Datasets:

LatestEval

/

arxiv-latest

like 0

Follow

LatestEval 1

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). A significant outcome concerning the convergence of bulk eigenvalues of H_N is discussed here. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What does the research reveal about the behavior of bulk eigenvalues of H_N under certain assumptions?

Under some assumptions on the potentials, it proves that the bulk eigenvalues of H_N converge to a Poisson process after appropriate re-scaling.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. An example of convergence to a specific distribution type for the top eigenvalue of H_N is mentioned here. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

Can you provide an example of a specific distribution that the top eigenvalue of H_N converges to under a given type of distribution for 𝔞?

Assuming 𝔞 has a doubly exponential type upper tail distribution, it is known that the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class.

example

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). Information regarding the phase transition verification in bulk statistics at a critical value of α is presented here. In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What has been verified concerning the phase transition in bulk statistics for a certain critical value of α?

The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. The basis for the derivation of the large deviation profile for matrices with sub-Gaussian entries is explained here. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What was the derivation of the large deviation profile of the largest eigenvalue of matrices with sub-Gaussian entries based on?

The large deviation profile of the largest eigenvalue of matrices with sub-Gaussian entries was derived from the collective behavior of random potentials.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. The intent behind the study of large deviation for heavily-tailed random variables is clarified here. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What is the purpose of studying the large deviation of spectral statistics for random variables with heavy tails?

The large deviation of spectral statistics for random variables with tails heavier than Gaussian was to account for deviations due to few outlier potentials taking very large values.

purpose

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. The behavior and distribution of the top eigenvalue when matrix entries have a finite fourth moment are summarized here. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What is the behavior of the top eigenvalue when matrix entries have a finite fourth moment in relation to the semicircle law and its distribution at the edge?

When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of the semicircle law, and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . A summarization of what influences the large deviation profile and the asymptotic distribution of the top eigenvalue of H_N^α is noted here. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

What factor governs both the large deviation profile and the asymptotic distribution of the top eigenvalue of H_N^α?

The large deviation profile and the asymptotic distribution of the top eigenvalue of H_N^α is always governed by outlier potentials of 𝔞 taking unusually large values.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. The paper's main results, focusing on the top eigenvalue's large deviation profile and asymptotic distribution, are outlined here. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

Can you summarize the main results presented in this paper?

The main results of the paper outline the large deviation profile, and the asymptotic distribution of the top eigenvalue of the random operator H_N^α, with detailed conditions and limiting behaviors.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. A comparison between the upper and lower tail estimates' precision is provided here. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. There has also been some recent investigations on the matrix model (), where a scaling limit of top eigenvalue (when α=3/2) was derived in , and the bulk localizing properties were investigated in when the random variables have heavy tails. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

How do the upper tail estimates compare to the lower tail estimates in terms of accuracy according to the paper?

The upper tail estimates in cases (1) and (2) are sharp, while the lower tail estimates are crude and may not be optimal.

summary

arxiv

2312.12146v1

INTRODUCTION The theory of random Schrödinger operators gained much success in describing various phenomena in mathematical physics. A 1-d random Schrödinger operator in its finite volume version can be described as the following operator defined on {0,1,⋯,N} H_Nφ(k)=φ(k-1)+φ(k+1)+𝔞(k)φ(k), φ(0)=φ(N)=0, where 𝔞(k),k=1,⋯,N, are certain random potentials. It corresponds to the matrix representation H_N=H_N^∞+Λ_N^0, where H_N^∞=[ 0 1 0 ⋯ 0; 1 0 1 ⋱ 0; ⋮ ⋱ ⋱ ⋱ ⋮; 0 ⋯ 1 0 1; 0 ⋯ 0 1 0 ], and Λ_N^0=diag(𝔞(1),⋯,𝔞(N)). Under some slight assumptions on the potentials 𝔞(·), proves that the bulk eigenvalues of H_N converge to a Poisson process after an appropriate re-scaling. Further, assuming that 𝔞 has a doubly exponential type upper tail distribution, it is known that λ_1(N), the top eigenvalue of H_N, converges to a Gumbel distribution of max-order class. To investigate the transition from localized phase to de-localized phase, a variant of H_N was introduced in (for a modified model) and where the potentials are vanishing with N, i.e. H_N^α=H_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)). The transition from localized to de-localized phase has been verified the bulk statistics at the critical value α=1/2, see for example . In this work we investigate the asymptotics of the top eigenvalue of H_N^α with decaying diagonal potentials (). When the potentials 𝔞 are not too heavy-tailed, it is easy to check that the top eigenvalue of H_N^α converges to 2 almost surely, and when 𝔞 has heavier tails, the top eigenvalue can be arbitrarily large with positive probability. We identify the large deviation profile of the top eigenvalue in the former light-tailed case, then identify the threshold of criticality, and finally derive the law of the top eigenvector in the critical case. When 𝔞(·) has even heavier tails, we verify that the top eigenvalue has a Poisson distribution. This work is partially inspired by the recent progress in large deviation of Wigner matrices. When the matrix has sub-Gaussian entries with a bound on its Laplace transform, the large deviation profile of the largest eigenvalue was derived via spherical integral in , where it was the collective behavior of random potentials, instead of the behavior of one outlier potential, that accounts the large deviation profile. When the random variables have a tail heavier than Gaussian, the large deviation of spectral statistics was derived in (and see for deviation of the largest eigenvalue in this regime) where the deviation is accounted for by few outlier potentials taking very large values. For heavy-tailed wigner matrices, the distribution of the top eigenvalue was also well studied by many authors. When matrix entries have finite fourth moment, the top eigenvalue sticks to the edge of semicircle law , and the rescaled top eigenvalue has a Tracy-Widom fluctuation at the edge. When the matrix entries do not have a finite fourth moment, it is verified in and that the largest eigenvalue has a Poisson distribution. Finally, when the density of the matrix entries have a tail decay x^-4dx, the distribution of the top eigenvalue is explicitly determined in . For the random Schrödinger operator H_N^α, we will show that the large deviation profile, and the asymptotic distribution of the top eigenvalue is always governed by outlier potentials of 𝔞 taking unusually large values. We also study a model of close relevance: consider the matrix G_N^∞= [ 0 √((N-1)/N) 0; √((N-1)/N) 0 √((N-2)/N); ⋱ ⋱; 0 √(2/N) 0; √(2/N) 0 √(1/N); 0 √(1/N) 0 ]. This model is closely related to the matrix model of beta ensembles, introduced by Domitriu and Edelman in . Indeed, G_N^∞ can be thought of the zero temperature model in this beta-parameterized matrix family (). We are also interested in random operators of the following form G_N^∞+Λ_N^α, Λ_N^α=diag(N^-α𝔞(1),⋯, N^-α𝔞(N)), but in order to foster better analytic treatment, we will instead work with an orthogonal invariant version G_N^αdef=G_N^∞+U_NΛ_N^α U_N^t, where U_N is a Haar distributed orthogonal matrix of size N, independent of the random variables 𝔞(·), and U_N^t denotes the transpose of U_N. We will verify that the large deviation profiles of the top eigenvalue of H_N^∞ and G_N^∞ are very similar, except that the precise expression of large deviation rate function is different. §.§ Statement of main results The main results of this paper are outlined as follows: Consider the random operator H_N^α defined in (), and denote by λ_1(N) the largest eigenvalue of H_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that for some β>0, some C>0, 0<C_1<C_2<∞ we have for any t>1, C_1exp(-Ct^β)≤ℙ(𝔞>t)≤ C_2exp(-Ct^β), ℙ(𝔞<-t)≤ C_2exp(-Ct^β), then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(√(λ^2-4))^β, and for any λ<2, we have the lower tail estimate: there exists some constants C_2,C_3>0, and when β<2 there exists some sufficiently small c>0 such that ℙ(λ_1(N)<λ)≤ C_2e^-C_3N^αβ+1, β∈ [2,∞), C_2e^-C_3N^αβ+c, β∈(0,2). The constant C_3>0 depends on λ, and can be written as an explicit function of λ. See Remark for a discussion. * (Sub-critical) Assume that for some β>1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(√(λ^2-4))^-β, and for any λ<2 we have the lower tail estimate: when β>2, for any λ<2 and any ϵ>0 we can find C_2=C_2(ϵ)>0 depending on ϵ and λ such that ℙ(λ_1(N)<λ)≤ (C_2N)^-(α(β-ϵ)+1) and when β∈(0,2], for any λ<2 we can find sufficiently small c>0 and some C_2=C_2(β,c,λ) such that ℙ(λ_1(N)<λ)≤ (C_2N)^-c. * (Critical regime) Assume there exists some C>0,C_1>0 such that for any t>1, ℙ(𝔞>t)=Ct^-1/α, ℙ(𝔞<-t)≤ C_1t^-1/α, then almost surely λ_1(N)law→√(ξ^2+4), N→∞, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. * (Randomness dominating) Assume that for some 0<β<1/α, C>0,C_1>0 we have for all t>1, ℙ(𝔞>t)= Ct^-β, ℙ(𝔞<-t)≤ C_1t^-β, then N^α-1/βλ_1(N)law→ξ, N→∞, where ξ is a non-negative valued random variable satisfying for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-β. The upper tail estimates in cases (1) and (2) are sharp, yet the lower tail estimates derived in cases (1) and (2) are crude and may not be optimal. A related example is the lower tail of KPZ equation with narrow wedge initial condition , where the exact lower tail estimate was derived via computing the exponential moment of the Airy-2 process, which admits an interpretation via the stochastic Airy operator. The stochastic Airy operator is closely related to the theme of this paper, yet it is defined via Brownian motion and thus a lot of analytical techniques (like the Girsanov transform) can be used. In our paper, we consider different moment assumptions on the random variables, so that the analytical techniques in the context of stochastic Airy operator are not available to us for deriving a sharper lower tail estimate. The rationale for recent research on the matrix model with heavy-tailed random variables is outlined here. We can also derive the limiting distribution of the second, third, fourth, etc. largest eigenvalues of H_N^α, denoted by λ_i(N),i≥ 2. All the derivations depend on the distribution of the largest value of 𝔞(1),⋯,𝔞(N) and the finite rank perturbation formula, so that explicit distributions of the second, third, fourth, etc. eigenvalues can be written out explicitly. In cases (1) and (2), we can check that the deviation probabilities of λ_i(N),i≥ 2 are negligible compared to e^-N^αβ and N^1-αβ respectively, and is thus negligible compared to deviation of λ_1(N). In cases (3) and (4) however, deviations of λ_i(N),i≥ 2 have the same magnitude as that of λ_1(N), and we have the following point process characterization: In case (3) of Theorem , for any ϵ>0, the largest eigenvalues of H_N^α in (2+ϵ,∞) converges to a random point process which is the image under the map x↦√(x^2+4) of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. In case (4) of Theorem , the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯, where λ_1(N)≥λ_2(N)≥⋯ denote the eigenvalue of H_N^α in decreasing order, converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0. The Poisson distribution claimed in (4) has been verified for several other random matrix models with heavy tail distribution, see . Meanwhile, in all these claims, the assumption that the tail estimate for 𝔞 holds for all t>1 can be easily changed to holding for all t larger than a fixed number. In the following we summarize the results on top eigenvalues concerning the random operator (). We first introduce a convenient notation: for any λ≥ 2, denote by f(λ) the solution to f(λ)+1/f(λ)=λ that satisfies f(λ)≥ 1. Consider the random operator G_N^α defined in (), where U_N is a Haar distributed unitary matrix independent of the random variables 𝔞(·). We denote by λ_1(N) the largest eigenvalue of G_N^α. Assume the random potentials 𝔞(1),⋯,𝔞(N) are i.i.d. distributed as a real-valued symmetric random variable 𝔞. Then * (Weibull distribution) Assume that the random variables 𝔞(·) satisfy Assumptions (), () Then for any λ>2, we have the upper tail estimate lim_N→∞log(ℙ(λ_1(N)>λ))/N^αβ=-C(f(λ))^β, and for any λ<2 we have the lower tail estimate as in (). * (Sub-critical) Assume that the random variables satisfy () and (). Then for any λ>2, we have the upper tail estimate: lim_N→∞ N^αβ-1ℙ(λ_1(N)>λ)=(f(λ))^-β, and for any λ<2 we have the lower tail estimates in exactly the same form as in (), () (though the numerical constants may be slightly different). * (Critical regime) Assume that the random variables satisfy (), () Then λ_1(N)law= 2, ξ≤ 1, ξ+1/ξ, ξ>2, where ξ is a non-negative-valued random variable that satisfies, for any λ>0, ℙ(ξ>λ)=1-e^-Cλ^-1/α. More generally, the largest eigenvalues of G_N^α in (2, ∞) converges to a random point process which is the image under the map x↦ x+1/x of the Poisson process on [0,∞) with intensity C/α x^1/α+1, x>0. * (Randomness dominating) Assume that the random variables satisfy (), (). Then the top eigenvalue λ_1(N) satisfies (). Meanwhile, the point process N^α-1/βλ_1(N)≥ N^α-1/βλ_2(N)≥⋯ converges to the Poisson point process on [0,∞) with intensity Cβ/ x^β+1, x>0, where λ_1(N)≥λ_2(N),⋯≥ denote the eigenvalues of G_N^α in decreasing order. The matrix model G_N^∞ () is closely related to the Gaussian beta ensemble. The small deviations of top eigenvalue of Gaussian beta ensembles were derived in . Note that the precise large deviation rate function was not derived in . In this paper we derive the precise large deviation rate function, but for a modified model with orthogonal invariance. From a different perspective, there has been a number of recent works concerning deviations of the top eigenvalue of Gaussian beta ensembles at the high temperature limit, including and . These works mainly consider the double scaling regime of the Gaussian beta ensemble in which β→ 0 and Nβ→∞, and is very different from the models we investigate in this paper where there is only one scaling. §.§ Plan of the paper In Section we prove Theorem and . Then in Section we prove Theorem .

Why were recent investigations carried out on the matrix model in relation to heavy-tailed random variables?

Recent investigations on the matrix model aimed to derive a scaling limit of top eigenvalue and investigate the bulk localizing properties when the random variables have heavy tails.

purpose

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. A summary explanation for Y Oph's exclusion from period-luminosity relation calibration is provided here. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

Why is Y Ophiuchi not usually included in the calibration of the period-luminosity relation?

Y Ophiuchi (Y Oph) is generally not taken into account in the calibration of the period-luminosity relation because of its unusual behaviour.

summary

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). The effect of color excess on the accuracy of distance measurements for Y Oph is discussed here. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

What could bias the distance measurements of Y Oph when based on photometric relations?

The observed color excess E(B-V)=0.650 of Y Oph near the galactic plane might bias the distance measurements based on photometric relations if the reddening is not well known.

purpose

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. A summary of Y Oph's distance determined through infrared interferometry is mentioned here. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

What is the distance to Y Oph as derived from infrared interferometry?

The derived distance from infrared interferometry of Y Oph is d=491±18pc.

summary

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. An instance of a significant discrepancy in distance measurements for Y Oph is cited here. The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

Can you give an example of a discrepancy in distance measurements for Y Oph?

The distance obtained from Gaia DR3 parallax for Y Oph is strongly discrepant, with d=742± 21pc.

example

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. An example of consistent parallax-based distance measurements is illustrated here. Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

Can you provide an example where RUWE and G-band magnitude measurements are consistent with other distance measurements?

T Vul has both comparable RUWE and magnitude in G-band with Y Oph, and its parallax-based distance is consistent with HST parallax and BW distance from the literature.

example

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . The reason for questioning the pulsation mode of Y Oph is provided here. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

Why is it questionable to assume that Y Oph pulsates in the first overtone mode?

It is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period.

purpose

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . An explanation for the small amplitude variations in Y Oph is given here. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

What is an example of a reason for the small amplitude variations in Y Oph that is not related to the first-overtone mode?

The small amplitudes of Y Oph are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip.

example

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. An example discussing the binary characteristics of Y Oph is mentioned here. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

Can you give an example of a discovered property of Y Oph pertaining to its binary nature?

Y Oph is a spectroscopic binary system having an orbital period of 2612days and its effect on photometry and radial velocity are likely marginal.

example

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . The objective for applying hydrodynamical pulsation models to Y Oph is outlined here. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. is a one-dimensional Lagrangian convective code that was already used for modelling RR Lyrae, BL Her and classical Cepheids . On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

What is the purpose of using hydrodynamical pulsation models to study Y Oph?

The purpose of constraining Y Oph properties with hydrodynamical pulsation models is to explain its low radial velocity and light curve amplitude, brightness, and distance discrepancy.

purpose

arxiv

2312.12046v1

INTRODUCTION The discovery of the Y Ophiuchi (Y Oph) variability, which has a pulsation period of about 17days, dates back to 130years ago . At the same epoch, the period-luminosity relation (hereafter PL relation) of Cepheids was discovered . Y Oph, however, is generally not taken into account in the calibration of the PL relation because of its unusual behaviour. Indeed, using Baade-Wesselink (BW) method, found that this star is 1mag fainter than what would be derived using PL relation. In other words, the luminosity derived by the BW method corresponds to a Cepheid with a pulsation period of 10days or less. Several variants of BW analysis inferred the distance of Y Oph, but they are scattered between about 400 and 650pc . Most of these methods used the Surface Brigthness Color Relation (SBCR). However, Y Oph is close to the galactic plane and exhibits an important color excess of about E(B-V)=0.650 which might bias the distance measurements based on photometric relations if the reddening is not well known. Among these BW variants, observed directly the angular diameter along the pulsation cycle with infrared interferometry. The advantage is to not use SBCRs which are sensitive to empirical calibrations and interstellar extinction. The derived distance from this method is d=491±18pc. On the other hand, the distance obtained from Gaia DR3 parallax is strongly discrepant with d=742± 21pc . The renormalized weight error (RUWE) of Y Oph indicates a good quality of the parallax (RUWE = 1.04<1.4), and, thus, can be used for astrometry. Nevertheless, Y Oph slightly saturates the detector with G=5.6mag. Indeed, warns that stars with G<6mag generally have inferior astrometry due to calibration issues. Despite small saturations we do not expect such large errors of measurements that would explain the difference obtained with BW methods. As an example, T Vul has both comparable RUWE and magnitude in G-band with Y Oph, but the distance inferred from its parallax d=581±20pc is consistent with HST parallax and BW distance from the literature . Moreover, the color variation of Y Oph is smaller than others Cepheids due to its small light curve amplitude, thus the chromaticity effects on Gaia detector are likely mitigated. Because of its low amplitude, Y Oph is sometimes supposed to be a first-overtone pulsator . However, it is questionable to assume that Y Oph pulsates in the first overtone since this mode is usually excited for Cepheids of much shorter pulsation period. From observations, the longest pulsation period for a securely identified first-overtone Cepheid is 7.57day in the case of V440 Per . In the OGLE Milky Way survey, the longest overtone period claimed is 9.437day . Another explanation for the small amplitudes is proposed by . The small amplitudes are not a direct result of the first-overtone mode, but are associated with the proximity to the blue edge of the fundamental mode instability strip. This might be the case for Y Oph which exhibits a mean effective temperature of about 5800K as measured by spectroscopic observations . This is a high temperature compared to the stars of the same pulsation period which have a mean temperature of about 5500K as in the the case for example for CD Cyg (P=17.07day) or SZ Aql (P=17.14day) . Several studies have shown spectroscopic binarity evidences of Y Oph, although its effect on photometric and radial velocity measurements remains negligible. suggested that Y Oph is a spectroscopic binary system having an orbital period of 2612days and concluded that its effect on photometry and radial velocity are likely marginal. extended the analysis of the mean radial velocity variation and derived an orbital period of 1222.5days. However, the companion was not detected in IUE spectra and NACO lucky imaging which yields a magnitude difference of at least 2.5mag in the K_s band . More recently, a third companion of much longer orbital period (about 28000days) is suspected from light-travel time effect observed in O-C diagram . In order to give physical insights to this intriguing star, in particular to explain its low radial velocity and light curve amplitude, brightness and distance discrepancy, we aim at constraining Y Oph properties with hydrodynamical pulsation models. Non-linear pulsation modeling offers a valuable approach for exploring the underlying physics of variable stars, as demonstrated for the first time by . Additionally, it provides a tool to accurately determine the mass of Cepheids, a crucial factor when considering the mass discrepancy associated with these stars . Pulsation models are also useful to determine the distance of Cepheids . In the case of Y Oph, a non-linear model was proposed to fit radial velocity measurements and V-band apparent magnitude . These authors found that a 7M_⊙ model is consistent with short distance and low absolute luminosity in V-band (respectively 423pc and M_V = -3.996mag). However, Y Oph is modeled with an average effective temperature by about 1000K lower than recent high-resolution spectroscopic temperature (T_eff=4720K<5800K). In this paper, we propose for the first time to constrain hydrodynamical pulsation models of Y Oph with an extensive set of observations. On the modelling side, the recently released Radial Stellar Pulsation (RSP) tool in Modules for Experiments in Stellar Astrophysics can be used to model non-linear pulsation of Y Oph. The definition and application of the Radial Stellar Pulsation (RSP) tool are described here. On the observational side, we benefit from a complete set of observations to constrain the models. To this end, we gathered from literature effective temperature and radial velocity curves, angular diameter from interferometric observations and light curves in the visible and the infrared. The paper is organized as follows. We first construct a grid of equilibrium static models and we perform a linear non-adiabatic analysis with in Sect. . Then, we use the solutions that have a positive growth rate for pulsation period of ≈17days to perform a nonlinear calculations for this star in Sect. . We then present our fitting strategy to a full set of observations in Sect. . We finally present the results in Sect. and we discuss the discrepancy between Gaia and BW distances in Sect. . We conclude in Sect. .

What is the Radial Stellar Pulsation (RSP) tool and for what has it been used?

The Radial Stellar Pulsation (RSP) tool is a one-dimensional Lagrangian convective code used for modelling RR Lyrae, BL Her, and classical Cepheids.

terminology

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. A method to control pests without using chemicals is discussed here. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What is one potential alternative strategy mentioned in the text for controlling pests without chemical measures?

One potential alternative strategy to chemical measures is releasing a natural enemy of the pest to control them.

purpose

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. An ecological concept describing the interaction between a natural enemy and a pest is explained here. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

How is the relationship between a natural enemy and a pest described in ecological terms?

In the context of predator-prey relationships, the predator population can eliminate the pest from the system via depredation.

terminology

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. A method to enhance the effectiveness of natural predators in controlling pests is mentioned here. Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

Can you provide an example of a method to enhance the rate of pest depredation by predators?

One way to boost the depredation rate is by providing pests with an alternative food source.

example

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. A summary of factors critical to the effectiveness of biological pest control is provided here. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What factors are critical for the success of biological pest control according to the passage?

The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets.

summary

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. A potential negative impact of uncontrolled predator population growth on ecosystem sustainability is discussed here. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What future sustainability issue is raised concerning unbounded growth of the introduced predator population?

Unbounded growth of the introduced predator population would negatively impact the sustainability of the ecosystem via non-target effects.

future

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. The definition of mutual interference in ecological systems is provided here. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What is mutual interference in the context of ecological interactions?

Mutual interference is defined as the behavioral interactions among feeding organisms that reduce the individual food consumption or acquisition time.

terminology

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. An example of a model that supposedly accurately represents predator-prey interactions is mentioned here. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What is an example of a model that is argued to closely represent true predator-prey dynamics?

Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), which is argued to represent true predator-prey dynamics closely.

example

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. A particular model of interference in predator-prey dynamics is defined here. The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

What does the Beddington-Deangelis functional response refer to?

The Beddington-Deangelis functional response is a model of interference in predator-prey dynamics.

terminology

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . The reason for replacing the Holling type II functional response with the Beddington-Deangelis functional response in the modified system is explained here. Prasad et al. investigate many interesting dynamical phenomena in this model, including stable coexistence at a lower abundance of prey, by varying the mutual interference term. Their analysis is focused on one coexistence state.

Why was the Holling type II functional response replaced in the modified system studied by Prasad et al.?

In the modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response.

purpose

arxiv

2312.11726v1

INTRODUCTION Biotic stressors such as pests, pathogens, and weeds have become a threat to the crop yield and efficacy of food products in various ways. For instance, unwanted harmful pests cause considerable losses of crop yield, both in quantity and quality. The main source to overcome this issue has been chemical measures; particularly including pesticides and insecticides. The consumption of synthetic pesticides such as DDT, BHC, aldrin, dieldrin, captan, and 2,4-D has accelerated by the 1940s (). However, these pesticides/insecticides have many harmful consequences and negative drawbacks both from an economical and an environmental point of view. Particularly, they contaminate the environment, remaining there for an indefinitely long period. Also, many of them are hazardous chemicals that can cause incredibly high risk even to human health. One potential alternative strategy to this is releasing a natural enemy of the pest to control them. From the modeling point of view, this is simply a predator-prey relationship where, the predator population can grow, reproduce and perhaps eliminate the pest from the system via depredation. They are primarily three ways to introduce the natural enemy concept in the field; conservation of existing natural enemies, introducing new natural enemies and establishing a permanent population (classical biological control), mass rearing, and periodic release, on a seasonal basis (,, ). A possible drawback of this approach tends to be incomplete elimination of pests due to an insufficient depredation rate. One way to boost the depredation rate is by providing them with an alternative food source (, , , , ). Some studies such as , , point out that providing non-prey food items would magnify the abundance, and life history traits, such as longevity, fecundity, or foraging behavior of predators; typically which would directly contribute to the efficacy of depredation. The provision of AF may or may not successfully control the target pest. The quality and quantity of alternative food sources play a vital role in the outcomes of biological pest control targets. In this context, effort has been put into exploring several strategies for integrated pest management programs. To this end mathematical models that incorporate AF for the predator in predator-prey systems with have been intensely studied (, , , , ). Although these show that pest extinction can be achieved , there are some possible complications with such an approach. One potential unrealistic feature of these models is the unbounded growth of the introduced predator population, in certain parameter regimes. This would negatively impact sustainability of the ecosystem via non-target effects . In addition, has pointed out that the provision of high quality supplementary food source can cause apparent competition between the two food sources. Consequently, a predator might reduce target prey consumption; perhaps even switch completely to the alternative food source. To avoid the outcome of unbounded predator growth, a possible solution is introducing a limiting factor to this growth (). The limiting behavior can be modeled into the predator dynamics via mutual interference. Mutual interference is defined as the behavioral interactions among feeding organisms, that reduce the time that each individual spends obtaining food, or the amount of food each individual consumes. It occurs most commonly where the amount of food is scarce, or when the population of feeding organisms is large. Early work on mutual/predator interference , modeled the interference term as (x/1+hx)y^m, where y is the predator density, x is prey density, h is the handling time, or the time it takes the predator to kill and ingest the prey, and 0 < m < 1 is the interference parameter. Others have modeled interference as (x/1+hx)^m_1 . One could consider this situation to model a predator with a greater feeding rate, or a more aggressive predator, than one in which m_1 = 1, the classical case. This is clear from simple comparison, p(x)|_m_1=1 < p(x)|_0 < m_1 < 1, ∀ x > 0. Other forms of interference consider the ratio dependent response, p(x) = (x/y+x), and it is argued that this response is closer to the true predator-prey dynamics of natural systems, mechanistically, empirically as well as behaviorally. Another form of interference is modeled via the Beddington-Deangelis functional response . The systematic study by Prasad et al. () extensively focuses on incorporating a predator mutual interference effect into Srinivasu's original model . In this modified system, the Holling type II functional response is replaced to incorporate mutual interference by adopting precisely a Beddington-Deangelis functional response. Prasad et al. A key finding regarding the effect of mutual interference on prey abundance is highlighted here. Their analysis is focused on one coexistence state.

What main finding did Prasad et al. uncover related to mutual interference and prey abundance?

Prasad et al. found that varying the mutual interference term could lead to stable coexistence at a lower abundance of prey.

summary

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A description of nonoverlapping domain decomposition methods is provided here. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

What are nonoverlapping domain decomposition methods, and why are they commonly used for numerical schemes in elliptic equations?

Nonoverlapping domain decomposition methods are a common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication.

terminology

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. An example of convergence analysis for specific domain decomposition methods is mentioned here. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

Can you give an example of an analysis made on the convergence properties of certain domain decomposition methods?

The convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation.

example

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic types of nonoverlapping decomposition methods are listed here. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

What are the basic nonoverlapping decomposition methods used for elliptic equations?

The basic nonoverlapping decomposition methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods.

terminology

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. The reason for the decomposition of domain Ω into subdomains is explained here. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

Why is the domain Ω decomposed into subdomains Ω_1 and Ω_2 when discussing domain decomposition methods?

The domain Ω is decomposed into subdomains Ω_1 and Ω_2 to introduce the methods of domain decomposition.

purpose

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. A summary of the Neumann–Neumann method's performance for certain types of equations is given here. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

Could you summarize the efficacies and convergence qualities of the Neumann–Neumann method when applied to semilinear elliptic equations and degenerate quasilinear equations?

The Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient for semilinear elliptic equations and has poor convergence properties for degenerate quasilinear equations.

summary

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . The purpose behind proposing modifications to the Neumann–Neumann method is outlined here. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

What is the purpose of proposing two modified Neumann–Neumann methods?

Two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one are proposed to obtain methods that require fewer solutions of nonlinear equations.

purpose

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. The two primary goals of the research study are highlighted here. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

What are the two main objectives of this study regarding the modified Neumann–Neumann methods?

The goals of the study are to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations and to rigorously prove linear convergence for these methods on general Lipschitz domains.

purpose

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. The study's future goals around convergence proofs are stated here. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. Finally, in we present some numerical results that verify our analysis. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

Regarding the convergence of modified Neumann–Neumann methods, what does the study aim to achieve in the future?

Future work includes the goal to prove convergence of the modified Neumann–Neumann method for nondegenerate quasilinear equations and the second modified method for semilinear equations.

future

arxiv

2312.11177v1

INTRODUCTION We consider a bounded Lipschitz domain Ω⊂^d for d=2, 3, with boundary ∂Ω, together with the quasilinear elliptic equation with homogeneous Dirichlet boundary conditions { -∇·α(x, u, ∇ u)+β(x, u, ∇ u) =f in Ω, u =0 on ∂Ω. . We will refer to the equation as semilinear if α(x, y, z) = α̃(x)z. A common approach to generate numerical schemes for elliptic equations that can be implemented in parallel and only rely on local communication is to employ nonoverlapping domain decomposition methods. The performance and convergence properties of these methods have been extensively studied for linear elliptic equations, as surveyed in , but less is known for the methods applied to quasilinear elliptic equations. Convergence results for overlapping Schwarz methods have been derived in . Note that even considers degenerate quasilinear equations, i.e., when α(x, y, z) = α̃(x, y, z)z and α̃(x, y, z)=0 for some y≠ 0. For nonoverlapping decompositions we are only aware of a few studies that derive rigorous convergence results. In the convergence of the Dirichlet–Neumann and Robin–Robin methods are analyzed for a one-dimensional quasilinear elliptic equation, and in our own paper we prove convergence without order for the Robin–Robin method applied to degenerate quasilinear equations with a p-structure. Moreover, a modified Dirichlet–Neumann method for nonlinear equations is studied in , and convergence is derived for one-dimensional problems. There are also some studies relating to quasilinear equations and decompositions with cross points , but without convergence results. We are unaware of any linear convergence results for nonoverlapping decompositions applied to quasilinear, or even semilinear, elliptic problems on general Lipschitz domains. The basic nonoverlapping decompositions methods are the already mentioned Dirichlet–Neumann, Neumann–Neumann, and Robin–Robin methods. In the context of linear elliptic equations the Neumann–Neumann method has better convergence properties on non-symmetric domain decompositions compared to the Dirichlet–Neumann method. Furthermore, for linear equations and after a finite element discretization one can prove that the Neumann–Neumann method converges linearly with an error reduction constant C that is uniformly bounded with respect to the mesh width h, whereas the Robin–Robin method obtains a deteriorating constant of the form C=1-𝒪(√(h)). With these considerations in mind, the Neumann–Neumann method is a natural starting point for developing linearly convergent methods for more general elliptic equations. In order to introduce the methods, we decompose the domain Ω into two subdomains Ω_1 and Ω_2 such that Ω=Ω_1∪Ω_2, Ω_1∩Ω_2=∅ and Γ=(∂Ω_1∩∂Ω_2)∖∂Ω, where ∂Ω_i denotes the boundary of Ω_i and Γ is the interface separating the subdomains. The subdomains can also be replaced by two families of nonadjacent subdomains, which is required in order to obtain a parallel method. Two possible decompositions are illustrated in . The Neumann–Neumann method for the nonlinear equation is then as follows. Let s_1,s_2>0 be method parameters and η^0 an initial guess on Γ. For n=0,1,2,…, find (u_1^n+1, u_2^n+1, w_1^n+1, w_2^n+1) such that { -∇·α(x, u^n+1_i, ∇ u^n+1_i)+β(x, u^n+1_i, ∇ u^n+1_i) =f in Ω_i, u^n+1_i =0 on ∂Ω_i∖Γ, u^n+1_i = η^n on Γ, for i=1,2, -∇·α(x, w^n+1_i, ∇ w^n+1_i)+β(x, w^n+1_i, ∇ w^n+1_i) =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, α(x, w^n+1_i, ∇ w^n+1_i)·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2, . with η^n+1=η^n-s_1w_1^n+1Γ-s_2w_2^n+1Γ. Here, ν_1 denote the outwards pointing unit normal on ∂Ω_1, and the computed quantities (u_1^n, u_2^n,η^n) approximate (u|_Ω_1, u|_Ω_2, u|_Γ), respectively. We will refer to the last three equalities in as the auxiliary problem. However, the Neumann–Neumann method requires the solution of four nonlinear problems in each iteration, which is inefficient even for semilinear elliptic equations. The method also displays surprisingly poor convergence properties for degenerate quasilinear equations. For example consider the semilinear equation -Δ u+|u|u=f and the p-Laplace equation -∇·(|∇ u|^p-2∇ u)+u=f, both given on the unit square Ω=(0,1)^2 with a decomposition into two L-shaped subdomains Ω_1 and Ω_2. The Neumann–Neumann method then requires a significantly larger amount of equation solves compared to other methods for the semilinear equation; see . Furthermore, it does not even generate a decreasing error with respect to the number of iterations for the p-Laplace equation; see . The details of the numerical experiments are given in . In order to obtain methods that require less solutions of nonlinear equations, we propose two modified Neumann–Neumann methods based on replacing the auxiliary problem with a linear one, i.e., we employ a linear preconditioner. For the first modified method we replace the auxiliary problem by { -Δ w^n+1_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, ∇ w^n+1_i·ν_1 = α(x, u_1^n+1, ∇ u_1^n+1)·ν_1-α(x, u_2^n+1, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . We also consider a second method where the auxiliary problem is iteration-dependent, in particular via the linearization of the auxiliary problem. For notational simplicity, we assume that α(x, y, z)=α(x, z) and β(x, y, z)=β(x, y). We then introduce J_α and J_β, the Jacobians of α and β with respect to z and y, respectively. The second modified method is then given by replacing the auxiliary problem with { -∇·(J_α(x, ∇ u_i^n+1)∇ w_i)+J_β(x, u_i^n+1)w_i =0 in Ω_i, w^n+1_i =0 on ∂Ω_i∖Γ, J_α(x, ∇ u^n+1_i)∇ w^n+1_i·ν_1 = α(x, ∇ u_1^n+1)·ν_1-α(x, ∇ u_2^n+1) ·ν_1 on Γ, for i=1,2. . The efficiency gains of the modified Neumann–Neumann methods over the standard Neumann–Neumann method are illustrated in for our semilinear example. The modified methods even perform well for the degenerate p-Laplace equation; see . Hence, there are two goals for this study. First, to provide numerical evidence of the advantages of the modified Neumann–Neumann methods for quasilinear equations, including degenerate ones. Secondly, to rigorously prove linear convergence for the modified methods when applied to semi- or quasilinear elliptic equations with domains and subdomains that are only assumed to be Lipschitz. The paper is organized as follows. In we give the abstract framework to analyze nonlinear domain decomposition methods and state the assumptions on our equations. In we then introduce the weak formulations and in we define the transmission problem and the nonlinear Steklov–Poincaré operators. In we describe the Steklov–Poincaré interpretations of the two modified Neumann–Neumann methods. We then prove convergence of the first modified method for nondegenerate quasilinear equations in and in we prove the convergence of the second modified method for semilinear equations. In we give a brief explanation how the same analysis can be applied to the Galerkin discretization of the elliptic equation. The role of the numerical results in the context of the paper is summarized here. To keep the notation simple we will denote generic constants by c, C>0, which may change from line to line.

What is the significance of the numerical results mentioned at the end of the paper?

Numerical results are presented to verify the analysis of the modified Neumann–Neumann methods' convergence properties.

summary

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. The distinction in methodology between DCRNN and GWNet regarding graph structures is elaborated here. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

What is the key difference between the approaches of DCRNN and GWNet in building graph structures for traffic prediction?

DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters.

terminology

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. The reason for the variability of spatial dependencies at different times is summarized here. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

Why do researchers find that spatial dependencies vary at different times in the context of traffic prediction?

The spatial dependence of different times is not the same due to the complex urban transportation environment.

summary

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. The rationale behind the effectiveness of data-driven approaches over self-adaptive approaches in adjusting graph structures is discussed here. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

Why might data-driven approaches be more effective than self-adaptive approaches for adjusting graph structures in dynamic traffic prediction?

Data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters, because traffic patterns can always change randomly.

purpose

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. An example highlighting the time-shift issue in urban traffic is provided here. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

Can you provide an example that illustrates the time-shift problem in urban transportation?

Time-shift is a common problem in urban transportation, for instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas.

example

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. The solution proposed to tackle the time-shift and data noise issues is summarized here. To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

What solution is proposed to address the problems of time-shift and data noise in modeling spatial dependencies of traffic data?

To solve the time-shift and data noise problems, a dynamic frequency domain graph convolutional network (DFDGCN) is proposed.

summary

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). The aim of using frequency domain information in DFDGCN is discussed here. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. Specifically, we summarize the main contributions of our work as follows: * We focus on the time-shift problem when mining spatial dependence of traffic data, and to the best of our knowledge we are the first to mitigate it by frequency domain analysis with the property of Fourier transform. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

For what purpose is the frequency domain information of traffic data used in the dynamic frequency domain graph convolutional network (DFDGCN)?

To mitigate the effect of time-shift, examining the traffic patterns through the frequency domain information of traffic data is proposed.

purpose

arxiv

2312.11933v1

INTRODUCTION With the increase of global urbanization and population concentration, complex transportation environments have become a major challenge for urban management, leading to the urgent need for Intelligent Transportation Systems (ITS) about AI to enhance the population carrying capacity of cities. Traffic prediction is an indispensable part of ITS, which aims to accurately predict future traffic data by learning historical traffic states and patterns obtained from sensors. However, due to the high correlation between sensors, future information of each sensor relies not only on its own historical information but also on the historical data of other sensors. Spatio-temporal graph neural networks are proposed to capture not only the temporal dependence of traffic data inherent in each sensor, but also to mine the spatial dependence among sensors, and this class of methods has been proven to outperform general time series forecasting methods such as ARIMA and LSTM with the main enhancement coming from mining the reliable spatial dependencies among sensors. Mining effective spatial dependencies, also known as graph structures, has been an important challenge in traffic prediction. Early schemes such as DCRNN relies on prior spatial information to calculate predefined graph structures, GWNet self-adaptively learns graph structures with learnable parameters. STGCN , AGCRN, and MTGNN all apply or improve upon these static graphs to achieve outstanding enhancements. In subsequent research, researchers find that the spatial dependence of different times is not the same due to the complex urban transportation environment. ASTGNN and DMSTGCN apply learnable parameters at each timestamp such as day of week or hour of day to self-adaptively capture spatial dependencies over different time periods. In contrast, the approach of applying changing historical traffic data to model dynamic graphs is called data-driven. Since traffic patterns can always change randomly, data-driven approaches that can adjust the graph structure based on historical information during the testing phase tend to be more effective than self-adaptive approaches that rely purely on learnable parameters. STFGCN, DGCRN, and DSTAGNN all utilize self-attention scores of time domain information or the spatial proximity between traffic data in time domain to construct the graph structures. However, we argue that data-driven modeling of spatial dependence faces two problems: Time-Shift and Data Noise. Time-shift is a common problem in urban transportation. For instance, during urban commuting, when people come off duty from industrial areas, the congested traffic can take half an hour or more to reach residential areas. Time-shift of traffic patterns due to similar situations is common in modern cities because of urban functional zones, and it can result in tidal wave-like delay variations between interrelated traffic data as shown in Fig. . However, time-shift can invalidate spatial proximity measures, which are widely applied to capture spatial dependence of traffic data. It's because schemes such as Euclidean distance and cosine similarity are computed in aligned dimensions, but changes in interrelated traffic data affected by time shifts do not occur at the same timestamp, which raises difficulties in examining proximity to capture the correlation between roads in time domain. Fig. represents the cosine similarity between traffic flows at the same observation time in one week, and we can observe that the similarity between traffic flows in time domain exhibits low differentiation and stacking. This is due to the misalignment of the temporal dimension caused by time-shift, which makes the traffic flows on related roads have similar similarity to those on uncorrelated roads, which makes it difficult to explore valid spatial dependencies through proximity. Noise in traffic data is an unavoidable problem as in general data-driven strategies, and is often triggered by emergencies such as traffic accidents or road construction. Therefore, traffic data with noise is not always reliable, and it is a realistic challenge to introduce additional information to reduce the effect of noise while learning the spatial relationships between sensors. To solve the above problems, we propose a dynamic frequency domain graph convolutional network(DFDGCN). To mitigate the effect of time-shift, we propose to examine the traffic patterns through the frequency domain information of traffic data. As shown in Fig., the frequency domain information between sensors exhibits better differentiation, it is because even if the traffic information is time-shifted, relying on the time-shift property of the Fourier transform, the frequency domain components of traffic data will still be in the same phase dimension, which will facilitate us to mine dynamic spatial dependence subsequently. In addition, we introduce identity embedding of sensors and time embedding when capturing traffic data for graph learning to minimize the effect of data noise. The primary contributions of this research work are summarized here. * We propose DFDGCN to learn dynamic graph structures among sensors by combining traffic data with identity embedding and time embedding to cope with data noise. * Extensive experiments on four publicly available datasets demonstrate that our proposed model outperforms baselines, and ablation experiments witness the validity of our model.

What are the main contributions of the work discussed in the passage?

The main contributions of the work are focusing on the time-shift problem in spatial dependence of traffic data and mitigating it by frequency domain analysis, and proposing DFDGCN to learn dynamic graph structures and cope with data noise.

summary

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. The definition of a latin square of order N is provided here. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

What defines a latin square of order N?

A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column.

terminology

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. The concept of a 2-realization is explained here. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

How is a 2-realization defined in the context of integer partitions and latin squares?

A 2-realization of an integer partition P = (h_1,h_2,…,h_n) of N, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n].

terminology

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. An example of when a (h_1h_2… h_n) exists for n=3 is mentioned here. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

Can you provide an example of when a (h_1h_2… h_n) exists for n=3?

A (h_1h_2… h_n) exists when n=3 if and only if h_1=h_2=h_3.

example

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. The conditions for the existence of a latin square with certain parameters are summarized here. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

Could you summarize the conditions for the existence of a latin square (a^ub^n-u) where a>b>0 and n>4?

For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b.

summary

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. A necessary future condition for the existence of a (h_1h_2… h_n) is stated here. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

What future condition is necessary for the existence of a (h_1h_2… h_n)?

If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i.

future

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. The description of a latin cube of order N is stipulated here. If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

How is a latin cube of order N described?

A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3).

terminology

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. The concept of a 3-realization is explicated here. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). From <Ref>, we know that a (1^2) does not exist. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

What does a 3-realization of an integer partition P represent in the context of latin cubes?

A 3-realization of an integer partition P = (h_1,h_2,…,h_n) of N, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n].

terminology

arxiv

2312.10981v1

INTRODUCTION L. Fuchs asked the following question : If n is any positive integer and n=n_1+n_2+…+n_k any fixed partition of n, is it possible to find a quasigroup (Q,∘) of order n which contains subquasigroups on the sets Q_1,Q_2,…,Q_k of orders n_1,n_2,…,n_k respectively, whose set theoretical union is Q? The problem of finding such a quasigroup is equivalent to finding a latin square of order n with disjoint subsquares of orders n_1,n_2,…,n_k (these subsquares must be row, column and symbol disjoint). It is easy to find partitions for which such a quasigroup does not exist, however the problem of existence is far from complete. Much work has been done by considering latin squares, which are equivalent structures to quasigroups, and such a latin square is known as a realization or partitioned incomplete latin square (PILS). For partitions with at most four parts, existence has been completely determined. Many cases have been studied for partitions with 5 or 6 parts, however the problem is not yet solved. For any partition, existence has been determined when the parts are of at most two distinct sizes. The same problem can be posed for m-ary quasigroups, which are equivalent to m-dimensional latin hypercubes. In this paper, we investigate this problem by first considering latin cubes and also by examining the relationships between latin hypercubes of different dimenions. A latin square of order N is an N× N array, L, on a set of N symbols such that each symbol occurs exactly once in each row and column. An entry of L is defined by its row and column position as L(i,j). An m× m sub-array of L that is itself a latin square of order m is a subsquare of L. Subsquares are disjoint if they share no row, column or symbol of L. For any positive integer a, we define [a] to be the set of integers {1,2,…,a}. Given another integer b, b+[a] is the set {b+1,b+2,…,b+a}. We will assume that all latin squares of order N use the symbols of [N]. The array in is a latin square of order 5, and the highlighted cells indicate a subquare of order 2 and three subsquares of order 1. An integer partition of N is a sequence of non-increasing integers P = (h_1,h_2,h_3,…) where ∑_i=1^∞ h_i = N. An integer partition P = (h_1,h_2,…,h_n) will be represented by (h_1h_2… h_n) or (h_i_1^α_1h_i_2^α_2… h_i_m^α_m), where there are α_j copies of h_i_j in the partition. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 2-realization of P, denoted (h_1h_2… h_n), is a latin square of order N with pairwise disjoint subsquares of order h_i for each i∈ [n]. We say that a 2-realization is in normal form if the subsquares are along the main diagonal and the i^th subsquare, H_i, is of order h_i. The latin square in Example is a 2-realization of the partition (2^11^3). The existence of 2-realizations for n=1,2,3,4 has been completely determined in . Take a partition (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0. Then a (h_1h_2… h_n) * always exists when n=1; * never exists when n=2; * exists when n=3 if and only if h_1=h_2=h_3; * exists when n=4 if and only if h_1=h_2=h_3, or h_2=h_3=h_4 with h_1≤ 2h_4. The problem has also been solved for partitions with subsquares of at most two distinct orders. Dénes and Pásztor gave the result for subsquares all of the same order and the result for two different orders was considered by Heinrich and finished by Kuhl and Schroeder . For n≥ 1 and a≥ 1, a (a^n) exists if and only if n≠ 2. The following theorem completes the problem for all partitions with at most two distinct orders, considering only the cases which haven't already been covered in the previous theorems. For a>b>0 and n>4, a (a^ub^n-u) exists if and only if u≥ 3, or 0< u < 3 and a≤ (n-2)b. Colbourn has determined two necessary, but not sufficient, conditions for the existence of a 2-realization. If a (h_1h_2… h_n) exists, then h_1≤∑_i=3^nh_i. If a (h_1h_2… h_n) exists, then N^2 - ∑_i=1^nh_i^2≥ 3( ∑_i∈ Dh_i)(∑_j∈Dh_j) for any D⊆{1,2,…,n} and D = {1,2,…,n}∖ D. Fuchs' problem considers subgroups of quasigroups with binary operations which are equivalent to latin squares, and so this problem can be extended to quasigroups with ternary operations by looking at latin cubes. These are also known as permutation cubes. A latin cube C of order N is an N× N× N array on a symbol set of size N, with entries defined by C(j_1,j_2,j_3), such that if (j_1,j_2,j_3) and (j'_1,j'_2,j'_3) differ in exactly one position, then C(j_1,j_2,j_3) ≠ C(j'_1,j'_2,j'_3). If C is instead an r× s× t array with r,s,t≤ N, then we call C a latin box of order r by s by t. We will consider a line of C to be a set of N cells with two fixed coordinates. A column, row or file is thus a line represented by ([N],j_2,j_3), (j_1,[N],j_3) or (j_1,j_2,[N]) respectively. A layer of C is the set of cells with only one fixed coordinate. Layers can thus be represented as (j_1,[N],[N]), ([N],j_2,[N]) or ([N],[N],j_3). Similar to latin squares, a subcube of C is an m× m× m sub-array which is itself a latin cube of order m. Subcubes are disjoint if they share no row, column, file or symbol. As with latin squares, we will assume that all latin cubes use the symbols of [N]. For an integer partition P = (h_1,h_2,…,h_n) of N with h_1≥ h_2≥…≥ h_n > 0, a 3-realization of P, denoted (h_1h_2… h_n), is a latin cube of order N with pairwise disjoint subcubes of order h_i for each i∈ [n]. As with 2-realizations, a 3-realization is in normal form if the subcubes are along the main diagonal (cells of the form (i,i,i) for all i∈ [N] and the additional cells needed to form the subsquares) and the i^th subcube, H_i, is of order h_i. The arrays in are the layers of a (2^21^1). The subcubes are highlighted. The problem of existence of a (h_1h_2… h_n) is different to that of a (h_1h_2… h_n). To see this, consider the latin cube in . This latin cube is a 3-realization of the partition P = (1,1), so a (1^2). An example showing the difference between 2-realization and 3-realization is provided here. Thus, there are partitions which can be realized in latin cubes but not in latin squares.

What is an example illustrating the difference in realization between 2-dimensional and 3-dimensional latin structures?

A (1^2) does not exist as a 2-realization, but a (1^2) can exist as a 3-realization in a latin cube.

example

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . The special property of physical phenomena is modeled by this equation. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

What does the KdVB equation represent in mathematical modeling?

The Korteweg-de Vries-Burgers (KdVB) equation is a mathematical model that exhibits both dissipation and dispersion.

terminology

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. Significant outcomes regarding the KdVB equation and the impact of distributed memory are mentioned here. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

What outcomes did the authors manage to establish for the KdVB equation according to the passage?

The authors established well-posedness and stability outcomes for the KdVB equation with a distributed memory.

summary

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. The functional contribution of the distributed memory term in the equation is discussed here. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

What role does the distributed memory term play in the KdVB equation according to the text?

The distributed memory term in the KdVB equation contributes to the stability of the system by playing the role of a dissipation mechanism.

purpose

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. A pertinent inquiry concerning the memory-type boundary control's effect on the KdVB equation is raised here. This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

What question is the text posing regarding the boundary control of memory-type in relation to the KdVB equation?

Understanding the impact of a boundary control of memory-type on the behavior of the KdVB equation’s solutions when applied is the natural question posed.

purpose

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. An exemplar function α meeting certain criteria is provided here. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

Can you provide an example of a function α that satisfies the specified conditions in the passage?

A typical function α that satisfies given conditions is one that converges exponentially to zero at infinity, like α(s) = d2 e^-d1 s, with specific positive constants d1 and d2.

example

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. An example of a function with a specific decay characteristic is detailed here. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

What is an example of a function α described in the passage that has a slower decay rate than the exponential one?

An example of a function α with a decay rate to zero at infinity weaker than the exponential one is α(s) = d2 (1+s)^-d1 for certain positive constants d1 and d2.

example

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). A condition for the system's well-posedness is articulated here. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

Under what condition is the problem shown to be well-posed in the study?

The problem is shown to be well-posed as long as the hypothesis (H) holds.

summary

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. The reason for the beneficial nature of the memory component in the boundary control is explored here. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

Why is the memory part of the boundary control considered beneficial according to the passage?

The memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α.

purpose

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. The adaptability of the utilized techniques to other types of equations is mentioned here. Finally, we present a numerical analysis by constructing a numerical scheme based on a finite differences method, and provide numerical examples illustrating our theoretical results. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

Which other type of equations can the techniques used for the KdVB system be applied to, as mentioned in the text?

The techniques used for the KdVB system can also be applied to the Kuramoto-Sivashinsky (KS) equation.

future

arxiv

2312.11950v1

INTRODUCTION It is well-known that numerous physical phenomena exhibit both dissipation and dispersion . This very special property is mathematically modeled by the Korteweg-de Vries-Burgers (KdVB) equation and hence has gained a considerable prominence. As a matter of fact, the KdVB and its close relatives, has been the subject of many studies (see for instance ). Instead of highlighting the contribution of each of these papers, the reader is referred to for a comprehensive discussion of this point. In turn, it is worth mentioning that in , the authors managed to establish well-posedness and stability outcomes for the KdVB equation with a distributed memory. In fact, it turned out that such a distributed memory term plays a role of a dissipation mechanism and hence contributes to the stability of the system. Nonetheless, boundary controls of memory-type are commonly used in practice and consequently the natural question is: when a boundary control of memory-type is applied, what is the impact on the behavior of the solutions of the KdVB equation? This is mainly the motivation of the present work. More precisely, the problem under consideration involves the third-order KdVB equation with a boundary infinite memory {[ ∂_t y(x,t) -ω_0 ∂_x^2 y(x,t) + ω_1 ∂_x^3 y(x,t) +ω_2 ∂_x y(x,t) + ω_3 y(x,t) ∂_x y(x,t)=0, x ∈ I, t >0,; y (0,t)=y(1,t)=0, t∈ℝ_+ ,; ∂_x y(1,t)=ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds, t∈ℝ_+ ,; y(x,0)=y_0 (x) , x∈ I,; ∂_x y(0,-t) =y_1 (t), t∈ℝ_+ , ]. where y:I×ℝ_+:=(0,1)× [0,∞) →ℝ is the amplitude of the dispersive wave at the spatial variable x and time t , ∂_u^n denotes the differential operator of order n with respect to u; that is ∂^n∂ u^n. In turn, y_0 :I→ℝ and y_1 :ℝ_+→ℝ are known initial data, α:ℝ_+→ℝ_+ is a given function and ω_i are real constants (physical parameters) satisfying the following hypothesis (H): 0,1truecm * The memory kernel α satisfies α∈ C^2 (ℝ_+), α^' < 0, α(0)>0, lim_s→∞α(s)=0 and -ξ (s) α^'(s)≤α^'' (s) ≤ -ξ_0 α^'(s), s∈ℝ_+ , for a positive constant ξ_0 and a function ξ :ℝ_+→ℝ_+^*:=(0,∞) such that ξ∈ C^1 (ℝ_+), ξ^'≤ 0 and ∫_0^∞ξ (s)ds =∞. * The constants ω_0, ω_1 and ω_4 satisfy ω_0≥ 0, ω_1 >0 and |ω_4| <1. * The following relationship between α, ω_1 and ω_4 holds ∫_0^∞-α^'(s)/ξ (s)dx:=α_0 < ω_1(1-ω_4^2)/ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (i) Typical functions α satisfying (), (), () and () are the ones which converge exponentially to zero at infinity like α(s)=d_2 e^-d_1 s, where d_1 and d_2 are positive constants satisfying (ξ_0 =ξ (s) =d_1) d_2/d_1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. But (), (), () and () allow α to have a decay rate to zero at infinity weaker than the exponential one like α(s)=d_2 (1+s)^-d_1, where d_1 >1 and d_2 >0 satisfying (ξ_0 =d_1 and ξ (s) =d_1 (1+s)^-1) d_2/d_1 -1 < ω_1(1-ω_4^2)ω_1^2 (1-ω_4^2)+(1+ω_1 ω_4)^2. 0,1truecm (ii) The expression ω_4 ∂_x y(0,t)+∫_0^∞α (s) ∂_x y(0,t-s) ds is viewed as a boundary control of memory-type. It is also relevant to note that ω_4 obeys the last condition in () and hence one may take ω_4=0. This means that, in this event, the boundary control is a purely memory-type one. 0,1truecm As mentioned earlier, our aim is to address the effect of the presence of the infinite memory term in the boundary control on the behavior of the solutions to (). To do so, we shall place the system in the so-called past history framework (see also for a further discussion about the history approach and for another methodology of treatment of systems with memory). The problem is shown to be well-posed as long as the hypothesis (H) holds. Then, we prove that the memory part of the boundary control is beneficial in decaying the energy of the system under different circumstances of the kernel α. 0,1truecm Before providing an overview of this article, we point out that this work goes beyond the earlier one in several respects. First, we deal with a boundary control in contrast to , where the control is distributed. As the reader knows, it is usually more delicate (mathematically speaking) to consider a boundary control than a distributed one. Second, we mange to show that the system under consideration () is well-posed and its solutions are stable despite the presence of the memory term. Moreover, the desired stability property remains attainable even if the memory term is the sole action of the boundary control; that is ω_4=0. The proofs are based on the multiplier method and a combination of the semigroups approach of linear operators and the fixed point theory. Third, we show that the techniques used for the KdVB system () can also be applied to another type of equations, namely, the Kuramoto-Sivashinsky (KS) equation. The approach taken for numerical analysis and the type of examples used to illustrate the results are specified here. 0,1truecm The paper comprises six sections excluding the introduction. In Section , we put forward some preliminaries. Section is devoted to establishing the well-posedness of the KdVB system. In Section , two estimates of the energy decay rate are provided depending on the feature of the memory kernel. Indeed, it is shown that the decay of the energy corresponding of KdVB solutions is basically a transmission of the decay properties of the memory kernel. In section , we treat the KS system (). In Section , we give a numerical analysis for both KdVB and KS systems. Lastly, brief concluding remarks are pointed out in Section .

How was the numerical analysis for the theoretical results presented in the passage executed?

A numerical analysis was carried out by constructing a numerical scheme based on a finite differences method, and numerical examples were provided to illustrate the theoretical results.

summary

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . The summary of interconnections in the standard model is mentioned here. In addition to the experimental searches for further degrees of freedom, theoretical studies of consistent quantum field theories have paved the way for new concepts, e.g., with a consistent quantization of interacting spin 3/2 particles requiring supersymmetry. In the present letter, we explore the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories. This type of fermions has been discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field . In recent solid-state research, these non-relativistic degrees of freedom find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points (e.g., inverted band gap semiconductors, pyrochlore iridates) ; such systems can give rise to interesting quantum critical phenomena . Also gauged versions have been studied recently in the context of quantum spin liquids . While the generalization of the underlying algebra to the relativistic case is, in principle, straightforward, we find that the construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

What does the standard model particularly demonstrate in terms of interconnections of fields?

The interconnections in the standard model are particularly obvious, containing spin 0,1/2,1 fields and all possible renormalizable interactions allowed by the symmetries.

summary

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . These interconnections are particularly obvious in the standard model containing spin 0,1/2,1 fields and accommodating all possible renormalizable interactions allowed by the symmetries. The purpose behind theoretical studies of consistent quantum field theories is discussed here. In the present letter, we explore the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories. This type of fermions has been discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field . In recent solid-state research, these non-relativistic degrees of freedom find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points (e.g., inverted band gap semiconductors, pyrochlore iridates) ; such systems can give rise to interesting quantum critical phenomena . Also gauged versions have been studied recently in the context of quantum spin liquids . While the generalization of the underlying algebra to the relativistic case is, in principle, straightforward, we find that the construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

What have theoretical studies of quantum field theories led to regarding new concepts?

Theoretical studies of consistent quantum field theories have paved the way for new concepts, such as the consistent quantization of interacting spin 3/2 particles requiring supersymmetry.

purpose

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . These interconnections are particularly obvious in the standard model containing spin 0,1/2,1 fields and accommodating all possible renormalizable interactions allowed by the symmetries. In addition to the experimental searches for further degrees of freedom, theoretical studies of consistent quantum field theories have paved the way for new concepts, e.g., with a consistent quantization of interacting spin 3/2 particles requiring supersymmetry. The exploration mentioned in the present letter is highlighted here. This type of fermions has been discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field . In recent solid-state research, these non-relativistic degrees of freedom find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points (e.g., inverted band gap semiconductors, pyrochlore iridates) ; such systems can give rise to interesting quantum critical phenomena . Also gauged versions have been studied recently in the context of quantum spin liquids . While the generalization of the underlying algebra to the relativistic case is, in principle, straightforward, we find that the construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

What is the primary focus of the present letter in terms of Luttinger fermions?

The present letter explores the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories.

purpose

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . These interconnections are particularly obvious in the standard model containing spin 0,1/2,1 fields and accommodating all possible renormalizable interactions allowed by the symmetries. In addition to the experimental searches for further degrees of freedom, theoretical studies of consistent quantum field theories have paved the way for new concepts, e.g., with a consistent quantization of interacting spin 3/2 particles requiring supersymmetry. In the present letter, we explore the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories. The discovery of Luttinger fermions is described here. In recent solid-state research, these non-relativistic degrees of freedom find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points (e.g., inverted band gap semiconductors, pyrochlore iridates) ; such systems can give rise to interesting quantum critical phenomena . Also gauged versions have been studied recently in the context of quantum spin liquids . While the generalization of the underlying algebra to the relativistic case is, in principle, straightforward, we find that the construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

Who discovered Luttinger fermions and in what context?

Luttinger fermions were discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field.

terminology

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . These interconnections are particularly obvious in the standard model containing spin 0,1/2,1 fields and accommodating all possible renormalizable interactions allowed by the symmetries. In addition to the experimental searches for further degrees of freedom, theoretical studies of consistent quantum field theories have paved the way for new concepts, e.g., with a consistent quantization of interacting spin 3/2 particles requiring supersymmetry. In the present letter, we explore the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories. This type of fermions has been discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field . An example of the application of Luttinger fermions in solid-state research is provided here. Also gauged versions have been studied recently in the context of quantum spin liquids . While the generalization of the underlying algebra to the relativistic case is, in principle, straightforward, we find that the construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

Can you provide an example of where non-relativistic Luttinger fermions are applied in solid-state research?

In solid-state research, non-relativistic Luttinger fermions find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points.

example

arxiv

2312.12058v1

INTRODUCTION One of the remarkable features of quantum field theories is given by the interconnection of fields as representations of the Lorentz group , their powercounting dimensionality and the renormalizability of interacting field theories in d=3+1-dimensional spacetime . These interconnections are particularly obvious in the standard model containing spin 0,1/2,1 fields and accommodating all possible renormalizable interactions allowed by the symmetries. In addition to the experimental searches for further degrees of freedom, theoretical studies of consistent quantum field theories have paved the way for new concepts, e.g., with a consistent quantization of interacting spin 3/2 particles requiring supersymmetry. In the present letter, we explore the possibility to construct relativistic versions of Luttinger fermions and perform first perturbative studies of corresponding interacting quantum field theories. This type of fermions has been discovered by Luttinger while searching for the most general form of a nonrelativistic Hamiltonian of a semi-conductor excitation in a magnetic field . In recent solid-state research, these non-relativistic degrees of freedom find extensive application in spin-orbit coupled materials with quadratic band touching/crossing points (e.g., inverted band gap semiconductors, pyrochlore iridates) ; such systems can give rise to interesting quantum critical phenomena . Also gauged versions have been studied recently in the context of quantum spin liquids . The requirements for constructing a fully relativistic action are summarized here.

What is required for the construction of a fully relativistic action according to the passage?

The construction of a fully relativistic action requires a reducible representation in terms of the related Dirac algebra with interesting consequences for the construction of interacting quantum field theories.

summary

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . The suitability of Graph Convolutional Networks (GCNs) for human motion prediction is described here. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

Why are Graph Convolutional Networks (GCNs) particularly suitable for human motion prediction?

Graph Convolutional Networks (GCNs) are used for human motion prediction due to the graph-like nature of human body joints and bones.

terminology

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Common representation forms for human motion data are mentioned here. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

In what form is human motion data commonly represented?

Human motion data are often in the form of 3D skeleton sequences.

example

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? The explanation of Universal Graph Convolution (UniGC) is given here. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

What is the Universal Graph Convolution (UniGC)?

Universal Graph Convolution (UniGC) is a novel graph convolution concept that includes different graph convolutions as its special cases.

terminology

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Various types of existing graph convolutions are listed here. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

What are some of the existing types of graph convolutions according to the passage?

Existing graph convolutions include spatial, temporal, spatial-temporal, dense, trajectory, space-time-separable (STS), and unified (G3D) graph convolutions.

example

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. The potential issue of stacking multiple layers of UniGC is discussed here. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

Why might stacking multiple layers of UniGC be problematic?

Stacking multiple layers of UniGC would require too many parameters, making computation less efficient and optimization more challenging.

purpose

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. The components that make up GCNext are outlined here. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

What are the key elements that constitute the GCNext framework?

GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block that chooses the best-fitting one for each sample.

terminology

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. A comparison of computational cost between GCNext and existing GCNs is provided here. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

How does GCNext compare to existing GCNs in terms of computational cost?

GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes.

summary

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. The significance of the Universal Graph Convolution (UniGC) proposal is summarized here. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

What is the significance of proposing the Universal Graph Convolution (UniGC)?

The proposal of UniGC represents a new concept and a new paradigm of designing graph convolutions.

summary

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. The innovative aspect of GCNext is described here. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. * We conduct extensive experiments on three benchmark datasets, Human3.6M, AMASS, and 3DPW, which show that GCNext can yield up to 9× lower computational cost than existing GCNs with comparable model sizes, besides achieving state-of-the-art performance.

What is the innovation behind the GCNext framework?

GCNext is a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise.

summary

arxiv

2312.11850v1

INTRODUCTION The idea of predictable human motion has attracted much research interest over the years across a wide range of applications such as human-robot interaction and autonomous driving. To generate plausible predictions, early successes were achieved with RNNs and CNNs . At present, human motion prediction is dominated almost exclusively by Graph Convolutional Networks (GCNs) , due to the innate graph-like nature of human body joints and bones. Human motion data are often in the form of 3D skeleton sequences, where each dimension corresponds to a different aspect—time, space, or channel, as shown in Fig. . To extract features, various graph convolutions are proposed. They either focus on aggregating information in a single dimension , or combine aggregations in two dimensions without considering the remaining one . There are yet a number of other graph convolution types unexplored by existing works. One is naturally motivated to ask: What makes an ideal type of graph convolution for human motion prediction? To answer this question, we propose Universal Graph Convolution (UniGC), a novel graph convolution concept that pushes the boundaries of existing knowledge by re-conceptualizing different graph convolutions, both existing and unexplored, as its special cases. In the most general case, UniGC uses 6D global adjacency to encode inter-relationships within and across space, time, and channels, which can be categorized into 7 types as shown in Fig. . The entry at position [i,k,m,j,ℓ,n] of the 6D adjacency represents the relationship between channel-m of joint-k at time-i and channel-n of joint-ℓ at time-j. Using two techniques—adjacency masking and sub-adjacency tying, UniGC can be specialized into various different special cases, which focus on different subsets of all the 7 types of relationships. Existing graph convolutions are either the vanilla form or variations of these special cases, such as spatial , temporal , spatial-temporal , dense , trajectory , space-time-separable (STS) , and unified (G3D) graph convolutions. This gives rise to the second question: What is the best way for UniGC to be leveraged network-wise to better address the task? An intuitive answer based on common practices would be to stack multiple layers of UniGC. However, its 6D global adjacency would require too many parameters, making computation less efficient and optimization more challenging. This paper gives a better answer by proposing GCNext, which exploits the versatility and adaptability of UniGC on network-level. In contrast to existing approaches which typically involve manually choosing one from all possible types of graph convolutions, stacking it multiple layers, and evaluating the resulting network, GCNext is structured to make the choice-making process learnable and sample-specific, where each layer dynamically chooses the optimal graph convolution type based on each different sample. Specifically, each layer in GCNext contains several graph convolutional blocks of different types in parallel and a light-weight selector block tasked with choosing the best-fitting one for each sample. Other network designs include a reinvention of update operations, and a replacement of batch normalization with layer normalization, both of which lead to performance improvement and efficiency boost. Combining these designs, GCNext can yield up to 9× reduction in computational cost compared to existing GCNs with comparable model sizes. GCNext offers multiple use cases. First, it can be trained from scratch to obtain a GCN combining the strengths of different graph convolutions both layer-wise and sample-wise. Second, it can be used to refine a preexisting GCN with a predetermined type of graph convolution. Our proposed UniGC and GCNext combined represent a module-to-network paradigm package, providing an effective approach that better addresses human motion prediction. In summary, our contributions are three-fold: [itemize,1]leftmargin=0.37cm * On concept-level, we propose UniGC, a novel graph convolution concept that re-conceptualizes different graph convolutions as its special cases. The specialization is achieved through two techniques: adjacency masking and sub-adjacency tying. UniGC represents not only a new concept of graph convolution, but also a new paradigm of designing graph convolutions. * On network-level, we propose GCNext, a novel GCN framework that dynamically determines the optimal GCN architecture both sample-wise and layer-wise. GCNext offers multiple use cases, including training a new GCN from scratch or refining a preexisting GCN. The advantages of GCNext demonstrated by experimental results are highlighted here.

What are the advantages of GCNext over existing GCNs according to the experimental results?

GCNext achieves up to 9× lower computational cost than existing GCNs with comparable model sizes and state-of-the-art performance on benchmark datasets.

summary

arxiv

2312.10794v1

The introduction of Transformers is mentioned here. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What are Transformers and who introduced them?

Transformers are a type of neural network architecture introduced by Vaswani et al. in 2017.

terminology

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. The significant impact of self-attention mechanisms is summarized here. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What has been the impact of self-attention mechanisms in the field of AI?

The introduction of self-attention mechanisms in Transformers has been instrumental in advancing AI in computer vision and NLP.

summary

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). An example of a dynamical system application in DNNs is provided here. In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

Can you give an example of how a dynamical system can be applied in the context of deep neural networks?

An example of a dynamical system is one where an input image evolves over time according to a time-varying velocity field.

example

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . The purpose of conceptualizing Transformers as flow maps is explained here. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

Why are Transformers modeled as flow maps on the space of probability measures?

Transformers are conceptualized as flow maps on the space of probability measures to better understand their functionality.

purpose

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. The behavior of particle clustering in Transformers is summarized here. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What is a notable observed behavior of particles in Transformers with respect to their roles in learning tasks?

Particles in Transformers tend to cluster, which is a significant observation for tasks such as next-token prediction.

summary

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. The discrepancy between theory and practice for Transformers output distribution is discussed here. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What discrepancy does the paper note between the theory and practice of Transformers regarding the output distribution?

The limiting distribution becoming a point mass is at odds with the fact that large language models are effective and diverse in practice.

summary

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. The resolution to the paradox of the limiting distribution is addressed here. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

How is the paradox concerning the limiting distribution in Transformers resolved according to the passage?

The paradox of the limiting distribution becoming a point mass is resolved by recognizing the existence of a long-time metastable state.

summary

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The twofold goal of the manuscript is outlined here. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What is the twofold goal of the manuscript as outlined in the passage?

The goal of the manuscript is to provide a mathematical framework to study Transformers and to propose research directions concerning the long-time clustering phenomenon.

purpose

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. Future research directions suggested by the paper are mentioned here. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. Our main result, Theorem , indicates that in high dimension d≥ n, a set of n particles randomly initialized on will cluster to a single point as t→+∞. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What future research directions does the paper suggest in relation to Transformers?

The paper establishes concrete connections to mathematical topics and suggests promising research directions related to the clustering phenomenon in Transformers.

future

arxiv

2312.10794v1

OUTLINE The introduction of Transformers in 2017 by Vaswani et al. marked a significant milestone in development of neural network architectures. Central to this contribution is self-attention, a novel mechanism which distinguishes Transformers from traditional architectures, and which plays a substantial role in their superior practical performance. In fact, this innovation has been a key catalyst for the progress of artificial intelligence in areas such as computer vision and natural language processing, notably with the emergence of large language models. As a result, understanding the mechanisms by which Transformers, and especially self-attention, process data is a crucial yet largely uncharted research area. A common characteristic of deep neural networks (DNNs) is their compositional nature: data is processed sequentially, layer by layer, resulting in a discrete-time dynamical system (we refer the reader to the textbook for a general introduction). This perspective has been successfully employed to model residual neural networks—see Section for more details—as continuous-time dynamical systems called neural ordinary differential equations (neural ODEs) . In this context, an input x(0)∈^d, say an image, is evolving according to a given time-varying velocity field as ẋ(t)=v_t(x(t)) over some time interval (0,T). As such, a DNN can be seen as a flow map x(0) ↦ x(T) from ^d to ^d. Even within the restricted class of velocity fields {v_t}_t≥0 imposed by classical DNN architectures, such flow maps enjoy strong approximation properties as exemplified by a long line of work on these questions . In this article we observe that Transformers are in fact flow maps on (^d), the space of probability measures over ^d. To realize this flow map from measures to measures, Transformers evolve a mean-field interacting particle system. More specifically, every particle (called a token in this context) follows the flow of a vector field which depends on the empirical measure of all particles. In turn, the continuity equation governs the evolution of the empirical measure of particles, whose long-time behavior is of crucial interest. In this regard, our main observation is that particles tend to cluster under these dynamics. This phenomenon is of particular relevance in learning tasks such as next-token prediction, wherein one seeks to map a given input sequence (i.e., a sentence) of n tokens (i.e., words) onto a given next token. In this case, the output measure encodes the probability distribution of the next token, and its clustering indicates a small number of possible outcomes. Our results indicate that the limiting distribution is actually a point mass, leaving no room for diversity or randomness, which is at odds with practical observations. This apparent paradox is resolved by the existence of a long-time metastable state. As can be seen from Figures and , the Transformer flow appears to possess two different time-scales: in a first phase, tokens quickly form a few clusters, while in a second (much slower) phase, through the process of pairwise merging of clusters, all tokens finally collapse to a single point. The goal of this manuscript is twofold. On the one hand, we aim to provide a general and accessible framework to study Transformers from a mathematical perspective. In particular, the structure of these interacting particle systems allows one to draw concrete connections to established topics in mathematics, including nonlinear transport equations, Wasserstein gradient flows, collective behavior models, and optimal configurations of points on spheres, among others. On the other hand, we describe several promising research directions with a particular focus on the long-time clustering phenomenon. The main results we present are new, and we also provide what we believe are interesting open problems throughout the paper. The rest of the paper is arranged in three parts. §.§.§ Part : Modeling We define an idealized model of the Transformer architecture that consists in viewing the discrete layer indices as a continuous time variable. This abstraction is not new and parallels the one employed in classical architectures such as ResNets . Our model focuses exclusively on two key components of the Transformers architecture: self-attention and layer-normalization. Layer-normalization effectively constrains particles to evolve on the unit sphere , whereas self-attention is the particular nonlinear coupling of the particles done through the empirical measure. (Section ). In turn, the empirical measure evolves according to the continuity partial differential equation (Section ). We also introduce a simpler surrogate model for self-attention which has the convenient property of being a Wasserstein gradient flow for an energy functional that is well-studied in the context of optimal configurations of points on the sphere. §.§.§ Part : Clustering In this part we establish new mathematical results that indicate clustering of tokens in the large time limit. The main mathematical result regarding particle clustering is described here. We complement this result with a precise characterization of the rate of contraction of particles into a cluster. Namely, we describe the histogram of all inter-particle distances, and the time at which all particles are already nearly clustered (Section ). We also obtain a clustering result without assuming that the dimension d is large, in another asymptotic regime (Section ). §.§.§ Part : Further questions We propose potential avenues for future research, largely in the form of open questions substantiated by numerical observations. We first focus on the case d=2 (Section ) and elicit a link to Kuramoto oscillators. We briefly show in Section how a simple and natural modification of our model leads to non-trivial questions related to optimal configurations on the sphere. The remaining sections explore interacting particle systems that allow for parameter tuning of the Transformers architectures, a key feature of practical implementations. PART: Modeling We begin our discussion by presenting the mathematical model for a Transformer (Section ). We focus on a slightly simplified version that includes the self-attention mechanism as well as layer normalization, but excludes additional feed-forward layers commonly used in practice; see Section . This leads to a highly nonlinear mean-field interacting particle system. In turn, this system implements, via the continuity equation, a flow map from initial to terminal distributions of particles that we present in Section .

What is the main mathematical result concerning particle clustering in Transformers presented in the manuscript?

The main mathematical result is that in high dimensions, a set of particles will cluster to a single point as time goes to infinity.

summary

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. The concept of Pareto optimal solutions is introduced here. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What are Pareto optimal solutions in the context of sustainable decision-making?

Pareto optimal solutions are compromises with different trade-offs when there is no single solution satisfying all conflicting objectives simultaneously.

terminology

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. The purpose of multi-objective optimisation methods is outlined here. They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What is the purpose of multi-objective optimisation methods?

Multi-objective optimisation methods are designed to support decision-makers by generating Pareto optimal solutions and aiding in trade-off analyses and decision-making.

terminology

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. The definition of deep uncertainty is provided here. Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

How is deep uncertainty defined in the context of decision-making?

Deep uncertainty is a situation where probabilities of plausible future scenarios are unknown or experts cannot agree upon.

terminology

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The definition of multi-scenario multi-objective optimisation is given here. Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What does multi-scenario multi-objective optimisation refer to?

Multi-scenario multi-objective optimisation problems consider multiple objectives under deep uncertainty.

terminology

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. An approach for coping with deep uncertainty is mentioned here. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What is the recommended approach for handling deep uncertainty, and why is it suggested?

The most recommended approach to cope with deep uncertainty is to monitor and adapt, meaning the system should always be ready to be adapted if a particular scenario manifests itself.

purpose

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. The summary of the proposed optimisation structure is provided here. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What structure is proposed in the passage to address deep uncertainty?

A two-stage multi-scenario, multi-objective optimisation structure is proposed to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions.

summary

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. The aim of the paper regarding the MS2MORO framework is outlined here. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What is the purpose of proposing the MS2MORO framework in this paper?

The paper aims to address existing gaps and contribute to multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework.

purpose

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Consequences of adding more stages in the multi-stage structure are summarized here. Indeed, it introduces a new trade-off between getting better performances and computation resources. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What are the implications of considering more stages in multi-stage optimisation?

Considering more stages in a multi-stage structure adds more uncertainty and complexity to the problem, which requires additional computation resources.

summary

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. A new trade-off introduced by more stages in optimisation is discussed here. Accordingly, in this paper, we also propose a two-stage moving horizon approach (within a T-stage structure) and compare it with the other proposed multi-stage approach. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What trade-off does the introduction of more stages in optimisation present?

Adding more stages in optimisation introduces a trade-off between achieving better performances and the demand for computation resources.

summary

arxiv

2312.11745v1

INTRODUCTION Sustainable decision-making requires simultaneous consideration of multiple conflicting objectives, e.g., economics versus environmental and social objectives. Typically, there is no single optimal solution satisfying all the conflicting objectives at the same time. Instead, there are several compromises called Pareto optimal with different trade-offs. Therefore, the decision-maker must find the most balanced Pareto optimal solution based on their preferences. Because the Pareto optimal solutions are unknown a priori, additional support is needed to generate those solutions (e.g., utilising mathematical models) and help the decision-maker in trade-off analyses and decision-making. Multi-objective optimisation methods have been designed to provide this kind of support to decision-makers (see, e.g., and references therein). They also applied in many application areas (e.g., ). Moreover, many real-life decision problems are tied to unpredictable events related to future and human behaviours, such as climate change, natural hazards, socioeconomic, pandemics, military/ cyber-attacks, political crises, and religious beliefs. Decision-makers need to make decisions with incomplete knowledge about the consequences, all available options, and the future state of the world (referred to as scenarios from now on) . Because of different sources of uncertainties and the lack of historical data, precise forecasting of future outcomes and transitions is impossible. In this case, the probabilities of plausible future scenarios are unknown, or various experts cannot agree upon one, classified as deep uncertainty . Also, the outcomes and performances of any decision may vary in different scenarios, making the decision-making process too convoluted . The class of optimisation problems considers multiple objectives under deep uncertainty, known as multi-scenario (or scenario-based) multi-objective optimisation . Scenarios, in this context, are utilised to structure and organise the present uncertainty by defining different plausible scenarios for uncertain statements of the problem. Using scenarios to represent uncertainty in a problem can allow the decision-maker to think, discuss, and plan for various plausible representations of an uncertain event without considering the probabilities of their occurrences. In contrast to a wide variety of methods developed for stochastic and probabilistic multi-objective optimisation problems (see, e.g., ), classified as mild uncertainty , only a few studies can be found in the literature investigating multi-objective optimisation under deep uncertainty (e.g., ), despite the need. Nonetheless, handling deep uncertainty has turned into a hot topic in various fields, particularly after worldwide crises in the third decade of the twenty-first century—i.e., the COVID-19 pandemic and the Russian invasion of Ukraine that led to economic and energy crises. The most recommended approach to cope with deep uncertainty is to monitor and adapt , meaning that the system should always be ready to be adapted if a particular scenario manifests itself. Indeed, one needs to identify various scenario-specific adaptation (or contingency) plans in advance and implement the right one after scenario realisation. This way, one can avoid over-conservatism by implementing the relevant adaptation plan after scenario realisation. proposed a two-stage multi-scenario, multi-objective optimisation structure to handle deep uncertainty and identify dynamic-robust Pareto optimal solutions. In the two-stage structure, the decision-making process has been divided into two stages: In the first stage, knowledge about some states of the problem is lacking, and this is when the initial decision needs to be made and implemented. The second stage occurs after unfolding the uncertain states when the adaptation (or recourse) decisions can be implemented. Therefore, a more reasonable initial decision, which is common among all plausible scenarios, and a set of adaptation decisions relating to different scenarios, one for every plausible scenario, is made. However, only the initial decision will be implemented in the first stage, while the implementation of the adaptation decision is postponed to the second stage after scenario realisation. Then, the relevant adaptation decision, which depends on the unfolded scenario, would be implemented. This approach has also been successfully applied in a South-African sugar industry case study . However, real processes are continuous and scenarios continuously unfold, generally with some dependency. Hence, decisions (or plans) must be frequently adapted in more than two stages. Note that, in this sense, “stages" are artificial to structure thinking of an indefinite future. Furthermore, although the two-stage structure lets us evaluate the consequences of the initial decision after achieving every plausible scenario, the long-term effects of the initial decision and the outcomes of the adaptation decisions need to be investigated in a more extended structure. Therefore, this paper aims to address the above-mentioned gaps and contributes to the multi-objective robust decision-making by proposing a multi-stage multi-scenario multi-objective robust optimisation (MS2MORO) framework for dynamic-robust decision-making. To develop a robust strategy for an indefinite future, one may consider the following options: (1) look multiple steps ahead—always plan as if it is a T-stage horizon; or (2) look one step ahead—always plan as if it is a two-stage planning window and then start another two-stage planning window after the first scenario realisation stage (two-stage moving horizon). Clearly, considering more stages adds more uncertainty and complexity to the problem, which naturally requires additional computation resources. Indeed, it introduces a new trade-off between getting better performances and computation resources. A comparison of different multi-stage optimisation approaches is highlighted here. Finally, we apply both proposed approaches to a sequential portfolio selection example and compare the results. This paper is organised as follows: Section briefly reviewed necessary concepts and notations. The proposed multi-stage multi-scenario multi-objective optimisation framework, relevant concepts and models, and the proposed moving horizon approach are discussed in Section . In Section , all the proposed structures and approaches are illustrated and compared in detail through a sequential portfolio selection application under deep uncertainty. Further robustness analysis and discussion are performed in Section , before we conclude in Section .

What new approach is proposed and compared within the context of multi-stage optimisation?

A two-stage moving horizon approach is proposed within a T-stage structure and is compared with another multi-stage approach.

summary