}{2} \cdot \vec{W}_{R}^\mu \, \chi_R - i\, \frac{{g'}}{2}B^\mu \chi_R \,, \label{cdtriplet}\end{aligned}$$ where $g_{L,R}$ and ${g'}$ are gauge coupling constants. The SU(2)$_{L,R}$ and U(1)$_{B-L}$ gauge fields are denoted by $\vec W_{L,R}^{\mu}$, and $B^\mu$, respectively, and $\vec \tau$ are the Pauli matrices. Notice that we will not impose any discrete symmetry forcing $g_R=g_L$, as in fully LR symmetric models. The SM gauge group is broken down to $\text{U}(1)_{\text{em}}$ by the VEV of a Higgs bidoublet $$\Phi=\left( \begin{array}{cc} \phi_1^0 &\phi_2^+\\ \phi_1^- &\phi_2^0 \end{array} \right)\sim (2,2,0)\;,\; \tilde{\Phi}=\tau_2 \Phi^\ast \tau_2\,, \label{Phidef}$$ to which corresponds the covariant derivative $$D^\mu \Phi = \partial^\mu \Phi- i g_{L} \,\vec {W}_{L}^\mu \cdot \frac{\vec{\tau}}{2}\,\Phi+ i g_{R}\, \Phi\,\frac{\vec{\tau}}{2} \cdot \vec {W}_{R}^\mu\,. \label{def5}$$ The vacuum configuration of $\Phi$ is $$\langle \Phi \rangle=\frac{1}{\sqrt{2}}\left( \begin{array}{cc} v_1 &0\\ 0 & e^{i\delta}v_2 \end{array} \right)\,,\label{def7}\,$$ where $v_1=v\cos\beta$ and $v_2=v\sin\beta$, with $\tan\beta=v_2/v_1$ and $v=246$ GeV. In principle, the phase $\delta$ could trigger spontaneous CP violation in the scalar sector [@Ecker:1983hz]. Although this is an interesting possibility, for the sake of simplicity of our analysis we set $\delta=0$. Gauge boson masses and gauge scalar interactions arise from the gauge-invariant scalar kinetic terms: $$\begin{aligned} \text{TLRM:}\;\mathcal{L}_{\Phi}&={\operatorname{Tr}}[(D_\mu \Phi)^\dag(D^\mu \Phi)]+{\operatorname{Tr}}[(D_\mu \Delta_R)^\dag(D^\mu \Delta_R)] \,, \notag \\ \text{DLRM:}\;\mathcal{L}_{\Phi}&={\operatorname{Tr}}[(D_\mu \Phi)^\dag(D^\mu \Phi)]+(D_\mu \chi_R)^\dag D^\mu \chi_R\,.\label{DLRMkin}\end{aligned}$$ The charged gauge boson mass eigenstates are the SM $W$ boson, and a new $W^\prime$ boson, which we identify as being the 2 TeV resonance $R$ in Fig. \[fig:topVY\]. From eqs. (\[DLRMkin\]) one has, in the limit $v\ll v_R$, $$M_{W} \simeq \frac{1}{2}g_L v \left[1-\frac{\epsilon^2}{2k^2}\sin^2 2\beta \right]\;,\; M_{W^\prime} \simeq \frac{k}{2}g_Rv_R\left[1+\frac{\epsilon^2}{2k^2}\right]\,, \label{Wmasses}$$ where $$\epsilon\equiv \frac{v}{v_R}\simeq k\,\frac{g_R}{\,g_L}\frac{M_W}{M_{W^\prime}}\ll 1 \label{epsdef}\,,$$ and $k=\sqrt{2}\,(1)$ for the TLRM (DLRM). In the above equation, the last inequality stems from $ M_{W^\prime} \gg M_{W}$ and $g_R \sim \mathcal{O}(g_L)$. The mixing between the $W$ and $W^\prime$ mass eigenstates, $$\left(\! \begin{array}{c} W_{L}^{\mu}\\ W_{R}^{\mu} \end{array} \! \right) = \left(\! \begin{array}{cc} \cos \zeta & -\sin \zeta \\ \sin \zeta & \cos \zeta \end{array}\! \right) \left(\! \begin{array}{c} W^{\mu}\\ W^{\prime\mu} \end{array}\! \right)\,, \label{Wmix}$$ is parameterised by an angle $\zeta$ for which $$\tan 2\zeta= \frac{g_L\,g_R\, \epsilon^2 }{k^2g_R^2+(g_R^2-g_L^2)\epsilon^2}\sin 2\beta \simeq \frac{g_R}{g_L}\frac{M_W^2}{M_{W^\prime}^2}\sin 2\beta \ll 1 \label{Wmixs}\,.$$ Except for $W' \to WZ$ decays, which are enhanced by $M_{W'}^2/M_W^2$ due to the longitudinal helicity components, we will neglect $W-W^\prime$ mixing, which is equivalent to considering $W^\pm\sim W_L^\pm$ and $W^{\prime\pm}\sim W_R^\pm$. The physical neutral gauge bosons $Z$, $Z^\prime$ and the photon $A$ are related to the weak SU(2)$_{L,R}$ and U(1)$_{B-L}$ states $W_{L}^{3\mu}$, $W_{R}^{3\mu}$ and $B^\mu$ by $$\left(\! \begin{array}{c} W_{L}^{3\mu}\\ W_{R}^{3\mu}\\ B^\mu \end{array} \! \right) \simeq \left(\! \begin{array}{ccc} c_W & \dfrac{\epsilon^2}{k^4}\cot\theta_W \cos\varphi \sin^3\varphi& s_W\\ -s_W \cos\varphi & -\sin\varphi& c_W \cos\varphi\\ -s_W \sin\varphi & \cos\varphi& c_W \sin\varphi \end{array}\! \right) \left(\! \begin{array}{c} Z^{\mu}\\ Z^{\prime\mu}\\ A^\mu \end{array}\!\right)\,, \label{Zmix}$$ where $c_W\equiv \cos\theta_W$ and $s_W\equiv \sin\theta_W$, $\theta_W$ being the weak mixing angle, and with a new mixing angle $\varphi$ given by $$\cos\varphi=\frac{g_L}{g_R}\tan\theta_W\,,\quad \sin\varphi=\frac{g_L}{{g'}}\tan\theta_W \,. \label{phiangle}\,$$ The tangent of the $Z-Z'$ mixing angle $\xi$ is given by the ratio of the $(1,2)$ and $(1,1)$ elements of the mixing matrix, $$\tan \xi = \frac{\cos \varphi \sin^3 \varphi}{s_W} \frac{\epsilon^2}{k^4} \ll 1 \,.$$ At zeroth order in the small parameter $\epsilon$, the mixing between the neutral gauge bosons is completely determined by the requirements that (i) the photon couples to the electric charge; and (ii) the $Z$ boson couplings to fermions deviate little from the SM prediction. This also sets a relation among the gauge couplings, $${g'}=\frac{g_Lg_R\tan\theta_W}{\sqrt{g_R^2-g_L^2\tan^2\theta_W}}\,,$$ implying $g_R > g_L \tan \theta_W \simeq 0.55 \, g_L$. At zeroth order in $\epsilon$, the masses of the neutral gauge bosons $Z$ and $Z^\prime$ are given by $$M_{Z} \simeq \frac{g_L v}{2 c_W} \simeq \frac{M_W}{c_W} \,,\quad M_{Z^\prime} \simeq \frac{k^2}{2} v_R \sqrt{g_R^2+{{g'}}^2} \simeq \frac{k}{\sin \varphi} M_{W'} \label{Zmasses}\,.$$ In both the TLRM and DLRM, the neutral scalar spectrum contains three CP-even scalars $h^0$ and $H_{1,2}^0$, and one pseudoscalar $A^0$. In the limit $v\gg v_R$ (or equivalently $\epsilon \ll 1$), and barring unnatural cancellations, the neutral complex scalar fields $\phi_{1,2}^0$ and $\Delta_R^0,\chi_R^0$ can be written in terms of the physical fields as $$\begin{aligned} \phi_1^0 &\simeq \frac{1}{\sqrt{2}}\left[-h^0\sin\alpha +H_1^0 \cos\alpha + i (A^0 \sin\beta+G_1^0\cos\beta)\right] \,, \notag \\ \phi_2^0 &\simeq \frac{1}{\sqrt{2}}\left[h^0\cos\alpha +H_1^0 \sin\alpha + i (A^0 \cos\beta-G_1^0\sin\beta) \right] \,, \notag \\ \Delta_R^0,\chi_R^0 &\simeq \frac{1}{\sqrt{2}}(H_2^0+G_2^0) \label{chir00}\,,\end{aligned}$$ where $G_{1,2}^0$ are the Goldstone bosons and the angle $\alpha$ is the $h_0-H_1^0$ mixing angle, in the notation of the two Higgs doublet model [@Branco:2011iw]. Notice that, in general, $\alpha$ depends on the parameters of the scalar potential (see section \[sec:5\]). Moreover, mixing among $h_0, H_1^0$ and $H_2^0$ could also occur. However, and since present experimental results seem to indicate that the properties of $h^0$ are those of the SM Higgs, we will only focus on scenarios which lead to a Higgs mixing pattern like the one given above, with $\alpha$ constrained to lay in the experimentally allowed ranges in the context of a two Higgs doublet model [@Aad:2015pla]. As for the charged scalar sector, both models include a pair of charged scalars $H^\pm$, which are related to the components of $\Phi$ and $\Delta_R$ (or $\chi_R$) by the relations $$\begin{aligned} \phi_1^\pm &= \frac{kH^\pm\sin\beta}{\sqrt{k^2+\epsilon^2\cos^2 2\beta }} + G_1^\pm\cos\beta - \frac{\epsilon\,G_2^\pm \sin\beta \cos 2\beta}{\sqrt{k^2+\epsilon^2\cos^2 2\beta }} \,, \notag \\ \phi_2^\pm &=\frac{kH^\pm\cos\beta}{\sqrt{k^2+\epsilon^2\cos^2 2\beta }} - G_1^\pm\sin\beta- \frac{\epsilon\,G_2^\pm \cos\beta \cos2\beta }{\sqrt{k^2+\epsilon^2\cos^2 2\beta }} \,, \notag \\ \Delta_R^\pm,\chi_R^\pm &=\frac{\epsilon\,H^\pm}{\sqrt{k^2+\epsilon^2\cos^2 2\beta }} + \frac{kG_2^\pm}{\sqrt{k^2+\epsilon^2\cos^2 2\beta }}\label{chirpm}\,,\end{aligned}$$ where and $G_{1,2}^\pm$ are the charged Goldstone bosons. In the case of the TLRM, there are two doubly-charged scalars $\Delta_R^{\pm\pm}$ that already are physical. Since we are not interested in the phenomenology related with $\Delta_R^{\pm\pm}$, we consider these states to be heavy enough to not play any significant role in our analysis. In the approximation of eqs. (\[chir00\]), and taking $\sqrt{k^2+\epsilon^2\cos^2 2\beta }\simeq k$ in eqs. (\[chirpm\]), the relevant couplings between two vector bosons and one scalar are: $$\begin{aligned} W^+ W^- h^0 [H_1^0] & : \quad g_L M_W \; \sin(\beta-\alpha) \; [ \cos(\beta-\alpha)] \,, \notag \\ W^{\prime\pm} W^\mp h^0 [H_1^0] & : \quad -g_R M_W \; \cos(\beta+\alpha) \; [\sin(\beta+\alpha)] \,, \notag \\ W^{\prime\pm}W^\mp A^0 &:\quad \pm i g_R M_W\cos 2\beta \,, \notag \\ Z Z h^0 [H_1^0] &: \quad \frac{g_LM_W}{2c_W^2} \; \sin(\beta-\alpha) \; \left[\cos(\beta-\alpha)\right]\notag \,,\\ Z Z^\prime h^0[H_1^0] &:\quad \;\frac{g_RM_W}{c_W}\sin\varphi \; \sin(\beta-\alpha) \; \left[\cos(\beta-\alpha) \right] \,, \notag \\ W^{\prime\pm} Z H^\mp & : \quad - \frac{g_R M_W}{c_W} \cos 2\beta \,.\end{aligned}$$ Notice that the $W^{\prime\pm} Z H^\mp$ interaction receives a contribution from the $\Delta_R$ ($\chi_R$) kinetic term. These contributions differ by a $\sqrt 2$ factor for the triplet and doublet, but the difference is compensated when going to the physical basis, namely eqs. (\[chirpm\]). The relevant couplings of one gauge boson to two scalars are $$\begin{aligned} W^\pm H^\mp h^0 [H_1^0] &: \quad \mp \frac{g_L}{2} (p_{h^0 [H_1^0]} - p_{H^\pm})^\mu \; \cos(\beta-\alpha) \left[ - \sin(\beta-\alpha) \right] \,, \notag \\ W^\pm H^\mp A^0 &: \quad -i \frac{g_L}{2} (p_{A^0} - p_{H^\pm})^\mu \,, \notag \\ W^{\prime \pm} H^\mp h^0 [H_1^0] &: \quad \mp \frac{g_R}{2} (p_{h^0 [H_1^0]} - p_{H^\pm})^\mu \; \sin(\beta+\alpha) \left[ - \cos(\beta+\alpha) \right] \,, \notag \\ W^{\prime \pm} H^\mp A^0 &: \quad i \frac{g_R}{2} (p_{A^0} - p_{H^\pm})^\mu \sin 2 \beta \,, \notag \\ \displaybreak Z A^0 h^0 [H_1^0] &: \quad - \frac{g_L}{2 c_W} (p_{h^0 [H_1^0]} - p_{A^0})^\mu \; \cos(\beta-\alpha) \left[ - \sin(\beta-\alpha) \right] \,, \notag \\ Z' A^0 h^0 [H_1^0] &: \quad - \frac{g_R}{2} \sin \varphi \, (p_{h^0 [H_1^0]} - p_{A^0})^\mu \; \cos(\beta-\alpha) \left[ - \sin(\beta-\alpha) \right] \,, \notag \\ Z' H^+ H^- &: \quad \frac{g_R}{2} \sin \varphi \, (p_{H^+} - p_{H^-})^\mu \,,\end{aligned}$$ with $p_X^\mu$ the flowing-in four-momentum of particle $X$. In both the TLRM and DLRM, the three lepton and quark families are placed in left- and right-handed doublets $$\begin{aligned} & Q_{Li} = \left(\! \begin{array}{c} u_i\\ d_i \end{array} \! \right)_L\sim (2,0,1/3) \,,\quad \ell_{Li}= \left(\! \begin{array}{c} \nu_i\\ e_i \end{array} \!\right)_L \sim(2,0,-1) \,, \notag \\[1mm] & Q_{Ri} = \left(\!\begin{array}{c} u_i\\ d_i \end{array} \! \right)_R \sim(0,2,1/3) \,,\quad \ell_{Ri}=\left(\! \begin{array}{c} \nu_i\\ e_i \end{array}\! \right)_R \sim (0,2,-1) \,, \label{def1}\end{aligned}$$ where $i=1,2,3$ is a family index. Gauge interactions among fermions and gauge fields are given by: $$\begin{aligned} \mathcal{L}_{\text{g}}=\sum_{\psi=Q,\ell} \bar{\psi}_L \gamma^\mu D_{\mu}^L \psi_L + (L\rightarrow R)\;,\; D_{\mu}^{L,R}= i\partial_\mu +g_{L,R} \,\frac{\vec{\tau}}{2} \cdot \vec{W}_{\mu L,\mu R}+\, \frac{{g'}}{2}B^\mu\,,\end{aligned}$$ while the most general Yukawa Lagrangian is: $$\begin{aligned} \mathcal{L}_{\text{Yuk}}=-\,\bar{\ell}_L (\,Y_\ell \Phi + \tilde{Y}_\ell \tilde{\Phi})\ell_R -\,\bar{Q}_L (\,Y_q \Phi + \tilde{Y}_q \tilde{\Phi})Q_R+\text{h.c.}\,, \label{Yuk}\end{aligned}$$ where $Y_{\ell,q}$ and $\tilde{Y}_{\ell,q}$ are general complex Yukawa matrices. In the case of the TLRM, the additional term $-\bar{\ell_R^c} (i\tau_2 \Delta_R)\ell_R$ can be involved in the neutrino mass generation. In general, LRSM models suffer from large flavour-changing neutral current (FCNC) effects due to non-diagonal couplings of the neutral scalars with leptons and quarks. Constraints coming from the analysis of $K_L-K_S$ mass difference require neutral scalar masses larger than $5-10$ TeV [@Mohapatra:1983ae]. This lower bound increases by approximately one order of magnitude if one considers contributions to the CP-violating parameter $\epsilon_{CP}$ coming from $\Delta S=2$ Higgs exchange [@Pospelov:1996fq]. Since in our framework we require that $H_1^0$, $H^\pm$ and $A^0$ are relatively light, the Yukawa interactions given above will, in general, lead to unacceptably large FCNC effects. We will therefore consider that the above couplings are somehow suppressed (perhaps due to some extra symmetry) and fermion masses arise from Yukawa interactions generated, for instance, by higher-order operators. Such possibility has been recently explored in Ref. [@Dobrescu:2015yba], where a Yukawa pattern of the Type II two Higgs doublet model has been reproduced by considering dimension-6 operators of the type $\bar{\psi}_L \tilde{\Phi}\Delta_R^\dag\Delta_R\psi_R$ and $\bar{\psi}_L \tilde{\Phi}\tilde{\Delta}_R^\dag\tilde{\Delta}_R\psi_R$, with $\tilde{\Delta}_R=\tau_2 \Delta_R \tau_2$. In the DLRM the same reasoning can be applied replacing $\Delta_R$ by the doublet combination $\chi_R\chi_R^\dag$, which transforms as a triplet under $\text{SU}(2)_R$. $W'$ multiboson decays {#sec:3} ====================== When kinematically allowed, the $W'$ decay widths into two bosons are $$\begin{aligned} & \Gamma(W' \to WZ) = \frac{g_R^2}{192 \pi } \sin^2 2\beta \, \frac{\lambda(M_{W'}^2,M_W^2,M_Z^2)^{3/2}}{M_{W'}^5} \notag \\ & \qquad \qquad \qquad \qquad \times \left( 1 + 10 \frac{M_W^2}{M_{W'}^2} + 10 \frac{M_Z^2}{M_{W'}^2} + \frac{M_W^4}{M_{W'}^4} + \frac{M_Z^4}{M_{W'}^4} + 10 \frac{M_W^2 M_Z^2}{M_{W'}^4} \right) \,, \notag \\ & \Gamma(W' \to H^\pm Z) = \frac{g_R^2}{192 \pi c_W^2} \cos^2 2 \beta \, \frac{\lambda(M_{W'}^2, M_Z^2, M_{H^\pm}^2)^{1/2}}{M_{W'}} \notag \\ & \qquad \qquad \qquad \qquad \times \left( 1 + 10 \frac{M_Z^2}{M_{W'}^2} - 2 \frac{M_{H^\pm}^2}{M_{W'}^2} + \frac{M_Z^4}{M_{W'}^4} + \frac{M_{H^\pm}^4}{M_{W'}^4} - 2 \frac{M_Z^2 M_{H^\pm}^2}{M_{W'}^4} \right) \,, \notag \\ & \Gamma(W' \to WS) = \frac{g_R^2}{192 \pi}x_S^2 \, \frac{\lambda(M_{W'}^2, M_W^2, M_S^2)^{1/2}}{M_{W'}} \notag \\ & \qquad \qquad \qquad \qquad \times \left( 1 + 10 \frac{M_W^2}{M_{W'}^2} - 2 \frac{M_S^2}{M_{W'}^2} + \frac{M_W^4}{M_{W'}^4} + \frac{M_S^4}{M_{W'}^4} - 2 \frac{M_W^2 M_S^2}{M_{W'}^4} \right) \,, \notag \\ & \Gamma(W' \to H^\pm S) = \frac{g_R^2}{192 \pi } (1-x_S^2)\, \frac{\lambda(M_{W'}^2, M_{H^\pm}^2,M_S^2)^{3/2}}{M_{W'}^5} \,, \label{ec:Wdec}\end{aligned}$$ with $x_S^2 = \cos^2 (\beta+\alpha),\, \sin^2 (\beta+\alpha),\, \cos^2 2\beta$ for $S=h^0,H_1^0,A^0$, respectively, and $$\lambda(x,y,z)=x^2+y^2+z^2-2xy-2xz-2yz \,.$$ The partial widths into two fermions are $$\begin{aligned} & \Gamma(W' \to f \bar f') = \frac{N_c \, g_R^2}{48\pi} \, \frac{\lambda(M_{W'}^2,m_f^2,m_{f'}^2)}{M_{W'}} \notag \\ & \qquad \qquad \qquad \qquad \times \left( 1 - \frac{m_f^2}{2 M_{W'}^2} - \frac{m_{f'}^2}{2 M_{W'}^2} - \frac{m_f^4}{2 M_{W'}^4} - \frac{m_{f'}^4}{2 M_{W'}^4} + \frac{m_f^2 m_{f'}^2}{M_{W'}^4} \right) \,,\end{aligned}$$ with $N_c$ a colour factor. In the limit that $M_{W'}$ is much larger than the other masses, the branching ratio into two bosons is around $8\%$. The scalars $S$ produced in $W'$ decays can further decay into two gauge bosons, a gauge boson plus a lighter scalar, or two fermions. We list here the partial widths, provided the channels are open. For the decay of the heavy neutral scalar they are $$\begin{aligned} & \Gamma(H_1^0 \to W W) = \frac{g_L^2}{64 \pi} \cos^2(\beta-\alpha) \frac{M_{H_1^0}^3}{M_W^2} \left( 1 - 4 \frac{M_W^2}{M_{H_1^0}^2} \right)^{1/2} \left( 1-4 \frac{M_W^2}{M_{H_1^0}^2} + 12 \frac{M_W^4}{M_{H_1^0}^4} \right) \,, \notag \\ \displaybreak & \Gamma(H_1^0 \to Z Z) = \frac{g_L^2}{128 \pi c_W^2} \cos^2(\beta-\alpha) \frac{M_{H_1^0}^3}{M_Z^2} \left( 1 - 4 \frac{M_Z^2}{M_{H_1^0}^2} \right)^{1/2} \left( 1-4 \frac{M_Z^2}{M_{H_1^0}^2} + 12 \frac{M_Z^4}{M_{H_1^0}^4} \right) \,, \notag \\ & \Gamma(H_1^0 \to Z A^0) = \frac{g_L^2}{64 \pi c_W^2} \sin^2(\beta-\alpha) \, \frac{\lambda(M_{H_1^0}^2,M_Z^2,M_{A^0}^2)^{3/2}}{M_{H_1^0}^3 M_Z^2} \,, \notag \\ & \Gamma(H_1^0 \to H^+ W^-) = \Gamma(H_1^0 \to H^- W^+) = \frac{g_L^2}{64 \pi } \sin^2(\beta-\alpha) \, \frac{\lambda(M_{H_1^0}^2,M_W^2,M_{H^\pm}^2)^{3/2}}{M_{H_1^0}^3 M_W^2} \,, \notag \\ & \Gamma(H_1^0 \to f \bar f) = \frac{N_c h_{ff}^2}{16 \pi} M_{H_1^0} \left( 1 - 4 \frac{m_f^2}{M_{H_1^0}^2} \right)^{3/2} \,, \label{Hdec}\end{aligned}$$ $f$ being a fermion with Yukawa coupling $h_{ff}$ to $H_1^0$. The $H_1^0 h^0 Z$ coupling vanishes and therefore the decay $H_1^0 \to h^0 Z$ does not take place. The heavy scalar can also decay into $SS = h^0 h^0,A^0 A^0$, with widths $$\Gamma(H_1^0 \to SS) = \frac{v^2 \lambda_{H_1^0 SS}^2}{32 \pi M_{H_1^0}} \left( 1-4 \frac{m_{S}^2}{M_{H_1^0}^2} \right)^{1/2} \,,$$ with $\lambda_{H_1^0 SS}$ dimensionless trilinear couplings of order unity, which depend on the coefficients in the scalar potential (see section \[sec:5\]) and the mixing in the scalar sector. We will not consider these decays, which are less important for heavier $H_1^0$. For the pseudoscalar the widths are $$\begin{aligned} & \Gamma(A^0 \to Z h^0) = \frac{g_L^2}{64 \pi c_W^2} \cos^2(\beta-\alpha) \, \frac{\lambda(M_{A^0}^2,M_Z^2,M_{h^0}^2)^{3/2}}{M_{A^0}^3 M_Z^2} \,, \notag \\ & \Gamma(A^0 \to Z H_1^0) = \frac{g_L^2}{64 \pi c_W^2} \sin^2(\beta-\alpha) \, \frac{\lambda(M_{A^0}^2,M_Z^2,M_{H_1^0}^2)^{3/2}}{M_{A^0}^3 M_Z^2} \,, \notag \\ & \Gamma(A^0 \to H^+ W^-) = \Gamma(A^0 \to H^- W^+) = \frac{g_L^2}{64 \pi } \,\frac{\lambda(M_{H^\pm}^2,M_W^2,M_{A^0}^2)^{3/2}}{M_{A^0}^3 M_W^2} \,, \notag \\ & \Gamma(A^0 \to f \bar f) = \frac{N_c (h'_{ff})^2}{16 \pi} M_{A^0} \left( 1 - 4 \frac{m_f^2}{M_{A^0}^2} \right)^{1/2} \,, \label{ec:Adec}\end{aligned}$$ with $h'_{ff}$ the Yukawa coupling to $A^0$ of the fermion $f$. For the charged scalar, $$\begin{aligned} & \Gamma(H^\pm \to W h^0) = \frac{g_L^2}{64 \pi } \cos^2(\beta-\alpha) \, \frac{\lambda(M_{H^\pm}^2,M_W^2,M_{h^0}^2)^{3/2}}{M_{H^\pm}^3 M_W^2} \,, \notag \\ & \Gamma(H^\pm \to W H_1^0) = \frac{g_L^2}{64 \pi } \sin^2(\beta-\alpha) \, \frac{\lambda(M_{H^\pm}^2,M_W^2,M_{H_1^0}^2)^{3/2}}{M_{H^\pm}^3 M_W^2} \,, \notag \\ & \Gamma(H^\pm \to W A^0) = \frac{g_L^2}{64 \pi } \, \frac{\lambda(M_{H^\pm}^2,M_W^2,M_{A^0}^2)^{3/2}}{M_{H^\pm}^3 M_W^2} \,, \notag \\ & \Gamma(H^\pm \to f \bar f') = \frac{3 h_{ff'}^2}{16\pi} \frac{\lambda(M_{H^\pm}^2,m_f^2,m_{f'}^2)}{M_{H^\pm}} \left( 1 -\frac{m_f^2}{M_{H^\pm}^2} - \frac{m_{f'}^2}{M_{H^\pm}^2} \right) \,,\end{aligned}$$ with $f$, $f'$ two fermions and $h_{ff'}$ their Yukawa coupling to $H^\pm$. The $H^\pm W^\mp Z$ coupling is absent. We remark that the partial widths into two bosons grow with the third power of the mass of the decaying scalar, therefore these decays dominate over the rest of decays as soon as there is phase space available. Depending on the scalar mass hierarchy, there is a plethora of possible $W'$ cascade decay chains yielding multiboson signals. We will focus on two simple cases: (i) an alignment scenario where $A^0$ is lighter than $H_1^0$ and $H^\pm$; (ii) a small misalignment, and the masses of the three new scalars close so that they decay into SM gauge or Higgs bosons. Notice that the constraints on a pseudoscalar [@Chatrchyan:2013qga; @Khachatryan:2014lpa; @Khachatryan:2015tha; @Khachatryan:2015baw] are very loose, and greatly depend on the couplings assumed to the different fermions. For the charged scalar, we take a mass safely above current limits [@Agashe:2014kda], which anyway depend strongly on the parameters of the model. The same applies to the heavy scalar $H_1^0$, which also has suppressed coupling to the $W$ and $Z$ bosons. SM-like Higgs scenario ---------------------- We first consider a scenario where $\beta - \alpha= \pi/2$, in which case $h^0$ has the properties of the SM Higgs boson, and with $A^0$ lighter than $H_1^0$ and $H^\pm$, assumed to have equal masses for simplicity. We plot in figure \[fig:Wdec-s1\] the partial widths for the $W'$ decays in eqs. (\[ec:Wdec\]), normalised to $g_R=1$, as a function of $\beta$. We take fixed masses $M_{A^0} = 100$ GeV, $M_{H_1^0} = M_{H^\pm} = 500$ GeV. For fixed parameters in the scalar potential, the scalar masses do change with $\beta$, therefore figure \[fig:Wdec-s1\] is intended to illustrate the functional dependence on $\beta$ of the different decay widths. (The dependence on the $H^\pm$, $H_1^0$ and $A^0$ masses is due to kinematics, and very mild when they are much lighter than $M_{W'}$.) In this scenario, the channels $H_1^0 \to Z A^0$ and $H^\pm \to W A^0$ are open and, as aforementioned, these decays are expected to dominate. For example, with the assumed values for the masses, the Yukawa couplings required to have $\Gamma (H_1^0 \to b \bar b / t \bar t) = \Gamma(H_1^0 \to Z A^0)$ are $h_{bb} = 1.04$, $h_{tt}=1.66$, respectively, and the coupling required to have $\Gamma (H^\pm \to t \bar b) = \Gamma(H^\pm \to W A^0)$ is $h_{tb} = 1.2$. We therefore neglect the decays of $H_1^0$ and $H^\pm$ into quarks, while $A^0$ is expected to decay into $b \bar b$. We collect in table \[tab:sc1\] the multiboson signals produced in $W'$ cascade decays, for the scenario here considered. ---------------- -- ---------------------------------- -- -------------------------------------- $W' \to WZ$ $W' \to H^\pm Z \to W A^0 Z$ $W' \to H^\pm H_1^0 \to W A^0 Z A^0$ $W' \to W h^0$ $W' \to W H_1^0 \to W Z A^0$ $W' \to W A^0$ $W' \to H^\pm h^0 \to W A^0 h^0$ $W' \to H^\pm A^0 \to W A^0 A^0$ ---------------- -- ---------------------------------- -- -------------------------------------- : Multiboson signals from $W'$ decays in an alignment scenario with $A^0$ lighter than $H_1^0$ and $H^\pm$.\[tab:sc1\] We present in figure \[fig:DvsT-s1\] (left) the total size of the $WZ$ diboson (blue) and $WZX$ triboson (red) signals as a function of $\beta$. On the right panel we do the same for the $Wh^0$ and $Wh^0 X$ signals. Additionally, we include the partial widths to $W A^0$ and $W A^0 X$. These final states could mimick the ones with a Higgs boson if $M_{A^0} \sim M_{h^0}$, as the mass window typically used for tagging fat jets as $h^0$ candidates is wide, for example $110 \leq m_J \leq 135$ GeV in ref. [@CMS:2015gla]. -- -- -- -- -- -- Higgs mixing scenario --------------------- Current limits on Higgs couplings [@Aad:2015pla] allow for small deviations from the SM prediction, in particular a small non-zero $\cos (\beta - \alpha)$. We parameterise these deviations introducing a small angle $\gamma$ so that $\beta - \alpha = \pi/2 -\gamma$. We consider a scenario where $H_1^0$, $H^\pm$ and $A^0$ have similar masses so that decays among them are kinematically forbidden (for sufficiently large mass splittings, decays with off-shell $W/Z$ bosons may be important). For simplicity, we take all their masses equal, $M_{H_1^0} = M_{A^0} = M_{H^\pm} = 500$ GeV. The dependence on the angle $\beta$ of the $W'$ decay widths into two bosons, normalised to $g_R = 1$, is plotted in figure \[fig:Wdec-s2\], taking a small misalignment $\sin \gamma = 0.1$. Notice that there is a small phase shift $\gamma/2$ with respect to figure \[fig:Wdec-s1\] in the partial widths for $W' \to Wh^0$, $W' \to WH_1^0$, $W' \to H^\pm h^0$, and $W' \to H^\pm H_1^0$. The small mixing $\cos (\beta - \alpha) = \sin \gamma$ allows decays into SM gauge or Higgs bosons, i.e. $H_1^0 \to W^+ W^-$, $H_1^0 \to ZZ$, $A^0 \to Z h^0$, $H^\pm \to W h^0$, although they compete with the decays into fermions. We classify in table \[tab:sc2\] the possible multiboson signals from $W'$ cascade decays. ---------------- -- ---------------------------------- -- ------------------------------------ $W' \to WZ$ $W' \to W H_1^0 \to W W W$ $W' \to H^\pm H_1^0 \to W h^0 W W$ $W' \to W h^0$ $W' \to W H_1^0 \to W Z Z$ $W' \to H^\pm H_1^0 \to W h^0 ZZ$ $W' \to W A^0 \to W Z h^0$ $W' \to H^\pm A^0 \to W h^0 Z h^0$ $W' \to H^\pm Z \to W h^0 Z$ $W' \to H^\pm h^0 \to W h^0 h^0$ ---------------- -- ---------------------------------- -- ------------------------------------ : Multiboson signals from $W'$ decays in the Higgs mixing scenario with $H_1^0$, $A^0$ and $H^\pm$ of similar mass, and non-zero $\cos(\beta-\alpha)$.\[tab:sc2\] In figure \[fig:DvsT-s2\] (left) the total size of the $WZ$ diboson (blue), $WZX$ (red) and $WWX$ (orange) triboson signals is plotted as a function of $\beta$. For triboson signals, the solid lines correspond to negligible Yukawa couplings. For the dashed lines, we have chosen $h_{bb} = h'_{bb}$, equal to the SM bottom quark Yukawa coupling; $h_{tt} = h'_{tt}$, equal to the SM top quark Yukawa coupling; and $h_{tb} = \sqrt{h_{bb} h_{tt}}$. On the right panel we present the $Wh^0$ and $Wh^0 X$ signals. -- -- -- -- -- -- Multiboson cross sections {#sec:4} ========================= So far we have considered the relative size of diboson and triboson signals in two simplified scenarios, and their dependence on the angle $\beta$. We now address the possible size of these signals for a $W'$ boson with a mass near 2 TeV. The next-to-leading order $W'$ cross section [@Duffty:2012rf] at a centre-of-mass (CM) energy of 8 TeV can be parameterised as $$\sigma_{W'}(\text{pb}) = 638 \, g_R^2 \times \exp \left[ -4.02 M -0.088 M^2 -0.073 M^3 \right] \,,$$ with $M$ the $W'$ mass in TeV. The total $W'$ width is nearly independent of $\beta$, $\Gamma = 167 \,g_R^2$ GeV in the alignment scenario and $\Gamma = 166.5 \,g_R^2$ GeV in the Higgs mixing scenario, with a negligible variation of $\pm 0.5$ GeV depending on $\beta$. The approximate $WZ$ diboson and $WZX/WWX$ triboson branching ratios are collected in table \[tab:23BR\]. ----------- ------------------------ ------------------------- ------------------------- $\text{Br}(W' \to WZ)$ $\text{Br}(W' \to WZX)$ $\text{Br}(W' \to WWX)$ alignment $0.02 \sin^2 2 \beta$ $0.039\cos^2 2 \beta$ $0$ mixing $0.02 \sin^2 2 \beta$ $0.044\cos^2 2 \beta$ $0.011\cos^2 2 \beta$ ----------- ------------------------ ------------------------- ------------------------- : Diboson and triboson branching ratios for the Higgs alignment and Higgs mixing scenarios.\[tab:23BR\] In both cases we include the decays of the $W'$ boson into the three generations of light leptons plus a heavy neutrino $N$, with a mass taken as 500 GeV. Notice that in the Higgs mixing scenario the triboson signals may be depleted by the $H_1^0$, $A^0$, $H^\pm$ decays into fermions. The maximum size of diboson plus triboson signals depends on the relative efficiencies of each one, which can only be obtained with a detailed simulation, out of the scope of this work. The possible size of the coupling $g_R$ is constrained by other processes. Searches for $W' \to t \bar b$ production by the CMS Collaboration yield a limit $\sigma (W' \to t\bar b) \leq 40~\text{fb}$ with a 95% confidence level (CL) [@Khachatryan:2015edz] for $W'$ masses between 1.9 and 2.2 TeV, where a sum of $t \bar b$ and $\bar t b$ final states is understood. Limits from the ATLAS Collaboration [@Aad:2014xra; @Aad:2014xea] are looser. In a flavour-diagonal scenario (with no $W'$ charged mixing), and independently of the presence of other decay channels, $\Gamma(W' \to WZ) /\Gamma(W' \to t \bar b) \sim \sin^2 2\beta/12$, therefore one has a maximum $\sigma(W' \to WZ) = 3.3$ fb, only one half of the cross section needed to explain the number of excess events at the 2 TeV peak [@Aguilar-Saavedra:2015rna]. Analogously, $\sigma(W' \to WZX+WWX)$ has a maximum of $6-9$ fb, also below the required cross section especially since the efficiency is smaller than for $WZ$. However, the constraint from $W' \to t \bar b$ can be softened or even evaded if a nearly diagonal $W'$ quark mixing matrix is not assumed. Another constraint results from dijet production. The ATLAS Collaboration sets a limit [@Aad:2014aqa] $\sigma(W' \to jj) \times \mathcal{A} \leq 60$ fb for $M_{W'} = 2$ TeV, with a 95% CL. With an acceptance $\mathcal{A} \simeq 0.45$ [@Aad:2014aqa], this constraint is translated into $\sigma(W') \leq 280$ fb, i.e. $g_R \leq 1.05$, if all decay channels are open. The CMS Collaboration sets a similar limit [@Khachatryan:2015sja], $\sigma(W' \to jj) \times \mathcal{A} \leq 100$ fb for $M_{W'} = 2$ TeV. Taking an approximate acceptance of 0.64 [@Khachatryan:2015sja] (for isotropic decays) yields a looser limit, $\sigma(W') \leq 330$ fb. Interestingly, the CMS Collaboration observes a $2 \sigma$ excess but at slightly smaller invariant masses, $m_{jj} \simeq 1.8$ TeV. A third constraint results from the non-observation of the heavy $Z'$ boson. The relation between the $W'$ and $Z'$ masses depends on the representation of the scalars that break $\text{SU}(2)_R$, and also on the coupling $g_R$. We plot in figure \[fig:Zp1\] (left) the ratio $M_{Z'}/M_{W'}$ as a function of $g_R/g_L$ in the two cases that $\text{SU}(2)_R$ is broken by a scalar doublet and a scalar triplet. On the right panel we plot the $Z' \to e^+ e^-$ branching ratio, as well as the branching ratio for the $Z'$ bosonic decay modes, as a function of $g_R/g_L$. The $Z'$ boson is taken much heavier than its decay products. -- -- -- -- -- -- Combining the cross section dependence on the mass and couplings, and the coupling dependence of the $Z'$ boson mass, we plot in figure \[fig:Zp2\] the total $Z'$ boson production cross section at leading order, as a function of $g_R/g_L$, as well as the $Z' \to e^+ e^-$ cross section, for a reference $W'$ mass of 2 TeV and CM energies of 8 TeV and 13 TeV. A $K$ factor of 1.16 [@Aad:2014cka] is included to approximately reproduce the NLO cross section [@Melnikov:2006kv]. For $Z'$ masses of $2-3$ TeV, the unobservation of a signal in the 8 TeV run by the ATLAS Collaboration [@Aad:2014cka] implies $\sigma(Z' \to e^+ e^-) \lesssim 0.2$ fb, assuming lepton universality. Therefore, for a fixed $W'$ mass of 2 TeV, $Z'$ boson searches imply $g_R/g_L \leq 1$ for the doublet, while they do not constrain the range of $g_R$ shown in the case of the triplet. For heavier $W'$ bosons, the limits are looser. -- -- -- -- -- -- We conclude this section by discussing possible low-energy and precision electroweak data constraints on the $W^\prime$ and $Z^\prime$ masses and mixings [@Langacker:1991zr; @Langacker:2008yv; @Rizzo:2006nw]. In the specific context of LR symmetric models, limits on $M_{W^\prime}$, $M_{Z^\prime}$ and their corresponding mixing angles have been obtained, for instance, in Refs. [@Langacker:1989xa; @Polak:1991pc; @Chay:1998hd; @Hsieh:2010zr]. For a small $W-W'$ mixing angle $\zeta$ we have, from eq. (\[Wmixs\]) and taking $M_{W'} = 2$ TeV, $$\zeta_g \equiv \frac{g_R}{g_L} \zeta \simeq \frac{1}{2} \frac{g_R^2}{g_L^2} \frac{M_W^2}{M_{W'}^2} \sin 2\beta \simeq 0.0008 \, \frac{g_R^2}{g_L^2} \sin 2\beta \,.$$ For $g_R$ of order unity, $\zeta_g$ is below the upper limits in ref. [@Langacker:1989xa], which are of the order $|\zeta_g| \lesssim (1-2) \times 10^{-3}$, depending on the assumptions about the mixing in the right-handed sector. For the same $W'$ mass, the $Z-Z'$ mixing angle is $$\xi \simeq \frac{0.0016}{k^2} \frac{(g_R^2-g_L^2 \tan^2 \theta_W)^{3/2}}{g_L g_R^2 c_W} \,,$$ with $k^2=2 \; (1)$ for the TLRM (DLRM). The second factor in the above equation is of order unity, [*e. g.*]{} it is approximately $1.5$ for $g_R=1$, therefore the neutral mixing is compatible with the constraints from low-energy and LEP data, $-0.00040 < \xi < 0.0026$ [@Chay:1998hd]. A similar analysis presented in ref. [@Hsieh:2010zr] shows that, for $M_{W^\prime} = 2$ TeV, $M_{Z^\prime}/M_{W^\prime} \gtrsim 1.6\,(1.2)$ for the TLRM (DLRM). According to figure \[fig:Zp1\], this implies $g_R/g_L \lesssim 1.2~(1.0)$. Although these bounds are slightly in tension with the cases we are interested in, they can be relaxed with the addition of extra matter content, which naturally appears in embeddings of the $\text{SU}(2)_L \times \text{SU}(2)_R$ in a larger group.[^2] Benchmark examples {#sec:5} ================== In this section we analyse benchmark scenarios that can account for multiboson production in the context of the TLRM and DLRM, providing some examples of the general behaviour discussed in section \[sec:3\]. This requires specifying the scalar potential $V$, which can be written as $$\begin{aligned} V=V_\Phi+V_{\Phi_R}+V_{\Phi_R,\Phi}\,, \label{V}\end{aligned}$$ where $V_\Phi$ and $V_{\Phi_R}$ contain only terms with $\Phi$ and $\Phi_R \equiv \Delta_R$ ($\chi_R$) in the DLRM (TLRM), respectively, and mixed terms involving both $\Phi$ and $\Phi_R$ are included in $V_{\Phi_R,\Phi}$. The most general gauge invariant scalar potential $V_\Phi$ is [@Deshpande:1990ip] $$\begin{aligned} V_\Phi & = & -\mu_1^2 \,{\operatorname{Tr}}(\Phi^\dag \Phi)-\mu_2^2 \,\left[{\operatorname{Tr}}(\tilde{\Phi} \Phi^\dag)+{\operatorname{Tr}}(\tilde{\Phi}^\dag \Phi)\right]+\lambda_1 {\operatorname{Tr}}(\Phi^\dag \Phi)^2+ \lambda_2 \left[{\operatorname{Tr}}(\tilde{\Phi}\Phi^\dag)^2 \right.\nonumber \\ && + \left.{\operatorname{Tr}}(\tilde{\Phi}^\dag\Phi)^2\right] +\lambda_3 {\operatorname{Tr}}(\tilde{\Phi} \Phi^\dag){\operatorname{Tr}}(\tilde{\Phi}^\dag \Phi)+\lambda_4 {\operatorname{Tr}}(\Phi^\dag \Phi)\left[{\operatorname{Tr}}(\tilde{\Phi} \Phi^\dag)+{\operatorname{Tr}}(\tilde{\Phi}^\dag \Phi)\right]\,, \label{Vphi}\end{aligned}$$ where $\mu_i$ are mass parameters, and $\lambda_i$ are dimensionless. (For simplicity we restrict ourselves to the case of real coefficients in the potential.) The pure $\Phi_R$ terms for the TLRM (DLRM) are $$\begin{aligned} V_{\Delta_R}&= &-\mu_3^2\, {\operatorname{Tr}}(\Delta_R^\dag \Delta_R)+\alpha_1 {\operatorname{Tr}}(\Delta_R^\dag \Delta_R)^2+\alpha_2 {\operatorname{Tr}}(\Delta_R \Delta_R) {\operatorname{Tr}}(\Delta_R^\dag \Delta_R^\dag) \,,\notag\\ V_{\chi_R}&=&-\mu_3^2\, \chi_R^\dag \chi_R+\alpha_1 (\chi_R^\dag \chi_R)^2\label{VCR}\,,\end{aligned}$$ while for the mixed terms $V_{\Phi_R,\Phi}$ one has $$\begin{aligned} V_{\Delta_R,\Phi}&= & \rho_1 {\operatorname{Tr}}(\Phi^\dag \Phi){\operatorname{Tr}}(\Delta_R^\dag \Delta_R)+ \rho_2 {\operatorname{Tr}}(\Phi^\dag \Phi\Delta_R \Delta_R^\dag) \nonumber\\ &&+\rho_3\left[{\operatorname{Tr}}(\Phi\tilde{\Phi}^\dag)+{\operatorname{Tr}}(\Phi^\dag\tilde{\Phi})\right]{\operatorname{Tr}}(\Delta_R^\dag \Delta_R)\,, \notag \\ V_{\chi_R,\Phi}&=& \rho_1 {\operatorname{Tr}}(\Phi^\dag \Phi) \chi_R^\dag \chi_R+ \rho_2\, \chi_R^\dag\Phi^\dag \Phi \chi_R+\rho_3 \chi_R^\dag (\tilde{\Phi}^\dag\Phi+\Phi^\dag\tilde{\Phi})\chi_R\,. \label{VCRphi}\end{aligned}$$ Notice that, in general, other invariant dimension-4 combinations of the fields can be included in $V$. However, it can be shown that those can always be written as linear combinations of the terms given above. Detailed analyses of the above potential have been presented in refs. [@Senjanovic:1978ev; @Deshpande:1990ip; @Gunion:1989in]. Here, in order to provide representative examples of the benchmark scenarios discussed in section \[sec:3\], it is sufficient to consider simpler cases where some of the parameters in the potential vanish. In the first one, labeled as benchmark A, we impose a discrete symmetry $\Phi\rightarrow i\Phi$ to the scalar potential, and set $v_2=0$ [@Senjanovic:1978ev; @Gunion:1989in]. This corresponds to having $\mu_2^2=0$ in eq. (\[Vphi\]) and $\rho_3=0$ in eqs. (\[VCRphi\]). In the second one, labeled as benchmark B, we set $\lambda_4=\rho_1=\rho_3=0$ which, although not motivated by any special symmetry, will allow us to reproduce analytically the Higgs-mixing scenario considered in section \[sec:3\]. In benchmark A the minimisation conditions $\partial V/\partial v_{1,R}=0$ allow to write the mass parameters $\mu_{1,3}^2$ as $$\begin{aligned} & \mu_1^2=\lambda_1 v^2+\frac{\rho_1}{2}v_R^2 \,,\quad \mu_3^2=\alpha_1 v_R^2+\frac{\rho_1}{2}v^2\,. \label{mudefA}\end{aligned}$$ Notice that in this case $v_1=v$ since $v_2=0$. Inserting the above equalities in $V$, one can obtain the neutral scalar masses, $$\begin{aligned} & m_{h^0}^2\simeq \frac{4\,\alpha_1\lambda_1-\rho_1^2}{2\alpha_1} \,v^2 \,, \quad \,\,\,m_{H_1^0}^2\simeq 2\, v^2 (2\lambda_2+\lambda_3)+ \frac{\rho_2}{2}\,v_R^2 \,, \notag \\ & m_{H_2^0}^2\simeq 2 \alpha_1 v_R^2+ \frac{\rho_1^2}{2\alpha_1} v^2 \,, \quad m_{A^0}^2= \frac{\rho_2}{2}v_R^2 +2v^2(\lambda_3-2\lambda_2) \,, \label{mH20A}\end{aligned}$$ as well as the charged scalar mass, $$m_{H^\pm}^2 = \dfrac{\rho_2}{2k^2}\left(k^2v_R^2+v^2\right) \,, \label{mHpmA}$$ where, as before, $k=\sqrt{2}\,(1)$ for the TLRM (DLRM). From these expressions we conclude that $\rho_2$ has to be positive and small in order to yield $m_{H^\pm} \ll v_R$. Also, since we are taking $m_{H_2^0}^2 \sim v_R$, we must have $\alpha_1 >0$. This implies $4\,\alpha_1\lambda_1>\rho_1^2$ to have a positive $m_{h^0}^2$. Inverting these equations, we can obtain approximate expressions for the potential parameters in terms of the scalar masses, $$\begin{aligned} &\lambda_1\simeq \frac{1}{4v^2}\left( m_{H_1^0}^2+2 m_{h^0}^2-\sqrt{m_{H_1^0}^4-4v_R^2 v^2 \rho_1^2}\right)\;,\; \lambda_2\simeq\frac{m_{H_1^0}^2-m_{A^0}^2}{8v^2}\,,\nonumber\\ &\lambda_3\simeq \frac{1}{4v^2}\left(m_{A^0}^2+m_{H_1^0}^2-2m_{H^\pm}^2\right)\;,\;\alpha_1 \simeq \frac{1}{4v^2}\left( m_{H_1^0}^2+\sqrt{m_{H_1^0}^4-4v_R^2 v^2 \rho_1^2}\right)\,,\nonumber\\ &\rho_2\simeq \frac{2m_{H^\pm}^2}{v_R^2}\,.\end{aligned}$$ Choosing a scalar spectrum similar to that considered in the previous section, $$m_{h^0}=125~\text{GeV} \,,\quad m_{H_1^0}=m_{H^\pm}=500~\text{GeV} \,, \quad m_{A^0}=100~\text{GeV} \,, \quad m_{H_2^0}=4~\text{TeV} \,, \label{spect}$$ and taking $M_{W^\prime}=2$ TeV and $g_R = 1$, we get for the TLRM and DLRM the parameters $$\begin{aligned} \text{TLRM}&:\;\lambda_1 \simeq 0.38 \,,\quad \lambda_2 \simeq 0.50 \,,\quad \lambda_3 \simeq -0.98 \,,\quad\alpha_1 \simeq 1.0\,,\quad \rho_2 \simeq 0.06\,, \notag \\ \text{DLRM}&:\;\lambda_1 \simeq 0.63 \,,\quad \lambda_2 \simeq 0.50 \,,\quad \lambda_3 \simeq -0.98 \,,\quad \alpha_1 \simeq 0.50 \,,\quad \rho_2 \simeq 0.03 \label{params}\,,\end{aligned}$$ for $\rho_1=1$. At first order in $\epsilon$ the neutral complex scalar fields $\phi_{1,2}^0$ and $\chi_R^0,\,\Delta_R^0$ can be written as $$\begin{aligned} & \phi_1^0 \simeq \frac{-h^0 + s_{13} H_2^0 + i G_1^0 }{\sqrt{2}} \,,\quad \phi_2^0 \simeq \frac{H_1^0 + i A^0 }{\sqrt{2}} \,, \notag \\ & \Delta_R^0,\,\chi_R^0 \simeq \frac{s_{13}h_0 +H_2^0+iG_2^0}{\sqrt{2}}\;, \label{chir0b}\end{aligned}$$ with $$s_{13} \simeq \frac{\epsilon}{2} \frac{\rho_1}{\alpha_1} \simeq \frac{2\epsilon\rho_1v_R^2}{m_{H_1^0}^2+\sqrt{m_{H_1^0}^4-4\,v_R^2v^2\rho_1^2}} \,.$$ When compared with eqs. (\[chir00\]), this leads to $\cos(\beta-\alpha)=\cos \alpha=0$, i.e. no Higgs mixing. Notice that the mixing with $H_2^0$ (parameterised by $s_{13}$) is always small, even if $\rho_1 \sim 1$. Besides, in this benchmark the trilinear couplings $\lambda_{H_1^0 h^0 h^0}$ and $\lambda_{H_1^0 A^0 A^0}$ identically vanish. In benchmark B, for which $\lambda_4=\rho_1=\rho_3=0$, the minimisation conditions with respect to $v_{1,2}$ and $v_R$ lead to $$\begin{aligned} \mu_1^2 &= & \lambda_1 v^2 -\frac{\rho_2v_R^2\sin^2\beta}{2\cos 2\beta }\,,\nonumber \\ \mu_2^2 &= & \left(\lambda_2+\frac{\lambda_3}{2}\right)v^2\sin 2\beta +\frac{\rho_2}{8}v_R^2\tan 2\beta \,,\nonumber \\ \mu_3^2 & = & \frac{\rho_2}{2}v^2\sin 2\beta +\alpha_1v_R^2 \,. \label{mudefs}\end{aligned}$$ The masses of the CP-even and CP-odd scalars are in this case $$\begin{aligned} m_{h^0}^2 & \simeq & 2\left[\lambda_1+(2\lambda_2+\lambda_3)\sin^2 2\beta \right]v^2 \,,\notag \\ m_{H_1^0}^2 &\simeq & \frac{\rho_2 v_R^2}{2\cos 2\beta }+2v^2(2\lambda_2+\lambda_3)\cos^2 2\beta \,,\notag \\ m_{H_2^0}^2 & \simeq & \frac{v_R^2}{2\alpha_1} \left(4\alpha_1^2+\epsilon^2\,\rho_2^2\sin^4\beta\right)\label{mH202B}\,,\notag \\ m_{A^0}^2 &= & \frac{\rho_2 v_R^2}{2\cos 2\beta }+2v^2(\lambda_3-2\lambda_2)\label{mA0B} \,,\end{aligned}$$ where the dependence on the angle $\beta$ is apparent. The charged scalar mass is $$\begin{aligned} m_{H^\pm}^2= \frac{\rho_2 }{2k^2 \cos 2\beta } [k^2v_R^2+v^2\cos^2 2\beta ] \,.\end{aligned}$$ Again, inverting these equations we can find the potential parameters in terms of the scalar masses, $$\begin{aligned} \lambda_1 &\simeq & \frac{1}{2v^2}\left[m_{h^0}^2+(m_{H^\pm}^2-m_{H_1^0}^2)\tan^2 2\beta \right] \,, \notag \\ \lambda_2 &\simeq & \frac{1}{8v^2\cos^2 2\beta }\left[m_{H_1^0}^2-m_{H^\pm}^2+(m_{H^\pm}^2-m_{A^0}^2)\cos^2 2\beta \right]\,, \notag \\ \lambda_3 &\simeq & \frac{1}{4v^2\cos^2 2\beta }\left[m_{H_1^0}^2-m_{H^\pm}^2-(m_{H^\pm}^2-m_{A^0}^2)\cos^2 2\beta \right]\,,\notag \\ \alpha_1 &\simeq & \frac{m_{H^0_2}^2}{2v_R^2} \,,\quad \rho_2\simeq \frac{2m_{H^\pm}^2\cos 2\beta }{v_R^2}\,.\label{alpha1B} \end{aligned}$$ In contrast with benchmark A, here the Higgs mixing pattern is non-trivial. In particular, the alignment condition $\cos(\beta-\alpha)=0$ is not automatically fulfilled since $$\cos(\beta-\alpha)\simeq \frac{4\epsilon^2}{\rho_2} (2\lambda_2+\lambda_3)\cos^2 2\beta \simeq \frac{\Delta m_H^2+\epsilon^2m_{H_1^0}^2\cos^2 2\beta}{m_{H^\pm}^2}\tan 2\beta\label{cosbma}\,,$$ which is still very small if $\Delta m_H^2 \equiv m_{H_1^0}^2-m_{H^\pm}^2=0$. However, by slightly lifting the degeneracy assumption between the $H_1^0$ and $H^\pm$ masses, one can in principle obtain a sizable $h^0-H_1^0$ mixing. Besides, mixing in the $1-3$ and $2-3$ CP-even neutral scalar sectors will be also generated[^3], $$\begin{aligned} \theta_{13}\simeq \frac{m_{H^\pm}^4\epsilon \sin\beta\sin(4\beta)}{2 m_{H_2^0}^2(m_{H^\pm}^2- m_{H_2^0}^2)}\;,\;\; \theta_{23}\simeq \frac{\epsilon\, m_{H^\pm}^2(m_{H_2^0}^2-m_{H^\pm}^2\sin^2\beta)\cos(2\beta)\sin\beta}{ m_{H_2^0}^2(m_{H_2^0}^2- m_{H^\pm}^2)}\,,\end{aligned}$$ but it is always very small because $\epsilon\ll 1$ and $m_{H^\pm}/m_{H_2}\ll 1$. As numerical example we take $\cos(\beta-\alpha)\simeq 0.1$ with $\beta=0.1\pi$, in which case $W'$ decays yield diboson plus triboson production. The spectrum is the same as in (\[spect\]), except for the charged-Higgs mass that we now take as $m_{H^\pm}=530$ GeV, in order to obtain a non-zero Higgs mixing. This spectrum results from the scalar parameters $$\begin{aligned} \text{TLRM}&:\;\lambda_1 \simeq 0.24 \,,\quad \lambda_2 \simeq 0.47 \,,\quad \lambda_3 \simeq -1.3 \,,\quad \alpha_1 \simeq 1.0 \,,\quad \rho_2 \simeq 0.06 \,, \notag \\ \text{DLRM}&:\;\lambda_1 \simeq 0.24 \,,\quad \lambda_2 \simeq 0.47 \,,\quad \lambda_3 \simeq -1.3 \,,\quad \alpha_1 \simeq 0.50 \,,\quad \rho_2 \simeq 0.03 \label{params2}\,.\end{aligned}$$ We note that in this benchmark we cannot obtain mixing $\cos(\beta-\alpha)\neq 0$ for $\beta = 0$, as it can be observed from eq. (\[cosbma\]). Discussion {#sec:6} ========== The ATLAS excess [@Aad:2015owa] in $JJ$ production near $m_{JJ} = 2$ TeV is kinematically compatible with the production of a heavy resonance decaying into two bosons $W/Z$ plus an extra particle $X$, with an intermediate resonance as in figure \[fig:topVY\]. As a possible realisation of this mechanism, in this paper we have considered a SM extension with an additional $\text{SU}(2)_R$, in which the new gauge boson $W'$ is the natural candidate to explain the $JJ$ excess. We have shown in two simple scenarios that, provided the additional scalars present in the model are lighter than the $W'$ boson, the decays $W' \to WZX$ can dominate over decays $W' \to WZ$, as their respective partial widths are proportional to $\cos^2 2 \beta$ and $\sin^2 2 \beta$. In case there is a strong hierarchy among the VEVs of the two neutral scalars that break the $\text{SU}(2)_L$ gauge symmetry, $W' \to WZ$ decays will be largely suppressed ($\sin 2 \beta \sim 0$) with the rate for $W' \to WZX$ reaching its apex ($\cos 2 \beta \sim 1$). If such a hierarchy does not exist, we will have a mixture of $WZ$ and $WZX$ production in general, unless the two VEVs are equal, in which case $WZX$ production is suppressed. The latter is the situation considered in previous literature [@spin1] explaining the excess as $W' \to WZ$ production. Besides the kinematics, one has to consider the size itself of the observed excess. For a $\text{SU}(2)_R$ coupling $g_R=1$ and $\cos 2\beta = 1$, the triboson cross section is $\sigma_{WZX} \gtrsim 10$ fb. (For comparison, the maximum diboson signal is one half of this value for the same $g_R$.) While in principle this cross section is of the magnitude needed to explain the excess in ref. [@Aad:2015owa], the efficiency for triboson signals is expected to be smaller [@Aguilar-Saavedra:2015rna]. A careful evaluation of this efficiency — which depends not only on the precise details of the boson tagging but also on the identity of the particle $X$ and its mass — is out of the scope of this work. In the absence of such a detailed simulation, several qualitative arguments suggest that the efficiency for triboson signals may be not too low so as to explain the ATLAS diboson excess. 1. The decrease in selection efficiency would be around a factor of six [@Aguilar-Saavedra:2015rna] if only the kinematical configurations where the extra particle $X$ is well separated from the $W$ and $Z$ bosons were to contribute to a “diboson” signal after the kinematical selection requirements of the ATLAS analysis [@Aad:2015owa]. However, it is expected that configurations where $X$ (or some of its decay products) merge with the bosons will also contribute to this signal. 2. In this respect, one of the boson tagging variables used by the ATLAS Collaboration is the jet mass $m_J$, which is required to lie in a suitable interval around the $W$ or $Z$ pole mass. Clearly, if $X$ merges with a boosted $W/Z$ boson, then $m_J$ will increase, thus reducing the boson tagging efficiency compared to the direct $W' \to WZ$ decay. Another tagging variable is the number of tracks $N_\text{trk}$ in the jet, required to be $N_\text{trk} \leq 30$ [@Aad:2015owa; @ATLASr2JJ]. Likewise, if $X$ merges with a boosted $W/Z$ boson, the number of tracks in the jet will be larger and the boson tagging efficiency will be correspondingly lower. As a consequence of these tagging requirements, for the kinematical configurations where $X$ merges with the $W/Z$ bosons one expects a reduced, but not zero, boson tagging efficiency. 3. In the run 2 $JJ$ search [@ATLASr2JJ], the ATLAS Collaboration has provided results for the $JJ$ invariant mass distribution when requirements on one of these boson tagging variables are dropped. Interestingly, when the $m_J$ or $N_\text{trk}$ cuts are not applied, slight bumps in the $m_{JJ}$ distributions are seen around $m_{JJ}$ = 2 TeV, which are not visible when the full boson tagging is performed. Although the dataset is still limited by statistics and definite conclusions cannot be drawn, this feature certainly deserves a more detailed investigation. 4. Additional processes may mimick $WZ$ or $WZX$ production, for example $WA^0$, $WA^0A^0$ and $WA^0 h^0$ production, if the new pseudoscalar $A^0$ has a mass similar to the $W/Z$ masses, thus also increasing the potential signal. On the other hand, the possibility that $g_R$ is larger than unity is in principle allowed, leading to larger triboson cross sections. In this respect, the $W'$ gauge couplings to the quarks can be reduced due to mixing with additional vector-like quarks, as suggested in ref. [@Collins:2015wua], thereby increasing the $W'$ branching ratios into multiboson final states. (The decrease in $W'$ cross section is compensated by a larger $g_R$.) We also note that direct $W' \to q \bar q$ decays, with the two quarks tagged as boson jets, have also been proposed as additional contributions to the ATLAS $JJ$ excess [@Dobrescu:2015jvn]. By considering the efficiency plots in ref.[@Aad:2015owa] and assuming for simplicity that the tagging variables $\sqrt y$, $N_\text{trk}$ and $m_J$ are uncorrelated, we estimate that the tagging efficiency for light jets ($\epsilon_j$) is $1/40$ of the efficiency for true boson jets ($\epsilon_V$). Therefore, the $W' \to q\bar q$ signal will be suppressed by a factor $(\epsilon_j/\epsilon_V)^2 = 1/1600$ and, likely, contributes negligibly to a possible signal. (The $W' \to jj$ signal would be comparable to $W' \to WZ$ if $\epsilon_j=\epsilon_V/5$.) The contribution of $W' \to t\bar b$ with the top and bottom quark jets mistagged as boson jets is expected to be subdominant, because $\sigma(W' \to t \bar b) \leq 40$ fb [@Khachatryan:2015edz], therefore if we assume a mistagging efficiency $\epsilon_b=\epsilon_V/40$ for $b$-quark jets the possible contribution is marginal. The ATLAS diboson excess remains an interesting hint for new physics at the LHC, and for sure new run 2 data will bring light on it, settling the issue of whether this peak, if a real effect, is a diboson resonance or something more complex. For a $W'$ mass of 2 TeV, the cross section at 13 TeV is approximately 7 times higher than at 8 TeV, making up for the smaller luminosity alrady collected in 2015. The new measurements at 13 TeV [@ATLASr2JJ; @ATLASr2lnJ; @ATLASr2vJ; @ATLASr2llJ; @CMS:2015nmz] are yet inconclusive (although they seem to disfavour the possibility of a $R \to JJ$ resonance), and more data and refined analyses are needed to draw a definite conclusion. Whatever the final outcome of the new measurements is, we have shown in this paper that the scalar sector of models with an extra $\text{SU}(2)_R$ provides a rich variety of multiboson signals that are worth exploring in collider experiments. Acknowledgelements {#acknowledgelements .unnumbered} ================== We thank J. Collins, M. Pérez-Victoria and J. Santiago for useful comments. This work has been supported by [*Fundação para a Ciência e a Tecnologia*]{} (FCT, Portugal) under the project UID/FIS/00777/2013, by MINECO (Spain) project FPA2013-47836-C3-2-P and by Junta de Andalucía project FQM 101. [99]{} G. Aad [*et al.*]{} \[ATLAS Collaboration\], arXiv:1506.00962 \[hep-ex\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], JHEP [**1408**]{} (2014) 173 \[arXiv:1405.1994 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], JHEP [**1408**]{} (2014) 174 \[arXiv:1405.3447 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**75**]{} (2015) 5, 209 \[Erratum Eur. Phys. J. C [**75**]{} (2015) 370\] \[arXiv:1503.04677 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**75**]{} (2015) 69 \[arXiv:1409.6190 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1506.01443 \[hep-ex\]. CMS Collaboration, Report CMS-PAS-EXO-14-010. J. A. Aguilar-Saavedra, JHEP [**1510**]{} (2015) 099 \[arXiv:1506.06739 \[hep-ph\]\]. J. Hisano, N. Nagata and Y. Omura, Phys. Rev. D [**92**]{} (2015) 5, 055001 \[arXiv:1506.03931 \[hep-ph\]\]; H. S. Fukano, M. Kurachi, S. Matsuzaki, K. Terashi and K. Yamawaki, Phys. Lett. B [**750**]{} (2015) 259 \[arXiv:1506.03751 \[hep-ph\]\]; D. B. Franzosi, M. T. Frandsen and F. Sannino, arXiv:1506.04392 \[hep-ph\]; L. Bian, D. Liu and J. Shu, arXiv:1507.06018 \[hep-ph\]; K. Cheung, W. Y. Keung, P. Y. Tseng and T. C. Yuan, arXiv:1506.06064 \[hep-ph\]; Y. Gao, T. Ghosh, K. Sinha and J. H. Yu, Phys. Rev. D [**92**]{} (2015) 5, 055030 \[arXiv:1506.07511 \[hep-ph\]\]; A. Thamm, R. Torre and A. Wulzer, arXiv:1506.08688 \[hep-ph\]; J. Brehmer, J. Hewett, J. Kopp, T. Rizzo and J. Tattersall, JHEP [**1510**]{} (2015) 182 \[arXiv:1507.00013 \[hep-ph\]\]; Q. H. Cao, B. Yan and D. M. Zhang, arXiv:1507.00268 \[hep-ph\]; G. Cacciapaglia and M. T. Frandsen, Phys. Rev. D [**92**]{} (2015) 055035 \[arXiv:1507.00900 \[hep-ph\]\]; T. Abe, R. Nagai, S. Okawa and M. Tanabashi, Phys. Rev. D [**92**]{} (2015) 5, 055016 \[arXiv:1507.01185 \[hep-ph\]\]; J. Heeck and S. Patra, Phys. Rev. Lett.  [**115**]{} (2015) 12, 121804 \[arXiv:1507.01584 \[hep-ph\]\]; T. Abe, T. Kitahara and M. M. Nojiri, arXiv:1507.01681 \[hep-ph\]; A. Carmona, A. Delgado, M. QuirÛs and J. Santiago, JHEP [**1509**]{} (2015) 186 \[arXiv:1507.01914 \[hep-ph\]\]; H. S. Fukano, S. Matsuzaki and K. Yamawaki, arXiv:1507.03428 \[hep-ph\]; L. A. Anchordoqui, I. Antoniadis, H. Goldberg, X. Huang, D. Lust and T. R. Taylor, Phys. Lett. B [**749**]{} (2015) 484 \[arXiv:1507.05299 \[hep-ph\]\]; K. Lane and L. Prichett, arXiv:1507.07102 \[hep-ph\]; A. E. Faraggi and M. Guzzi, arXiv:1507.07406 \[hep-ph\]; M. Low, A. Tesi and L. T. Wang, Phys. Rev. D [**92**]{} (2015) 8, 085019 \[arXiv:1507.07557 \[hep-ph\]\]; P. Arnan, D. Espriu and F. Mescia, arXiv:1508.00174 \[hep-ph\]; P. S. Bhupal Dev and R. N. Mohapatra, Phys. Rev. Lett.  [**115**]{} (2015) 18, 181803 \[arXiv:1508.02277 \[hep-ph\]\]; A. Dobado, F. K. Guo and F. J. Llanes-Estrada, arXiv:1508.03544 \[hep-ph\]; F. F. Deppisch, L. Graf, S. Kulkarni, S. Patra, W. Rodejohann, N. Sahu and U. Sarkar, arXiv:1508.05940 \[hep-ph\]; U. Aydemir, D. Minic, C. Sun and T. Takeuchi, arXiv:1509.01606 \[hep-ph\]; R. L. Awasthi, P. S. B. Dev and M. Mitra, arXiv:1509.05387 \[hep-ph\]; T. Li, J. A. Maxin, V. E. Mayes and D. V. Nanopoulos, arXiv:1509.06821 \[hep-ph\]. P. Ko and T. Nomura, arXiv:1510.07872 \[hep-ph\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], arXiv:1512.05099 \[hep-ex\]. ATLAS collaboration, ATLAS-CONF-2015-073. ATLAS collaboration, ATLAS-CONF-2015-075. ATLAS collaboration, ATLAS-CONF-2015-068. ATLAS collaboration, ATLAS-CONF-2015-071. CMS Collaboration \[CMS Collaboration\], CMS-PAS-EXO-15-002. C. W. Chiang, H. Fukuda, K. Harigaya, M. Ibe and T. T. Yanagida, JHEP [**1511**]{} (2015) 015 \[arXiv:1507.02483 \[hep-ph\]\]; G. Cacciapaglia, A. Deandrea and M. Hashimoto, Phys. Rev. Lett.  [**115**]{} (2015) 17, 171802 \[arXiv:1507.03098 \[hep-ph\]\]; V. Sanz, arXiv:1507.03553 \[hep-ph\]; C. H. Chen and T. Nomura, Phys. Lett. B [**749**]{} (2015) 464 \[arXiv:1507.04431 \[hep-ph\]\]; Y. Omura, K. Tobe and K. Tsumura, Phys. Rev. D [**92**]{} (2015) 5, 055015 \[arXiv:1507.05028 \[hep-ph\]\]; W. Chao, arXiv:1507.05310 \[hep-ph\]; C. Petersson and R. Torre, arXiv:1508.05632 \[hep-ph\]; C. H. Chen and T. Nomura, arXiv:1509.02039 \[hep-ph\]; D. Aristizabal Sierra, J. Herrero-Garcia, D. Restrepo and A. Vicente, arXiv:1510.03437 \[hep-ph\]. B. C. Allanach, B. Gripaios and D. Sutherland, Phys. Rev. D [**92**]{} (2015) 5, 055003 \[arXiv:1507.01638 \[hep-ph\]\]; D. Kim, K. Kong, H. M. Lee and S. C. Park, arXiv:1507.06312 \[hep-ph\]; S. P. Liew and S. Shirai, arXiv:1507.08273 \[hep-ph\]; H. Terazawa and M. Yasue, arXiv:1508.00172 \[hep-ph\]; D. GonÁalves, F. Krauss and M. Spannowsky, Phys. Rev. D [**92**]{} (2015) 5, 053010 \[arXiv:1508.04162 \[hep-ph\]\]; S. Fichet and G. von Gersdorff, arXiv:1508.04814 \[hep-ph\]; L. Bian, D. Liu, J. Shu and Y. Zhang, arXiv:1509.02787 \[hep-ph\]; A. Sajjad, arXiv:1511.02244 \[hep-ph\]; B. Bhattacherjee, P. Byakti, C. K. Khosa, J. Lahiri and G. Mendiratta, arXiv:1511.02797 \[hep-ph\]. B. C. Allanach, P. S. B. Dev and K. Sakurai, arXiv:1511.01483 \[hep-ph\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**73**]{} (2013) 1, 2263 \[arXiv:1210.4826 \[hep-ex\]\]. S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Phys. Rev. Lett.  [**110**]{} (2013) 14, 141802 \[arXiv:1302.0531 \[hep-ex\]\]. T. Aaltonen [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett.  [**111**]{} (2013) 3, 031802 \[arXiv:1303.2699 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**747**]{} (2015) 98 \[arXiv:1412.7706 \[hep-ex\]\]. F. F. Deppisch, T. E. Gonzalo, S. Patra, N. Sahu and U. Sarkar, Phys. Rev. D [**90**]{} (2014) 5, 053014 \[arXiv:1407.5384 \[hep-ph\]\]. M. Heikinheimo, M. Raidal and C. Spethmann, Eur. Phys. J. C [**74**]{} (2014) 10, 3107 \[arXiv:1407.6908 \[hep-ph\]\]. J. A. Aguilar-Saavedra and F. R. Joaquim, Phys. Rev. D [**90**]{} (2014) 11, 115010 \[arXiv:1408.2456 \[hep-ph\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Eur. Phys. J. C [**74**]{} (2014) 11, 3149 \[arXiv:1407.3683 \[hep-ex\]\]. C. H. Chen and T. Nomura in ref. [@spin0] G. Aad [*et al.*]{} \[ATLAS Collaboration\], arXiv:1509.00672 \[hep-ex\]. J. de Blas, M. Chala, M. Perez-Victoria and J. Santiago, JHEP [**1504**]{} (2015) 078 \[arXiv:1412.8480 \[hep-ph\]\]. J. C. Pati and A. Salam, Phys. Rev. D [**10**]{}, 275 (1974) \[Erratum-ibid. D [**11**]{}, 703 (1975)\]; R. N. Mohapatra and J. C. Pati, Phys. Rev. D [**11**]{}, 2558 (1975); G. Senjanovic and R. N. Mohapatra, Phys. Rev. D [**12**]{}, 1502 (1975). G. Senjanovic, Nucl. Phys. B [**153**]{} (1979) 334. G. Ecker, W. Grimus and H. Neufeld, Nucl. Phys. B [**247**]{} (1984) 70. G. C. Branco, P. M. Ferreira, L. Lavoura, M. N. Rebelo, M. Sher and J. P. Silva, Phys. Rept.  [**516**]{} (2012) 1 \[arXiv:1106.0034 \[hep-ph\]\]. R. N. Mohapatra, G. Senjanovic and M. D. Tran, Phys. Rev. D [**28**]{} (1983) 546; F. J. Gilman and M. H. Reno, Phys. Lett. B [**127**]{} (1983) 426; Phys. Rev. D [**29**]{} (1984) 937; G. Ecker and W. Grimus, Nucl. Phys. B [**258**]{} (1985) 328. M. E. Pospelov, Phys. Rev. D [**56**]{} (1997) 259. B. A. Dobrescu and Z. Liu, JHEP [**1510**]{} (2015) 118. S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**722**]{} (2013) 207 \[arXiv:1302.2892 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1510.01181 \[hep-ex\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1511.03610 \[hep-ex\]. K. A. Olive [*et al.*]{} \[Particle Data Group Collaboration\], Chin. Phys. C [**38**]{} (2014) 090001. D. Duffty and Z. Sullivan, Phys. Rev. D [**86**]{} (2012) 075018 \[arXiv:1208.4858 \[hep-ph\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], arXiv:1509.06051 \[hep-ex\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Eur. Phys. J. C [**75**]{} (2015) 4, 165 \[arXiv:1408.0886 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Lett. B [**743**]{} (2015) 235 \[arXiv:1410.4103 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Rev. D [**91**]{} (2015) 5, 052007 \[arXiv:1407.1376 \[hep-ex\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Rev. D [**91**]{} (2015) 5, 052009 \[arXiv:1501.04198 \[hep-ex\]\]. G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Rev. D [**90**]{} (2014) 5, 052005 \[arXiv:1405.4123 \[hep-ex\]\]. K. Melnikov and F. Petriello, Phys. Rev. D [**74**]{} (2006) 114017 \[hep-ph/0609070\]. P. Langacker, M. x. Luo and A. K. Mann, Rev. Mod. Phys.  [**64**]{} (1992) 87. P. Langacker, Rev. Mod. Phys.  [**81**]{} (2009) 1199 \[arXiv:0801.1345 \[hep-ph\]\]. T. G. Rizzo, hep-ph/0610104. P. Langacker and S. U. Sankar, Phys. Rev. D [**40**]{} (1989) 1569. J. Polak and M. Zralek, Nucl. Phys. B [**363**]{} (1991) 385. J. Chay, K. Y. Lee and S. h. Nam, Phys. Rev. D [**61**]{} (2000) 035002 \[hep-ph/9809298\]. K. Hsieh, K. Schmitz, J. H. Yu and C.-P. Yuan, Phys. Rev. D [**82**]{} (2010) 035011 \[arXiv:1003.3482 \[hep-ph\]\]. J. A. Aguilar-Saavedra, R. Benbrik, S. Heinemeyer and M. Pérez-Victoria, Phys. Rev. D [**88**]{} (2013) 9, 094010 \[arXiv:1306.0572 \[hep-ph\]\]. N. G. Deshpande, J. F. Gunion, B. Kayser and F. I. Olness, Phys. Rev. D [**44**]{} (1991) 837. J. F. Gunion, J. Grifols, A. Mendez, B. Kayser and F. I. Olness, Phys. Rev. D [**40**]{} (1989) 1546. J. H. Collins and W. H. Ng, arXiv:1510.08083 \[hep-ph\]. B. A. Dobrescu and P. J. Fox, arXiv:1511.02148 \[hep-ph\]. [^1]: Recently, the tension between the $JJ$ excess and the SM-like results in the rest of ATLAS diboson searches has been numerically quantified [@Aad:2015ipg], and amounts to 2.9 standard deviations. Preliminary results from the second run at 13 TeV leave no significant excess either [@ATLASr2JJ; @ATLASr2lnJ; @ATLASr2vJ; @ATLASr2llJ; @CMS:2015nmz], with a mild $1\,\sigma$ enhancement over the SM prediction near 2 TeV in the ATLAS $JJ$ search. [^2]: According to ref. [@Hsieh:2010zr], the measurements that mainly drive the limits for this model are the $b$-quark forward-backward asymmetry at LEP and the $Z$ boson hadronic width. These two quantities are also modified when vector-like fermions mix with the third generation [@Aguilar-Saavedra:2013qpa]. [^3]: Here, $\theta_{13}$ and $\theta_{23}$ are defined according to the standard parameterisation of a unitary $3\times 3$ matrix [@Agashe:2014kda].

--- abstract: 'We show that every planar graph $G$ has a 2-fold 9-coloring. In particular, this implies that $G$ has fractional chromatic number at most $\frac92$. This is the first proof (independent of the 4 Color Theorem) that there exists a constant $k<5$ such that every planar $G$ has fractional chromatic number at most $k$. We also show that every $n$-vertex planar graph has an independent set of size at least $\frac{3n}{13}$. This improves on Albertson’s bound of $\frac{2n}9$.' author: - 'Daniel W. Cranston' - Landon Rabern bibliography: - '45ct.bib' title: ---

--- abstract: 'We prove that the the free orthogonal and free unitary quantum groups ${\mathbb{F}}O_{N}^{+}$ and ${\mathbb{F}}U_{N}^{+}$ are weakly amenable and that their Cowling-Haagerup constant is equal to $1$. This is achieved by estimating the completely bounded norm of the projections on the coefficients of irreducible representations of their compact duals. An argument of monoidal equivalence then allows us to extend this result to quantum automorphism groups of finite spaces and even yields some examples of weakly amenable non-unimodular discrete quantum groups with the Haagerup property.' address: 'Univ. Paris Diderot, Sorbonne Paris Cité, UMR 7586, 8 place FM/13, 75013, Paris, France' author: - Amaury Freslon bibliography: - '../quantum.bib' title: Examples of weakly amenable discrete quantum groups --- Introduction ============ The *free orthogonal* and *free unitary* quantum groups were constructed by A. Van Daele and S. Wang in [@van1996universal; @wang1995free]. They are defined as universal C\*-algebras generalizing the algebras of continuous functions on the classical orthogonal and unitary groups, together with some additional structure turning them into *compact quantum groups*. From then on, these compact quantum groups have been studied from various points of view : probabilistic, geometric and operator algebraic. In particular, their reduced C\*-algebras and von Neumann algebras form interesting classes of examples somehow in the same way as those arising from discrete groups. After the first works of T. Banica [@banica1996theorie; @banica1997groupe], it appeared that these operator algebras are closely linked to free group algebras. This link was made more clear by the results of S. Vaes and R. Vergnioux [@vaes2007boundary] on exactness and factoriality and those of M. Brannan [@brannan2011approximation] on the Haagerup property and the metric approximation property. Weak amenability for locally compact groups was originally defined by M. Cowling and U. Haagerup in [@cowling1989completely] and studied in the context of real simple Lie groups by J. de Cannière, M. Cowling and U. Haagerup in [@cowling1989completely; @de1985multipliers]. In the discrete setting, many examples were provided by N. Ozawa’s result [@ozawa2007weak] stating that all Gromov hyperbolic groups are weakly amenable. Weak amenability has recently attracted a lot of attention since it is a key ingredient in some of S. Popa’s deformation/rigidity techniques, see for example [@ozawa2010class; @ozawa2010classbis]. Another feature of this approximation property is that it provides a numerical invariant which carries to the associated operator algebras and may thus give a way to distinguish them. An introduction to approximation properties for classical discrete groups can be found in [@brown2008finite Chapter 12], though no knowledge on this subject will be required afterwards. The results mentionned in the first paragraph naturally raise the issue of weak amenability for free quantum groups. It has been strongly suspected for some time that they have a Cowling-Haagerup constant equal to $1$, and this is what we prove in the present paper. To do this, we show that the completely bounded norm of the projections on coefficients of a fixed irreducible representation (i.e. on “words of fixed length”) grows polynomially. This fact can then easily be combined with M. Brannan’s proof of the Haagerup property to yield weak amenability when $F = \operatorname{Id}$, or more generally when $F$ is unitary. In the other (non-unimodular) cases, we are unable to give a complete answer to the problem, due to the lack of a proof of the Haagerup property. Let us now briefly outline the organization of the paper. In Section \[sec:preliminaries\], we recall some basic facts about compact and discrete quantum groups and we fix notations. We also give some fundamental definitions and results concerning free quantum groups. We then introduce weak amenability for discrete quantum groups in Section \[sec:quantumwa\]. Section \[subsec:blocks\] contains the first technical part of our result, reducing the problem to controlling the norms of certain blocks of “operator-valued functions” on the discrete quantum groups considered. This can be interpreted as an analogue of the operator-valued Haagerup inequality for free groups. Another technical result is worked out in Section \[subsec:recursion\] to obtain a suitable bound on the completely bounded norm of the projection on some fixed irreducible representation. Combining these results then easily yields our main result in Section \[subsec:final\]. Using monoidal equivalence, we can then transfer our estimate to quantum automorphism groups of finite-dimensional C\*-algebras and we can then extend our weak amenability result to this part of this class. Acknowledgments {#acknowledgments .unnumbered} --------------- We are deeply indebted to R. Vergnioux for the time he spent discussing the arguments of this paper and the knowledge on free quantum groups he was kind enough to share. We would also like to thank M. Brannan, M. de la Salle, P. Fima, P. Jolissaint and S. Vaes for discussions on topics linked to quantum groups and approximation properties at various stages of this project. Preliminaries {#sec:preliminaries} ============= Notations --------- All inner products will be taken linear in the second variable. For two Hilbert spaces $H$ and $K$, ${\mathcal{B}}(H, K)$ will denote the set of bounded linear maps from $H$ to $K$ and ${\mathcal{B}}(H):={\mathcal{B}}(H, H)$. In the same way we will use the notations ${\mathcal{K}}(H, K)$ and ${\mathcal{K}}(H)$ for compact linear maps. We will denote by ${\mathcal{B}}(H)_{*}$ the predual of ${\mathcal{B}}(H)$, i.e. the Banach space of all normal linear forms on ${\mathcal{B}}(H)$. On any tensor product $H\otimes H'$ of Hilbert spaces, we define the flip operator $$\Sigma : \left\{\begin{array}{ccc} H\otimes H' & \rightarrow & H'\otimes H \\ x\otimes y & \mapsto & y\otimes x \end{array}\right.$$ We will use the usual leg-numbering notations : for an operator $X$ acting on a tensor product, we set $X_{12}:=X\otimes1$, $X_{23}:=1\otimes X$ and $X_{13}:=(\Sigma\otimes 1)(1\otimes X)(\Sigma\otimes 1)$. The identity map of an algebra $A$ will be denoted ${\imath}_{A}$ or simply ${\imath}$ if there is no possible confusion. For a subset $B$ of a topological vector space $C$, $\operatorname{\overline{span}}B$ will denote the *closed linear span* of $B$ in $C$. The symbol $\otimes$ will denote the *minimal* (or spatial) tensor product of C\*-algebras or the topological tensor product of Hilbert spaces. The spatial tensor product of von Neumann algebras will be denoted $\overline{\otimes}$ and the algebraic tensor product (over ${\mathbb{C}}$) will be denoted $\odot$. Compact and discrete quantum groups ----------------------------------- Discrete quantum groups will be seen as duals of compact quantum groups in the sense of Woronowicz. We briefly review the basic theory of compact quantum groups as introduced in [@woronowicz1995compact]. Another survey, encompassing the non-separable case, can be found in [@maes1998notes]. Emphasis has been put on the explicit description of the associated $L^{2}$-space since this will prove crucial in the sequel. A *compact quantum group* ${\mathbb{G}}$ is a pair $(C({\mathbb{G}}), \Delta)$ where $C({\mathbb{G}})$ is a unital C\*-algebra and $\Delta : C({\mathbb{G}})\rightarrow C({\mathbb{G}})\otimes C({\mathbb{G}})$ is a unital $*$-homomorphism such that $$\begin{aligned} (\Delta\otimes {\imath})\circ\Delta & = & ({\imath}\otimes\Delta)\circ\Delta \\ \operatorname{\overline{span}}\{\Delta(C({\mathbb{G}}))(1\otimes C({\mathbb{G}}))\} & = & C({\mathbb{G}})\otimes C({\mathbb{G}}) \\ \operatorname{\overline{span}}\{\Delta(C({\mathbb{G}}))(C({\mathbb{G}})\otimes 1)\} & = & C({\mathbb{G}})\otimes C({\mathbb{G}})\end{aligned}$$ The main feature of compact quantum groups is the existence of a Haar state which is both left and right invariant (see [@woronowicz1995compact Thm 1.3]). Let ${\mathbb{G}}$ be a compact quantum group. There is a unique *Haar state* on ${\mathbb{G}}$, that is to say a state $h$ on $C({\mathbb{G}})$ such that for all $a\in C({\mathbb{G}})$, $$\begin{aligned} ({\imath}\otimes h)\circ {\Delta}(a) = h(a).1 \\ (h\otimes {\imath})\circ {\Delta}(a) = h(a).1\end{aligned}$$ Let $(L^{2}({\mathbb{G}}), \pi_{h}, \xi_{h})$ be the associated GNS construction and let $C_{\text{red}}({\mathbb{G}})$ be the image of $C({\mathbb{G}})$ under the GNS map $\pi_{h}$. It is called the *reduced C\*-algebra* of ${\mathbb{G}}$. Let $W$ be the unique unitary operator on $L^{2}({\mathbb{G}})\otimes L^{2}({\mathbb{G}})$ such that $$W^{*}(\xi\otimes \pi_{h}(a)\xi_{h}) = (\pi_{h}\otimes \pi_{h})\circ\Delta(a)(\xi\otimes \xi_{h})$$ for $\xi \in L^{2}({\mathbb{G}})$ and $a\in C({\mathbb{G}})$, and let ${\widehat}{W} := \Sigma W^{*}\Sigma$. Then $W$ is a *multiplicative unitary* in the sense of [@baaj1993unitaires], i.e. $W_{12}W_{13}W_{23} = W_{23}W_{12}$ and we have the following equalities : $$C_{\text{red}}({\mathbb{G}}) = \operatorname{\overline{span}}\{({\imath}\otimes {\mathcal{B}}(L^{2}({\mathbb{G}}))_{*})(W)\} \text{ and } \Delta(x) = W^{*}(1\otimes x)W.$$ Moreover, we can define the *dual discrete quantum group* ${\widehat}{{\mathbb{G}}} = (C_{0}({\widehat}{{\mathbb{G}}}), {\widehat}{\Delta})$ by $$C_{0}({\widehat}{{\mathbb{G}}}) = \operatorname{\overline{span}}\{({\mathcal{B}}(L^{2}({\mathbb{G}}))_{*}\otimes{\imath})(W)\} \text{ and } {\widehat}{\Delta}(x) = \Sigma W(x\otimes 1)W^{*}\Sigma.$$ The two von Neumann algebras associated to these quantum groups are then $$L^{\infty}({\mathbb{G}}) = C_{\text{red}}({\mathbb{G}})'' \text{ and } \ell^{\infty}({\widehat}{{\mathbb{G}}})=C_{0}({\widehat}{{\mathbb{G}}})''$$ where the bicommutants are taken in ${\mathcal{B}}(L^{2}({\mathbb{G}}))$. The coproducts extend to normal maps on these von Neumann algebras and one can prove that $W\in L^{\infty}({\mathbb{G}})\overline{\otimes} \ell^{\infty}({\widehat}{{\mathbb{G}}})$. The Haar state of ${\mathbb{G}}$ extends to a state on $L^{\infty}({\mathbb{G}})$. Irreducible representations and the GNS construction {#subsec:gns} ---------------------------------------------------- We will need in the sequel an explicit description of the GNS construction of the Haar state $h$ using the following notion of irreducible representation of a compact quantum group. A *representation* of a compact quantum group ${\mathbb{G}}$ on a Hilbert space $H$ is an operator $u\in L^{\infty}({\mathbb{G}})\overline{\otimes} {\mathcal{B}}(H)$ such that $({\Delta}\otimes {\imath})(u) = u_{13}u_{23}$. It is said to be *unitary* if the operator $u$ is unitary. Let ${\mathbb{G}}$ be a compact quantum group and let $u$ and $v$ be two representations of ${\mathbb{G}}$ on Hilbert spaces $H_{u}$ and $H_{v}$ respectively. An *intertwiner* (or *morphism*) between $u$ and $v$ is a map $T\in {\mathcal{B}}(H_{u}, H_{v})$ such that $v(1\otimes T) = (1\otimes T)u$. The set of intertwiners between $u$ and $v$ will be denoted $\operatorname{Mor}(u, v)$. A representation $u$ will be said to be *irreducible* if $\operatorname{Mor}(u, u) = {\mathbb{C}}.\operatorname{Id}$ and it will be said to be a *subrepresentation* of $v$ if there is an isometric intertwiner between $u$ and $v$. We will say that two representations are *equivalent* (resp. *unitarily equivalent*) if there is an intertwiner between them which is an isomorphism (resp. a unitary). Let us define two fundamental operations on representations. Let ${\mathbb{G}}$ be a compact quantum group and let $u$ and $v$ be two representations of ${\mathbb{G}}$ on Hilbert spaces $H_{u}$ and $H_{v}$ respectively. The *direct sum* of $u$ and $v$ is the diagonal sum of the operators $u$ and $v$ seen as an element of $L^{\infty}({\mathbb{G}})\otimes {\mathcal{B}}(H_{u}\oplus H_{v})$. It is a representation denoted $u\oplus v$. The *tensor product* of $u$ and $v$ is the element $u_{12}v_{13}\in L^{\infty}({\mathbb{G}})\otimes {\mathcal{B}}(H_{u}\otimes H_{v})$. It is a representation denoted $u\otimes v$. The following generalization of the classical Peter-Weyl theory holds (see [@woronowicz1995compact Section 6]). Every representation of a compact quantum group is equivalent to a unitary one. Every irreducible representation of a compact quantum group is finite dimensional and every unitary representation is unitarily equivalent to a sum of irreducible ones. Moreover, the linear span of the coefficients of all irreducible representations is a dense Hopf $*$-subalgebra of $C({\mathbb{G}})$ denoted $\operatorname{Pol}({\mathbb{G}})$. Let $\operatorname{Irr}({\mathbb{G}})$ be the set of isomorphism classes of irreducible unitary representations of ${\mathbb{G}}$. If $\alpha\in \operatorname{Irr}({\mathbb{G}})$, we will denote by $u^{\alpha}$ a representative of the class $\alpha$ and by $H_{\alpha}$ the finite dimensional Hilbert space on which $u^{\alpha}$ acts. There are isomorphisms $$C_{0}({\widehat}{{\mathbb{G}}}) = \bigoplus_{\alpha\in \operatorname{Irr}({\mathbb{G}})}{\mathcal{B}}(H_{\alpha})\text{ and }\ell^{\infty}({\widehat}{{\mathbb{G}}}) = \prod_{\alpha\in \operatorname{Irr}({\mathbb{G}})}{\mathcal{B}}(H_{\alpha}).$$ The minimal central projection in $\ell^{\infty}({\widehat}{{\mathbb{G}}})$ corresponding to the identity of ${\mathcal{B}}(H_{\alpha})$ will be denoted $p_{\alpha}$. We now proceed to describe explicitely the GNS representation of the Haar state using the irreducible representations. For any $\alpha\in \operatorname{Irr}({\mathbb{G}})$, there is a unique (up to unitary equivalence) irreducible representation, called the *contragredient representation* of $\alpha$ and denoted $\overline{\alpha}$, such that $\operatorname{Mor}(\varepsilon, \alpha\otimes \overline{\alpha})\neq \{0\} \neq \operatorname{Mor}(\varepsilon, \overline{\alpha}\otimes \alpha)$, $\varepsilon$ denoting the trivial representation. This yields an antilinear isomorphism $$j_{\alpha} : H_{\alpha}\rightarrow H_{\overline{\alpha}}.$$ The matrix $j_{\alpha}^{*}j_{\alpha}\in {\mathcal{B}}(H_{\alpha})$ is unique up to multiplication by a real number. We will say that $j_{\alpha}$ is *normalized* if $\operatorname{Tr}(j_{\alpha}^{*}j_{\alpha}) = \operatorname{Tr}((j_{\alpha}^{*}j_{\alpha})^{-1})$ (this only determines $j_{\alpha}$ up to some complex number of modulus one, but this is of no consequence in our context). In that case we will set $Q_{\alpha} = j_{\alpha}^{*}j_{\alpha}$, $\dim_{q}(u^{\alpha}) = \operatorname{Tr}(Q_{\alpha}) = \operatorname{Tr}(Q_{\alpha}^{-1})$ and $t_{\alpha}(1) = \sum j_{\alpha}(e_{i})\otimes e_{i}$, where $(e_{i})$ is some fixed orthonormal basis of $H_{\alpha}$. We will also set $u^{\alpha}_{i, j} = ({\imath}\otimes e_{i}^{*})u^{\alpha}({\imath}\otimes e_{j})$. Note that by construction, $t_{\alpha} : {\mathbb{C}}\rightarrow H_{\overline{\alpha}}\otimes H_{\alpha}$ is a morphism of representations. Let us define a map $$\psi_{\alpha} : \left\{\begin{array}{ccc} H_{\overline{\alpha}}\otimes H_{\alpha} & \rightarrow & C_{\text{red}}({\mathbb{G}}) \\ \eta\otimes \xi & \mapsto & \pi_{h}[(1 \otimes j_{\overline{\alpha}}(\eta)^{*})u^{\alpha}(1 \otimes\xi)] \end{array}\right.$$ According to [@woronowicz1995compact Eq. 6.8] we have, for any $z, z'\in H_{\overline{\alpha}}\otimes H_{\alpha}$, $$h(\psi_{\alpha}(z')^{*}\psi_{\alpha}(z)) = \frac{1}{\dim_{q}(\alpha)}\langle z', z\rangle.$$ and $\Psi = \oplus_{\alpha}\sqrt{\dim_{q}(\alpha)}\psi_{\alpha}.\xi_{h} : \oplus_{\alpha} (H_{\overline{\alpha}}\otimes H_{\alpha}) \rightarrow L^{2}({\mathbb{G}})$ is an isometric isomorphism of Hilbert spaces. If we let $E_{i, j}$ denote the operator on $H_{\alpha}$ sending $e_{i}$ to $e_{j}$ and the other vectors of the basis to $0$, we can define another map $$\Phi_{\alpha} : \left\{\begin{array}{ccc} H_{\overline{\alpha}}\otimes H_{\alpha} & \longrightarrow & {\mathcal{B}}(H_{\alpha}) \\ j_{\overline{\alpha}}(e_{i})\otimes e_{j} & \mapsto & E_{i, j} \end{array}\right.$$ Now, we observe that $\Theta_{\alpha} = \psi_{\alpha}\circ\Phi_{\alpha}^{-1} : {\mathcal{B}}(H_{\alpha})\rightarrow C_{\text{red}}({\mathbb{G}})$ sends $E_{i, j}$ to $\pi_{h}(u^{\alpha}_{i, j})$ and that $$\begin{aligned} h(\Theta_{\alpha}(E_{i, j})^{*}\Theta_{\alpha}(E_{k, l})) & = & \frac{1}{\dim_{q}(\alpha)}\langle (\Phi_{\alpha}^{-1}(E_{i, j}), \Phi_{\alpha}^{-1}(E_{k, l})\rangle \\ & = & \frac{1}{\dim_{q}(\alpha)}\langle j_{\alpha}(e_{i})\otimes e_{j}, j_{\alpha}(e_{k})\otimes e_{l}\rangle \\ & = & \frac{\delta_{j, l}}{\dim_{q}(\alpha)}\langle Q_{\alpha}e_{i}, e_{k}\rangle \\ & = & \frac{1}{\dim_{q}(\alpha)}\operatorname{Tr}(Q_{\alpha}E_{i, j}^{*}E_{k, l}).\end{aligned}$$ Thus, if we endow ${\mathcal{B}}(H_{\alpha})$ with the scalar product $\langle A, B\rangle_{\alpha} = \dim_{q}(\alpha)^{-1}\operatorname{Tr}(Q_{\alpha}A^{*}B)$, we get an isometric isomorphism of Hilbert spaces $$\Theta = \oplus_{\alpha} \Theta_{\alpha}.\xi_{h} : \oplus_{\alpha} {\mathcal{B}}(H_{\alpha}) \rightarrow L^{2}({\mathbb{G}}).$$ Note that the duality map $S_{\alpha} : A \mapsto \langle A, .\rangle_{\alpha}$ being bijective on the finite dimensional space ${\mathcal{B}}(H_{\alpha})$, one can endow $\oplus_{\alpha} {\mathcal{B}}(H_{\alpha})_{*}$ with a Hilbert space structure making it isomorphic to $L^{2}({\mathbb{G}})$ via $\Theta\circ (\oplus_{\alpha}S_{\alpha}^{-1})$. This isomorphism is “natural” since it sends $\omega\in {\mathcal{B}}(H_{\alpha})_{*}$ to $\pi_{h}[({\imath}\otimes\omega)(u^{\alpha})].\xi_{h}$. Let $u^{\alpha}$ and $u^{\beta}$ be two irreducible representations of ${\mathbb{G}}$ and assume, for the sake of simplicity, that *every irreducible subrepresentation of $u^{\alpha}\otimes u^{\beta}$ appears with multiplicity one*. This is no restriction in the case of free quantum groups that we will be considering (see Theorem \[thm:freefusion\]). Let $v_{\gamma}^{\alpha, \beta} : H_{\gamma}\rightarrow H_{\alpha}\otimes H_{\beta}$ be an isometric intertwiner. Note that $v_{\gamma}^{\alpha, \beta}Q_{\gamma} = (Q_{\alpha}\otimes Q_{\beta}) v_{\gamma}^{\alpha, \beta}$. We have, $$\begin{aligned} ({\imath}\otimes \omega)(u^{\alpha})({\imath}\otimes \omega')(u^{\beta}) & = & ({\imath}\otimes \omega\otimes \omega')(u^{\alpha}_{12}u^{\beta}_{13}) \\ & = & ({\imath}\otimes \omega\otimes \omega')(u^{\alpha}\otimes u^{\beta}) \\ & = & ({\imath}\otimes \omega\otimes \omega')(\sum_{\gamma\subset \alpha\otimes \beta}({\imath}\otimes v_{\gamma}^{\alpha, \beta})u^{\gamma}({\imath}\otimes v_{\gamma}^{\alpha, \beta})^{*}) \\ & = & \sum_{\gamma\subset \alpha\otimes \beta}({\imath}\otimes[\omega\otimes \omega']^{\gamma})(u^{\gamma})\end{aligned}$$ where $\omega^{\gamma}(x) = \omega(v_{\gamma}^{\alpha, \beta}\circ x\circ(v_{\gamma}^{\alpha, \beta})^{*})$ for $\omega\in {\mathcal{B}}(H_{\alpha}\otimes H_{\beta})_{*}$. Using the duality map $S_{\alpha}^{-1}$, we can write the map induced on $C_{\text{red}}({\mathbb{G}})$ by the product under our identification : for $A\in {\mathcal{B}}(H_{\alpha})$ and $B\in {\mathcal{B}}(H_{\beta})$, $$\Theta_{\alpha}(A).\Theta_{\beta}(B) = \sum_{\gamma\subset \alpha\otimes \beta} \Theta_{\gamma}((v_{\gamma}^{\alpha, \beta})^{*}(A\otimes B)v_{\gamma}^{\alpha, \beta}).$$ We can now give an explicit formula for the GNS representation $\pi_{h}$. Let $x$ be a coefficient of $u^{\alpha}$ and let $\xi\in p_{\beta}L^{2}({\mathbb{G}})\simeq {\mathcal{B}}(H_{\beta})$. Identify $x$ with $\pi_{h}(x)\xi_{h}$, which is an element of $p_{\alpha}L^{2}({\mathbb{G}})\simeq {\mathcal{B}}(H_{\alpha})$. Making the identification by $\Theta$ implicit, we have $$\label{eq:gns} \pi_{h}(x)\xi = \sum_{\gamma\subset\alpha\otimes \beta} (v_{\gamma}^{\alpha, \beta})^{*}(x\otimes \xi)v_{\gamma}^{\alpha, \beta} = \sum_{\gamma\subset\alpha\otimes \beta} \operatorname{Ad}(v_{\gamma}^{\alpha, \beta})(x\otimes \xi).$$ Free quantum groups ------------------- We will be concerned in the sequel with the free unitary and free orthogonal quantum groups. They were first defined by A. Van Daele and S. Wang in [@van1996universal; @wang1995free] and the definition was later slightly modified by T. Banica in [@banica1996theorie]. This section is devoted to briefly recalling the definition and main properties of these free quantum groups. If $A$ is a C\*-algebra and if $u = (u_{i, j})$ is a matrix with coefficients in $A$, we set $\overline{u} = (u_{i, j}^{*})$. \[de:freeqgroups\] Let $N\in {\mathbb{N}}$ and let $F\in GL_{N}({\mathbb{C}})$ be such that $F\overline{F} \in {\mathbb{R}}.\operatorname{Id}$. We denote by $A_{u}(F)$ the universal unital C\*-algebra generated by $N^{2}$ elements $(u_{i, j})$ such that the matrices $u=(u_{i, j})$ and $F\overline{u}F^{-1}$ are *unitary*. Similarly, we denote by $A_{o}(F)$ the universal unital C\*-algebra generated by $N^{2}$ elements $(v_{i, j})$ such that the matrix $v=(v_{i, j})$ is *unitary* and $v=F\overline{v}F^{-1}$. One can easily check that there is a unique coproduct ${\Delta}_{u}$ (resp. ${\Delta}_{o}$) on $A_{u}(F)$ (resp. $A_{o}(F)$) such that for all $i, j$, $$\begin{aligned} {\Delta}_{u}(u_{i, j}) = \sum_{i, j = 0}^{N}u_{i, k}\otimes u_{k, j} \\ {\Delta}_{o}(v_{i, j}) = \sum_{i, j = 0}^{N}v_{i, k}\otimes v_{k, j}\end{aligned}$$ A pair $(A_{u}(F), {\Delta}_{u})$ is called a *free unitary quantum group* and will be denoted $U^{+}(F)$. A pair $(A_{o}(F), {\Delta}_{o})$ is called a *free orthogonal quantum group* and will be denoted $O^{+}(F)$. Their discrete duals will be denoted respectively ${\mathbb{F}}U^{+}(F)$ and ${\mathbb{F}}O^{+}(F)$. The restriction on the matrix $F$ in the definition is equivalent to requiring the fundamental representation $v$ of $O^{+}(F)$ to be irreducible. That assumption is necessary in order to get a nice description of the representation theory of $O^{+}(F)$. Any *compact matrix pseudogroup* in the sense of [@woronowicz1987compact Def. 1.1] is a compact quantum subgroup of a free unitary quantum group. Moreover, if its fundamental corepresentation is equivalent to its contragredient, then it is a compact quantum subgroup of a free orthogonal quantum group. In this sense, we can see $U^{+}(F)$ and $O^{+}(F)$ as quantum generalizations of the usual unitary and orthogonal groups. The representation theory of free orthogonal quantum groups was computed by T. Banica in [@banica1996theorie]. \[thm:freefusion\] The equivalence classes of irreducible representations of $O^{+}(F)$ are indexed by the set ${\mathbb{N}}$ of integers ($u^{0}$ being the trivial representation and $u^{1}=u$ the fundamental one), each one is isomorphic to its contragredient and the tensor product is given (inductively) by $$u^{1}\otimes u^{n} = u^{n+1}\oplus u^{n-1}.$$ Moreover, if $N=2$, then $\dim_{q}(u^{n})=n+1$. Otherwise, $$\dim_{q}(u^{n}) = \frac{q^{n+1} - q^{-n-1}}{q - q^{-1}},$$ where $q + q^{-1}= \operatorname{Tr}(Q_{1})$ and $0\leqslant q\leqslant 1$. We will use the shorthand notation $D_{n}$ for $\dim_{q}(u^{n})$ in the sequel. The following inequality always holds : $q+q^{-1} \geqslant N$. The representation theory of $U^{+}(F)$ was also explicitely computed by T. Banica in [@banica1997groupe]. However, we will only need the following result [@banica1997groupe Thm. 1] (see [@wang1995free] for the definition of the free product of discrete quantum groups). \[thm:freesubgroup\] The discrete quantum group ${\mathbb{F}}U^{+}(F)$ is a quantum subgroup of ${\mathbb{Z}}\ast {\mathbb{F}}O^{+}(F)$. The following lemma summarizes some standard calculations which will be used several times in the sequel. \[lem:quantumdimension\] Let $a>b$ be integers, then $D_{a-b}^{-1}\leqslant D_{b}/D_{a}\leqslant q^{a-b}$. Moreover, for any integer $c$, $q^{c}D_{c}\leqslant (1-q^{2})^{-1}$. Let $n\in {\mathbb{Z}}$ such that $n\geqslant -b$. Decomposing $u^{b+n}\otimes u^{a+n+1}$ and $u^{b+n+1}\otimes u^{a+n}$ into sums of irreducible representations yields $$D_{b+n}D_{a+n+1} = D_{a-b+1} + \dots + D_{a+b+2n+1} \leqslant D_{a-b-1} + \dots + D_{a+b+2n+1} = D_{b+n+1}D_{a+n}$$ This inequality means that the sequence $(D_{b+n}/D_{a+n})_{n\geqslant -b}$ is increasing, thus any term is greater than its first term $D_{a-b}^{-1}$ and less than its limit $q^{a-b}$. The second part of the lemma is obvious since $q^{c}D_{c} = (1-q^{2c+2})/(1-q^{2})$. Weak amenability for discrete quantum groups {#sec:quantumwa} ============================================ We now give some definitions and properties concerning weak amenability for discrete quantum groups. It is based on the notion of multipliers associated to bounded functions. Let ${\widehat}{{\mathbb{G}}}$ be a discrete quantum group and $a\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$. The *left multiplier* associated to $a$ is the map $m_{a} : \operatorname{Pol}({\mathbb{G}}) \rightarrow \operatorname{Pol}({\mathbb{G}})$ defined by $$(m_{a}\otimes {\imath})(u^{\alpha}) = (1\otimes ap_{\alpha})u^{\alpha},$$ for any irreducible representation $\alpha$ of ${\mathbb{G}}$. This definition relies on the identification of $\ell^{\infty}({\widehat}{{\mathbb{G}}})$ with $\prod {\mathcal{B}}(H_{\alpha})$ which is specific to the case of discrete quantum groups. However, since $W$ reads as $\prod u^{\alpha}$ in this identification, we can equivalently define the multiplier $m_{a}$ in the following way : for any $\omega\in {\mathcal{B}}(L^{2}({\mathbb{G}}))_{*}$, $m_{a}(({\imath}\otimes \omega)(W)) = ({\imath}\otimes \omega)((1\otimes a)W)$. This definition makes sense in a more general setting and corresponds to the definition of J. Kraus and Z.J. Ruan in [@kraus1999approximation] for Kac algebras and to the definition of M. Junge, M. Neufang and Z.J. Ruan in [@junge2009representation] (see also [@daws2011multipliers]) for locally compact quantum groups. A net $(a_{i})$ of elements of $\ell^{\infty}({\widehat}{{\mathbb{G}}})$ is said to *converge pointwise* to $a\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$ if $a_{i}p_{\alpha} \rightarrow ap_{\alpha}$ for any irreducible representation $\alpha$ of ${\mathbb{G}}$. An element $a\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$ is said to have *finite support* if $ap_{\alpha}$ is non-zero only for a finite number of irreducible representations $\alpha$. The key point to get a suitable definition of weak amenability is to have an intrinsic characterization of the completely bounded norm of a multiplier. Such a characterization is given by the following theorem [@daws2011multipliers Prop 4.1 and Thm 4.2]. \[theorem:quantumgilbert\] Let ${\widehat}{{\mathbb{G}}}$ be a discrete quantum group and $a\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$. Then $m_{a}$ extends to a competely bounded multiplier on ${\mathcal{B}}(L^{2}({\mathbb{G}}))$ if and only if there exists a Hilbert space $K$ and two maps $\xi, \eta \in {\mathcal{B}}(L^{2}({\mathbb{G}}), L^{2}({\mathbb{G}})\otimes K)$ such that $\|\xi\|\|\eta\| = \|m_{a}\|_{cb}$ and $$\label{eq:quantumgilbert} (1\otimes \eta)^{*}{\widehat}{W}_{12}^{*}(1\otimes \xi){\widehat}{W} = a\otimes 1.$$ Moreover, we then have $m_{a}(x) = \eta^{*}(x\otimes 1)\xi$. Notice that thanks to this theorem, the completely bounded norm of $m_{a}$ is the same when it is extended to $C_{\text{red}}({\mathbb{G}})$, $L^{\infty}({\mathbb{G}})$ or ${\mathcal{B}}(L^{2}({\mathbb{G}}))$. Denoting by $\|m_{a}\|_{cb}$ this norm, we can give a definition of weak amenability. \[de:quantumwa\] A discrete quantum group ${\widehat}{{\mathbb{G}}}$ is said to be *weakly amenable* if there exists a net $(a_{i})$ of elements of $\ell^{\infty}({\widehat}{{\mathbb{G}}})$ such that - $a_{i}$ has finite support for all $i$. - $(a_{i})$ converges pointwise to $1$. - $K:=\limsup_{i} \|m_{a_{i}}\|_{cb}$ is finite. The lower bound of the constants $K$ for all nets satisfying these properties is denoted $\Lambda_{cb}({\widehat}{{\mathbb{G}}})$ and called the *Cowling-Haagerup constant* of ${\widehat}{{\mathbb{G}}}$. By convention, $\Lambda_{cb}({\widehat}{{\mathbb{G}}})=\infty$ if ${\widehat}{{\mathbb{G}}}$ is not weakly amenable. It is clear on the definition that a discrete group $G$ is weakly amenable in the classical sense (see e.g. [@brown2008finite Def. 12.3.1]) if and only if the associated discrete quantum group is weakly amenable (and the constants are the same). We recall the following notions of weak amenability for operator algebras. A C\*-algebra $A$ is said to be *weakly amenable* if there exists a net $(T_{i})$ of linear maps from $A$ to itself such that - $T_{i}$ has finite rank for all $i$. - $\|T_{i}(x)-x\|\rightarrow 0$ for all $x\in A$. - $K :=\limsup_{i}\|T_{i}\|_{cb}$ is finite. The lower bound of the constants $K$ for all nets satisfying these properties is denoted $\Lambda_{cb}(A)$ and called the *Cowling-Haagerup constant* of $A$. By convention, $\Lambda_{cb}(A) = \infty$ if the C\*-algebra $A$ is not weakly amenable. A von Neumann algebra $N$ is said to be *weakly amenable* if there exists a net $(T_{i})$ of normal linear maps from $N$ to itself such that - $T_{i}$ has finite rank for all $i$. - $(T_{i}(x)-x) \rightarrow 0$ ultraweakly for all $x\in N$. - $K :=\limsup_{i}\|T_{i}\|_{cb}$ is finite. The lower bound of the constants $K$ for all nets satisfying these properties is denoted $\Lambda_{cb}(N)$ and called the *Cowling-Haagerup constant* of $N$. By convention, $\Lambda_{cb}(N) = \infty$ if the von Neumann algebra $N$ is not weakly amenable. See [@kraus1999approximation Thm. 5.14] for a proof of the following result. \[thm:quantumwa\] Let ${\widehat}{{\mathbb{G}}}$ be a *unimodular* (i.e. the Haar state on ${\mathbb{G}}$ is tracial) discrete quantum group, then $$\Lambda_{cb}({\widehat}{{\mathbb{G}}}) = \Lambda_{cb}(C_{\text{red}}({\mathbb{G}})) = \Lambda_{cb}(L^{\infty}({\mathbb{G}})).$$ Block decomposition and Haagerup inequality {#subsec:blocks} =========================================== Our aim is to prove a polynomial bound for the completely bounded norm of the projection on the linear span of coefficients of an irreducible representation $u^{d}$ in $C_{\text{red}}(O^{+}(F))$. Let us give some motivation for this. First note that this projection is simply the multiplier $m_{p_{d}}$ associated to $p_{d}\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$. If we choose for every integer $k$ and real number $t$ a scalar coefficient $b_{k}(t)$, we can define a net of (radial) elements $$a_{i}(t) = \sum_{k=0}^{i}b_{k}(t)p_{k}\in \ell^{\infty}({\widehat}{{\mathbb{G}}}).$$ If the $b_{k}(t)$ have sufficiently nice properties and if the completely bounded norm of the operators $m_{p_{d}}$ can be controlled, the net $(a_{i}(t))$ will satisfy all the hypothesis in Definition \[de:quantumwa\] and ${\mathbb{F}}O^{+}(F)$ will be weakly amenable. Our strategy to obtain the polynomial bound is inspired from the proof of U. Haagerup’s estimate for the completely bounded norm of projections on words of fixed length in free groups. The original proof is unpublished but the argument is detailed in G. Pisier’s book [@pisier2003introduction]. Following the scheme of the proof, we will first, in this section, prove an operator-valued analogue of the Haagerup inequality. From now on, we fix an integer $N > 2$ and a matrix $F\in GL_{N}({\mathbb{C}})$ satisfying $F\overline{F} \in {\mathbb{R}}.\operatorname{Id}$. We will write ${\mathcal{H}}$ for the Hilbert space $L^{2}(O^{+}(F))$ which is identified to $\oplus_{k} {\mathcal{B}}(H_{k})$ as explained in Subsection \[subsec:gns\] ($H_{k}$ being the carrier Hilbert space of the $k$-th irreducible representation). Let $H$ be a fixed Hilbert space and let $X\in {\mathcal{B}}(H)\odot \operatorname{Pol}(O^{+}(F))$ (it is enough to control the norm on this dense subalgebra), chose $d\in {\mathbb{N}}$ and set $X^{d} = ({\imath}\otimes m_{p_{d}})(X)$. These objects should be thought of as “operator-valued functions with finite support” on ${\mathbb{F}}O^{+}(F)$. Our aim is to control the norm of $X^{d}$ using the norm of $X$. Recall from [@vergnioux2007property] that there is a natural length function on $\operatorname{Irr}(O^{+}(F))$ such that the irreducible representation $u_{d}$ has length $d$. Using this notion, one could give a rigorous definition of “operator-valued functions with support in the words of length $d$”. This, however, will not be needed here. We start by decomposing the operators into more elementary ones. For any two integers $a$ and $b$, we set $$\begin{aligned} B_{a, b}(X) & := & ({\imath}\otimes p_{a})X({\imath}\otimes p_{b}) \\ B_{a, b}(X^{d}) & := & ({\imath}\otimes p_{a})X^{d}({\imath}\otimes p_{b})\end{aligned}$$ This is simply $X$ (resp. $X^{d}$) seen as an operator from ${\mathcal{B}}(H_{b})$ to ${\mathcal{B}}(H_{a})$ and obviously has norm less than $\|X\|$ (resp. $\|X^{d}\|$). We will call it a *block*. The operator $X^{d}$ admits a particular decomposition with respect to these blocks. \[lem:blockdecomposition\] Set $X^{d}_{j} = \displaystyle\sum_{k=0}^{+\infty}B_{d-j+k, j+k}(X^{d})$, then $X^{d} = \displaystyle\sum_{j=0}^{d}X^{d}_{j}$. Clearly, $X^{d} = \sum_{a, b}B_{a, b}(X^{d})$. If we decompose $X^{d}$ as $\sum_{i} T_{i}\otimes x_{i}$, with $x_{i}$ a coefficient of $u^{d}$ and $T_{i}\in {\mathcal{B}}(H)$, we see that $X^{d}$ sends $H\otimes (p_{b}{\mathcal{H}})$ into $\oplus_{c} (H\otimes (p_{c}{\mathcal{H}}))$ where the sum runs over all irreducible subrepresentations $c$ of $d\otimes b$. Thus, we can deduce from Theorem \[thm:freefusion\] that $B_{a, b}(X^{d})$ vanishes as soon as $a$ is not of the form $d+b-2j$ for some $0\leqslant j\leqslant \min(d, b)$. Consequently, $$X^{d} = \sum_{b=0}^{+\infty}\sum_{j=0}^{\min(d, b)}B_{d+b - 2j, b}(X^{d}) = \sum_{j=0}^{d}\sum_{b=j}^{+\infty}B_{d+b - 2j, b}(X^{d}).$$ Writing $b = k+j$, we get the desired result. This should be thought of as a decomposition according to the “number of deleted letters” in the action of $X^{d}$. Thanks to the triangle inequality, we can restrict ourselves to the study of $\|X^{d}_{j}\|$. Proposition \[prop:blocks\] further reduces the problem to the study of only one specific block in $X^{d}_{j}$. Before stating and proving it, we have to introduce several notations and elementary facts. Recall from Subsection \[subsec:gns\] that for $\gamma\subset \alpha\otimes \beta$, $v_{\gamma}^{\alpha, \beta} : H_{\gamma} \mapsto H_{\alpha}\otimes H_{\beta}$ denotes an isometric intertwiner and let $M_{k}^{+} : {\mathcal{H}}\otimes {\mathcal{B}}(H_{k}) \rightarrow {\mathcal{H}}$ be the orthogonal sum of the operators $\operatorname{Ad}(v_{l+k}^{l, k})$. Under our identification of ${\mathcal{H}}$ with $\oplus {\mathcal{B}}(H_{k})$, the restriction of $M_{k}^{+}$ to ${\mathcal{B}}(H_{l})\otimes {\mathcal{B}}(H_{k})$ is just the map induced by the product composed with the orthogonal projection onto ${\mathcal{B}}(H_{l+k})$. If we endow ${\mathcal{B}}(H_{k})$ with the scalar product $\langle ., .\rangle_{k}$, it can be seen as a subspace of ${\mathcal{H}}$ and we can compute the norm of the restriction of $M_{k}^{+}$ to ${\mathcal{B}}(H_{l})\otimes {\mathcal{B}}(H_{k})$ with respect to the Hilbert structure on ${\mathcal{H}}\otimes {\mathcal{H}}$. Let $x\in {\mathcal{B}}(H_{l})\otimes {\mathcal{B}}(H_{k})$, then $$\begin{aligned} \|M_{k}^{+}(x)\|^{2} & = & \frac{1}{D_{l+k}}\operatorname{Tr}(Q_{k+l}M_{k}^{+}(x)^{*}M_{k}^{+}(x)) \\ & = & \frac{1}{D_{l+k}}\operatorname{Tr}(Q_{k+l}(v^{l, k}_{l+k})^{*}x^{*}v^{l, k}_{l+k}(v^{l, k}_{l+k})^{*}x v^{l, k}_{l+k}) \\ & \leqslant & \frac{1}{D_{l+k}}\operatorname{Tr}(Q_{l+k}(v^{l, k}_{l+k})^{*}x^{*}x v^{l, k}_{l+k}) \\ & = & \frac{1}{D_{l+k}}\operatorname{Tr}(v^{l, k}_{l+k}Q_{l+k}(v^{l, k}_{l+k})^{*}x^{*}x) \\ & \leqslant & \frac{1}{D_{l+k}}(\operatorname{Tr}\otimes \operatorname{Tr})((Q_{l}\otimes Q_{k})x^{*}x) \\ & = & \frac{D_{l}D_{k}}{D_{l+k}}\|x\|^{2}\end{aligned}$$ i.e. $\|M_{k}^{+}(p_{l}\otimes {\imath})\|^{2} = D_{l}D_{k}/D_{l+k}$ (the norm is attained at $x=v^{l, k}_{l+k}(v^{l, k}_{l+k})^{*}$). Note that this computation also proves that $\|M_{k}^{+}\|^{2} = \displaystyle\frac{1-q^{2k+2}}{1-q^{2}}$ and in particular that $\|M_{1}^{+}\|^{2} = 1+q^{2}\leqslant 2$. \[rem:adjoint\] The computation of the adjoint of $M_{k}^{+}$ is similar to the computation of the norm. One has $(M_{k}^{+})^{*}p_{l+k} = (D_{l}D_{k}/D_{l+k})\operatorname{Ad}((v_{l+k}^{l, k})^{*})$. Let us now state and prove the main result of this section. \[prop:blocks\] For integers $a, b$ and $c$, set $$N_{a, b}^{c} = 1 - \frac{D_{(a-b+c)/2}D_{(b-a+c)/2-1}}{D_{a+1}D_{b}}$$ whenever this expression makes sense. Then, if we set $$\chi_{j}^{d}(k) = \sqrt{\frac{D_{d-j}D_{j+k}}{D_{d-j+k}D_{j}}}\prod_{i=0}^{j-1}(N_{d-j+i, k+i}^{d-j+k})^{-1},$$ we have, for all $k$, $\|B_{d-j+k, j+k}(X^{d})\|\leqslant \chi_{j}^{d}(k)\|B_{d-j, j}(X^{d})\|$. Let us first focus on the one-dimensional case. Let $x$ be a coefficient of $u^{d}$ seen as an element of ${\mathcal{B}}(H_{d})$ and choose an integer $k$. Let us compare the two operators $$\begin{aligned} A & = & [M_{k}^{+}(p_{d-j}xp_{j}\otimes {\imath}) (M_{k}^{+})^{*}](\xi) \\ B & = & (p_{d-j+k}xp_{j+k})(\xi)\end{aligned}$$ for $\xi\in p_{j+k}{\mathcal{H}}= {\mathcal{B}}(H_{j+k})$. Setting $V = ({\imath}\otimes v_{j+k}^{j, k})^{*}(v_{d-j}^{d, j}\otimes {\imath})v_{d-j+k}^{d-j, k}$, we have an intertwiner between $u^{d-j+k}$ and $u^{d\otimes (j+k)}$. Since that inclusion has multiplicity one, there is a complex number $\mu_{j}^{d}(k)$ such that $$V = \mu_{j}^{d}(k)v_{d-j+k}^{d, j+k}.$$ Now, using Equation (\[eq:gns\]) and Remark \[rem:adjoint\], we have $$\begin{aligned} B & = & (v_{d-j+k}^{d, j+k})^{*}(x\otimes \xi)v_{d-j+k}^{d, j+k} \\ A & = & V^{*}(x\otimes \xi)\left(\frac{D_{j}D_{k}}{D_{j+k}}\right)V\end{aligned}$$ and consequently $B = \lambda^{d}_{j}(k) A$, with $\lambda_{j}^{d}(k) = (D_{j}D_{k}/D_{j+k})^{-1}\vert \mu_{j}^{d}(k) \vert^{-2}$. Let us compute $\vert\mu_{j}^{d}(k)\vert$. If we set $v_{+}^{a, b}=(v_{a+b}^{a, b})^{*}$ and define two morphisms of representations $$\begin{aligned} \mathcal{T}_{A} & = & (v_{+}^{d-j, j}\otimes v_{+}^{j, 0} \otimes {\imath}_{k})({\imath}_{d-j}\otimes t_{j}\otimes {\imath}_{k}) v_{d-j+k}^{d-j, k} \\ \mathcal{T}_{B} & = & (v_{+}^{d-j, j}\otimes v_{+}^{j, k})({\imath}_{d-j}\otimes t_{j}\otimes {\imath}_{k}) v_{d-j+k}^{d-j, k}\end{aligned}$$ we have, up to some complex numbers of modulus $1$, $$\mathcal{T}_{A} = \|\mathcal{T}_{A}\|(v_{d-j}^{d, j}\otimes {\imath})v_{d-j+k}^{d-j, k}\text{ and }\mathcal{T}_{B} = \|\mathcal{T}_{B}\|v_{d-j+k}^{d, j+k}.$$ Since moreover $({\imath}\otimes v_{j+k}^{j, k})^{*}\mathcal{T}_{A} = \mathcal{T}_{B}$, we get $\vert \mu_{j}^{d}(k)\vert^{2}=\|\mathcal{T}_{B}\|^{2}/\|\mathcal{T}_{A}\|^{2}$. Thanks to [@vergnioux2005orientation Prop. 2.3] and [@vergnioux2007property Lem. 4.8], we can compute the norms of $\mathcal{T}_{A}$ and $\mathcal{T}_{B}$ and obtain $$\vert \mu_{j}^{d}(k) \vert^{2} = \prod_{i=0}^{j-1}\frac{N^{d-j+k}_{d-j+i, k+i}}{N^{d-j}_{d-j+i, i}} = \prod_{i=0}^{j-1}N^{d-j+k}_{d-j+i, k+i}.$$ Note that for $j=0$, the above product is not defined. However, $\lambda_{0}^{d}(k)=1$ since $\mathcal{T}_{A} = \mathcal{T}_{B}$ in that case. As $\lambda_{j}^{d}(k)$ does not depend on $\xi$, we have indeed proved the following equality in ${\mathcal{B}}({\mathcal{H}})$ : $$p_{d-j+k}xp_{j+k} = \lambda_{j}^{d}(k)[M_{k}^{+}(p_{d-j}xp_{j}\otimes {\imath})(M_{k}^{+})^{*}].$$ Now we go back to the operator-valued case. We have $X^{d} = \sum_{i} T_{i}\otimes x_{i}$, where $x_{i}\in \operatorname{Pol}(O^{+}(F))$ is a coefficient of $u^{d}$ and $T_{i}\in {\mathcal{B}}(H)$, hence $$\lambda_{j}^{d}(k) [({\imath}\otimes M_{k}^{+})(B_{d-j, j}(X^{d})\otimes {\imath})({\imath}\otimes M_{k}^{+})^{*}] = B_{d-j+k, j+k}(X^{d}).$$ Using the norms of the restrictions of $M_{k}^{+}$ computed above, we get $$\|B_{d-j+k, j+k}(X^{d})\| \leqslant \lambda_{j}^{d}(k) \|({\imath}\otimes M_{k}^{+})B_{d-j, j}(X^{d})({\imath}\otimes M_{k}^{+})^{*}\| \leqslant \chi_{j}^{d}(k)\|B_{d-j, j}(X^{d})\|.$$ \[cor:blocks\] There is a constant $K(q)$, depending only on $q$, such that for any $d\in {\mathbb{N}}$ and $0\leqslant j \leqslant d$, $\|X_{j}^{d}\|\leqslant K(q)\|B_{d-j, j}(X^{d})\|$. According to Lemma \[lem:quantumdimension\], we have $$\frac{D_{d-j}D_{k-1}}{D_{d-j+i+1}D_{k+i}} \leqslant q^{i+1}q^{i+1} = q^{2i+2},$$ thus $(N_{d-j+i, k+i}^{d-j+k})^{-1}\leqslant (1-q^{2i+2})^{-1}$. Again by Lemma \[lem:quantumdimension\], $D_{d-j}/D_{d-j+k}\leqslant q^{k}$ and $D_{j+k}/D_{j}\leqslant D_{k}$, hence $$\chi_{j}^{d}(k)\leqslant \sqrt{q^{k}D_{k}}\prod_{i=0}^{j-1}\frac{1}{1-q^{2i+2}}\leqslant \frac{1}{\sqrt{1-q^{2}}}\prod_{i=0}^{+\infty}\frac{1}{1-q^{2i+2}} = K(q).$$ As a summary of what has been worked out in this section, we can state an analogue of the Haagerup inequality for “operator-valued functions” on ${\mathbb{F}}O^{+}(F)$. There exists a constant $K(q)$ depending only on $q$ such that $$\max_{0\leqslant j\leqslant d}\{\|B_{d-j, j}(X^{d})\|\} \leqslant \|X^{d}\| \leqslant K(q)(d+1)\max_{0\leqslant j\leqslant d}\{\|B_{d-j, j}(X^{d})\|\}$$ The first inequality comes from the fact that $\|B_{d-j, j}(X^{d})\|\leqslant \|X^{d}\|$ and the second one from the triangle inequality combined with Proposition \[prop:blocks\]. This inequality should be compared to A. Buchholz’s inequality [@buchholz1999norm Thm 2.8] and to [@pisier2003introduction Eq 9.7.5] in the free group case. The completely bounded norm of projections {#subsec:recursion} ========================================== We now want to find some polynomial $P$ such that $\|X^{d}\|\leqslant P(d)\|X\|$. Thanks to Proposition \[prop:blocks\], the problem reduces to finding a polynomial $Q$ such that $\|B_{d-j, j}(X^{d})\|\leqslant Q(d)\|X\|$. This will be done using the following recursion formula. \[prop:recursion\] Set $$N_{1}^{+} = \bigoplus_{l}\frac{D_{l+1}}{D_{1}D_{l}}M_{1}^{+}(p_{l}\otimes {\imath}).$$ According to Remark \[rem:adjoint\], $(N_{1}^{+})^{*}$ is the sum of the operators $\operatorname{Ad}((v_{l+k}^{l, k})^{*})$. There are coefficients $C_{j}^{d}(s)$ such that for $0\leqslant j\leqslant d$, $$\begin{aligned} B_{d-j+1, j+1}(X) & - & ({\imath}\otimes M_{1}^{+})(B_{d-j, j}(X)\otimes {\imath})({\imath}\otimes N_{1}^{+})^{*} \\ = B_{d-j+1, j+1}(X^{d+2}) & + & \sum_{s=0}^{\min(j, d-j)}C_{j}^{d}(s)B_{d-j+1, j+1}(X^{d-2s})\end{aligned}$$ The idea of the proof is similar to the one used in the proof of Proposition \[prop:blocks\]. We first consider the one-dimensional case. Let $x$ be a coefficient of $u^{l}$ seen as an element of ${\mathcal{B}}(H_{l})$. Fix an element $\xi\in p_{j+1}{\mathcal{H}}$. Again, the operators $$\begin{aligned} A & = & [M_{1}^{+}(p_{d-j}xp_{j}\otimes {\imath})(N_{1}^{+})^{*}](\xi) \\ B & = & (p_{d-j+1}x_{l}p_{j+1})(\xi)\end{aligned}$$ are proportional. Note that if $l>d+2$, $l<\vert d-2j\vert$ or $l-d$ is not even, both operators are $0$. Note also that if $l=d+2$, $A = 0$. The other values of $l$ can be written $d-2s$ for some positive integer $s$ between $0$ and $\min(j, d-j)$. In that case, the existence of a scalar $\nu_{j}^{d}(s)$ such that $B = \nu^{d}_{j}(s) A$ follows from the same argument as in the proof of Proposition \[prop:blocks\]. Let us compute $\nu^{d}_{j}(s)$, noticing that thanks to the normalization of $N_{1}^{+}$, the constant $\nu_{j}^{d}(s)$ only corresponds to the “$\mu$-part” of the constant $\lambda$ of Proposition \[prop:blocks\]. This time we have to set $$\begin{aligned} \mathcal{T}_{A} & = & (v_{+}^{d-s-j, j-s}\otimes v_{+}^{j-s, s}\otimes {\imath}_{1})({\imath}_{d-j-s}\otimes t_{j-s}\otimes {\imath}_{s+1}) v_{d-j+1}^{d-j-s, s+1} \\ \mathcal{T}_{B} & = & (v_{+}^{d-s-j, j-s}\otimes v_{+}^{j-s, s+1})({\imath}_{d-j-s}\otimes t_{j-s}\otimes {\imath}_{s+1}) v_{d-j+1}^{d-j-s, s+1}\end{aligned}$$ Again, applying [@vergnioux2005orientation Prop. 2.3] and [@vergnioux2007property Lem. 4.8] yields $$\nu^{d}_{j}(s) = \frac{\|\mathcal{T}_{A}\|^{2}}{\|\mathcal{T}_{B}\|^{2}} = \prod_{i=0}^{j-s-1}\frac{N^{d-j}_{d-j-s+i, s+i}}{N^{d-j+1}_{d-j-s+i, s+i+1}}.$$ Like in the proof of Proposition \[prop:blocks\], we can now go back to the operator-valued case. We have $$X = \sum_{l}\sum_{i=0}^{k(l)} T_{l}^{(i)}\otimes x_{l}^{(i)}$$ where $x_{l}^{(i)}\in \operatorname{Pol}(O^{+}(F))$ are coefficients of $u^{l}$ and $T_{l}^{(i)}\in {\mathcal{B}}(H)$. Setting $$X^{l} = \sum_{i=0}^{k(l)}T_{l}^{(i)}\otimes x_{l}^{(i)},$$ we have $$B_{d-j+1, j+1}(X^{l}) = \nu^{d}_{j}(s) ({\imath}\otimes M_{1}^{+})(B_{d-j, j}(X^{l})\otimes {\imath})({\imath}\otimes N_{1}^{+})^{*}$$ and setting $C_{j}^{d}(s) = 1 - \nu_{j}^{d}(s)^{-1}$ concludes the proof. The last result we need is a control on the coefficients $C_{j}^{d}(s)$ and $\chi^{d}_{j}(s)$. \[lem:constantsum\] For any $0\leqslant j\leqslant d$, $\displaystyle\sum_{s=0}^{\min(j, d-j)}\vert C_{j}^{d}(s)\vert\chi^{d-2s}_{j-s}(s+1) \leqslant 1$. We first give another expression of $\vert C_{j}^{d}(s)\vert$. Decomposing into sums of irreducible representations yields $$\begin{aligned} D_{d-s-j+i+1}D_{s+i+1} - D_{d-s-j}D_{s} = D_{d-j+2} + \dots + D_{d-j+2i+2} = D_{i}D_{d-j+i+2} \\ D_{d-s-j+i+1}D_{s+i} - D_{d-s-j}D_{s-1} = D_{d-j+1} + \dots + D_{d-j+2i+1} = D_{i}D_{d-j+i+1}\end{aligned}$$ which implies that $$N^{d-j+1}_{d-j-s+i, s+i+1} = \frac{D_{i}D_{d-j+i+2}}{D_{d-s-j+i+1}D_{s+i+1}}\text{ and }N^{d-j}_{d-s-j+i, s+i} = \frac{D_{i}D_{d-j+i+1}}{D_{d-s-j+i+1}D_{s+i}}.$$ Hence $$\nu_{j}^{d}(s) = \prod_{i=0}^{j-s-1}\frac{N^{d-j}_{d-s-j+i, s+i}}{N^{d-j+1}_{d-j-s+i, s+i+1}} = \prod_{i=0}^{j-s-1}\frac{D_{d-j+i+1}D_{s+i+1}}{D_{s+i}D_{d-j+i+2}} = \frac{D_{j}D_{d-j+1}}{D_{s}D_{d-s+1}}.$$ Again, noticing that $D_{j}D_{d-j+1} - D_{s}D_{d-s+1} = D_{d-j-s}D_{j-s-1}$ yields $$\vert C_{j}^{d}(s)\vert = \vert 1 - \nu_{j}^{d}(s)^{-1}\vert = \frac{D_{d-j-s}D_{j-s-1}}{D_{d-j+1}D_{j}}.$$ According to Lemma \[lem:quantumdimension\], we thus have $$\vert C_{j}^{d}(s)\vert \leqslant q^{s+1}q^{s+1} = q^{2s+2}$$ Now we turn to $\chi_{j-s}^{d-2s}(s+1)$. In fact, we are going to bound $\chi_{j}^{d}(s+1)$ independantly of $d$ and $j$. Decomposing into sums of irreducible representations, we get $$D_{d-j+i+1}D_{k+i} - D_{d-j}D_{k-1} = D_{d-j+k+1} + \dots + D_{d-j+k+2i+1} = D_{i}D_{d-j+k+i+1},$$ which implies that $N_{d-j+i, k+i}^{d-j+k} = D_{i}D_{d-j+k+i+1}/D_{d-j+i+1}D_{k+i}$. Now we can compute $$\begin{aligned} \frac{\chi_{j}^{d}(s+1)}{\chi_{j}^{d}(s)} & = & \sqrt{\frac{D_{j+s+1}D_{d-j+s}}{D_{j+s}D_{d-j+s+1}}}\prod_{i=0}^{j-1}\frac{D_{s+1+i}D_{d-j+s+i+1}}{D_{s+i}D_{d-j+s+i+2}} \\ & = & \sqrt{\frac{D_{j+s+1}D_{d-j+s}}{D_{j+s}D_{d-j+s+1}}}\frac{D_{j+s}D_{d-j+s+1}}{D_{s}D_{d+s+1}} \\ & = & \frac{\sqrt{D_{j+s}D_{d-j+s+1}D_{d-j+s}D_{j+s+1}}}{D_{s}D_{d+s+1}}.\end{aligned}$$ Using Lemma \[lem:quantumdimension\] again, we get $$\frac{\chi_{j}^{d}(s+1)}{\chi_{j}^{d}(s)} \leqslant \sqrt{q^{j}D_{j}q^{d-j}D_{d-j}} \leqslant \frac{1}{1-q^{2}}.$$ Since $\chi_{j}^{d}(1) \leqslant (1-q^{2})^{-1}$, we have proved that $\chi_{j}^{d}(s+1)\leqslant (1-q^{2})^{-s-1}$. This bound is independant of $d$ and $j$, thus it also works for $\chi_{j-s}^{d-2s}(s+1)$. Combining this with our previous estimate we can compute $$\sum_{s=0}^{\min(j, d-j)}\vert C_{j}^{d}(s)\vert\chi_{j-s}^{d-2s}(s+1) \leqslant \sum_{s=0}^{+\infty}\left(\frac{q^{2}}{1-q^{2}}\right)^{s+1} = \frac{q^{2}}{1-2q^{2}}.$$ The last term is less than $1$ as soon as $q\leqslant 1/\sqrt{3}$, hence in particular for any $q$ such that $q+q^{-1}\geqslant 3$. Gathering all our results will now give the estimate we need. To make things more clear, we will proceed in two steps. First we bound the norms of the blocks of $X^{d}$. \[prop:polynomialblocks\] There exists a polynomial $Q$ such that for any integer $d$ and $0\leqslant j\leqslant d$, $\|B_{d-j, j}(X^{d})\| \leqslant Q(d)\|X\|$. First note that $B_{d, 0}(X^{d}) = B_{d, 0}(X)$ and $B_{0, d}(X^{d}) = B_{0, d}(X)$, hence we only have to consider the case $1\leqslant j \leqslant d-1$. Moreover, applying the triangle inequality to the recursion relation of Proposition \[prop:recursion\] yields $$\begin{aligned} \|B_{d-j+1, j+1}(X^{d+2})\| & \leqslant & (1+\|M_{1}^{+}\|\|N_{1}^{+}\|)\|X\| \\ & + & \sum_{s=0}^{\min(j, d-j)}\vert C_{j}^{d}(s)\vert \|B_{d-j+1, j+1}(X^{d-2s})\|.\end{aligned}$$ We proceed by induction, with the following induction hypothesis $H(d)$ : “For any integer $l \leqslant d$ and any $0\leqslant j\leqslant l$, $\|B_{l-j, j}(X^{l})\| \leqslant Q(l)\|X\|$ with $Q(X) = 2X+1$.” Because of the remark at the beginning of the proof, $H(0)$ and $H(1)$ are true. Knowing this, we just have to prove that for any $d$, $H(d)$ implies the inequality for $d+2$. Indeed, this will prove that assuming $H(d)$, both the inequalities for $d+1$ (noticing that $H(d)$ implies $H(d-1)$) and $d+2$ are true, hence $H(d+2)$ will hold. Assume $H(d)$ to be true for some $d$ and apply the recursion formula above. The blocks in the right-hand side of the inequality are of the form $B_{d-j+1, j+1}(X^{d-2s})$. By Proposition \[prop:blocks\] and $H(d)$, $$\begin{aligned} \|B_{d-j+1, j+1}(X^{d-2s})\| & = & \|B_{(d-2s)-(j-s)+s+1, (j-s)+s+1}(X^{d-2s})\| \\ & \leqslant & \chi_{j-s}^{d-2s}(s+1)\|B_{(d-2s) - (j-s), (j-s)}(X^{d-2s})\| \\ & \leqslant & \chi_{j-s}^{d-2s}(s+1)Q(d-2s)\|X\|.\end{aligned}$$ Then, bounding $Q(d-2s)$ by $Q(d)$ and using Lemma \[lem:constantsum\] yields $$\|B_{d-j+1, j+1}(X^{d+2})\| \leqslant 3\|X\| + Q(d)\|X\| \leqslant Q(d+2)\|X\|.$$ Since $\|B_{d-j+1, j+1}(X^{d+2})\| = \|B_{(d+2)-(j+1), j+1}(X^{d+2})\|$, the inequality is proved for $1\leqslant j+1\leqslant d+1$. In other words, we have $\|B_{d-J, J}(X^{d+2})\| \leqslant Q(d+2)\|X\|$ for any $1\leqslant J\leqslant d+1$. As noted at the beginning of the proof, this is enough to get $H(d+2)$. Secondly we bound the norm of $X^{d}$ itself. \[thm:completelyboundedprojection\] Let $F\in GL_{N}({\mathbb{C}})$ be such that $F\overline{F}\in {\mathbb{R}}.\operatorname{Id}$ and let $0\leqslant q\leqslant 1$ be the real number defined in Theorem \[thm:freefusion\]. Then, if $q\leqslant 3^{-1/2}$ (in particular if $N\geqslant 3$), there exists a polynomial $P$ such that for all integers $d$, $$\|m_{p_{d}}\|_{cb} \leqslant P(d).$$ We use the notations of Proposition \[prop:polynomialblocks\]. We know from Corollary \[cor:blocks\] that $\|X^{d}_{j}\| \leqslant K(q)\|B_{d-j, j}(X^{d})\|$, thus $\|X^{d}_{j}\|\leqslant K(q)Q(d)\|X\|$. If we set $P(X) = K(q)(X+1)Q(X)$, we get $\|X^{d}\|\leqslant P(d)\|X\|$ by applying the triangle inequality to the decomposition of Lemma \[lem:blockdecomposition\]. One could slightly improve this bound by noticing that since we can replace $Q(d)$ by $1$ when $j=0$ or $d$, $\|X^{d}\|\leqslant K(q)(2d^{2}-d+1)\|X\|$. When $q=1$, we get the usual compact group $SU(2)$. It was proved in [@vergnioux2007property] that this group (or rather its discrete quantum dual) has the Rapid Decay property, and that consequently $m_{p_{d}}$ grows at most polynomially. Since any bounded map from a C\*-algebra into a *commutative* C\*-algebra is completely bounded with same norm (e.g. by [@conway2000course Prop 34.6]), Theorem \[thm:completelyboundedprojection\] also works in that case. This of course suggests that it holds for $SU_{q}(2)$ for any value of $q$, though the majorizations of Lemma \[lem:constantsum\] are not good enough to provide such a statement. It is proved in [@pisier2003introduction Thm 9.7.4] that in the free group case, the completely bounded norm of the projections on words of fixed length grows at most linearly. Our technique cannot determine whether such a result still holds in the quantum case but proves the slightly weaker fact that the growth is at most quadratic. However, we can prove that it is also at least linear. Let us first recall that the sequence $(\mu_{k})$ of (dilated) Chebyshev polynomials of the second kind is defined by $\mu_{0}(X) = 1$, $\mu_{1}(X) = X$ and $$X\mu_{k}(X) = \mu_{k-1}(X) + \mu_{k+1}(X)$$ Let $F\in GL_{N}({\mathbb{C}})$ be such that $F\overline{F}\in {\mathbb{R}}.\operatorname{Id}$. Then, there exists a polynomial $R$ of degree one such that $$\|m_{p_{d}}\|_{cb}\geqslant R(d).$$ Since $\|m_{p_{d}}\|_{cb} \geqslant \|m_{p_{d}}\|$, we will simply prove a lower bound for this second norm. Let $\chi_{n}\in \operatorname{Pol}({\mathbb{G}})$ be the character of the representation $u^{n}$, i.e. $$\chi_{n} = ({\imath}\otimes \operatorname{Tr})(u_{n}).$$ Our aim is to prove that looking at the action of $m_{p_{d}}$ on $\chi_{d+2} - \chi_{d}$ is enough to get the lower bound. It is known (see [@banica1996theorie]) that sending $\chi_{n}$ to the restriction to $[-2, 2]$ of $\mu_{n}$ yields an isomorphism between the sub-C\*-algebra of $C_{\text{red}}(O^{+}(F))$ generated by the elements $\chi_{n}$ and $C([-2, 2])$. Moreover, the restriction of these polynomials to the interval $[-2, 2]$ form a Hilbert basis with respect to the scalar product associated to the semicircular law $$d\nu = \frac{\sqrt{4-t^{2}}}{2\pi}dt.$$ Let us denote by $\pi :C([-2, 2]) \rightarrow {\mathcal{B}}(L^{2}([-2, 2], d\nu))$ the faithful representation by multiplication operators. What precedes means precisely that we have, for any finitely supported sequence $(a_{n})$, $$\left\|\sum_{n} a_{n}\chi_{n}\right\|_{C_{\text{red}}(O_{N}^{+})} = \left\|\sum_{n} a_{n}\mu_{n\vert[-2, 2]}\right\|_{\infty} = \left\|\sum_{n} a_{n}\pi(\mu_{n})\right\|_{{\mathcal{B}}(L^{2}([-2, 2], d\nu))}.$$ Let $e_{i}$ denote the image of $\mu_{i}$ in $L^{2}([-2, 2], d\nu)$ and denote by $T_{n}$ the operator sending $e_{i}$ to $e_{i+n}$ for $n\geqslant 0$. Letting $E_{j}$ denote the linear span of the vectors $e_{i}$ for $0\leqslant i\leqslant j$, we can also define operators $T_{-n}$ which are $0$ on $E_{n-1}$ and send $e_{i}$ to $e_{i-n}$ for $i\geqslant n$. The last operator we need, denoted $S_{n}$, sends $e_{i}\in E_{n}$ to $e_{n-i}$ and is $0$ on $E_{n}^{\perp}$. These translation operators obviously have norm $1$. Moreover, a simple computation using Theorem \[thm:freefusion\] (or equivalently the recursion relation of the Chebyshev polynomials) shows that $$\pi(\mu_{n+2} - \mu_{n}) = T_{n+2} - S_{n} - T_{-(n+2)}.$$ Thus $\|\chi_{n+2} - \chi_{n}\| = \|\pi(\mu_{n+2} - \mu_{n})\|\leqslant 3$. On the other hand, it easily seen that $\mu_{n}(2) = n+1$. In fact, this is true for $\mu_{1}(X) = X$ and $\mu_{2}(X) = X^{2}-1$ and we have the recursion relation $$2\mu_{n}(2) = \mu_{n+1}(2) + \mu_{n-1}(2).$$ This implies that $\|\chi_{n}\| = \|\mu_{n}\|_{\infty} \geqslant n+1$. Combining these two facts yields $$\|m_{p_{d}}\|\geqslant \frac{\|m_{p_{d}}(\chi_{d+2} - \chi_{d})\|}{\|\chi_{d+2} - \chi_{d}\|} = \frac{\|-\chi_{d}\|}{\|\chi_{d+2} - \chi_{d}\|} \geqslant \frac{d+1}{3}$$ and setting $R(X) = (X+1)/3$ concludes the proof. Monoidal equivalence and weak amenability {#subsec:final} ========================================= All the results proved so far hold in great generality, i.e. at least for any ${\mathbb{F}}O^{+}(F)$ with $F$ of size at least $3$ satisfying $F\overline{F}\in {\mathbb{R}}.\operatorname{Id}$. However, we will need in the proof of Theorem \[thm:maintheorem\] a result of M. Brannan proving that some specific multipliers are completely positive. That assertion has up to now only been proved in the case $F = I_{N}$, hence our restriction in Theorem \[thm:maintheorem\]. We will discuss this issue later on. Let us first deduce weak amenability of free quantum groups from the preceding sections. \[thm:maintheorem\] Let $N\geqslant 2$ be an integer, then the discrete quantum groups ${\mathbb{F}}O_{N}^{+}$ and ${\mathbb{F}}U_{N}^{+}$ are weakly amenable and their Cowling-Haagerup constant is equal to $1$. For $N=2$, this result is already known by amenability of the discrete quantum group ${\mathbb{F}}O_{2}^{+} = \widehat{SU_{-1}(2)}$. Thus, we will assume $N>2$. We are going to use a net of elements in $\ell^{\infty}({\mathbb{F}}O_{N}^{+})$ introduced by M. Brannan in [@brannan2011approximation] to prove the Haagerup property and the metric approximation property. For $t\in [0, N]$, set $b_{k}(t) = \mu_{k}(t)/\mu_{k}(N)$ and $$a_{i}(t) = \sum_{k=0}^{i} b_{k}(t)p_{k}\in \ell^{\infty}({\mathbb{F}}O_{N}^{+}).$$ This is a net of finite rank elements converging pointwise to the identity and we now have to prove that the completely bounded norms of the associated multipliers satisfy the boundedness condition. If we fix some $2 < t_{0} < 3$, then [@brannan2011approximation Prop. 4.4] asserts the existence of a constant $K_{0}$, depending only on $t_{0}$, such that for any $t_{0}\leqslant t < N$, $0<b_{k}(t)<K_{0}(t/N)^{k}$. According again to [@brannan2011approximation Prop. 4.4], the multipliers associated to the elements $a(t) = \sum_{k} b_{k}(t)p_{k}$ (where the sum runs from $0$ to infinity) are unital and completely positive. Moreover, for any $t_{0}\leqslant t < N$, $$\begin{aligned} \|m_{a(t)} - m_{a_{i}(t)}\|_{cb} & \leqslant & \sum_{k > i} K_{0}\left(\frac{t}{N}\right)^{k}\|m_{p_{k}}\|_{cb}. \\\end{aligned}$$ This sum tends to $0$ as $i$ goes to infinity since Theorem \[thm:completelyboundedprojection\] implies that it is the rest of an absolutely converging series. This implies that $\limsup \|m_{a_{i}(t)}\|_{cb} = 1$. In other words, $\Lambda_{cb}({\mathbb{F}}O_{N}^{+}) = 1$. By [@freslon2012note Thm. 4.2], we also have $\Lambda_{cb}({\mathbb{Z}}\ast {\mathbb{F}}O_{N}^{+}) = 1$, hence $\Lambda_{cb}({\mathbb{F}}U_{N}^{+}) = 1$ by Theorem \[thm:freesubgroup\]. Let us point out that the above results, [@freslon2012note Thm. 4.2] and the isomorphisms of [@weber2012classification Prop 3.2] (see also [@raum2012isomorphisms Thm 4.1]) imply that the free bistochastic quantum groups $B_{N}^{+}$ and their symmetrized versions $(B_{N}^{+})'$ and $(B_{N}^{+})^{\sharp}$ have the Haagerup property and are weakly amenable with Cowling-Haagerup constant equal to $1$. Let us now explain how this technique can be extended to other families of discrete quantum groups. We will use the notion of monoidal equivalence and we refer the reader to [@bichon2006ergodic] for the relevant definitions and properties. We are thankful to S. Vaes for suggesting the following argument. \[prop:monoidalequivalence\] Let ${\mathbb{G}}_{1}$ and ${\mathbb{G}}_{2}$ be two monoidally equivalent compact quantum groups. Then, if we index the equivalence classes of irreducible representations of both quantum groups by the same set through the monoidal equivalence, and if we denote by $p^{i}_{x}$ the projection in $\ell^{\infty}({\widehat}{{\mathbb{G}}}_{i})$ corresponding to $x\in \operatorname{Irr}({\mathbb{G}}_{i})$ for $i=1, 2$, we have $$\|m_{p^{1}_{x}}\|_{cb} = \|m_{p^{2}_{x}}\|_{cb}.$$ Let $B$ be the linking algebra given by [@bichon2006ergodic Thm 3.9]. There is an action $\alpha : C({\mathbb{G}}_{1})\rightarrow C({\mathbb{G}}_{1})\otimes B$ such that $$(m_{p^{1}_{x}}\otimes {\imath})\circ \alpha = \alpha\circ Q_{x},$$ where $Q_{x}$ denotes the projection in $B$ onto the spectral subspace associated to the irreducible representation $u^{x}$. The injective $*$-homomorphism $\alpha$ being completely isometric, we deduce $$\|Q_{x}\|_{cb} \leqslant \|m_{p^{1}_{x}}\|_{cb}.$$ Now, we know from the proof of [@vaes2007boundary Thm 6.1] that there is an injective $*$-homomorphism $\theta : C({\mathbb{G}}_{2})\rightarrow B\otimes B^{op}$ such that $$(Q_{x}\otimes {\imath})\circ\theta = \theta\circ m_{p^{2}_{x}},$$ yielding $$\|m_{p^{2}_{x}}\|_{cb} \leqslant \|Q_{x}\|_{cb}.$$ Proposition \[prop:monoidalequivalence\] means in particular that proving the polynomial bound in the case of $SU_{q}(2)$ (say at least for $q\leqslant 3^{-1/2}$) would give an alternative proof of Theorem \[thm:completelyboundedprojection\]. However, it is not clear to us that such a computation would really be easier. We can now give a second class of examples of weakly amenable discrete quantum groups. We refer the reader to [@banica1999symmetries] and [@banica2002quantum] for the definition of quantum automorphism groups. Let $B$ be a finite-dimensional C\*-algebra with $\dim(B)\geqslant 6$ and let $\sigma$ be the $\delta$-trace on $B$. Then, the compact quantum automorphism group of $(B, \sigma)$ is weakly amenable and has Cowling-Haagerup constant equal to $1$. Let us first prove a more general statement. Consider the sub-C\*-algebra $C(SO_{q}(3))$ of $C(SU_{q}(2))$ generated by the coefficients of $u^{\otimes 2}$, where $u$ denotes the fundamental representation of $SU_{q}(2)$. The restriction of the coproduct turns this C\*-algebra into a compact quantum group, called $SO_{q}(3)$, which can be identified with the compact quantum automorphism group of $M_{2}({\mathbb{C}})$ with respect to a $(q+q^{-1})$-form (see [@soltan2010quantum]). Its irreducible representations can be identified with the even irreducible representations of $SU_{q}(2)$, and re-indexing them by ${\mathbb{N}}$ gives the $SO(3)$-fusion rules $u^{1}\otimes u^{n} = u^{n-1}\oplus u^{n}\oplus u^{n+1}$. Consequently, we have $\|m_{p_{d}}\|_{cb}\leqslant P(2d)$ as soon as $q\leqslant 3^{-1/2}$. We know from [@de2010actions Thm 4.7] that if $\sigma$ is any $\delta$-form on $B$, the compact quantum automorphism group of $(B, \sigma)$ is monoidally equivalent to $SO_{q}(3)$ if and only if $q+q^{-1} = \delta$. Thus, by Proposition \[prop:monoidalequivalence\], we have $$\|m_{p_{d}}\|_{cb}\leqslant P(2d)$$ for any integer $d$ as soon as $\delta$ is big enough. Direct computation shows that $\delta \geqslant \sqrt{6}$ is a sufficient condition. If now $\sigma$ is the $\delta$-trace, $\delta = \sqrt{\dim(B)}$ and it was proved in [@brannan2012reduced Thm 4.2] that the elements $$a(t) = \sum_{k=0}^{+\infty} \frac{\mu_{2k}(\sqrt{t})}{\mu_{2k}(\sqrt{\dim(B)})}p_{k}$$ implement the Haagerup property. Applying the same proof as in Theorem \[thm:maintheorem\] then yields weak amenability with Cowling-Haagerup constant $1$. A particular case of the previous theorem is the *quantum permutation groups* $S_{N}^{+}$ (for $N\geqslant 6$) defined by S. Wang in [@wang1998quantum]. We can also deduce the Haagerup property and weak amenability with Cowling-Haagerup constant equal to $1$ for its symetrized version $(S_{N}^{+})'$. If $\dim(B)\leqslant 4$, the quantum automorphism group of $(B, \sigma)$ is amenable and is therefore weakly amenable with Cowling-Haagerup constant equal to $1$. Hence, the only case which is not covered by the previous theorem is the case of quantum automorphism groups of five-dimensional C\*-algebras. There are two such C\*-algebras, namely ${\mathbb{C}}^{5}$ and $M_{2}({\mathbb{C}})\oplus{\mathbb{C}}$. The quantum automorphism groups of these spaces are known to have the Haagerup property and we of course believe that they are weakly amenable with Cowling-Haagerup constant equal to $1$, though we do not have a proof of this fact. Let us further comment the consequences of Proposition \[prop:monoidalequivalence\]. Let us say that an element $a\in \ell^{\infty}({\widehat}{{\mathbb{G}}})$ is *central* if it is of the form $$a = \sum_{\alpha\in \operatorname{Irr}({\mathbb{G}})} b_{\alpha}p_{\alpha}$$ where $b_{\alpha}\in {\mathbb{C}}$ (i.e. $a$ belongs to the centre of $\ell^{\infty}({\widehat}{{\mathbb{G}}})$). Making linear combinations in the proof of Proposition \[prop:monoidalequivalence\], we see that if ${\mathbb{G}}_{1}$ and ${\mathbb{G}}_{2}$ are monoidally equivalent compact quantum groups, then any central element $a$ in ${\widehat}{{\mathbb{G}}}_{1}$ gives rise to a central element $a'$ in ${\widehat}{{\mathbb{G}}}_{2}$, the multipliers of which have the same completely bounded norm. Thus, weak amenability transfers through monoidal equivalence as soon as it can be implemented by central elements. Assume moreover that $m_{a}$ is u.c.p., then $m_{a'}$ is unital and has completely bounded norm $1$. Since any unital linear map of norm $1$ between two C\*-algebras is positive (see e.g. [@conway2000course Prop 33.9]), we can conlude that $m_{a'}$ is also u.c.p. This gives us the first examples of non-amenable, non-unimodular discrete quantum groups having approximation properties. \[prop:monoidalequivalencebis\] Let ${\mathbb{G}}$ be compact quantum group which is monoidally equivalent to $O^{+}_{N}$ or $U_{N}^{+}$ for some $N$ or to the compact quantum automorphism group of $(B, \sigma)$ for some finite-dimensional C\*-algebra $B$ of dimension at least $6$ endowed with its $\delta$-trace $\sigma$, then ${\widehat}{{\mathbb{G}}}$ has the Haagerup property and is weakly amenable with Cowling-Haagerup constant equal to $1$. Using these arguments we can in fact recover [@brannan2012reduced Thm 4.2] directly from [@brannan2011approximation Thm 4.5]. Note that under the conditions of Proposition \[prop:monoidalequivalencebis\], the linking algebra $B$ giving the monoidal equivalence also has the Haagerup property relative to the unique invariant state and is weakly amenable with Cowling-Haagerup constant $1$. We end with some comments on the following very natural question : when is it possible to implement an approximation property by multipliers associated to central elements on a discrete quantum group ? Let us make this question more formal. Let $A$ be an approximation property and say that a discrete quantum group has *central A* if there are central multipliers implementing the property $A$. On the positive part of the problem, we have the following obvious facts : 1. If ${\widehat}{{\mathbb{G}}}$ has central $A$, then it has $A$ (the converse is true for any discrete group since any element is central). 2. If ${\widehat}{{\mathbb{G}}}_{1}$ has central $A$ and if ${\mathbb{G}}_{1}$ is monoidally equivalent to ${\mathbb{G}}_{2}$, then ${\widehat}{{\mathbb{G}}}_{2}$ has central $A$. 3. In the previous case, the linking algebra also has central $A$ (with respect to the projections on spectral subspaces). 4. If ${\widehat}{{\mathbb{G}}}$ has the Haagerup property and is unimodular, then it has the central Haagerup property (this is a consequence of M. Brannan’s averaging technique used in the proof of [@brannan2011approximation Thm 3.7]). On the negative part of the problem, we can make two remarks : 1. Amenability of $SU_{q}(2)$ (or $SO_{q}(3)$) cannot be implemented by central multipliers (whereas there is an obvious way to implement it with general multipliers), otherwise all the compact quantum groups $O^{+}(F)$ (or all the compact quantum automorphism groups of finite-dimensional C\*-algebras with respect to some $\delta$-form) would be amenable. 2. If $q+q^{-1}$ is an integer, $SU_{q}(2)$ and $SO_{q}(3)$ have the central Haagerup property and are centrally weakly amenable with constant $1$. Hence these central approximation properties cannot be deduced from amenability.

--- abstract: 'This paper investigates a three-node multiple-input multiple-output relay system suffering from co-channel interference (CCI) at the multi-antenna relay. Contrary to the conventional relay networks, we consider the scenario that the relay is an energy harvesting (EH) node and has no embedded energy supply. But it is equipped with a rechargeable battery such that it can harvest and accumulate the harvested energy from RF signals sent by the source and co-channel interferers to support its operation. Leveraging the inherent feature of the considered system, we develop a novel accumulate-then-forward (ATF) protocol to eliminate the harmful effect of CCI. In the proposed ATF scheme, at the beginning of each transmission block, the relay can choose either EH operation to harvest energy from source and CCI or information decoding (ID) operation to decode and forward source’s information while suffering from CCI. Specifically, ID operation is activated only when the accumulated energy at the relay can support an outage-free transmission in the second hop. Otherwise, EH operation is invoked at the relay to harvest and accumulate energy. By modeling the finite-capacity battery of relay as a finite-state Markov Chain (MC), we derive a closed-form expression for the system throughput of the proposed ATF scheme over mixed Nakagami-m and Rayleigh fading channels. Numerical results validate our theoretical analysis, and show that the proposed ATF scheme with energy accumulation significantly outperforms the existing one without energy accumulation.' author: - | Yifan Gu, He Chen, Yonghui Li, and Branka Vucetic\ School of Electrical and Information Engineering, The University of Sydney, Sydney, NSW 2006, Australia\ Email: {yifan.gu, he.chen, yonghui.li, branka.vucetic}@sydney.edu.au bibliography: - 'References.bib' title: --- Conclusions =========== In this paper, we proposed an accumulate-then-forward scheme for a wireless energy harvesting (WEH) MIMO relay network suffering from co-channel interference. We modeled the dynamic charging and discharging behaviors of the finite-capacity relay battery by a a finite-state Markov chain (MC). We evaluated the transition probability matrix of the MC and derived a closed-form expression of system throughput over mixed Nakagami-m and Rayleigh fading channels. Numerical results validated the theoretical analysis and demonstrated the impact of co-channel interference on the system performance. Specifically, different from the conventional communication networks where CCI is always harmful, the CCI may benefit the performance of a WEH relay network. Results also showed that the proposed ATF scheme with energy accumulation outperforms the existing one without energy accumulation significantly.

--- abstract: 'The phonon dispersion, density of states, Grüneisen parameters, and the lattice thermal conductivity of single- and multi-layered boron nitride were calculated using first-principles methods. For the bulk [*h*]{}-BN we also report the two-phonon density of states. We also present simple analytical solutions to the acoustic vibrational mode-dependent lattice thermal conductivity. Moreover, computations based on the elaborate Callaway-Klemens and the real space super cell methods are presented to calculate the sample length and temperature dependent lattice thermal conductivity of single- and multi-layered hexagonal boron nitride which shows good agreement with experimental data.' author: - 'Ransell D’Souza' - Sugata Mukherjee title: 'Length dependent lattice thermal conductivity of single & multi layered hexagonal boron nitride: A first-principles study using the Callaway-Klemens & real space super cell methods' --- Introduction ============ Single and multilayered boron nitride are $sp^2$ bonded boron and nitrogen atoms arranged in a hexagonal honeycomb lattice arranged in ABAB stacking in multilayered and bulk materials. In spite of the fact that they are isomorphic to the multilayered graphene and graphite with similar lattice constants, unit cell masses and Van der Waals type bonding between the layers, their phonon properties are quite different. Consequently their physical properties such as the lattice thermal conductivity derived from the phonon dispersion should shed light on the fundamental physics of phonon transport of such two-dimensional (2D) nanomaterials. These nanomaterials in the form of semiconductor multilayers and other superstructures are promising candidates of materials with enhanced thermoelectrical properties and have been a topic of intensive research in recent years[@duana16]. In contrast to a large amount of theoretical and experimental work carried out on electron transport, only few studies on phonon transport have been reported. For example, using density functional theory (DFT) with quantum transport device simulation based on non-equilibrium Green’s function (NEGF), Fiori $et\ al$ [@fiori11] have proposed and investigated 2-D graphene transistors based on lateral heterobarriers. $Ab\ initio$ atomistic simulations on vertical heterobarrier graphene transistors have been analysed [@sciambi11; @mehr12]. Britnell $et\ al$ [@britnell12] have modelled graphene heterostructures devices with atomically thin boron nitride as a vertical transport barrier. Performance of any thermoelectric material is characterised by a dimensionless parameter termed as figure of merit, denoted by $ZT$, which is inversely proportional to total lattice thermal conductivity, including contributions due to electrons and phonons. However, for the materials investigated here, electron contribution is negligible compared to that of phonons owing to a considerable electronic band gap between their conduction and valence bands. Experiments to study the effects of grain-boundaries on the thermal transport properties of graphene have been carried out by a few groups [@xu14; @ma17]. These experiments show that a smaller sample length decreases the thermal conductivity, a necessity for a good thermoelectric material. Graphene has a higher thermal conductivity compared to graphite due to the long mean free path (MFP) of the phonons in the 2D lattices. The MFP can thus be reduced by creating defects in the sample. Recent studies by Malekpour [*el al.*]{} [@malekpour16] has shown that vacancies reduces the lattice thermal conductivity in graphene. Lattice thermal conductivity of a material is highly correlated to the thickness of the sample (or number of layers). For example, graphene has a much larger thermal conductivity than bilayer graphene and graphite. [@RDSM17; @hongyang2014]. Recently reported lattice thermal conductivity for In$_2$Se$_3$ exhibits [@zhou16] a strong dependence on the thickness or number of layers, with a value of 4 W/mK for a thickness of 5nm which increases to 60 W/mK for the sample with thickness 35nm. These results suggest that in order to manipulate the lattice thermal conductivity $\kappa_L$, a proper understanding of its dependence on the grain size, temperature and thickness dependence is essential. However, not many experiments on grain size, temperature and thickness have been carried out so far in single and multilayered boron nitride. We believe our present work will motivate experiments in the direction of tuning $\kappa_L$ in such 2D materials. Heat flow in single and multilayered boron nitride (SLBN and MLBN) is of great significance not only for fundamental understanding of such materials in terms of lattice thermal conductivity or thermoelectrics but also for technological applications. Single and MLBN are extremely atomically stable materials and can be easily supported between two leads. Besides, these materials exhibit a comparatively lower $\kappa_L$ in bulk than in single and multi-layered graphene. This makes SLBN and MLBN a good testing ground to study the length and temperature dependence of thermal conductivity. Manipulating the lattice thermal conductivity by varying its temperature and dimensions (through grain size engineering) will shed light on the fundamental understanding of thermoelectricity in such 2D materials and help in designing new novel materials for technological applications. Hexagonal boron nitride (h-BN) is relatively inert as compared to graphene due to its strong, in-plane, ionic bonding of its planar lattice structure and hence is a favourable substrate dielectric to improve graphene based devices [@dean10]. Although h-BN has appealing thermal properties, most studies, both experimentally and theoretically, are confined to single and multi-layered graphene [@chen2011; @chen2012; @balandin08; @ghosh08; @cai10; @jauregui10; @hongyang2014; @nika2009; @RDSM17]. Some experiments on lattice thermal conductivity ($\kappa_L$) have been reported by Jo $et\ al$ [@jo13] for multi-layered boron nitride (MLBN). Also, theoretical studies on thermal conductivity ($\kappa_L$) [@lindsay11] and conductance [@RDSM16] on such materials have been carried out using Tersoff empirical interatomic potential. However, first principle theoretical studies of $\kappa_L$ such as using the Boltzmann transport equations (BTE) for phonons from density functional perturbation theory (DFPT) on SLBN and MLBN are apparently not available. In this paper, we investigate numerically the sample length and temperature dependence of the thermal conductivity ($\kappa_L$) of single and multilayer h-BN by solving the phonon BTE beyond the relaxation time approximation (RTA) using the force constant derived from a real space super cell method, and also by solving the phonon BTE in the RTA using the Callaway-Klemens approach. A long standing puzzle has been to answer which acoustic phonon mode dominates the total lattice thermal conductivity for such 2D materials [@nika11]. There have been arguments on whether the out-of-plane ZA vibrational mode contributions to $\kappa_L$ are the most dominant or the least in comparison to the other acoustic modes. Owing to the selection rules restricting the phase space for phonon-phonon scattering in ideal graphene [@lindsay2010; @lindsay10; @seol10] and boron nitride [@lindsay11], the ZA mode seem to be the most dominant. In a rather sharp contrast, references [@nika2009; @nika09] suggest that since in the long wavelength limit ($q \rightarrow 0$), the phonon dispersion of the ZA modes seem to be flat thus making the phonon velocities small, and also the fact that the Grüneisen parameters are large, would make the ZA contributions to $\kappa_L$ the least in comparison to other acoustic modes. Here, using the Callaway-Klemens approach, we examine this discrepancy from analytical solutions to the phonon BTE for each of the acoustic modes using a closed form for the scattering rate for the three-phonon processes derived by Roufosse [*et. al.*]{} [@roufosse73] and an exact numerical solution for the phonon BTE beyond the relaxation time approximation (RTA) in which the phonon lifetimes are formed in terms of a set of coupled equations and solved iteratively. We also examined the sample length ($L$) dependence of $\kappa_L$ and found this to be very sensitive to $L$, which may justify the application of multilayered h-BN in thermoelectric devices by manipulating $\kappa_L$. In the next section we describe the theoretical framework and the first-principles DFT based calculational methods of $\kappa_L$. This is followed by the results obtained using the real space supercell method and the Callaway-Klemens method and a summary in the subsequent sections. Theoretical framework and Method of calculation =============================================== Electronic and phonon bandstructure calculations ------------------------------------------------ First-principles DFT and DFPT calculations were carried out on a hexagonal supercell for the monolayer, bilayer and bulk boron nitride, whereas an orthorhombic supercell was used for five layers h-BN sample, using the plane wave pseudopotential method as implemented in the QUANTUM ESPRESSO code [@giannozzi09]. We have used 2 atoms in the unit cell for SLBN, 20 atoms for five layered BN and 4 atoms in both bilayer and bulk boron nitride. To prevent interactions between the layers, a vacuum spacing of 20 Åwas introduced along the perpendicular direction to the layers ($z$-axis) mimicking an infinite BN sheet in the $xy$ plane. For MLBN and bulk-[*h*]{}BN, the Van der Waals interaction as prescribed by Grimme [@grimme1], was used between the layers. For the electronic structure calculations, Monkhorst-Pack grids of $16 \times 16 \times 1$ and $16 \times 16 \times 4$ were chosen for SLBN and MLBN, resepectively, for the $k$-point sampling. Self-consistent calculations with a 40 Ry kinetic energy cut-off and a 160 Ry charge density energy cutoff were used to solve the Kohn-Sham equations with an accuracy of 10$^{-9}$ Ry for the total energy. We used ultrasoft pseudopotential to describe the atomic cores with exchange-correlation potential kernel in the local density approximation [@rrkj90]. The electronic structure and total enrgy calculations were used to obtain the groundstate geometry before persuing the phonon calculations. For the phonon bandstructure calculations, the $q$-grid used in the calculations were $6 \times 6 \times 1$ for SLBN, $6 \times 6 \times 2$ for BLBN and bulk h-BN and $4 \times 4 \times 2$ for 5-layer BN, respectively. The density functional perturbation theory (DFPT) [@dfpt87], as implemented in the plane wave method [@giannozzi09], was used to calculate the phonon dispersion and phonon density of states (DOS) along the high-symmetric $q$-points. Calculation of the lattice thermal conductivity ----------------------------------------------- The calculation of lattice thermal conductivity $\kappa_L$ involves evaluation of second-order harmonic interatomic force constants (IFCs) as well as the third-order anharmonic IFCs. We have used first a real space supercell method which evaluates the third-order IFCs in a real space grid using DFT [@ShengBTE], whereas the second-order IFCs are obtained from the DFPT method [@giannozzi09; @dfpt87]. Secondly, using the Callaway-Klemens method [@callaway59; @klemens58], the relaxation times were obtained from the Grüeisen parameters. Finally, the length, thickness and temperature dependence of $\kappa_L$ were studied. ### Real space super cell approach In this method the third order anharmonic IFCs are calculated from a set of displaced supercell configurations depending on the size of the system, their symmetry group and the number of nearest neighbour interactions. A $4 \times 4 \times 2$ supercell including upto third nearest neighbour interactions were used to calculate the anharmonic IFCs for all the structures, generating 128 configurations for single and bulk BN, 156 for bilayer BN (BLBN) and 828 for five-layered BN (5LBN). The third order anharmonic IFCs are constructed from a set of third-order derivatives of energy, calculated from these configurations using the plane wave method [@giannozzi09]. The phonon lifetimes are calculated from the phonon BTE which are limited by phonon-phonon, isotropic impurity and boundary scattering [@lindsay11]. The three-phonon scattering rates are incorporated in this method, as implemented in the the ShengBTE code [@ShengBTE]. Elaborate details on the work-flow of the three-phonon scattering rates can be found in reference [@ShengBTE] while Lindsay [@lindsay11] specifically discuses this for bulk h-BN. The thermal conductivity matrix $\kappa_L^{\alpha \beta}$ is given as, $$\begin{aligned} \label{kl} \kappa_L^{\alpha \beta}=\frac{1}{k_BT^2\Omega N}\sum_{s}f_0(f_0+1)(\hbar \omega_s)^2v_{s}^{\alpha} \tau_{s}^0 (v_s^{\beta}+\Delta_s^{\beta}).\end{aligned}$$ $\kappa_L^{\alpha \beta}$ is then diagonalized to obtain the scalar lattice thermal conductivity $\kappa_L$ in a preferred direction in the $xy$ plane. In Eq. \[kl\] $\Omega$ is the volume of the unit cell, $N$ denotes the number of $q$-points in the Brillouin zone sampling. $f_0 = {1 / (e^{\hbar \omega_s/k_B T} - 1)}$ is the Bose-Einstein distribution function, $\tau_s^0$ is the relaxation time for the mode $s$ with phonon frequency $\omega_s$, $v_s$ is the phonon group velocity, and $\Delta_s$ denotes the measure of how much associated heat current deviates from the relaxation time approximation. Mathematically, $\Delta_s$ and $\tau_{\lambda}^0$ is expressed as [@ShengBTE], $$\begin{aligned} \Delta_{\lambda} &=& \frac{1}{N}\sum_{i=+,-}\sum_{\lambda^{'} \lambda^{''}} \Gamma^{i}_{\lambda \lambda^{'} \lambda^{''} } (\xi_{\lambda \lambda^{''}}F_{\lambda^{''}}-\xi_{\lambda \lambda^{'}}F_{\lambda^{'}}) \nonumber \\ &+& \frac{1}{N}\sum_{\lambda^{'}}\Gamma_{\lambda \lambda^{'}}\xi_{\lambda \lambda^{'}}F_{\lambda^{'}} \\ \frac{1}{\tau_{\lambda}^0} &=& \frac{1}{N}(\sum_{\lambda^{'} \lambda^{''}}^{+}\Gamma_{\lambda \lambda^{'} \lambda^{''}}^{+} + \frac{1}{2}\sum_{\lambda^{'} \lambda^{''}}^{-}\Gamma_{\lambda \lambda^{'} \lambda^{''}}^{-} + \sum_{\lambda^{'}}\Gamma_{\lambda \lambda^{'}})\end{aligned}$$ here $\lambda$($\lambda^{'}$,$\lambda^{''}$) represents the phonon branch index $s$($s^{'}$,$s^{''}$) and wave vector $q$($q^{'}$,$q^{''}$) while $\xi_{\lambda \lambda^{'}}$ and $F_{\lambda}$ is short-hand for $\frac{\omega_{\lambda^{'}}}{\omega_{\lambda}}$ and $\tau_{s}^0 (v_s^{\beta}+\Delta_s^{\beta})$ respectively. The three-phonon scattering rates denoted by $\Gamma^{i}_{\lambda \lambda^{'} \lambda^{''} }$($i = +,-$) and the scattering probabilities due to isotopic disorder denoted by $\Gamma_{\lambda \lambda^{'}}$ have the following expressions, $$\begin{aligned} \hspace{-2em} \Gamma^{\pm}_{\lambda \lambda^{'} \lambda^{''}} &=& \frac{\hbar \pi}{4 \omega_\lambda \omega_{\lambda^{'}}\omega_{\lambda^{'}}} \Big[\substack{f_0(\omega_{\lambda^{'}})-f_0(\omega_{\lambda^{''}}) \label{ae} \\ f_0(\omega_{\lambda^{'}})+f_0(\omega_{\lambda^{''}}+1)}\Big] \nonumber \\ &\times& \big|V_{\lambda \lambda^{'} \lambda^{''}}\big|^2\delta(\omega_\lambda \pm \omega_{\lambda^{'}} + \omega_{\lambda^{''}}) \\ \Gamma_{\lambda \lambda^{'}} &=& \frac{\pi \omega^2}{2}\sum_{i}f_s(i)\bigg[1-\frac{M_s(i)}{\overline{M}(i)}\bigg]^2 \nonumber \\ &\times& \big|e^{*}_{\lambda}\cdot e_{\lambda}\big|^2 \delta (\omega_{\lambda} - \omega_{\lambda^{'}}).\end{aligned}$$ Where $V^{\pm}$ is the scattering matrix element and is expressed in terms of the anharmonic IFCs ($\Phi$), eigen functions ($e$) and mass ($M$) of an atom as $$\begin{aligned} V_{\lambda \lambda^{'} \lambda^{''}} = \sum_{i,j,k}\sum_{\alpha \beta \gamma} \frac{\Phi_{ijk}^{\alpha \beta \gamma}e_{\lambda}^{\alpha}e_{\lambda^{'}}^{\beta}e_{\lambda^{''}}^{\gamma}}{\sqrt{M_i M_j M_k}}.\end{aligned}$$ In the above expression, $i,j,k$ run over the atomic indices and $\alpha, \beta, \gamma$ are the Cartesian coordinates. $\overline{M} = \sum_s f_s(i) M_s(i)$ is the average of masses ($M_s(i)$) of isotopes $s$ of the atoms $i$ having a relative frequency $f_s$. $\Gamma^{+(-)}$ represents the absorption (emission) processes. A phonon which is a result of the absorption process is a combined energy of two incident phonons, [*i.e.*]{} $\omega_{\lambda} + \omega_{\lambda^{'}} = \omega_{\lambda^{''}}$. Similarly, the emission process depicts the energy of an incident phonon being separated among two phonons, $\omega_{\lambda} = \omega_{\lambda^{'}} + \omega_{\lambda^{''}}$. Therefore in eq.\[ae\] it is easy to see that the Dirac delta function, $\delta(\omega_\lambda \pm \omega_{\lambda^{'}} + \omega_{\lambda^{''}})$, imposes the conservation of energy in the absorption and emission processes. It should be noted that the relaxation times is calculated in the ShengBTE code using an iterative approach by solving the phononBTE starting with the zeroth-order approximation, $\Delta_{\lambda} = 0$, also known as the RTA solution. These iterations continue till two successive values of $\kappa_L$ differ by $10^{-5}$ Wm$^{-1}$K$^{-1}$. The interatomic third-order force constants (IFCs) are calculated using a real space supercell approach. Length dependent thermal conductivity is then calculated by taking into account only phonons with a mean free path (MFP) below a certain threshold value. This is done by calculating the cumulative lattice thermal conductivity with respect to the allowed MFP. Furthermore, there have been recent advanced experimental techniques proposed to measure the cumulative $\kappa_L$ as a function of phonon mean free path [@minnich11; @regner13; @johnson13]. In order to compare our calculations to the lengths corresponding to experimental measurements, we fit the cumulative thermal conductivity in the form [@ShengBTE], $$\begin{aligned} \label{cum-k} \kappa_L(L) = \frac{\kappa_{L_{max}}}{1+\frac{L_0}{L}}.\end{aligned}$$ where $L_0$ is a fitting parameter. $\kappa_L$ corresponding to a given length is calculated over a temperature range using Eq. \[cum-k\] and the thermodynamic limit of the thermal conductivity ($\kappa_{L_{max}}$) is the value of $\kappa_L$ as $L \rightarrow \infty$. ### Callaway-Klemens approach (Analytical and numerical solutions) In the Callaway-Klemens’s [@callaway59; @klemens58] approach which has been modified by Nika [*el al*]{} [@nika2009], the expression for thermal conductivity along $x$ and $y$ directions for two-dimensional layered materials, according to the relaxation time approximation (RTA) to BTE and isotropic approximation to phonon dispersion is given by, $$\begin{aligned} \label{k} \kappa &=& {1\over 4\pi k_B T^2 N \delta} \nonumber \\ &\times& {\sum\limits_{s} \int\limits_{q_{min}}^{q_{max}}[\hbar \omega_s(q)]^2 v_s^2(q) \tau_{U,s}(q)\frac{e^{\frac{\hbar \omega_s(q)}{k_B T}}}{[e^{\frac{\hbar \omega_s(q)}{k_B T}}-1]^2} q dq},\end{aligned}$$ where $k_B$ is the Boltzmann constant, $\hbar$ is the reduced Planck constant, $T$ is the absolute temperature, $N$ is the number of layers, $\delta$ is the distance between two consecutive layers, $\omega_s (q)$ and $v_s(q)$ are the phonon frequency and velocity corresponding to the branch $s$ at phonon wave vector $q$. The wave vector corresponding to the Debye frequency and low cut-off frequency are denoted by $q_{max}$ and $q_{min}$, respectively. The method to calculate the low cut-off frequency will be discussed shortly. $\tau_{U,s}$ is the three-phonon Umklapp scattering corresponding to branch $s$ at the wave vector $q$ expressed as, $$\begin{aligned} \label{tau} \tau_{U,s} = \frac{Mv_s^2(q)\omega_{D,s}}{\gamma_s^2(q) k_B T \omega_s(q)^2}.\end{aligned}$$ Here, $M$ is total mass of the atoms in the unit cell, $\gamma_s(q)$ is the mode and wave vector dependent Grüneisen parameter. The validity of the form of relaxation time in the Umklapp scattering in eq. \[tau\] for a 2D and 3D material was originally proposed by Klemens [*et. al.*]{}[@klemens94], where phonons were treated by a two-dimensional Debye model. This sets up a mode for the thermal conductivity in terms of a 2D phonon gas. On the basis of the phonon frequency dependence of the specific heat and mean free path, the form of $\tau_{U,s}$ in eq. \[tau\] is valid for both 2D and 3D. Moreover, the calculations by Shen [*et. al.*]{} [@shen14] use the same form to describe the relaxation time of the Umklapp process for graphene and their results, when $\tau_{U,s}$ is multiplied by a factor of 3, are consistent with the paper of Lindsay [*et. al.*]{} [@lindsay11] which solves the phonon BTE beyond the RTA. Since eq. \[tau\] cannot determine whether the U-processes are forbidden or not, the factor of 3 is added due to the symmetries seen in graphene which is explained in detail later. Grüneissen parameter ($\gamma_s(q)$) and the Debye frequency ($\omega_{D,s}$) corresponding to the branch $s$ is calculated by solving, $$\begin{aligned} \label{wD} \frac{A}{2\pi}\int\limits_0^{\omega_{D,s}}q\Big|\frac{dq}{d\omega}\Big|d\omega = 1,\end{aligned}$$ where $A$ is the area of the unit cell. The acoustic branches for in-plane modes for SLBN, BLBN, 5LBN and Bulk-[*h*]{}BN are linear whereas the out-of-plane acoustic mode have a quadratic behavior and hence for a simplified analytical solution we express the phonon frequencies as $$\begin{aligned} \omega_s(q) &=& v_s q \Rightarrow [s = {\rm LA, TA}] \label{wLATA} \\ &=& \alpha q^2 \Rightarrow [s = {\rm ZA}] \label{wZA}\end{aligned}$$ Substituting these values in Eq. \[wD\], we find the Debye frequency is given by $$\begin{aligned} \ \omega_{D,s} &=& 2v_s\sqrt{\frac{\pi}{A}} \Rightarrow [s = {\rm LA, TA}] \\ &=& \frac{4\pi \alpha}{A} \Rightarrow [s = {\rm ZA}]\end{aligned}$$ The mode dependent anharmonic (Grüneissen) parameters were calculated by applying a biaxial strain of $\pm$ 0.5% to each of the structures. Fig. \[gp\] shows that the Grüneisen parameter for the in-plane modes have a slight deviation from its average value along the $\Gamma$ to K direction. Therefore assuming a constant value for $\gamma_s$ ($s$=LA,TA), Nika [*et al*]{} [@nika2009] have derived the following analytical solution for $\kappa$ associated with a particular mode $s$. $$\begin{aligned} \label{LATA} \kappa_s=\frac{M\omega_{D,s} v_s^2}{4\pi T (N\delta) \gamma_s^2}\ [{\rm ln}(e^x-1)+\frac{x}{1-e^x}-x]\Bigg|^\frac{\hbar \omega_{D,s}}{k_B T}_{\frac{\hbar \omega_{min,s}}{k_B T}}\end{aligned}$$ Since there is no ZO$'$ branch in SLBN, the low bound cut-off frequency cannot be introduced in analogy to that of bulk graphite. One can however avoid the logarithmic divergence by restricting the phonon mean free path on the boundaries of the sheets [@nika09]. This is accomplished by selecting the mode dependent low cut-off frequency ($\omega_{s,min}$) from the condition that the mean free path cannot be greater in size than physical length $L$ of the sheet, [*i.e,*]{} $$\begin{aligned} \label{wmin} \omega_{s,min}=\frac{v_s}{\gamma_s}\sqrt{\frac{Mv_s\omega_{D,s}}{k_B T L}}\end{aligned}$$ In the spirit of in-plane thermal conductivity study we extend our calculations to find an analytical form to the flexural phonons modes since the contribution from these branches are vital to the total thermal conductivity. Unlike for the case of in-plane modes, the Grüneisen parameters for the acoustic out-of-plane ZA modes have a strong $q$-dependence. From Fig. \[gp\] it can be seen that the expression $$\begin{aligned} \label{gZA} \gamma_{ZA}=\frac{\beta}{q^2},\end{aligned}$$ is a very good fit to the actual wave vector dependent Grüneisen parameters. Substituting eq. \[gZA\] and eq. \[wZA\] into eq. \[k\] and making a transformation, $x=\frac{\hbar \omega}{k_B T}$, the analytical form for $\kappa_{ZA}$ is given by $$\begin{aligned} \label{ZA} \kappa_{ZA} &=& \frac{2M\omega_D k_B^3T^2}{\pi N \delta \beta^2 \hbar^3 \alpha} \int\limits_0^{\frac{\hbar \omega_D}{k_B T}}x^4 \frac{e^x}{[e^x-1]^2}dx \nonumber \\ &=& \frac{2M\omega_D k_B^3T^2}{\pi N \delta \beta^2 \hbar^2 \alpha}\; G\Big(\frac{\hbar \omega_D}{k_BT}\Big), \end{aligned}$$ where the function $G(z)$ is expressed as $$\begin{aligned} G(z) &=& \frac{-4\pi^4}{15} + \frac{e^z z^4}{1-e^z} + 4z^3{\rm ln}(1-e^z) \nonumber \\ &+& 12z[z {\rm Li}_2(e^z)- 2{\rm Li}_3(e^z)] + 24{\rm Li}_4(e^z).\end{aligned}$$ Here, the polylogarithm function is defined as, ${\rm Li}_n(z)=\sum\limits_{i=1}^{\infty} \frac{z^i}{i^n}$. Results and discussions ======================= Phonon dispersion and density of states --------------------------------------- Accurate calculations of the harmonic second order IFCs are necessary for a precise description and understanding of the thermal conductivity. Deviations due to numerical artifacts from the expected behavior of acoustic modes can lead to incorrect results especially for 2D marterials [@jesus16]. The full structural relaxation of SLBN, BLBN, 5LBN and Bulk-[*h*]{}BN yield a lattice constant ($a_0$) of 2.49 Å. The interlayer spacing ($c$) for MLBN is found to be 3.33 Å. The experimentally measured $a_0$ is 2.50 Å [@kern99] and the ratio of interlayer spacing and the lattice constant ($\frac{a_0}{c}$) is 1.332 [@kern99] which is in excellent agreement with our calculated value of 1.337. The calculated phonon dispersion and phonon density of states are shown in Fig. \[phdos\] for (a) SLBN, (b) BLBN, (c) 5LBN and (d) Bulk-[*h*]{}BN along the high symmetric $q$-points in the irreducible hexagonal and orthogonal Brillouin zone (BZ) together with some available experimental data for Bulk-[*h*]{}BN [@serrano07]. As usually seen for acoustic modes, the in-plane longitudinal (LA) and transverse (TA) modes show a linear $q$ dependence at the long-wavelength limit while the out-of-plane (ZA) mode shows a quadratic ($q^2$) dependence. This quadratic dependence, which is a typical feature of layered crystals, is due the rotational symmetries of the out-of-plane phonon modes. For SLBN, there are six modes for each wave vector, three acoustic (LA,TA,ZA) and three optical (LO,TO,ZO). At the $\Gamma$ point the optical LO and TO modes are degenerate. For BLBN, if the two SLBN layers are far apart, effects due to their interlayer coupling can be neglected and the phonon dispersion will be exactly as what is seen in SLBN. However, when these two SLBN come closer, due to the interlayer coupling, the two-fold degeneracy is removed giving rise to in-plane and out-of-plane phase modes. The LA and TA modes are not perturbed much implying that the main effect of the interlayer interactions is due to the ZA modes. This is because the transverse motion of atoms in both the layers associated with these modes interact strongly with each other. The same reasons hold on why 5LBN has one zero and four raised frequencies at the $\Gamma$ point. In Bulk-[*h*]{}BN, there are four atoms per unit cell and the two atoms in each layer are now inequivalent therefore doubling each of the acoustic and optical modes. The acoustic modes at the zone boundaries fold back to the zone centre as two rigid layer modes [@tan12], [*viz*]{}, an optically Raman inactive and an Raman active mode. The Raman active LA$_2$ and TA$_2$ modes are doubly degenerate at the $\Gamma$ point having a finite value mentioned in Table \[pd\]. The layered breathing modes for MLBN are denoted by ZO$^{'}$ for BLBN and Bulk-[*h*]{}BN and ZO$^{'}_{i}$ ($i=1,2,3,4$) for 5LBN. ![\[phdos\] The calculated phonon dispersion (left) and phonon density of states (right) of (a) SLBN, (b) BLBN, (c) 5LBN and (d) Bulk-[*h*]{}BN along with experimental data (orange circles) [@serrano07]. The phonon dispersion were calculated along the high-symmetry points of the 2D Brillouin zone ($q_z = 0$) corresponding to the hexagonal cell for SLBN, BLBN, and Bulk-[*h*]{}BN and orthorhombic cell for 5LBN. We also plot in (d) the two-phonon DOS shown for Bulk-[*h*]{}BN in red dashed line. The cyan, magenta and green curves in (a,b,c,d) are the best linear and quadratic fit to the phonon dispersion referring to LA, TA and ZA modes, respectively.](layered-phonon-dos-hbn.pdf) The symmetries of SLBN, BLBN, 5LBN and Bulk-[*h*]{}BN structures at $\Gamma$ can be described using the character table shown in table \[ct\]. Using a standard group theoretical technique (see Appendix), it can be shown that for Bulk-[*h*]{}BN and BLBN that the 12 phonon modes are decomposed into the following irreducible representations: 2(A$_{\rm 2u}$ + B$_{\rm 1g}$ + E$_{\rm 2g}$ + E$_{\rm 1u}$) and 2(A$_{\rm 2u}$ + E$_{\rm g}$ + A$_{\rm 1g}$ + E$_{\rm u}$). Similarly for SLBN, the irreducible representation is A$_{\rm 2u}$ + B$_{\rm 1g}$ + E$_{\rm 2g}$ + E$_{\rm 1u}$ for the six phonon modes and 5LBN has an irreducible presentation given by 4(A$_1^{'}$ + E$^{''}$) + 6(A$_2^{''}$ + E$^{'}$). Transitions corresponding to the basis $x,y,z$ ($xy,yz,z^2$, etc.) are Infrared (Raman) active. Those that are neither Infrared or Raman are the silent modes. Due to the momentum conservation requirement ($q=0$), the first-order Raman scattering process is limited to the phonons at the center of the Brillouin zone. We therefore compare our calculated frequencies at the $\Gamma$ point corresponding to A$_{\rm 2u}$, E$_{\rm 1u}$, A$_{\rm 2}^{''}$, E$^{'}$, and E$_{\rm u}$ to the infrared experimental data and E$_{\rm 2g}$, E$^{''}$, A$_{\rm 1g}$, E$_{\rm g}$, and A$_{\rm 1}^{''}$ to the Raman experimental data as shown in table \[pd\]. [lcccc]{} Mode & --------------------------------------- Expt. (& Prev. calculated$^{\rm a}$ ) $\omega$ (cm$^{-1}$) --------------------------------------- & ---------------- Bulk-[*h*]{}BN (Sym.) ---------------- & -------- BLBN (Sym.) -------- & -------- SLBN (Sym.) -------- \ LA$_2$ & TA$_2$ & 51.62$^{\rm b}$ (52.43) & 58.55 (E$_{{\rm 2g}}$) & 25.73 (E$_{{\rm g}}$) & -\ ZO$'$ & Silent (120.98) & 85.01 (B$_{{\rm 1g}}$) & 66.54 (A$_{{\rm 1g}}$) & -\ ZO & 783.16$^{\rm c}$ (746.87) & 784.05 (A$_{{\rm 2u}}$) & 803.01 (A$_{{\rm 2u}}$) & 819.37 (A$_{{\rm 2u}}$)\ ZO$_2$ & Silent (809.78) & 823.17 (B$_{{\rm 1g}}$) & 818.25 (A$_{{\rm 1g}}$) & -\ LO & 1366.30$^{\rm b}$, 1370.33$^{\rm c}$, 1363.88$^{\rm d}$ (1379.20) & 1363.80 (E$_{{\rm 2g}}$) & 1364.45 (E$_{{\rm g}}$) & 1363.88 (E$_{{\rm 2g}}$)\ TO & 1367.10$^{\rm c}$ (1378.4) & 1366.95 (E$_{{\rm 1u}}$) & 1365.66 (E$_{{\rm u}}$) & 1363.88 (E$_{{\rm 1u}}$)\   & ------------------------ LA & TA (cm$^{-1}$) (Point Group Symmetry) ------------------------ & ----------------- LO (cm$^{-1}$) (P.G. Symmetry) ----------------- & ----------------- TO (cm$^{-1}$) (P.G. Symmetry) ----------------- & ----------------- ZO (cm$^{-1}$) (P.G. Symmetry) ----------------- \  & 14.60 (E$^{''}$) & 1409.46 (E$^{'}$) & 1405.23 (E$^{'}$) & 817.59 (A$^{'}$)\  & 31.10 (E$^{'}$) & 1408.91 (E$^{''}$) & 1404.94 (E$^{''}$) & 814.58 (A$^{''}$)\ 5LBN & 38.95 (E$^{''}$) & 1408.71 (E$^{'}$) & 1404.81 (E$^{'}$) & 812.58 (A$^{'}$)\  & 47.43 (E$^{'}$) & 1408.36 (E$^{''}$) & 1404.49 (E$^{''}$) & 810.27 (A$^{''}$)\  & & 1405.57 (E$^{''}$) & 1404.40 (E$^{'}$) & 803.27 (A$^{'}$)\ $^{\rm a}$ From [*ab initio*]{} dispersion calculations, Ref. [@serrano07].\ $^{\rm b}$ Experimental Raman data, Ref. [@nemanich81].\ $^{\rm c}$ Experimental Raman and Infrared data, Ref. [@geick66].\ $^{\rm d}$ Experimental Raman data, Ref. [@reich05]. Raman spectroscopy is the most adaptable tool that offers a direct probe for multi-layered samples [@tan12]. Table \[pd\] shows the transitions corresponding to the Inflared (E$^{'}$ and A$^{''}$) and Raman (E$^{''}$ and A$^{'}$) active modes in the case of 5LBN. Further experiments for layered boron nitride would be required to verify the correctness of calculations. However, LDA with VdW interaction have shown to accurately describe the phonon dispersions for layered graphene when the geometry ([*i.e.*]{} interlayer distance) is represented correctly even though the local or semi-local exchange correlation functionals may not represent the interactions correctly [@tan12]. Another experimental technique to analyse the modes of a system is the second-order Raman spectroscopy in which the peaks are seen over the entire frequency range. Most of these peaks are in agreement with the phonon density of states when the frequency is scaled by a factor of $2\,$ [@serrano07; @kern99]. We have hence plotted, to the right of our phonon dispersion, the frequency scaled DOS. However, as pointed out by Serrano [*et al.*]{}, peaks which are absent in the DOS can be seen in the second-order spectroscopy because the DOS does not take both overtones, [*i.e.*]{} summation of modes having the same frequencies, into account. The two phonon density of states (DOS$_{2ph}$) are also essential for the understanding of phonon anharmonic decay [@cusco16]. Experiments on the second-order Raman spectrum of h-BN has been performed by Reich [*et al*]{} [@reich05]. We show in Fig. \[phdos\](d) the two-phonon DOS [@esfarjani11], $$\begin{aligned} \label{tpdos} \hspace{-2em}{\rm DOS}_{2ph}(\omega) = \sum_{i,j} \delta(\omega - \omega_i - \omega_j) + \delta(\omega - \omega_i + \omega_j), \end{aligned}$$ for Bulk-[*h*]{}BN using our calculated harmonic interactions. The peaks seen experimentally [@reich05] at 1639.4 cm$^{-1}$, 1809.907 cm$^{-1}$ and 2289.8068 cm$^{-1}$ are absent in the DOS. However, these large spectral features are now observed at 1680.4 cm$^{-1}$, 1821.2 cm$^{-1}$ and 2306.7 cm$^{-1}$, due to two phonon DOS (DOS$_{2ph}$). Thermal conductivity calculated using real space supercell approach ------------------------------------------------------------------- ![\[kfig\] Calculated thermal conductivity of single and multilayer BN shown as a function of (a) temperature and (b) length, using the real space approach. In (a) the curves refer to the thermodynamic limit ($L\rightarrow \infty$). In (b) the sample length is in logarithmic scale. The square and triangle data points refer to experimental measurements for BLBN [@wang2016] and 5LBN [@jo13], respectively.](layered-kappa-L.pdf) In Fig. \[kfig\] (a) and (b) we show the variation of thermal conductivity as a function of temperature ($T$) and sample length, respectively. The sample length is measured along the direction of the heat flow. The theoretical computation was carried out using the interatomic force constants obtained from the real space approach and an iterative method in calculating the relaxation times as implemented in the ShengBTE code [@ShengBTE]. To have a broad understanding of the thermal conductivity, we study different types of possible unit cells, [*i.e.*]{}, MLBN considered here have even, odd and infinite number of layers since each unit cell has a different character table. Calculations were done using orthogonal cell for 5LBN and hexagonal cells for SLBN, BLBN and bulk-[*h*]{}BN. The study was carried out over a wide range of sample lengths between 0.01 $\mu$m and 1000 $\mu$m with 0.1$\mu$m grid. The temperature of each sample was varied between 10 K to 1000 K with a grid of 10 K. On plotting the thermodynamic limit ($L \rightarrow \infty$) for each of the system we find that $\kappa_L$ is practically independent of length for lengths greater than 100 $\mu$m. Our recent results of $\kappa_L$ in the thermodynamic limit ($L\rightarrow \infty$) for monolayer and bilayer graphene [@RDSM17] are in excellent agreement with the recent experimental work of Li [*et. al*]{} [@hongyang2014], whereas the thermodynamic limit for MLBN is much larger than some recent experimental measurements [@jo13; @wang2016]. Sample lengths used by Li [*et. al.*]{} were of the order of millimetres for the measurement of single and bilayer graphene while Jo [*et. al.*]{} and Wang [*et. al.*]{} have used sample lengths of 5 $\mu$m and 2 $\mu$m for 5LBN and BLBN, respectively. As mentioned earlier, $\kappa_L$ does not vary much for lengths larger than 100 $\mu$m but is extremely sensitive when the lengths are between 1 and 100 $\mu$m. Not surprising therefore, our thermodynamic limit of $\kappa_L$ are in good agreement for graphene but not for MLBN. In order to compare our calculations to that of experiments, we calculate the cumulative lattice thermal conductivity at lengths corresponding to the sample lengths used in the experiments. The cumulative $\kappa_L$ was calculated in the temperature range 10-1000 K. Fig. \[kfig\] (b) shows the cumulative thermal conductivity at room temperature (RT). ![\[kLexpt\] Calculated thermal conductivity of single and multilayer BN shown as a function of temperature at a constant length, using the real space approach. The square, circle, triangle data points refer to experimental measurements for BLBN [@wang2016], Bulk [*h*]{}-BN [@sichel76] and 5LBN [@jo13], respectively.](kappa-L-expt.pdf) The curves in Fig. \[kLexpt\], are the calculated values of $\kappa_L$ at constant lengths which are compared with the experimental observations [@wang2016; @sichel76; @jo13]. For the lengths used in the experiments the magnitudes of $\kappa_L$ for Bulk-[*h*]{}BN and Bi-layer lie in between SLBN and 5LBN with SLBN (5LBN) being the highest(lowest). The maxima of $\kappa_L$ of $\sim$ 500 Wm$^{-1}$K$^{-1}$ for Bulk-[*h*]{}BN is found in the temperature range 250-300 K and tends to saturate to a value $\sim$ 450 Wm$^{-1}$K$^{-1}$. Experimentally [@sichel76] the maxima is found between 150-200 K and tends to saturate to a value $\sim$ 400 Wm$^{-1}$K$^{-1}$. Lindsay [*et. al.*]{} [@lindsay11] varies the sample length and finds an excellent fit with the experimental data for $L=1.4 \mu$m. It must be noted that the sample length is not mentioned in the experimental reference [@sichel76] for Bulk-[*h*]{}BN. As the length of the sample increases, the maxima of $\kappa_L(T)$ shifts towards the left, [*i.e.*]{} the maxima is found at a lower temperature. Therefore for BLBN and 5LBN, where the lengths used in the experiments are larger than 1.4 $\mu$m, the maxima would be at lower temperatures, in total disagreement with the experiments [@wang2016; @jo13]. Our calculations for BLBN and 5LBN are in excellent agreement with experiments for the same lengths. Even though our calculated values diverge from the experimental measurements by Sichel [*et. al.*]{} [@sichel76] at higher temperatures, we believe that the behavior of $\kappa_L$ as calculated by us for bulk-[*h*]{}BN is correct. However, further experiments should throw more light on these discrepancies. It is our conjecture that $\kappa_L$ of Bulk-[*h*]{}BN should be similar to that of BLBN since the phonon dispersions in the two cases are very similar. ![\[kLmode\] Contribution to the thermal conductivity of single and multilayer BN from the acoustic modes; (a) ZA, (b) TA and (c) LA, shown as a function of temperature in thethermodynamic limit ($L\rightarrow \infty$), and as a function sample length at $T=300$K (d,e,f) using the real space approach.](mode-dependent-k-vs-L-T.pdf) In Fig. \[kLmode\] we show the acoustic mode dependent contributions to the total thermal conductivity for SLBN and MLBN by solving the phonon BTE beyond the RTA. The out of plane mode is clearly seen to contribute the most to the lattice thermal conductivity for all the mentioned structures. For SLBN the contributions from the ZA, TA and LA modes to $\kappa_L$ at room temperature are $\sim$ 86.1 %, 7.4 % and 6.5 %. A similar trend is observed in graphene [@lindsay2010]. Qualitatively one can understand why the ZA mode contributes the most to $\kappa_L$ by calculating the number of modes per frequency for each of the acoustic mode. Now the number of modes per frequency is proportional to the 2D density of phonon modes, $D_s(\omega) \propto \frac{q}{2\pi}\frac{dq}{d\omega}$, and hence the ratio of $D_{ZA}(\omega)$ and $D_{TA(LA)}(\omega)$ would give a measure of the contribution of the respective phonon modes. Assuming a quadratic fit to the ZA dispersion, $\omega_{ZA}=\alpha q_{ZA}^2$, and a linear fit to the in-plane TA and LA phonon dispersion, $\omega_{TA(LA)}= v_{TA(LA)}q_{TA(LA)}$, the ratio of the density of phonon modes is $\frac{D_{ZA}}{D_{TA(LA)}} = \frac{v_{LA(TA)}^2}{2\alpha\omega_{LA(TA)}}$. Here $\alpha$ and $v_{LA(TA)}$ are fitting parameters to the phonon dispersions shown in Fig. \[phdos\] and their values are shown in table \[para\]. Substituting these values, it is evident that $\frac{D_{ZA}}{D_{TA(LA)}} \gg 1$ at the long wavelength limit suggesting that the major contributions to the lattice thermal conductivity are due to the out of plane modes. Representing the ZA contribution of the thermal conductivity of MLBN at room temperature with respect to SLBN, we observe that $\kappa^{SLBN}_{ZA} = 1.28\kappa^{BLBN}_{ZA} = 2.17\kappa^{5LBN}_{ZA}$, suggesting that the significant decrease of $\kappa_L$ from SLBN to MLBN is because of the additional raised frequencies of the ZA layered breathing modes. Kong [*et. al.*]{} [@kong2009] reported that the lattice thermal conductivity of single layer graphene and bilayer are similar, $\kappa_L^{graphene} \approx \kappa_L^{bilayer}$, while Lindsay [*et. al*]{} [@lindsay2011] reported $\kappa_L^{graphene} \approx 1.37\kappa_L^{bilayer}$. The difference in their methodologies is that the latter has taken graphene symmetry into account, which is discussed in detail by Seol [*et.al.*]{} [@seol10] and Lindsay [*et.al.*]{} [@lindsay2011]. Besides the contribution due to the layer breathing out of plane modes, a decrease in $\kappa_L$ is also due to the violation of the selection rule [@seol10; @lindsay2011] which is incorporated in the formalism in the super-cell real space approach. In Fig. \[kLmode\] (d,e,f), we show the mode dependent $\kappa_L$ at room temperature as a function of sample length. At any given length, the maximum difference in $\kappa_L$ contributed from LA and TA modes for all the mentioned structure is $\sim$ 47 and 65 Wm$^{-1}$K$^{-1}$ respectively while that from the ZA mode is $\sim$ 750 Wm$^{-1}$K$^{-1}$, an order of magnitude larger, implying that the contribution from the in-plane thermal conductivity is almost independent of the number of layers. This characteristic has been seen using a Tersoff potential in the case of single and multilayered graphene and boron nitride [@lindsay2011; @lindsay12]. This rapid decrease in $\kappa_L$ by increasing the number of layers, which is mainly due to the ZA mode, suggests that the interlayer interactions are short ranged, [*i.e.*]{}, the BN layers only interact with neighbouring BN layers [@lindsay2011]. In all of the structures, the contribution to $\kappa_L$ from the ZA mode have a stronger $L$ dependence as compared to the TA and LA modes, [*i.e.*]{}, the contributions from the in-plane modes saturate to their thermodynamic limit at a lower $L$ value as compared to the contributions from the out-of-plane modes. This is due to the larger intrinsic scattering times allowing the ZA phonons to travel ballistically and the relatively smaller scattering time which reflects the diffusive transport of the TA and LA phonons [@lindsay2011]. Calculations based on the mode dependent contributions to $\kappa_L$ as a function of mean free path and recent advanced experimental techniques [@minnich11; @regner13; @johnson13] should motivate further studies in these directions. The in-plane phonon contributions having a small $L$ dependence in comparison to the contributions from the out of plane has been calculated for graphene recently using the Tersoff potential [@lindsay14] and their calculated cumulative mode dependent thermal conductivity behavior is in good agreement with our calculations for SLBN. Grüneisen parameter ------------------- Besides providing important information on the phonon relaxation time, the Grüneisen parameter ($\gamma$) also provides information on the degree of phonon scattering and anharmonic interactions between lattice waves. Therefore, an accurate calculation of the lattice thermal conductivity ($\kappa_L$) would require a precise calculation of $\gamma$ since anharmonic lattice displacements play a vital role in calculations of $\kappa_L$. Fig. \[gp\] displays the mode dependent $\gamma$ for SLBN, BLBN, 5LBN and Bulk-[*h*]{}BN along the high symmetric $q$ points. The anharmonic lattice displacements are carried out by dilating the unit cell by applying a biaxial strain of $\pm$ 0.5 % and is expressed as, $$\begin{aligned} \label{gamma} \begin{split} \gamma_s(q) & = \frac{-a_0}{2\,\omega_s(q)}\frac{\delta \omega_s(q)}{\delta a} \\ & \approx \frac{-a_0}{2\,\omega_s(q)} \Big[\frac{w_+ - w_-}{da}\Big] \end{split}\end{aligned}$$ This method has been used previously for graphite [@marzari05], single and bi-layer graphene [@RDSM17] and MoS$_2$ [@cai10]. ![\[gp\] Grüenisen parameters of each mode for (a) SLBN, (b) BLBN, (c) 5LBN and (d) Bulk-[*h*]{}BN. The colour representation of each mode and fit are shown on the right. The magenta curves are the best fit to the ZA mode along the direction in the BZ chosen to calculate the lattice thermal conductivity.](layered-grueneisen.pdf) where $\omega_s(q)$, $\omega_+$ and $\omega_-$ are the wave vector dependent phonon frequency of mode $s$, phonon frequency under positive and negative biaxial strain respectively. $a_0$ and $d a$ are respectively the relaxed lattice constant and difference in lattice constants when under positive and negative biaxial strain. We find that the acoustic modes correspond to the lowest Grüneisen parameters which is in-line with experimentally measured $\gamma$ [@sanjurjo83]. As in the case of graphene, the out-of-plane acoustic transverse mode has the largest negative $\gamma$ parameters. Positive (negative) Grüneisen parameters suggest a decrease (increase) in phonon frequencies as the lattice constant increases. Near the long-wavelength limit, $\gamma_{ZO^{'}}$ for 5LBN is positive but becomes negative as we move along the $\Gamma$ to Y direction in the BZ. $\gamma_{ZO^{'}}$, associating with the layer breathing mode suggests that due to the additional layers the atom vibrations along the perpendicular direction between them lose their coherence and hence decreases the phonon frequencies when the system is under a biaxial strain. As described in table \[ct\], E$_{\rm 2g}$, E$^{''}$, A$_{\rm 1g}$, E$_{\rm g}$, and A$_{\rm 1}^{''}$ are Raman active and hence in principle their Grüneisen parameter can be calculated experimentally using Raman spectroscopy. There exist experimental data for bulk [*h*]{}-BN but to the best of our knowledge there does not exist experimental data for single or MLBN. We therefore compare our results to that of bulk-[*h*]{} BN. The lowest Grüneisen parameters along the $\Gamma$-K-M directions for the TO and LO modes were found to be 1.72 and 1.59, respectively. Our calculations for these modes are only $\sim$ 1.1% and $\sim$ 1.3% larger than the experiment values of Sanjurjo [*et al.*]{} [@sanjurjo83] who have obtained the Grüneisen parameters by measuring the pressure dependence of Raman lines. The slight deviance from the experimental measured value could be because the measured values are for Zinc-blende-BN and not hexagonal BN. Analytical solutions to the Callaway-Klemens’s Approach ------------------------------------------------------- In order to compare the results obtained from the real space super cell approach (ShengBTE), we now study the mode, temperature and length dependence of single and MLBN calculated using the Callaway-Klemens’s approach as described earlier. We first obtain analytical solutions for each acoustic mode of the Phonon BTE by making some reasonable approximations to understand the basic behavior of temperature and length dependence of $\kappa_L$. In order to compare with the experimental results, we resort to exact numerical computation. We have carried out all the length dependent calculations at a constant temperature for MLBN at RT. The corresponding length dependent curves for MLBN are plotted in Fig. \[kaLATAZA\] (e). The parameters used in our study are shown in Table \[para\]. [c|c|c|c|c|c|c]{} System & ---------- $v_{LA}$ (m/s) ---------- : \[para\] Relevant parameters used in the calculations for the analytical solutions of the lattice thermal conductivity. & ---------- $v_{TA}$ (m/s) ---------- : \[para\] Relevant parameters used in the calculations for the analytical solutions of the lattice thermal conductivity. & $\gamma_{LA}$ & $\gamma_{TA}$ & ----------------------------- $\alpha$ $\times$ 10$^{-7}$ (m$^2$/s) ----------------------------- : \[para\] Relevant parameters used in the calculations for the analytical solutions of the lattice thermal conductivity. & ----------------------------- $\beta$ $\times$ 10$^{-20}$ (1/m$^2$) ----------------------------- : \[para\] Relevant parameters used in the calculations for the analytical solutions of the lattice thermal conductivity. \ SLBN & 17020.1 & 11599.8 & 1.546 & 0.452 & 3.99 & -6.827\ BLBN & 16379.4 & 11474.9 & 1.585 & 0.5673 & 3.75 & -6.086\ 5LBN & 21095 & 11420.6 & 1.48 & 0.424 & 4.2 & -6.348\ Bulk-[*h*]{}BN & 16379.4 & 11474.9 & 1.57 & 0.59 & 3.72 & -7.18\ Equations \[wLATA\] and \[wZA\] are plotted in Fig. \[phdos\] and Equation \[gZA\] is plotted in Fig. \[gp\] to compare the analytical fit to the actual phonon dispersion and Grüenisen parameters. ![\[kaLATAZA\] Acoustic modes and temperature dependence of lattice thermal conductivity for (a) SLBN (b) BLBN (c) 5LBN and (d) Bulk-[*h*]{}BN at a constant length. The theoretical calculations are carried out by using Eq. \[LATA\] for the LA and TA modes while Eq. \[ZA\] was used for the ZA mode. The parameters used in our calculations are shown in Table \[para\]. The colour representation for each mode are shown on the right. The black dots are the experimental measurements [@sichel76; @jo13; @wang2016]. Length dependence is worked out by varying $L$ in Eq. \[wmin\]](kappa-analytical-L.pdf) The individual contributions of each of the acoustic modes LA, TA, ZA and the sum of these, [*i.e.*]{} $\kappa_L$, for single and multilayered [*h*]{}-BN are shown in Fig. \[kaLATAZA\] (a,b,c,d). The variation of $\kappa_L$ values for BLBN and Bulk-[*h*]{}BN are quite similar but are lower for 5LBN. This is in good agreement with experiments [@sichel76; @jo13; @wang2016]. In all cases it is observed that amongst the acoustic modes the TA contribution is the largest, ZA to be the least whereas LA contribution is somewhere in between. It has been quite controversial as to which acoustic mode contributes the most to the total lattice thermal conductivity. For example, while some reports [@kuang16; @lindsay11; @lindsay2011; @seol10; @lindsay2010] show that the contributions from ZA to be the most dominant, there are many other reports [@shen14; @alofi13; @kong2009; @aksamija11; @nika2009; @nika11; @nika12; @wei14] that show exactly the opposite. Our analytical results concur with the latter, [*i.e.*]{} the contribution from the ZA mode is the least. The thermal conductivity for two-dimensional layered materials given by Eq. \[k\] is derived assuming both phonon energy dispersions and phonon scattering rates are weakly dependent on the direction of the Brillouin zone [@nika2009]. The calculation of $\kappa_L$ should be independent of the direction chosen resulting in an isotropic in-plane scalar $\kappa_L$. Calculation of $\kappa_L$ should therefore be independent of direction chosen. We move along the $\Gamma$ to K direction for systems in which a hexagonal unit cell is used and along $\Gamma$ to Y in the case of an orthorhombic unit cell. SLBN has the highest calculated $\kappa_L$, 5LBN has the least while $\kappa_L$ lies in between BLBN and Bulk-[*h*]{}BN. From Fig. \[kCK\] it can be easily seen that for temperatures below 100K, the contribution to the total $\kappa_L$ is mainly due to the flexural ZA modes. As in the case of graphene, SLBN can have a total of 12 processes involving the flexural phonons. However, Seol [*el al*]{} [@seol10] obtained a selection rule for the three-phonon scattering. This rule states that only an even number of ZA phonons is allowed to be involved in each process. Shen [*et al*]{} [@shen14] have listed four flexural allowed processes. Hence, the scattering rate of the Umklapp phonon-phonon process for the acoustic flexural branch is multiplied by a factor of $\frac{4}{12}$ and the relaxation time for the ZA mode becomes 3 times of that of Eq. \[tau\]. Therefore besides having a larger velocity and a smaller averaged Grüneisen parameters compared to the other systems, the major contribution for an increased $\kappa_L$ is due to the symmetry of the ZA mode. Phonon dispersions and Grüneisen parameters for BLBN and Bulk-[*h*]{}BN are very similar which explains why their calculated $\kappa_L$ have the same magnitude. In the case of 5LBN, there are additional five low frequency modes (also termed as layer-breading modes), which arise due to the interlayer moment. Due to this change in phonon dispersion, more phase-space states become available for phonon scattering and therefore decreases $\kappa_L$ [@balandin11]. Numerical solutions to the Callaway-Klemens’s Approach ------------------------------------------------------ Numerical calculations are carried out using the exact form of the phonon dispersion and Grüneisen parameters as displayed in Fig. \[phdos\] and Fig. \[gp\] rather than the analytical form of the acoustic modes and averaged Grüneisen parameters. We numerically solve Eq. \[k\] for each of the modes at a constant sample length varying temperature as well as at a constant temperature varying lengths between 0.1 to 10 $\mu$m. These results are compared with experimental data [@sichel76; @jo13; @wang2016] and shown in Fig. \[kCK\]. Numerically calculated values of $\kappa_L$ are in better agreement with the experimental data as compared to the analytical form. We find the contribution from the ZA modes to dominate at lower temperatures but rapidly decreases as the temperature increases making the flexural modes contribute the least at relatively higher temperatures. This is in line with previous theoretical calculations [@aksamija11]. ![\[kCK\] Acoustic modes and temperature dependence of lattice thermal conductivity for (a) SLBN (b) BLBN (c) 5LBN and (d) Bulk-[*h*]{}BN at a constant length. The theoretical calculations are carried out by solving Eq. \[k\] numerically for each of the modes. The colour representation for each mode are shown on the right. The black dots are the experimental measurements [@sichel76; @jo13; @wang2016]. Length dependence is worked out by varying $L$ in Eq. \[wmin\].](kappa-numerical-L.pdf) Summary ======= Phonon dispersions using a LDA pseudopotential with vdW interactions, density of states (DOS), the Grüneisen parameters and the lattice thermal conductivity have been calculated by the Callaway-Klemens and Real space super cell approach for SLBN, BLBN, 5LBN and Bulk-[*h*]{}BN. Additionally, in the case of Bulk[*h*]{}-BN, we calculate the two-phonon DOS. Irreducible representation using the character table at the $\Gamma$ point in the BZ for each of the systems have been derived in order to compare the symmetry modes with those obtained from Raman and infrared spectroscopy experiments. Results from the investigations by EELS data, Raman, second-order Raman and Infrared spectroscopy are found to be in excellent agreement with the theoretical calculations based on the phonon dispersion, DOS and two-phonon DOS which rely on the harmonic second order inter atomic force constants. Further, we have calculated the sample length and temperature dependence of lattice thermal conductivity by the real space super cell approach with the help of the second order IFCs calculated by DFPT. Lattice thermal conductivity at the thermodynamic limit for each system has a maxima between the 110-150 K. For sample sizes in the range 1-5 $\mu$m, $\kappa_L$ does not have a maxima. However with increase in temperature it tends to saturate at a value which is an order smaller than the thermodynamic limit. Our mode dependent calculations using the real space method suggests that the majority of the contribution to the thermal conductivity are due to the ZA phonons for all of the structures. The substantial decrease in $\kappa_L$ from single to MLBN is because of the additional layer breathing modes but mainly due to the fact that the interlayer interactions breaks the SLBN selection rule resulting in suppressing the ZA phonons contributions to $\kappa_L$ in MLBN. Contribution to $\kappa_L$ from the in-plane modes are not sensitive to the number of layers and have a lower $L$ dependence compared to the out of plane modes. This reduction in $\kappa_L$ from SLBN to MLBN which is mainly due to the ZA phonons indicate that the interlayer interactions are short ranged. The $L$ dependence of the TA and LA contributions to $\kappa_L$ saturate to their thermodynamic limit faster than that of the contribution from the ZA phonons implying that the ZA phonons travel ballistically along the sample while the TA phonons travel diffusively. Grüneisen ($\gamma$) parameters were obtained from first principle calculation by applying a positive and negative biaxial strain. For the in-plane acoustic modes, we find that $\gamma$ does not vary much from its mean value but the out-of-plane modes have a strong $q$-dependence. Our calculated $\gamma$ values for Bulk-[*h*]{}BN at the $\Gamma$ point is $\sim$ 1% larger than those obtained from experiments which measures the pressure dependence of Raman lines. $\gamma$ parameters for 5LBN suggest that due to the layer breathing modes, atoms along the perpendicular direction lose their coherence between each layer and decrease the phonon frequencies when under a biaxial strain. In comparison to the real space super cell approach, lattice thermal conductivity has been calculated, both analytically and numerically, using Callaway-Klemens formalism. To obtain analytical solution of the phonon, we make a linear fit to the LA and TA modes, a quadratic fit to the ZA mode, and use an averaged value for its Grüneisen parameters for the $\gamma$ parameters corresponding to the in-plane acoustic modes and an inverse square wave-vector dependence $\gamma$ for the out-of-plane modes. Theoretical results for sample length and temperature dependence of $\kappa_L$ are in good agreement with experimental observation. The phonon BTE is then solved analytically and numerically for SLBN, BLBN, 5LBN and Bulk-[*h*]{} BN to calculate $\kappa_L$ for a constant length over a wide range of temperatures and [*vice versa*]{} again in good agreement with available experimental results. Both the theoretical approaches, [*i.e.*]{} real space super cell and Callaway-Klemens, show the same magnitude for $\kappa_L$ but the temperature dependence by the two methods are different. The lattice thermal conductivity for these materials are practically length independent for sample lengths greater than 100 $\mu$m which tends to their thermodynamic limit. Calculated values for $\kappa_L$ for BLBN and 5LBN agree very well with experiments when calculated by the real space approach rather than by the Callaway-Klemens method. This may be because the experimental behavior of $\kappa_L$ for both BLBN and 5LBN tend to saturate at higher temperatures instead of having a maxima. However, the Callaway-Klemens method agrees better with available experimental data for Bulk-[*h*]{} BN. Further experiments could resolve this discrepancy. Mode dependent numerical calculations using the Callaway-Klemens formalism suggest that ZA modes dominate only at very low temperatures and have the least contribution as the temperature is increased. This is in stark conflict with our calculations based on real space super cell approach. Since the velocities and Grüneisen parameters are extremely similar for single and bi layer boron nitride, one would expect $\kappa_L$ for both the systems to be similar. However, in the case of graphene, we have a significant reduction in $\kappa_L$ which is seen both experimentally [@hongyang2014] and theoretically [@lindsay2011; @RDSM17]. The larger $\kappa_L$ in SLBN in comparison to BLBN using the Callaway-Klemens method was due to the symmetry put by hand and not a consequence of the theory. This implies that the closed form of the relaxation time used in Callaway-Klemens method is a poor approximation having little predictive value and one must solve the BTE beyond the RTA. Our calculations suggests that for an enhanced figure of merit, $ZT$, in such materials, the sample length must be in the $\mu$m range or smaller and should be stacked on top of each other. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Dr. Jesús Carrete of the Technical University, Vienna, for his insightful correspondence on the ShengBTE code for the calculation of the mode dependence contribution to the total thermal conductivity. We also thank D.L. Nika of Moldova State University and A.A. Balandin of the University of California, Riverside, for their helpful correspondence based on their recent publications. All calculations were performed in the High Performance Cluster platform at the S.N. Bose National Centre for Basic Sciences. RD acknowledges support through a Senior Research Fellowship of the S.N. Bose National Centre for Basic Sciences. Derivation of Irreducible representations ========================================= We define the reducible representation ($\Gamma_{\rm red}$) by placing three vectors on each atom in the unit cell which will obey the following rules when operated by a symmetry transformation. - If a vector is not moved (reversed) by an operation, it contributes 1 (-1) to $\chi$. - If a vector is moved to a new location by an operation, it contributes 0 to $\chi$. where $\chi$ is the character in the reducible representation. Our reducible representations ($\Gamma_{red}$) are shown in the column before every new point group representation in table \[ct\]. Using the reduction formula, $a_i = \frac{1}{g} \sum \chi_R \chi_{IR}$, where $a_i$ is the number of times an irreducible representation contributes to the reducible representation, $g$ is the total number of symmetry operations for a particular point group and $\chi_R$ ($\chi_{IR}$) is the corresponding character in the reducible (irreducible) representation, we derive the irreducible representations. D$_{{\rm 6h}}$ E 2C$_{6}$ 2C$_3$ C$_2$ 3C$_2^{'}$ 3C$_2^{''}$ i 2S$_6$ 2S$_3$ $\sigma_{\rm h}$ 3$\sigma_{\rm v}$ 3$\sigma_{\rm d}$ Basis ------------------------------------- ---- ---------- -------- ------- ------------ ------------- ------------------ -------- -------- ------------------ ------------------- ------------------- ---------------- A$_{\rm 2u}$ 1 1 1 1 -1 -1 -1 -1 -1 -1 1 1 $z$ B$_{\rm 1g}$ 1 -1 1 -1 1 -1 1 1 -1 -1 -1 1 $yz(3x^2-y^2)$ B$_{\rm 2g}$ 1 -1 1 -1 -1 1 1 1 -1 -1 1 -1 $xz(x^2-3y^2)$ E$_{\rm 2g}$ 2 -1 -1 2 0 0 2 -1 -1 2 0 0 {$x^2-y^2,xy$} E$_{\rm 1u}$ 2 1 -1 -2 0 0 -2 1 -1 2 0 0 {$x,y$} $\Gamma^{bulk-{\it h}BN}_{\rm red}$ 12 0 0 0 0 -4 0 -8 0 4 4 0   $\Gamma^{SLBN}_{\rm red}$ 6 0 0 0 -2 0 0 -4 0 2 0 2   D$_{\rm 3d}$ E 2C$_3$ 3C$_2^{'}$ i 2S$_6$ 3$\sigma_{\rm d}$ A$_{\rm 2u}$ 1 1 -1 -1 -1 1 $z$ A$_{\rm 1g}$ 1 1 1 1 1 1 $z^2$ E$_{\rm g}$ 2 -1 0 2 -1 0 {$xz, yz$} E$_{\rm u}$ 2 -1 0 -2 1 0 {$x, y$} $\Gamma^{BLBN}_{\rm red}$ 12 0 4 0 0 0 D$_{\rm 3h}$ E 2C$_3$ 3C$_2^{'}$ $\sigma_{\rm h}$ 2S$_3$ 3$\sigma_{\rm v}$ A$_1^{'}$ 1 1 1 1 1 1 $z^2$ A$_2^{''}$ 1 1 -1 1 1 -1 $z$ E$^{'}$ 2 -1 0 2 -1 0 {$x, y$} E$^{''}$ 2 -1 0 -2 1 0 {$xz, yz$} $\Gamma^{5LBN}_{\rm red}$ 30 0 10 2 -4 -2 [63]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\ 12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****,  ()]{} @noop [ ()]{} @noop [****,  ()]{} @noop [****, ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****, ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****, ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} in @noop [**]{}, Vol. ,  (, , ) p.  @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{} @noop [****,  ()]{}

--- author: - | Chris Cummins[^1]\ School of Informatics\ University of Edinburgh\ `c.cummins@ed.ac.uk`\ Zacharias V. Fisches^\*^\ Department of Computer Science\ ETH Zurich\ `zfisches@student.ethz.ch`\ Tal Ben-Nun\ Department of Computer Science\ ETH Zurich\ `talbn@inf.ethz.ch`\ Torsten Hoefler\ Department of Computer Science\ ETH Zurich\ `htor@inf.ethz.ch`\ Hugh Leather\ School of Informatics\ University of Edinburgh\ `hleather@inf.ed.ac.uk`\ bibliography: - 'refs.bib' title: | ProGraML: Graph-based Deep Learning for\ Program Optimization and Analysis --- =1 [^1]: Both authors contributed equally